80522 Volume 26 • Number 3 • 2012 ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE) THE WORLD BANK ECONOMIC REVIEW Volume 26 • 2012 • Number 3 THE WORLD BANK ECONOMIC REVIEW Is There a Metropolitan Bias? The relationship between poverty and city size in a selection of developing countries Céline Ferré, Francisco H.G. Ferreira, and Peter Lanjouw Impact of SMS-Based Agricultural Information on Indian Farmers Marcel Fafchamps and Bart Minten Crises, Food Prices, and the Income Elasticity of Micronutrients:Estimates from Indonesia Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman Economic Geography and Economic Development in Sub-Saharan Africa Maarten Bosker and Harry Garretsen The Decision to Import Capital Goods in India: Firms’ Financial Factors Matter Maria Bas and Antoine Berthou Coffee Market Liberalisation and the Implications for Producers in Brazil, Guatemala and India Bill Russell, Sushil Mohan, and Anindya Banerjee Pages 351–555 Implications of COMTRADE Compilation Practices for Trade Barrier Analyses and Negotiations Alexander J. Yeats www.wber.oxfordjournals.org 2 THE WORLD BANK ECONOMIC REVIEW editors Alain de Janvry and Elisabeth Sadoulet, University of California at Berkeley assistant to the editor Marja Kuiper editorial board Harold H. Alderman, World Bank (retired) Caroline Freund, World Bank Chong-En Bai, Tsinghua University, China Paul Glewwe, University of Minnesota, Pranab K. Bardhan, University of California, USA Berkeley Philip E. Keefer, World Bank Thorsten Beck, Tilburg University, Norman V. Loayza, World Bank Netherlands William F. Maloney, World Bank Johannes van Biesebroeck, K.U. Leuven, David J. McKenzie, World Bank Belgium Jaime de Melo, University of Geneva Maureen Cropper, University of Maryland, Ugo Panizza, UNCTAD USA Nina Pavcnik, Dartmouth College, USA Asli Demirgüç-Kunt, World Bank Vijayendra Rao, World Bank Jean-Jacques Dethier, World Bank Martin Ravallion, World Bank Quy-Toan Do, World Bank Jaime Saavedra-Chanduvi, World Bank Frédéric Docquier, Catholic University of Claudia Paz Sepúlveda, World Bank Louvain, Belgium Jonathan Temple, University of Bristol, Eliana La Ferrara, Università Bocconi, Italy UK Francisco H. G. Ferreira, World Bank Dominique Van De Walle, World Bank Augustin Kwasi Fosu, United Nations Christopher M. Woodruff, University of University, WIDER, Finland California, San Diego The World Bank Economic Review is a professional journal used for the dissemination of research in development economics broadly relevant to the development profession and to the World Bank in pursuing its development mandate. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantita- tive development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. Consistency with World Bank policy plays no role in the selection of articles. The Review is managed by one or two independent editors selected for their academic excellence in the field of development economics and policy. The editors are assisted by an editorial board composed in equal parts of scholars internal and external to the World Bank. World Bank staff and outside researchers are equally invited to submit their research papers to the Review. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 26 † 2012 † Number 3 Is There a Metropolitan Bias? The relationship between poverty and city size in a selection of developing countries 351 ´ line Ferre Ce ´ , Francisco H.G. Ferreira, and Peter Lanjouw Impact of SMS-Based Agricultural Information on Indian Farmers 383 Marcel Fafchamps and Bart Minten Crises, Food Prices, and the Income Elasticity of Micronutrients:Estimates from Indonesia 415 Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman Economic Geography and Economic Development in Sub-Saharan Africa 443 Maarten Bosker and Harry Garretsen The Decision to Import Capital Goods in India: Firms’ Financial Factors Matter 486 Maria Bas and Antoine Berthou Coffee Market Liberalisation and the Implications for Producers in Brazil, Guatemala and India 514 Bill Russell, Sushil Mohan, and Anindya Banerjee Implications of COMTRADE Compilation Practices for Trade Barrier Analyses and Negotiations 539 Alexander J. Yeats SUBSCRIPTIONS:A subscription to The World Bank Economic Review (ISSN 0258-6770) comprises 3 issues. Prices include postage; for subscribers outside the Americas, issues are sent air freight. Annual Subscription Rate (Volume 26, 3 Issues, 2012): Institutions—Print edition and site-wide online access: £176/ $264/E264, Print edition only: £162/$242/E242, Site-wide online access only: £147/$221/E221; Corporate—Print edition and site-wide online access: £264/$395/E395, Print edition only: £242/$362/E362, Site-wide online access only: £219/$330/E330; Personal—Print edition and individual online access: £43/$64/E64. US$ rate applies to US & Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary subscriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as firm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con- sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. The current year and two previous years’ issues are available from Oxford University Press. Previous BACK ISSUES: volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oup.com. Tel: þ44 (0)1865 353907. Fax: þ 44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oup.com. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, Tokyo, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: custserv.jp@oup.com. Tel: þ 81 3 5444 5858. Fax: þ 81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in February, June, and October, by Oxford University Press for the International Bank for Reconstruction and Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals postage paid at Cary, NC and at additional mailing offices. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to working with the global community to bring the highest quality research to the widest possible audience. Oxford Journals will protect the environment by implementing environmentally friendly policies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on environmental and ethical policies. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. ADVERTISING: Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales, Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ 44 (0)1865 354767; Fax: þ 44(0)1865 353774; E-mail: jnlsadvertising@oup.com. DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the respective authors and contributors and not of the International Bank for Reconstruction and Development/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. The World Bank Economic Review is printed on acid-free paper that meets the minimum require- PAPER USED: ments of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT # 2012 The International Bank for Reconstruction and Development/THE WORLD BANK All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA. Is There a Metropolitan Bias? The relationship between poverty and city size in a selection of developing countries Ce ´ , Francisco H.G. Ferreira, and Peter Lanjouw1 ´ line Ferre This paper provides evidence from eight developing countries of an inverse relation- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ship between poverty and city size. Poverty is both more widespread and deeper in very small and small towns than in large or very large cities. This basic pattern is gen- erally robust to the choice of poverty line. The paper shows, further, that for all eight countries, a majority of the urban poor live in medium, small or very small towns. Moreover, it is shown that the greater incidence and severity of consumption poverty in smaller towns is generally compounded by similarly greater deprivation in terms of access to basic infrastructure services, such as electricity, heating gas, sewerage and solid waste disposal. We illustrate for one country – Morocco – that inequality within large cities is not driven by a severe dichotomy between slum dwellers and others. Robustness checks are performed to assess whether the findings in the paper hinge on a specific definition of “urban area”; are driven by differences in the cost of living across city-size categories; by reliance on an income-based concept of well- being; or by the application of small-area estimation techniques for estimating poverty rates at the town and city level. JEL Codes: I32, O18, R12 In the late 1970s and in the 1980s, there was much discussion of “urban bias” in development circles. Following Lipton (1977), development economists increasingly recognized a widespread tendency among (almost always urban- based) governments to pursue policies that – explicitly or implicitly – taxed agriculture and transferred resources to industry and other urban activities. 1. Celine Ferre´ is an independent consultant based in Amsterdam, the Netherlands; her email address is celinef@gmail.com. Francisco Ferreira is a lead economist in the Development Research Group at the World Bank and a research fellow at IZA; his email address is fferreira@worldbank.org. Peter Lanjouw (corresponding author) is the manager of the Poverty and Inequality Group at the Development Research Group at the World Bank; his email address is planjouw@worldbank.org. We are grateful to Johan Mistiaen for setting us off on this project. We are also much indebted to Victoria Fazio, Philippe George Leite, Ericka Rasco ´ n, Martin Ravallion, and Timothy Thomas for advice and numerous contributions. Marianne Fay and participants at a World Bank Workshop on Urban Poverty in June 2007, the World Bank/InWent Development Policy Forum meeting in Berlin, 2007, and the LCSPP Seminar in November 2010, provided useful comments. These are the views of the authors and they should not be attributed to the World Bank. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 351 –382 doi:10.1093/wber/lhs007 Advance Access Publication February 14, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 351 352 THE WORLD BANK ECONOMIC REVIEW The motivation was not exclusively urban self-interest. There was a widespread belief, based on the influential early views of Rosenstein-Rodan (1943), Prebisch (1950), and others, that development was to a large extent synonymous with industrialization – and that industrialization inevitably implied urbanization. As markets could not solely be relied upon to allocate resources to that most dynamic sector, government was required to provide a “big push” to help economies along the righteous path of urban growth. Against that view, Lipton and his followers argued that urban bias implied a “sacrifice of efficient and equitable growth to rapid urban advance” ( p.310). By distorting relative prices and the “intersectoral terms of trade”, such policies Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 induced an inefficient allocation of resources that could lead to perpetually infant industries, at the expense of farmers, many of whom were the poorest people in the land.2 That was a time when an estimated 80-90% of the world’s poor lived in rural areas, and an important part of the argument against urban bias was that, in addition to distorting the allocation of capital and other resources, these policies were also anti-poor.3 In 2012, the situation is somewhat different. Urbanization has proceeded apace in the last quarter century such that the world’s urban population is now as large as its rural population. Extreme poverty remains a predominantly rural phenomenon, with some 75% of those who subsist on expenditures below $1-a-day still residing in rural areas in 2002, even when higher cost of living in urban areas is taken into account.4 But urban poverty has been falling more slowly than rural poverty – in part because urbanization has been a key driver behind rural poverty reduction, but some of those who migrate to urban areas remain poor. Urban poverty therefore accounts for a growing share of global poverty: Ravallion et al. (2007) estimate that the urban share of total extreme poverty rose from 19% in 1993 to 25% in 2002. In some regions, like Latin America (76.2%); Eastern Europe and Central Asia (63.5%) and Middle-East and North Africa (55.8%), urban poverty is already dominant. Poverty is expected to continue to urbanize, in the sense that the share of the total number of poor who live in urban areas is expected to continue to grow (with some exceptions, notably in Eastern Europe). Some expect that the urban share of $1-a-day ($2-a-day) poverty may reach 40% (51%) around 2030. Urban poverty is also thought to be accompanied by a different set of characteristics and challenges, including health and sanitation problems in urban slums, unemployment, and a greater incidence of violent crime. In response, strategies to fight urban poverty – and its specific peculiarities – are growing in importance, both at the national and at the international level. 2. A classic study by Bates (1981) documented the use of price regulation and marketing boards in Ghana, Nigeria and Zambia to extract surplus from farmers to the benefit of urban food consumers. 3. Ravallion et al. (2007) produced arguably the first global poverty statistics that cover the majority of the world’s population and disaggregate between urban and rural areas. They estimate that the urban share of the world’s extreme poor in 1993 was 19%. 4. See Ravallion et al. (2007). ´ , Ferreira and Lanjouw Ferre 353 Yet urban poverty is far from a homogeneous phenomenon, even within a single country. It is often remarked that poverty is spatially heterogeneous. Usually this is stated with reference to a marked rural-urban dichotomy in measured poverty. But there is also considerable spatial heterogeneity among urban areas, and one important dimension of that heterogeneity is across city sizes. In Brazil, for instance, while most anecdotal discussion of urban poverty focuses on the sprawling slums of Rio de Janeiro or Sa ˜ o Paulo, over 50% of the country’s urban poor live in towns with fewer than 50,000 inhabitants. Only around 10% live in cities with populations greater than a million. In Kazakhstan, the incidence of poverty in smaller towns is six times larger than Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 in Almaty. And there are large differences in access to local public goods and services too: in Morocco, average access to sewerage is over 80% in cities greater than a million, but less than 50% in the smallest towns. A greater understanding of how poverty – both in terms of incomes or consumption expenditures and in terms of access to public services – varies across different types of cities should help inform the discussion of appropriate poverty reduction strategies in most countries. Yet, the evidence base needed for this disaggregated analysis is seldom available, since household surveys – on which most poverty assessments are based – are seldom representative at the level of any but the largest metropolitan areas in the developing world. They are certainly not representative for smaller towns and cities, and informa- tion is not usually disaggregated along these lines.5 In this paper, we draw upon insights generated by small area poverty estima- tion (based on the combination of welfare estimates from household surveys with “sample” sizes from National Censuses) to investigate the relationship between poverty and city size in eight developing countries, namely Albania, Brazil, Kazakhstan, Kenya, Mexico, Morocco, Thailand and Sri Lanka. We find substantial variation in the incidence and depth of consumption poverty across city sizes in seven of the eight countries. For all seven countries where the data permits some kind of disaggregation of the incidence of public service access, there is also considerable variation across city sizes. In all cases, poverty is lowest and service availability is greatest in the largest cities – precisely those where governments, the middle-classes, opinion-makers and airports are disproportionately located. At a minimum this “poverty gradient” across city sizes needs to be borne in mind whenever considering options and priorities for addressing urban poverty. More speculatively, this evidence might leads us to ask whether, 5. It is rare for household surveys to include identifiers for the specific town or city within which the survey respondent resides – unless the city constitutes a specific stratum. In those settings where such identifiers are included, one can estimate urban poverty rates for different city size intervals (see for example, the poverty profile of Brazil by Ferreira, Lanjouw and Neri (2003) and of India by World Bank (2011)). While city size intervals constructed from survey data are able to reveal differential poverty outcomes across the urban spectrum, they do not convey the differences that may exist between individual towns and cities (see further below). 354 THE WORLD BANK ECONOMIC REVIEW alongside Lipton’s original urban bias, there exists also a “metropolitan bias” in the allocation of resources (including policy attention) to larger cities, at the expense of smaller towns, where most of the urban poor are located. There are a number of caveats which require that our results be treated with care. First, although our samples within countries are representative, our sample of countries is not. Although these eight countries are located in all six regions into which the World Bank routinely divides the developing world, they are not random draws.6 They are countries where there was an early interest in (and the data required for) constructing a poverty map. Second, we use national, rather than international, poverty lines. This has the advantage Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 that poverty is measured in the terms which each particular country’s residents feel is appropriate. But it has the disadvantage that poverty does not mean the same living standard across the eight countries. Third, we do not apply a uniform definition of “urban area” across our eight countries. Instead, we rely on the definition of urban settlements that each respective country reports in its census documents. Such administrative definitions are likely to vary across countries, and may not correspond closely to economic definitions of towns and cities (linked to population density).7 This could have a bearing on our results. We probe the robustness of our conclusions, in the context of Brazil, by re-defining urban settlements so as take explicit account of population density. A fourth caveat concerns our inability to systematically adjust for cost-of-living differences between cities. These differences might be expected to be smaller than those between urban and rural areas, but they may still matter, and we report a second robustness test with respect to cost-of-living differences in the only country in our sample for which data permits it, namely Brazil. Fifth, we do not gauge our findings to variation in equivalence scales – which would matter if family sizes and composition varied systematically across city sizes. Sixth, we assess poverty in a fairly restrictive way: focusing on the share of the population with incomes or consumption levels below the poverty line. It is widely acknowledged that poverty can be viewed more broadly, reflecting multiple dimensions of wellbeing. We seek to mitigate this concern by reporting the association between city size and access to various publicly-provided services and, in one instance, by looking separately at a health outcome (child malnutrition). Nevertheless, it should be acknowledged that the patterns observed in these spaces need not be repeated when other dimensions of poverty are considered. 6. The World Bank divides the developing world into sub-Saharan Africa (AFR), East Asia and the Pacific (EAP), Eastern Europe and Central Asia (ECA), Latin America and the Caribbean (LAC), the Middle-East and North Africa (MENA); and South Asia (SAR). 7. As noted by Lanjouw and Lanjouw (2001), definitions of rural and urban locality vary widely across countries. For example, in Malawi and Zimbabwe settlements with populations of 3000 and 2500 inhabitants, respectively, are defined as urban while in Taiwan a settlement with less than 250,000 inhabitants is designated as rural. ´ , Ferreira and Lanjouw Ferre 355 Each of the limitations of the analysis presented in this paper points to the need for additional research. This paper simply documents the existence of systematic differences in the breadth and depth of poverty (and access to ser- vices) across city sizes in eight geographically diverse developing countries. The paper is structured as follows. Section I provides an overview of the poverty mapping methodology which was used in each of the six countries, in order to generate reliable poverty estimates for every urban area captured in the population census. Section II describes the data sources to which this method was applied, in each country. Section III presents the consumption poverty pro- files by city size in each country. Section IV turns to the evidence on access to Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 publicly provided services across city sizes. Section V looks at differences in poverty (and inequality) within specific cities in Morocco, focusing in particular on poverty and inequality differences between slums and non-slum areas in larger towns. Section VI subjects our findings to some robustness checks. We first look to data from Brazil, in order to gauge the sensitivity of our findings to changes in the definition of urban areas and city sizes. We also use Brazilian data to probe for evidence of spatial cost of living variation across city-sizes. A subsequent robustness check is applied to the case of Mexico, to examine whether an alternative dimension of deprivation, namely child malnutrition, exhibits the same gradient across city size categories as income poverty. In a final robustness check we show that, in India, a poverty-city size gradient can be observed both directly from survey data and from small area estimation techni- ques. We conclude that the inverse relationship is thus not driven exclusively by our reliance on a particular estimation method. Section VII offers tentative con- clusions and discusses some of the questions that this descriptive paper raises for further research into urban poverty. I. ME T H O D O LO GY The economic analysis of the distribution of living standards in developing coun- tries relies almost entirely on household surveys. If their samples are selected appropriately, these surveys can collect detailed information from a relatively small number of households (perhaps 0.1% of the country’s total population), and yet generate information that is representative of the population as a whole. The law of large numbers ensures that the uncertainty about the population which results from sampling (the ‘sampling error’) becomes very small at sample sizes that are still cost effective. This enables researchers the world over to ask detailed questions from small groups of people, at a fraction of the cost that would be required if entire populations needed to be polled. But there is one drawback: the samples that are designed to be representative of large populations are not, in general, representative of specific non-random sub-divisions of that population. Indeed, the typical (nationally representative) household survey is not representative of sub-national units, such as states, pro- vinces or districts. There are exceptions, mostly in large countries, such as 356 THE WORLD BANK ECONOMIC REVIEW China, India or Brazil. But even in those countries, the problem is simply shifted down one level: living standards will vary enormously across different localities of (or towns and cities in) the states of Uttar Pradesh or Minas Gerais; but the Indian National Sample Survey and the Brazilian Pesquisa Nacional por Amostra de Domicı ´lios are not representative at those levels. A number of small-area estimation techniques have been developed to seek to address this missing data problem. In this paper, we rely on applications of the “poverty mapping” approach developed by Elbers, Lanjouw and Lanjouw (2002, 2003). This approach typically involves a household survey and a popu- lation census as data sources. First, the survey data are used to estimate a Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 prediction model for either consumption or incomes. The selection of explana- tory variables is restricted to those variables that can also be found in the census (or some other large dataset) or in a tertiary dataset that can be linked to both the census and survey. The parameter estimates are then applied to the census data, expenditures are predicted, and poverty (and other welfare) statis- tics are derived. The key assumption is that the models estimated from the survey data apply to census observations. Let W be a welfare indicator based on the distribution of a household level variable of interest, yh. Using a detailed household survey sample, we estimate the joint distribution of yh and observed correlates xh. By restricting the explanatory variables to those that also occur at the household level in the population census, parameter estimates from this “first stage” model can be used to generate the distribution of yh for any target population in the census conditional on its observed characteristics and, in turn, the conditional distri- bution of W. Elbers et al. (2002, 2003) study the precision of the resulting esti- mates of W and demonstrate that prediction errors will fall (or at least not rise) with the number of households in the target population, and will also be affected by the properties of the first stage models, in particular the precision of parameter estimates. A general rule of thumb is that welfare estimates obtained on this basis will be estimated fairly precisely as long as the target population comprises at least 1,000-5,000 households. The first-stage estimation is carried out using household survey data.8 The empirical models of household consumption allow for an intra-cluster correlation in the disturbances (see Elbers, Lanjouw and Lanjouw, 2002, 2003, Elbers, Lanjouw and Leite, 2008, and Demombynes et al., 2007, for more details). Failing to take account of spatial correlation in the disturbances would result in underestimated standard errors in the final poverty estimates (Tarozzi and Deaton, 2009). Different models are estimated for each region and the specifica- tions include census mean variables and other aggregate level variables in order to capture latent cluster-level effects. All regressions are estimated with household 8. These surveys are stratified at the region or state level, as well as for rural and urban areas. Within each region there are further levels of stratification, and also clustering. At the final level, a small number of households (a cluster) are randomly selected from a census enumeration area. ´ , Ferreira and Lanjouw Ferre 357 weights and with parsimonious specifications to be cautious about overfitting. Heteroskedasticity is also modeled in the household-specific part of the residual. Parameter estimates from all the first-stage models are then taken, in the second stage, to the population census. Since predicted household-level per capita consumption in the census is a function not only of the parameter esti- mates from the first stage consumption models estimated in the survey, but also of the precision of these estimates and of those parameters describing the disturbance terms in the consumption models, we do not produce just one pre- dicted consumption level per household in the census. Rather, a reasonably large number of predicted expenditures are simulated for each household (typic- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ally around 100 simulations). The full set of simulated household-level per capita expenditures are then used to calculate the welfare estimates for each target population. Demombynes, Elbers, Lanjouw and Lanjouw (2007) describe a variety of simulation approaches that are available and document that these all yield closely similar welfare estimates. Validation studies of the poverty mapping methodology remain rare; in settings where one can rigorously check the method, it is likely that it was not needed in the first place. However, a few such studies have been conducted and have yielded encouraging findings (see for example Demombynes, et al., 2007, and Elbers et al., 2008). We examine in Section VI whether there are grounds for suspecting that our broad findings con- cerning the relationship between urban poverty and city size are due to our employment of small-area estimates of poverty as opposed to direct measures. I I . D ATA Poverty-mapping exercises based on the methodology just described have now been conducted in a number of countries. We have selected eight of these coun- tries, on the basis of the availability of the micro-data files and of regional coverage, for analysis in this paper. Table 1 lists the total urban population (at the Census year) in the eight countries, both in absolute numbers and as a share of the total. The sample includes a wide variety of countries, from the relatively small (e.g. Albania) to the relatively large (e.g. Brazil), and from the predominantly rural (e.g. Sri Lanka) to the highly urbanized (e.g. Brazil). The Table also indicates which household survey was used for the estimation of the household expenditure model, including year and sample size. The year of the nearest available population Census, which was used to generate the small-area welfare estimates, is also included.9 9. Further details about the poverty maps analyzed here can be found in respectively (INSTAT, 2004) for Albania, IBGE (2003) for Brazil, Kenya Central Bureau of Statistics (2003) for Kenya, Lo´ pez-Calva et al. (2005) for Mexico, Haut Commissariat au Plan (2005) for Morocco, Healy and Jitsuchon (2007) for Thailand, Department of Census and Statistics (2005) for Sri Lanka. In the case of Kazakhstan, the poverty map for that country was produced on a pilot basis in collaboration with the Agency of Statistics of the Republic of Kazakhstan. The results of this exercise have not been placed in the public domain. 358 T A B L E 1 : Data Sources Albania Brazil Kazakh-stan Kenya Mexico Morocco Thailand Sri Lanka Urban Population 1.3m 125m 8.2m 5.0m 54.5m 12.7m 18.5m 2.2m Urban Population share 0.42 0.83 0.57 0.19 0.60 0.51 0.31 0.12 Census Year 2001 2000 1999 1999 2000 1994 2000 2001 Survey Year 2002 2002-3 2001 1997 2000 1998 2000 2002 Survey Name LSMS POF HBS WMS III ENIGH ENNVM SES HIES Survey Sample Size 3,600 48,470 11,883 10,874 10,108 5,184 24,747 20,100 THE WORLD BANK ECONOMIC REVIEW Poverty Line1 ALL 4,891 BRL 100 KZT 3,157 KES 2,648 PES 768 DHS 283 BAH 1370 LKR 1423 Poverty Line (2005 PPP$) 85 83 63 147 126 57 88 45 Equivalence Scale No No No Yes No No No No 1 All poverty lines are monthly, and displayed using national currencies. International acronyms apply. Note: LSMS: Living Standard Measurment Survey; POF: Pesquisa de Orc ¸ amentos Familiares; HBS: Household and Budget Survey; WMS: Welfare Monitoring Survey; ENIGH: Encuesta Nacional de Ingresos y Gastos de los Hogares; ENNVM: Enque ´ nages; ˆ te Nationale sur les Niveaux de Vie des Me SES: Socio-Economic Survey; HIES: Household Income and Expenditure Survey. Source: The above surveys and National Censuses. Consumer price indices and PPP exchange rates used for poverty line conversions come from PovcalNet. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ´ , Ferreira and Lanjouw Ferre 359 Finally, the Table also lists the poverty line used in each country, both in national currency (in survey year prices) and US dollars (in 2006 prices and at PPP exchange rates)10. As noted in the introduction, we have opted to use national poverty lines, which better capture the meaning of poverty in each specific country. This has the drawback that poverty measures are not defined with reference to comparable standards of living across countries. The alterna- tive of imposing a constant poverty line across countries, however, would have an even greater disadvantage. Had we selected a low internationally compar- able poverty line, such as $1-a-day, we would be comparing traces of poverty driven largely by measurement error and transitory shocks in the richer coun- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 tries (such as Albania and Brazil) with real poverty in Kenya and Sri Lanka. Had we instead selected a higher line, like those used in Albania, Brazil or Morocco, we would be comparing “reasonable” poverty incidences in the richer countries, with the bulk of the population in the poorer countries. Since this paper is largely about the relative extent of poverty in larger and smaller towns, the absolute level of the poverty line is of limited importance. We do, nevertheless, examine the sensitivity of our results to varying the poverty line in some of the countries, in the next section. Further caveats relate to the fact that we make no attempt to apply a uniform definition of urban, or to systematically correct for differences in the cost-of-living across different urban categories. In some settings, these differ- ences may be substantial, and future research should attempt to take them into account.11 In addition, with the exception of Kenya, we have used consump- tion expenditure per capita as the individual welfare indicator throughout. If there are substantive differences in family size or composition across different urban categories, one might like to investigate the robustness of the results with respect to different assumptions regarding equivalence scales. Note that with respect to both cost of living differences and equivalence scales, our find- ings will be sensitive to systematic differences between large cities and smaller towns. We are not as vulnerable, here, to differences that might exist between urban areas, generically, and rural areas. It is an important empirical question just how much variation there is between cities of different sizes in terms of prices, consumption patterns, and demographic characteristics. III. CONSUMPTION POVERTY BY CITY SIZE Table 2 presents our estimates of the three standard FGT poverty measures (as well as population shares and the share of the poor) for each country as a 10. Each poverty line is per capita per month. 11. Section VI reports on a robustness check indicating that our findings for Brazil are not overturned after refining the definition of urban we employ and correcting by cost of living differences across city-size categories. 360 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 : Poverty measures and shares for different city sizes in eight countries Population share1 FGT0 FGT1 FGT2 Share of the Poor2 Albania 0.25 Rural 0.30 Urban 0.42 0.18 0.04 0.01 0.31 M 0.15 0.18 0.04 0.02 0.11 S 0.13 0.18 0.04 0.01 0.09 XS 0.14 0.20 0.05 0.02 0.11 Brazil 0.22 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Rural 0.37 Urban 0.83 0.19 0.07 0.04 0.72 XL 0.22 0.09 0.03 0.01 0.09 L 0.07 0.17 0.06 0.03 0.06 M 0.24 0.15 0.05 0.03 0.17 S 0.01 0.19 0.07 0.04 0.01 XS 0.28 0.30 0.11 0.06 0.39 Kazakhstan 0.18 Rural 0.23 Urban 0.57 0.14 0.04 0.01 0.43 XL 0.08 0.03 0.01 0.00 0.01 M 0.29 0.13 0.04 0.01 0.21 S 0.05 0.18 0.05 0.02 0.05 XS 0.15 0.19 0.05 0.02 0.15 Kenya 0.51 Rural 0.52 Urban 0.19 0.47 0.17 - 0.16 XL 0.07 0.44 0.14 - 0.06 L 0.02 0.44 0.16 - 0.02 M 0.03 0.46 0.17 - 0.03 S 0.02 0.55 0.22 - 0.02 XS 0.04 0.49 0.21 - 0.04 Mexico 0.32 Rural 0.52 Urban 0.60 0.19 0.06 0.03 0.39 XL 0.27 0.18 0.06 0.03 0.16 L 0.13 0.14 0.04 0.02 0.06 M 0.11 0.19 0.05 0.03 0.07 S 0.04 0.25 0.07 0.04 0.03 XS 0.06 0.31 0.09 0.05 0.07 Morocco 0.17 Rural 0.23 Urban 0.51 0.11 0.03 0.01 0.34 XL 0.12 0.04 0.01 0.00 0.03 L 0.09 0.14 0.04 0.02 0.07 M 0.27 0.13 0.03 0.01 0.20 S 0.03 0.16 0.04 0.02 0.03 (Continued ) ´ , Ferreira and Lanjouw Ferre 361 TABLE 2: Continued Population share1 FGT0 FGT1 FGT2 Share of the Poor2 XS 0.01 0.12 0.03 0.01 0.01 Sri Lanka 0.23 Rural 0.25 Urban 0.12 0.09 0.02 0.01 0.05 L 0.03 0.08 0.02 0.01 0.01 M 0.03 0.07 0.02 0.01 0.01 S 0.02 0.09 0.03 0.01 0.01 XS 0.04 0.12 0.03 0.00 0.02 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Thailand 0.14 Rural 0.17 Urban 0.31 0.08 - - 0.17 XL 0.12 0.02 - - 0.01 M 0.03 0.04 - - 0.01 S 0.02 0.09 - - 0.01 XS 0.14 0.14 - - 0.13 1 Proportion of the population living in each category: urban, XL,L,M,S,XS. 2 Proportion of the country’s poor living in each category: urban, XL,L,M,S,XS. XL: . 1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: , 50 (‘000 inhabitants). Source: Authors’ analysis based on data described in the text. whole, and then for their urban areas, first as an urban aggregate, and then dis- aggregated into five size categories: towns smaller than 50,000 (“very small” or XS); between 50,000 and 100,000 (“small” or S); between 100,000 and 500,000 (“medium” or M); between 500,000 and 1 million (“large” or L) and above 1 million (“metropolitan areas” or XL). Two countries have no metro- politan areas: Albania and Sri Lanka. Two countries have no large cities: Albania and Kazakhstan. In all eight countries, both poverty incidence (FGT(0)) and depth (FGT(1)) are highest in either the very small (Albania, Brazil, Kazakhstan, Mexico, Sri Lanka, Thailand) or the small (Kenya and Morocco) categories.12 This pattern is particularly pronounced in the larger, more urbanized countries of Brazil, Kazakhstan, Mexico, Morocco, and Thailand where FGT(0) in the very small cities is up to six times larger than in the metropolitan areas, and often only slightly lower than in rural areas. In these countries, the more distribution- sensitive poverty measures paint a similar picture: FGT(2) is six times larger for very small towns than for metropolitan areas in Brazil; FGT(1) is five times larger in Kazakhstan. 12. FGT(a) denotes a member of the Foster, Greer and Thorbecke (1984) family of poverty indices, with parameter a. 362 THE WORLD BANK ECONOMIC REVIEW In the other three countries – Albania (heavily urbanized, but small in total population and area), Kenya and Sri Lanka ( predominantly rural) – the pattern is less pronounced, but it is still present. In fact, the coefficient on population in a simple OLS regression of poverty on city size is negative in all cases, and significant at the 10% level in five (four) out of eight cases for FGT(0) (FGT(1)). See Table 3.13 The inverse relation between poverty and city size can also be discerned in Figure 1, which presents the distribution of poverty incidence within each size category, by means of box-plots. The box-plots indicate that there is much greater variance in poverty rates among smaller towns, as one might expect Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 from their sheer number. But the median poverty rate falls markedly and consistently with city size in Brazil and Kazakhstan. It also falls in Albania, although less markedly. In Mexico, the gradient is clear across all city-size categories except for the metropolitan areas for which the median poverty rate is slightly higher than for the large city-size category (but below all other categories). In Morocco, the negative correlation detected in Table 3 is driven by much lower poverty in metropolitan areas, with no clear pattern among the other size categories. In Sri Lanka and Kenya, the relationship owes to greater poverty incidence in small and very small towns, with no clear pattern among medium and larger towns. Similarly in Thailand it is noteworthy that the over- whelming majority of urban centers belong to the extra small category, with metropolitan Bangkok representing the one very large exception. (Chiang Mai, Thailand’s second largest city, had a population of only just over 280,000 in 2008). The median poverty rate in Thailand’s smallest towns is markedly higher than in all other city-size categories. These patterns are also clearly visible in Figure 2, which presents the non-parametric regressions of FGT(0) on the logarithm of city size for each country. Here again, it is least visible in Kenya and Sri Lanka. In Morocco, as we have seen, the negative relationship is driven by markedly lower poverty in Casablanca. With the exception of Kenya and Mexico, metropolitan poverty incidence is less than half of the average urban poverty in every country in our sample that has at least one metropolitan area. The data underpinning Figures 1 and 2 are of interest not only in providing evidence of a poverty gradient with respect to city size in our eight countries, but also in documenting the great heterogeneity in poverty across towns within a given city size category. Thus, while in Brazil or Thailand the median poverty rate amongst towns in the XS town-size category is clearly higher than in the other categories, there evidently are very small towns that also enjoy very low poverty rates (as well as small towns with near universal poverty). Indeed, in 13. These regression coefficients are presented as illustrative of correlations only. City size is clearly endogenous, and there are evidently many omitted variables, so no inference of causality is possible. Some countries do not display the full set of regressions for lack of data (the Kenyan census for instance being very short, no information is available on infrastructure access). T A B L E 3 : Simple regressions of poverty indicators on city size, OLS Dependent variable: FGT0 FGT1 Water Electricity Sewer Country coeff p-value coeff p-value coeff p-value coeff p-value coeff p-value Albania 2 0.17 0.21 2 0.04 0.27 2 2 2 2 2 2 Brazil 2 0.10 0.03 2 0.04 0.03 0.03 0.02 0.01 0.03 0.03 0.04 Kazakhstan 2 0.20 0.00 2 0.06 0.00 2 2 0.00 0.06 0.92 0.02 Kenya 2 0.02 0.73 2 0.04 0.38 2 2 2 2 2 2 Mexico 2 0.02 0.06 2 0.004 0.28 0.007 0.24 0.002 0.07 0.007 0.16 Morocco 2 0.03 0.01 2 0.01 0.02 0.06 0.01 0.05 0.02 0.08 0.23 Thailand 2 0.04 0.15 2 0.01 0.22 1.00 0.04 0.06 0.00 Sri Lanka 2 0.08 0.08 2 0.03 0.05 0.40 0.23 2 0.03 0.54 2 2 Ferre Explanatory variable: city size in million of inhabitants (‘000,000). All poverty indicators take values between 0 and 1. Source: Authors’ analysis based on data described in the text. 363 ´ , Ferreira and Lanjouw Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 364 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ calculations based on the household surveys and censuses listed in Table 1. Note: XL: . 1,000, L: 500– 1,000, M: 100 – 500, S: 50 – 100, XS: , 50 (thousands inhabitants). Albania the evidence indicates that the lowest estimated urban poverty rates are found amongst towns in the XS category. It is thus important to bear in mind that while on average poverty rates in smaller towns tends to be higher than in medium and large cities, this is far from a general rule. To investigate the robustness of the inverse poverty-city size relationship to variations in the poverty line, we plotted the cumulative distribution function ´ , Ferreira and Lanjouw Ferre 365 F I G U R E 2. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ calculations based on the household surveys and censuses listed in Table 1. Note: fitted with Lowess regression – bandwidth ¼ 1. separately by size category for each country.14 Poverty is always higher in the smallest towns (XS), for any poverty line, in Albania, Brazil, Sri-Lanka and Thailand (up to the 90th percentile). It is generally lowest for metropolitan areas in the vicinity of the national poverty lines, but this ranking is not every- where robust to larger changes in the poverty line. Figures 3 and 4 illustrate two polar cases: Brazil and Morocco. Figure 3 shows that metropolitan areas 14. Again with the exception of Kenya, for which we do not have the disaggregated poverty mapping data. 366 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Brazilian poverty map, constructed from the POF survey and the 2000 census. Note: XL: . 1,000, L: 500– 1,000, M: 100 – 500, S: 50 – 100, XS: , 50 (thousands inhabitants). F I G U R E 4. Source: Moroccan poverty map, constructed from the ENNVM survey and the 1994 census. Note: XL: . 1,000, L: 500– 1,000, M: 100 – 500, S: 50 – 100, XS: , 50 (thousands inhabitants). first-order stochastically dominate all other size categories in Brazil: poverty is lower in these large cities than in any other type of town, by any poverty line. Conversely, very small towns are first-order stochastically dominated by every other size grouping: poverty is higher in this size category than in any other, by any poverty line. ´ , Ferreira and Lanjouw Ferre 367 A very different picture (in terms of dominance relationships) is that of Morocco, shown in Figure 4. The poverty ranking between metropolitan areas and large towns which is observed at the country’s poverty line of Dhs 3,400 reverses at higher poverty lines (above Dhs 8,000). Similarly, there is no domin- ance relationship among very small, small and medium towns in Morocco: their cumulative distribution functions cross many times. Even in Morocco, however, which displays the largest number of cumulative distribution function crossings in our sample, there is still one broad regularity: taken as a group, large and very large cities (L, XL) do provide a lower envelope for the smaller towns (XS, S, M). There is no strict stochastic dominance but it is evident that, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 for almost every poverty line one could think of, poverty is lower in the group of larger cities than in other urban settings. It is possible, of course, that poverty is both more widespread and deeper in smaller towns, but that population is so concentrated in large cities that the bulk of the poor live there. If this were the case, greater attention to (and resources for) metropolitan poverty might be justified on the basis that the share of poverty is greatest there. But Table 2 shows that this is nowhere the case. In fact, the share of the poor is lower than the population share in every country that has a metropolitan area: the difference is relatively small in Kenya, but very substantial elsewhere. In Brazil, although 22% of the popula- tion live in cities greater than 1 million, only 9% of the country’s poor do. In Kazakhstan, 14% of the population lives in Almaty, but only 3% of the poor. In Mexico, 27% of the population resides in Mexico City and the other very large metropolitan areas of the country, but only 16% of the poor live in these conurbations. In Morocco, 12% of the population lives in Casablanca but only 3% of the poor. At the other end of the size distribution, a majority of the country’s urban poor live in small or very small towns in four of our eight countries: Albania, Brazil, Sri Lanka and Thailand. If we add medium towns to the list, this rises to seven of the eight countries, including Kenya. And even in the case of Mexico, where the population weight of metropolitan areas is particularly large, the share of the urban poor in medium or smaller sized cities exceeds 40%. I V. A C C E S S TO SERVICES BY CITY SIZE Even though people are poorer in smaller towns than in large cities and even though a greater number of the poor live in those smaller towns, one might think that, due to the higher population densities in metropolitan areas, per- capita availability of publically provided basic services was lower there. This does not appear to be the case, however. Table 4 presents the proportion of households with access to various basic infrastructure services by city size, in seven of our eight countries.15 15. Kenya is once again omitted for data reasons. 368 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 : Access to services for different city sizes in seven developing countries Water Electricity Sewer Gas Garbage Fridge Electric Heat Albania Urban 0.88 0.62 M 0.91 0.68 S 0.87 0.57 XS 0.87 0.60 Brazil Urban 0.96 0.99 0.92 0.86 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 XL 0.98 1.00 0.94 0.92 L 0.97 1.00 0.91 0.89 M 0.97 1.00 0.93 0.91 S 0.96 0.99 0.94 0.89 XS 0.92 0.98 0.90 0.76 Kazakhstan Urban 1.00 0.68 0.55 XL 1.00 0.73 0.81 M 1.00 0.80 0.62 S 1.00 0.67 0.36 XS 1.00 0.40 0.31 Mexico Urban 0.93 0.99 0.94 0.83 XL 0.95 0.99 0.97 0.84 L 0.93 0.99 0.93 0.87 M 0.92 0.98 0.93 0.81 S 0.91 0.98 0.91 0.78 XS 0.89 0.98 0.90 0.76 Morocco Urban 0.77 0.82 0.87 XL 0.84 0.87 0.87 L 0.86 0.87 0.80 M 0.71 0.79 0.91 S 0.73 0.78 0.91 XS 0.75 0.78 0.45 Thailand Urban 0.65 0.16 0.86 XL 0.87 0.29 0.88 M 0.76 0.23 0.90 S 0.61 0.14 0.87 XS 0.50 0.07 0.84 Sri Lanka Urban 0.57 0.89 L 0.53 0.86 M 0.68 0.90 S 0.60 0.92 XS 0.51 0.89 XL: . 1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: , 50 (‘000 inhabitants). Source: Authors’ analysis based on data described in the text. ´ , Ferreira and Lanjouw Ferre 369 Access to piped water is generally quite high in Brazil, but it declines from 98% in metropolitan areas, to 92% in very small towns. In Mexico and Thailand the comparable figures are 95% to 89%, and 87% to 50%, respect- ively. In Morocco and Sri Lanka, the picture is less clear. In Morocco, as for income poverty, access to piped water is higher in the two largest size categories (L, XL), than in the other three (M, S, XS). In Sri Lanka, there is an inverted U curve, with access lowest in large and very small towns. Similar pat- terns hold in each of these countries with respect to access to electricity, although overall access rates tend to be higher. Access increases monotonically with city size in Brazil and Mexico; it is higher in L and XL cities than in S, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 XS and M towns in Morocco (but with no clear pattern within these two blocks), and it follows an inverted U in Sri Lanka. Access to networked sanitation and sewerage facilities is on average scarcer than piped water or electricity in most developing countries. And in our sample of countries, there is also a clear positive association between city size and access to networked sewerage services. In all five countries that report data on this service (Brazil, Kazakhstan, Mexico, Morocco and Thailand), very small towns have the lowest access rates – in two cases just barely half the rates observed in larger towns. Interestingly, however, in both Kazakhstan and Morocco, medium-sized towns report higher access rates than metropolitan areas. Access to piped natural gas is an important infrastructure service in Kazakhstan (for cooking and heating). Access is clearly and monotonically increasing with city size. The differences are quite sizable, with 81% of con- nected households in Almaty, but only 31% in very small towns. A similar pattern attains for electric heating apparatus in Albania. Access to organized solid waste disposal (garbage collection) is only reported for Brazil, where it is once again highest in metropolitan areas, and lowest in very small towns.16 V. L O O K I N G WITHIN CITIES: THE CASE OF MOROCCO A further plausible argument for focusing one’s poverty-reduction efforts on metropolitan areas might be that – even if poverty is less widespread or intense there; even if a smaller share of the poor live there; and even if they already enjoy superior access to services – these very large urban centers are deeply divided between rich and poor. If relative incomes matter for well-being, then the stark contrast between the crowded and steep hillsides of Rocinha and the neighboring verdant gardens of Ga ´ vea in Rio de Janeiro may be so inherently objectionable as to raise the priority that should be accorded to fighting poverty in large cities. 16. Although the relationship for intermediate size categories is not monotonic, and there is very little difference between large, medium and small towns in this respect. 370 THE WORLD BANK ECONOMIC REVIEW There may well be something to the argument that stark local inequalities may have greater costs than geographically diffuse inequality. There is some evidence that relative incomes in one’s vicinity do affect well-being directly (Luttmer, 2005), and that local inequality may lead to increased property violence (Demombynes and O ¨ zler, 2005). But there is much less evidence that inequality is indeed so much greater in metropolitan areas than in smaller towns. Although this is a popular notion, it is one for which very limited statis- tical backing exists – in large part for previously mentioned reasons: household surveys are not representative at the level of smaller towns, and so we know very little about local inequality in them. It may be that the accumulation of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 anecdotal evidence of large inequalities in developing country metropolises is itself simply another reflection of metropolitan bias: Journalists and photogra- phers, like most economists and policy analysts, prefer to visit Casablanca than Figuig17, and Rio de Janeiro than Bertolı ´nia.18 To shed some additional light on this matter, we now turn to some evidence from Morocco. Table 5 presents FGT(0) and three inequality measures (the Gini coefficient and the two Theil indices) for each of the five largest cities in the country, as well as the aggregate inequality for three city size categories. Overall intra-city inequality does not appear to be positively correlated with city size in this small sample, but this is not the main point. Taking advantage of the fine spatial disaggregation made possible by a poverty map, we calcu- lated inequality measures for various individual neighborhoods within each of these cities. We further classified these neighborhoods into slums and non- slums19. We then decomposed the two Theil indices of inequality20 for each of the five cities, into a component due to inequality within each of the two groups of neighborhoods, and a component between the two. For the argument that within-city inequality is egregiously large in metropolitan areas and large cities to hold (in Morocco), it would be necessary (but not sufficient) that the between-group shares reported in the last two columns of Table 5 be substan- tial. In the event, it appears that most inequality in the five largest cities in Morocco is not due to some great divide between slum areas and other parts of the town. Inequality appears to be considerably more widely dispersed within these two broad groups. 17. Casablanca is the biggest agglomeration of Morocco (2.9 million inhabitants), Figuig is a small town in L’Oriental (49,000). 18. Bertolinia is a small town in the Brazilian state of Piauı´, with fewer than 40,000 inhabitants. 19. A district (smallest level of disaggregation after the census track) was considered as a slum if less than 10% of the population had access to water and less than 10% of the population had access to electricity. 20. GE(0), or mean log deviation, is the Theil-L index. GE(1) is the Theil-T index. Both are perfectly decomposable into within- and between-group components, in the sense that the decomposition has no residual. GE(0) weighs within-group inequalities by population shares, while GE(1) weighs them by incomes shares. ´ , Ferreira and Lanjouw Ferre 371 T A B L E 5 : The role of metropolitan slums: Poverty and inequality decompositions within five cities in Morocco 2 groups: slums and All Urban Areas non-slums1 Morocco Population FGT0 GINI GE02 GE12 W03 W13 B03 B13 Casablanca 2,875,326 0.05 0.39 0.25 0.29 0.24 0.29 0.01 0.01 Rabat 604,680 0.04 0.41 0.27 0.30 0.27 0.29 0.01 0.01 Sale´ 575,600 0.12 0.38 0.24 0.26 0.22 0.25 0.02 0.02 Marrakech 503,802 0.07 0.37 0.23 0.23 0.23 0.23 0.00 0.00 `s Fe 501,592 0.23 0.47 0.38 0.41 0.38 0.41 0.00 0.00 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 200,000-500,000 4,930,980 0.12 0.36 0.21 0.23 0.21 0.22 0.00 0.00 , 200,000 2,736,390 0.15 0.22 0.13 0.14 0.13 0.11 0.00 0.00 All urban 4 12,728,370 0.11 0.37 0.23 0.25 0.23 0.24 0.01 0.01 1 A slum is a district (smallest disaggregation above the census tract) where less than 10% of the population has access to water and less than 10% has access to electricity. 2 GE0, or mean log deviation, is the Theil-L index. GE1 is the Theil-T index. Both are perfect- ly decomposable into within- and between-group components, in the sense that the decompos- ition has no residual. GE0 weighs within-group inequalities by population shares, while GE1 weighs them by incomes shares. 3 W0 and W1 display within-group inequality associated with the GE0 and GE1 measures re- spectively; B0 and B1 display the corresponding between-group inequality component. 4 Each index presented here was computed at the city level and then aggregated into each cat- egory (all urban, etc). Source: Authors’ calculations based on the ENNVM: Enque ˆ te Nationale sur les Niveaux de Vie des Me ´ nages. VI. FOUR ROBUSTNESS CHECKS We noted at the outset several important caveats attached to the broad findings in this paper. While we are unable to subject all of them to exhaustive scrutiny, we attempt, in this section, to probe several with a view to gauging the robust- ness of our findings. We assess, first, our reliance on official, country-specific, definitions of what constitutes an urban setting in our eight countries. In most countries, the designation of an area as urban is based on several criteria, combining both administrative functions and population characteristics. As a starting point, many countries designate as urban any area, irrespective of population size or density, that happens to have been declared a municipality or corporation by state law, or that has a Cantonment Board or notified town area committee. In general such administrative criteria are combined with additional criteria such as a minimum population (e.g. 5000 inhabitants), a minimum density of population (e.g. at least 400 persons per sq. km), and/or a minimum degree of economic diversification out of agriculture (e.g. at least 75% of male working population engaged in non-agricultural pursuits). As mentioned in the introduction, we take in this paper the definition of urban 372 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 : Poverty and City Size in Urban Brazil: Modified Definition of Towns and Cities based on Contiguous Census Tracts Total Population Share of Urban Share of Urban City Size (000’s) FGT0 Population Poor 25,000-100,000 21,261 0.22 0.20 0.29 100,000-500,000 22,384 0.16 0.21 0.22 . 500,000 61,655 0.12 0.59 0.49 All Urban 105,300 0.15 1.00 1.00 Source: Thomas (2008). Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 employed by each respective country, and do not try to impose a uniform definition across the countries in our sample. The question thus arises whether our finding of a town size–poverty gradient would survive the adoption of an alternative definition of urban. In particular, it is of interest to know whether using a more data driven, and less administrative, definition would affect our conclusions. To explore this concern we take the Brazil census data and derive our own urban area definitions for that country.21 We start with census tracts (group- ings of roughly 100 geographically contiguous households) that have been designated as either rural or urban by IBGE, the Brazilian National Statistical Office. We then assemble these census tracts into towns and cities by joining contiguous urban census tracts together. In some cases these form huge units. For example, a coastal area near Rio de Janeiro is 10 km wide and over 170 km long. Compared to the GRUMP (Global Rural Urban Mapping Project at the Columbia University Center for International Earth Science Information Network) settlement area database, this approach would appear to join together several distinct cities into megacities. Because these seem too large, we cut them up using Brazilian micro-regions (officially recognized agglomeration of municipios, but subunits of states) which divides up these “super agglomera- tions” into areas that closely approximate what GRUMP considers independent cities in Brazil, together with their immediate metropolitan areas. While the definition of a town or city, under this approach, is not tied strictly to a popu- lation density criterion ( probably the most economically logical way of defining a town) it does employ a single, uniform, criterion and is likely to correlate reasonably well with a purely density-based definition. Table 6 recalculates poverty across city size categories when cities have been defined in this alternative fashion. We also drop from consideration, here, the large subset of very small towns with less than 25,000 inhabitants22. One might 21. We draw here on work undertaken by Timothy Thomas in support of the project described in this paper (Thomas, 2008). 22. There are nearly 10,000 such settlements in Brazil. ´ , Ferreira and Lanjouw Ferre 373 argue that such very small towns are hard to distinguish clearly from rural areas, particularly where those rural areas are reasonably densely populated.23 We examine whether our city size – poverty gradient for Brazil survives the exclusion of these very smallest towns. We see from Table 6 that amongst all urban centers with a population of at least 25,000 inhabitants, poverty in small towns (with a population of 25,000 to 100,000) is markedly higher than in the medium sized cities (100,000-500,000) which in turn record greater poverty than the large cities of 500,000 inhabitants or more. In terms of poverty shares, our earlier findings also survive: while the inhabitants of Brazil’s largest cities represent 59% of the urban population (as per our new definition), they represent just 50% of the Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 urban poor. Conversely, small towns of 25,000 to 100,000 inhabitants account for just 20% of the urban population but nearly 30% of the urban poor. Our second robustness check assesses the possibility that there may be important cost of living differences between urban settlements of different sizes. The findings reported above have not attempted to adjust for such cost-of-living differences, because spatial price indices across city-size categories are not generally available. It is well recognized in the literature, however, that observed differences in poverty rates between urban and rural areas can be significantly attenuated once one corrects for the fact that the cost of living in urban areas may be much higher than in rural areas (generally because of the higher cost of food, housing and transport services). The possi- bility exists that our broad findings of greater poverty in small towns than in metropolitan areas might also be driven, at least in part, by our failure to allow for a higher cost of living in metropolitan areas. While household survey datasets are not generally large enough in sample size to permit the construction of a cost-of-living index across different city-size categories, our survey data for Brazil constitute an important exception. We are able to draw on the 2002 POF data (see Table 1) to construct a cost-of-living index across the broad city size categories employed in this paper, and can check whether our findings for Brazil, reported in previous sec- tions, are robust to this correction. There are many ways in which spatial price indices can be constructed. We follow here the approach applied by Ferreira, Lanjouw and Neri (2003) to the construction of a regional price index (that distinguished also between urban and rural areas) using 1996 PPV data for Brazil. This approach was subse- quently applied in World Bank (2007) to produce a regional price index based on the 2002 POF data, and is based on unit-value information provided in the 23. Nevertheless, the large number of small towns in developing countries, and the high poverty rates that characterize them (Table 2), raises the interesting question of the role of such towns in the classic rural-urban debates in development economics (for a recent overview, see Christiaensen, Demery and Kuhl, 2011). Lanjouw and Murgai (2009) suggest, in the context of India, that growth in such small towns has promoted rural non-farm diversification which, in turn, contributes to rural poverty reduction by both providing new employment opportunities for the poor and by raising agricultural wages via agricultural labor market tightening (see also World Bank, 2011). 374 THE WORLD BANK ECONOMIC REVIEW T A B L E 7 : Spatial Price Indices across City Size Categories in Urban Brazil Laspeyres Price Indices Based on the Cost of Food and Housing Region City Size Category Laspeyres Price Index North Large ( . 500,000) 0.94 Medium (100,000-500,000) 0.75 Small ( , 100,000) 0.68 North-East Large ( . 500,000) 0.66 Medium (100,000-500,000) 0.72 Small ( , 100,000) 0.60 South-East Large ( . 500,000) 0.55 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Medium (100,000-500,000) 0.49 Small ( , 100,000) 0.84 ˜ o Paulo Sa 1.00 South Large ( . 500,000) 0.76 Medium (100,000-500,000) 0.65 Small ( , 100,000) 0.62 Center-West Large ( . 500,000) 0.86 Medium (100,000-500,000) 0.80 Small ( , 100,000) 0.64 ¸ amentos Familiares. Source: Authors’ calculations based on the POF: Pesquisa de Orc POF survey on food items, as well as a hedonic model of rent. We re-apply the method here, but focus solely on urban areas and construct a Laspeyres price index that captures price differences across city-size categories (and regions). Because of the limited sample size of even the unusually large POF household survey we employ a three-way city size breakdown, distinguishing between cities larger than 500,000 persons, large towns with a population between 100,000 and 500,000, and towns with fewer than 100,000 inhabitants. For reasons of data availability our index captures only food and housing price differentials.24 The reference basket employed in our Laspeyres index is based on the consump- tion patterns of the second quintile of the national urban per capita consumption distribution. Table 7 presents our Laspeyres spatial price index based on the cost of food and housing in urban Brazil. Relative to the reference region of metropolitan Sa ˜o Paulo, the cost of living in other regions and city size categories of Brazil is gener- ally lower. In the regions of the North, South and Center West there is evidence of lower cost of living as towns become smaller. This pattern is less clear-cut in the North East, where the cost of living appears particularly high in large towns, and the South East, where those living in small towns appear to face the highest cost of living in the region (although still well below metropolitan Sa˜o Paulo). 24. As noted above, transport costs are also likely to vary substantially across city size categories. Among the poor, these costs typically account for a lower share of expenditures than food and housing. Nevertheless, the sensitivity of our results to spatial variations in the cost of living that include transport costs remains an issue for future work. ´ , Ferreira and Lanjouw Ferre 375 T A B L E 8 : Poverty measures for different city sizes in Brazil Checking for Robustness to Cost of Living Differences Population share1 FGT0 (nominal expenditure) FGT0 (real expenditure) Urban 0.83 0.19 0.18 XL 0.22 0.09 0.06 L 0.07 0.17 0.10 M 0.24 0.15 0.10 S 0.01 0.19 0.11 XS 0.28 0.30 0.19 XL: . 1,000, L: 500-1,000, M: 100-500, S: 50-100, XS: , 50 (‘000 inhabitants). Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ analysis based on data described in the text. We next take the price indices reported in Table 7 and apply them to the small-area based estimates of per capita consumption for each household in the population census. We then re-calculate poverty rates across region and city- size categories. Does this cost of living adjustment overturn our conclusion that urban poverty in smaller towns is significantly higher than in the large and metropolitan areas? Table 8 indicates that it does attenuate the “gradient” between poverty and city size somewhat, but is far from sufficient to negate or overturn our broad finding. In Brazil, it remains the case that the incidence of poverty in the smallest towns is roughly three times higher than in Metropolitan centers. In a third robustness check we investigate whether the finding of a negative gradient between poverty and city-size is somehow an artifact of the focus, in this paper, on income poverty as opposed to a broader conceptualization of deprivation. While the broad pattern of higher poverty and lower access to ser- vices in small towns was found to be quite robust across our eight countries, an important potential caveat to this assessment concerns health outcomes. It has been suggested in the literature that health outcome indicators in large cities in the developing world may lag behind those in smaller towns. For example, Chattopadhyay and Roy (2005) demonstrate that a variety of indica- tors of child mortality are more pronounced in the large cities of India than in towns and medium sized cities. This study finds that while infant mortality amongst the wealthiest classes in large cities are particularly low, infant mortality rates amongst the poorest classes are quite pronounced – and indeed are higher than amongst the poorer segments in small and medium sized towns. These are suggestive findings and may be related to the particularly unhealthy living conditions in over-crowded slum areas of large cities. However, evidence on health outcomes across city sizes categories remains scarce and there does not appear to be a broad consensus in the literature on the relatively higher health risks in large cities. For example, Kapadia-Kundu and Kanitkar (2002) argue, also with reference to India, that urban public health services generally place greater emphasis on mega-cities and metro- centers, to the relative neglect of smaller cities and towns. 376 THE WORLD BANK ECONOMIC REVIEW We probe this concern by examining the gradient of child anthropometric outcomes across city size categories in Mexico. We draw on a small area esti- mation effort undertaken by Rasco ´ n (2010) that parallels the work reported in preceding sections, but focuses on anthropometric outcomes rather than income poverty. Rasco ´ n combines the Mexican National Survey of Health and Nutrition 2006 with the Second Count of the Population and Dwellings 2005 in order to apply a variant of the Elbers et al. (2003) small area estimation procedure to the incidence of stunting and underweight amongst children aged 5 and below in Mexico.25 Lanjouw and Rasco ´ n (2010) examine the correlation of child health outcomes in urban areas with city size. Table 9 summarizes Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 their results and documents that the incidence of low height for age (stunting) and low weight for age (underweight) amongst children displays a similar gradient across city sizes as we have seen for income poverty. In Mexican cities that are larger than 500,000 inhabitants, the incidence of stunting and of underweight amongst children is 9%. In the case of stunting the incidence rises monotonically as cities decline in size. Amongst the smallest cities (of less than 10,000 inhabitants) the incidence is as high as 16%. In the case of under- weight, the incidence also rises, but less markedly: from 9% in the largest cities to 11% in the small cities. The higher incidence of child malnutrition in small towns also translates into more malnourished children: 27% of stunted chil- dren in urban areas are found in the largest cities, while 29% are in towns with less than 15,000 inhabitants. Similarly, 27% of underweight urban chil- dren reside in the largest cities, but 30% reside in the smallest towns. In a final robustness check, we ask whether our observed gradient between poverty and city-size is driven by our reliance on small area estimates of poverty for each city rather than direct measures of poverty for such local- ities.26 We have already noted that direct measures of poverty for individual towns and cities are not generally available in developing country settings. The household surveys that underpin poverty analysis in these countries do not generally cover sufficiently large samples to permit poverty measurement at this detailed level. As was described in Section 2, the small area estimation procedure applied in the present paper combines household survey with unit-record population census data in an effort to circumvent this small sample problem. The approach takes advantage of the full population coverage of the 25. Fujii (forthcoming) adapts the Elbers et al. procedure for the estimation of anthropometric outcomes and applies this methodology to Cambodia. Rasco ´ n adapts this procedure further to apply it to Mexican data. 26. Tarozzi and Deaton (2007) have recently expressed a concern that the small area estimation procedure employed by ELL (2002, 2003) may overstate the precision of local level poverty estimates. They base their argument on Monte Carlo simulation results. (See also Molina and Rao, 2010) Elbers, Lanjouw and Leite (2008) examine this issue with data for the state of Minas Gerais in Brazil, and find little evidence in that specific setting for concern. It remains true, though, that the ELL procedure estimates poverty, rather than directly measuring it, and as such there is interest in assessing whether the findings reported in this paper would also hold had poverty been directly measured. ´ , Ferreira and Lanjouw Ferre 377 T A B L E 9 : Child malnutrition estimates for different city sizes in Mexico Small area estimates of malnutrition amongst children under 5 in urban areas Stunting Underweight Share of Share of Share of Share of Locality size Urban National Urban National (inhabitants) Incidence Population Population Incidence Population Population L 0.09 0.27 0.15 0.09 0.27 0.18 M 0.11 0.44 0.24 0.09 0.43 0.28 S 0.16 0.29 0.15 0.11 0.30 0.19 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 L: . 500, M: 15-500, S: 2.5 – 15 (’000 inhabitants). Source: Lanjouw and Rasco ´ n (2010). population census and then applies statistical techniques to insert into the census an indicator of per capita expenditure or income for each household. This is necessary because in most developing (and developed) countries, the population census fails to collect detailed income or expenditure data. India offers an opportunity to probe the contention that our findings are merely an artifact of the methods we have employed. The Indian National Sample Survey Organization (NSSO) fields a very large sample survey every five years with a sample size that is sufficiently large to permit a breakdown of urban areas into city-size categories.27 Table 10 draws on a World Bank study (World Bank, 2011) to illustrate that at the national level for the years 1983, 1993 and 2004/5, National Sample Survey data show a clear gradient in poverty by city size. This gradient holds both at the national level, as well as at the level of individual states.28 A recent study applies the small area estimation methodology used here to estimate poverty at the local level in three states of India in 2004/5 (Gangopadhyay et al., 2010). The study confirms that in West Bengal, Orissa and Andhra Pradesh the poverty-city size gradient observed from NSS data also emerges from estimates derived out of the small-area esti- mation procedure (Table 11). Thus, at least in India, the finding of an inverse poverty-city size gradient is robust to alternative empirical methods. This pro- vides some support to the claim that the findings reported in preceding sections are not driven by our reliance on small-area estimation techniques. 27. Every five years the NSS fields a “thick round” with a sample size of around 120,000 households, The “thin rounds” fielded in the other years have sample sizes of around 30-40,000 households. 28. World Bank (2011) also shows that the pattern of differential per capita access to public services across city size categories is skewed in India, with small towns faring more poorly than large cities. 378 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 0 : Poverty in India’s Small Towns Exceeds Poverty in the Large Cities: Direct Evidence from the NSS. 1983 1993-94 2004-05 Rural 0.465 0.368 0.281 Urban: 0.423 0.328 0.258 Small towns 0.497 0.434 0.300 Medium towns 0.423 0.315 Large towns 0.290 0.202 0.147 Notes: Poverty rates based on NSS 1983, 1993 and 2004/5 surveys using Uniform Reference Period consumption and official poverty lines. Small , 50K, Medium 50K-1m, Large . ¼ 1m. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: World Bank (2011). T A B L E 1 1 : Small area estimates reveal high poverty in small towns in three Indian states West Bengal Orissa Andhra Pradesh No. Share Share No. Share Share No. Share Share City of of of of of of of of of Size towns Pop Poor FGT0 towns Pop Poor FGT0 towns Pop Poor FGT0 XL 1 0.20 0.08 0.05 - - - - 1 0.18 17 0.23 L 1 0.05 0.04 0.12 2 0.21 0.20 0.34 3 0.13 7 0.14 M 54 0.48 0.46 0.13 6 0.22 0.19 0.31 37 0.39 37 0.24 S 28 0.09 0.12 0.17 15 0.19 0.19 0.36 40 0.15 20 0.33 XS 298 0.18 0.31 0.23 121 0.38 0.42 0.39 104 0.15 18 0.31 Note: XL . 1m; L: 500K-1m; M: 100K-500K; S: 50K-100K; XS , 50K. Source: Gangophadyay et al. (2010) and World Bank (2010). VII. CONCLUDING REMARKS Using highly disaggregated poverty map data from eight countries drawn from all six regions of the developing world, we have shown evidence of a common – although not universal – inverse relationship between poverty and city size. In all countries in our sample, poverty is both more widespread (higher FGT(0)) and deeper (higher FGT(1)) in very small and small towns (those with a population below 100,000) than in large or very large cities (those with a population greater than 0.5 million). Metropolitan poverty, in particular, is considerably lower than poverty in other urban areas in all countries in our sample, except for Kenya and Mexico. Dominance analysis of cumulative distribution functions indicates that the basic pattern is generally robust to the choice of poverty line. Neither is it true that, because of sheer population size, most poor people in these countries live in large cities. In fact, in all eight countries, a majority of the urban poor lives in medium, small or very small towns. In four of them ´ , Ferreira and Lanjouw Ferre 379 (Albania, Brazil, Sri Lanka and Thailand), a majority of the urban poor lives in towns smaller than 100,000 people. The greater incidence and severity of consumption poverty in smaller towns is compounded by similarly greater deprivation in terms of access to basic infra- structure services, such as electricity, heating gas, sewerage and solid waste disposal. This pattern is not absolute. It does vary by type of service and across countries. Access rates seldom increase strictly monotonically with city size, but they do generally increase, so that for most services and in most countries, large cities and metropolitan areas have higher coverage rates than smaller towns. Finally, we have also shown for one particular country – Morocco – that Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 inequality within large cities is not driven by a severe dichotomy between slum dwellers and others. The notion of a single cleavage between slum residents and well-to-do burghers as the driver of urban inequality in the developing world appears to be unsubstantiated – at least in this case. Perhaps more important than the highly visible inequalities within our large cities are the less obvious differ- ences between them and smaller urban settlements. In countries like those studied here (with the possible exception of Kenya), poverty is greater and deeper in smaller towns, in both income and (at least some) non-income dimensions. While the findings reported in this paper are suggestive, we have also pointed to a number of caveats that merit further attention in order to gauge the robustness of the poverty - city size gradient, as well as its applicability in other developing countries. To begin with, while our household samples are representative of the populations within countries, our sample of countries is not statistically representative of the developing world as a whole. In a number of cases, we have been able to perform tentative robustness checks for a subset of these caveats. First, we have indicated that the criteria used to define urban areas vary across countries. It is not uncommon to observe large numbers of small towns designated as urban on the basis of some administrative criterion rather than one linked to population characteristics (e.g. population size, density or economic diversification). We have found, in the case of Brazil, that the inverse relationship between city size and poverty is robust to an alternative, non-administrative definition of urban areas. But it is important to probe more generally whether the gradient we observe in other countries might be a construct of the way cities are defined. Second, differences in cost of living between cities of different sizes can be significant. Failure to correct for such differences could generate a misleading sense of higher living standards in larger cities. Again in Brazil, we have found that adjusting for cost of living variation attenuates, but does not overturn, our observed city size-poverty gradient. Third, we have noted a contrasting literature suggesting that health outcomes might be worse in large cities than small towns. We have explored this conjecture by examining child nutritional outcomes across Mexican towns and cities, and have found that in this context a gradient for city size-child health mirrors, rather than offsets, the city size-poverty gradient uncovered in this paper. The general 380 THE WORLD BANK ECONOMIC REVIEW applicability of this finding also merits further attention.29 Finally, we acknowledge that our findings in this paper are based on small area estimates of poverty that combine survey-based prediction models with population census data. It is important to probe the findings of a city size-poverty gradient also with directly mea- sured poverty data for different cities. We have illustrated that, in the case of India, direct measures and small area estimates yield the same conclusion. It is extremely difficult to find survey data that is representative at sufficient levels of disaggregation to measure poverty directly in small towns – which is why we have used poverty mapping data. Nevertheless, to the extent that additional comparisons with direct survey-based measures are possible elsewhere, they are definitely of interest. A full Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 research agenda also presents itself with respect to uncovering possible reasons for the city size-poverty gradient observed in our sample of countries. We have, some- what provocatively, postulated that higher poverty rates in small towns might be the consequence of a “metropolitan bias” among policy makers, resulting in a dispro- portionate allocation of resources to large cities. Such a bias would be consistent with the evidence we have presented of differential per-capita availability of a variety of infrastructure and public services across the city-size spectrum. However, there are many other plausible explanations, including the possibility that the cost of infra- structure provision may be lower in larger towns and cities. More generally, we have noted in Figures 1 and 2, the considerable heterogen- eity in poverty outcomes amongst towns and cities within all of the size- categories we have considered. Thus, even amongst small towns there are some with low poverty and others with high poverty rates. Important questions arise as to what explains this variation. The “new” economic geography literature and also the recent World Development Report, entitled “Reshaping Economic Geography” devote considerable attention to the mechanisms through which concentration of population and economic activities can generate various kinds of externalities (for example, Krugman 1999, Henderson, Shalizi and Venables, 2001, World Bank, 2009). It is possible that the inverse association between urban poverty and city size reflects primarily urban agglomeration effects. An additional possibility is that the location of towns matters: small towns located near major urban centers may experience low poverty rates while those in remote areas are poorer. There may also be a tendency for towns located near major metropolitan areas to tend to be larger in size.30 Although beyond 29. A potentially interesting exercise is to directly estimate infant mortality rates from census data and to examine differentials of this welfare outcome across the urban city-size spectrum. We leave this exercise to future work. 30. An examination of these questions in the Indian state of West Bengal reveals that when towns are split into three groups – within a 100-kilometer radius of Kolkata, in a radius of 100-200 kilometers, and more than 200 kilometers – poverty sharply rises with distance from Kolkata when the towns are within a 100 kilometer radius (World Bank 2011). The relationship is weaker in the second group, and completely absent in the third, most distant, group. Thus in the first group the agglomeration effect that really matters is the one generated by Kolkata. However, in the second group, and even more strongly so in the third group, a separate agglomeration effect ( proxied by the town’s own size) is discernible. ´ , Ferreira and Lanjouw Ferre 381 the scope of this paper, it is clear that a great deal of additional work is needed to better understand how the empirical patterns we have uncovered come about, and how they might best be addressed. What is also clear, though, is that wherever a clear city size-poverty gradient holds, it should be acknowl- edged and confronted in any strategy for urban poverty reduction. REFERENCES Bates, Robert H. 1981. Markets and States in Tropical Africa. Berkeley, C.A.: University of California Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Press. Chattopadhyay, A., and T.K. Roy. 2005. “Are urban poor doing better than their rural counterpart in India? A study of fertility, family planning and health.” Demography India 34(2):299–312. Christiaensen, Luc, Lionel Demery, and Jesper Kuhl. 2011. “The (evolving) role of agriculture in poverty reduction – An empirical perspective”, Journal of Development Economics 96: 239– 254. Demombynes, Gabriel, and Berk O ¨ zler. 2005. “Crime and Local Inequality in South Africa.” Journal of Development Economics 76 (2):265– 92. Demombynes, Gabriel, Chris Elbers, Jean O. Lanjouw, and Peter Lanjouw (2007) “How Good a Map? Putting Small Area Estimation to the Test” Policy Research Working Paper No. 4155, The World Bank. Department of Census and Statistics, Sri Lanka. 2005. A Poverty Map for Sri Lanka: Findings and Lessons (Colombo: Department of Census and Statistics). Elbers, Chris, Jean O. Lanjouw, and Peter Lanjouw. 2002. “Micro-Level Estimation of Welfare”, Policy Research Working Paper No. 2911, The World Bank. ———. 2003. “Micro-level Estimation of Poverty and Inequality”, Econometrica, 71 (1), pp. 355–364. ¨ zler, and Kenneth Simler. 2004. “On the Unequal Elbers, Chris, Peter Lanjouw, Johan Mistiaen, Berk O Inequality of Poor Communities.” World Bank Economic Review 18(3):401– 21. Elbers, Chris, Peter Lanjouw, and Phillippe Leite. 2008. ‘Brazil within Brazil: Testing the Poverty Mapping Methodology in Minas Gerais’ World Bank Policy Research Working Paper No. 4513, The World Bank. Ferreira, Francisco, H.G., Peter Lanjouw, and Marcelo Neri. 2003. “A Robust Poverty Profile for Brazil Using Multiple Data Sources”, Revista Brasileira de Economia, 57 (1), pp. 59– 92. Foster, James, Joel Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty Measures”, Econometrica 52 (3): 761– 766. Fujii, Tomoko. Forthcoming. “Micro-level Estimation of Child Undernutrition Indicators in Cambodia”, World Bank Economic Review. Gangopadhyay, S., P. Lanjouw, T. Vishwanath, and N. Yoshida 2010. Identifying Pockets of Poverty: Insights from Poverty Mapping Experiments in Andhra Pradesh, Orissa and West Bengal, forthcom- ing Indian Journal of Human Development. Henderseon, J.V., Z Shalii, and A.J. Venables. 2001. “Geography and Development” Journal of Economic Geography 1:81– 105 Haut Commissariat au Plan, Maroc. 2005. Pauvrete ´ veloppement Humain et De ´ , De ´ veloppement Social au Maroc: Donne ´ es Cartographiques et Statistiques d’apre ´ ne ` s le Recensement Ge ´ ral de la Population et de l’Habitat (Rabat: Haut Commissariat au Plan). Healy, A., and S. Jitsuchon 2007. ‘Finding the Poor in Thailand’ Journal of Asian Economies 18(5), 739–759. Instituto Brasileiro de Geografia e Estatı ´stica, IBGE. 2003. ‘Mapa de Pobreza e Desigualdade Municı ´pios Brasileiros’ DVD ROM available from IBGE at: http://www.ibge.gov.br/english/ presidencia/noticias/noticia_visualiza.php?id_noticia=1293&id_pagina=1 382 THE WORLD BANK ECONOMIC REVIEW INSTAT (2004). Poverty and Inequality Mapping in Albania. (Tirana, INSTAT). Kapadia-Kundu, N., and T. Kanitkar. 2002. “Primary Healthcare in Urban Slums”. Economic and Political Weekly, December 21. Kenya Central Bureau of Statistics. 2003. Geographic Dimensions of Well-Being in Kenya (Regal Press Kenya, Nairobi). Krugman, Paul. 1999. ‘The Role of Geography in Development’ International Regional Science Review, 22(2):142–61. Lanjouw, Peter, and Jean O. Lanjouw. 2001. ‘The Rural Non-Farm Sector: Issues and Evidence from Developing Countries’, Agricultural Economics 26 (2001) 1–23. Lanjouw, Peter, and Rinku Murgai. 2009. ‘Agricultural Wage Labour, Non-Farm Employment and Poverty in Rural India’, Agricultural Economics 40 (2009) 243 –263. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ´ n. 2010. ‘Town Size and Malnutrition in Mexico’, mimeo, DECRG, Lanjouw, Peter, and Erika Rasco the World Bank. Lipton, Michael. 1977. Why Poor People Stay Poor: Urban Bias and World Development. London: Temple Smith. ´ pez-Calva, L., A. Melendez, E. Rasco Lo ´ n, L. Rodrı ´ kely (2005) ‘Poniendo ´guez-Chammusy, and M. Sze al Ingreso de los Hogares en el Mapa de Me ´ xico’ Working Paper EGAP-2005– 04, Tecnolo ´ gico de Monterrey, Campus Ciudad de Mexico. Luttmer, Erzo F.P. 2005. “Neighbors As Negatives: Relative Earnings And Well-Being,” Quarterly Journal of Economics, V120 (3,Aug), 963– 1002. Molina, I., and J.N.K. Rao (2010) ‘Small Areas Estimation of Poverty Indicators’ The Canadian Journal of Statistics 38 (3): 369–385 Prebisch, Raul. 1950. “The Economic Development of Latin America and its Principal Problems”, CEPAL. Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2007. “New Evidence on the Urbanization of Global Poverty”. Population and Development Review 33(4), 667–701. ´ n, Erika. 2010. “Cracking Malnutrition: an Resolved Challenge for Mexico?: Micro-Level Rasco Estimation of Child Malnutrition Indicators”, mimeo, Development Economics Research Group, the World Bank. Rosenstein-Rodan, Paul. 1943. “Problems of Industrialization in Eastern and Southeastern Europe”, Economic Journal, 53, pp. 202–12. Tarozzi, A., and A. Deaton 2009. ‘Using Census and Survey Data to Estimate Poverty and Inequality for Small Areas’ Review of Economics and Statistics 91(4) 773 –792. Thomas, T. (2008) ‘Is There Evidence of an Urban-Rural Spillover? An Empirical Analysis Using Spatially Explicit Data from Three Countries’, mimeo, Development Economics Research Group, the World Bank. World Bank (2007) Measuring Poverty Using Household Consumption Brazil Poverty Assessment, Poverty Reduction and Economic Management Network, The World Bank. ——— (2009) World Development Report: Reshaping Economic Geography (Washington D.C.: World Bank). ——— (2011) Perspectives on Poverty in India: Stylized Facts from Survey Data, India Poverty Assessment, Poverty Reduction and Economic Management Network, the World Bank. Impact of SMS-Based Agricultural Information on Indian Farmers Marcel Fafchamps and Bart Minten This study estimates the benefits that Indian farmers derive from market and weather information delivered to their mobile phones by a commercial service called Reuters Market Light (RML). We conduct a controlled randomized experiment in 100 villages Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 of Maharashtra. Treated farmers associate RML information with a number of deci- sions they have made, and we find some evidence that treatment affected spatial arbi- trage and crop grading. But the magnitude of these effects is small. We find no statistically significant average effect of treatment on the price received by farmers, crop value-added, crop losses resulting from rainstorms, or the likelihood of changing crop varieties and cultivation practices. Although disappointing, these results are in line with the market take-up rate of the RML service in the study districts, which shows small numbers of clients in aggregate and a relative stagnation in take-up over the study period. JEL codes: O13, Q11, Q13. The purpose of this study is to ascertain whether agricultural information dis- tributed through mobile phones generates economic benefits to farmers. We im- plement a randomized controlled trial of a commercial service entitled Reuters Market Light (RML) offered by the largest and best-established private pro- vider of agricultural price information in India at the time of the experiment. Operating in Maharashtra and other Indian states, RML distributes price, weather, and crop advisory information through SMS messages. We offered a one-year free subscription to RML to a random sample of farmers to test whether they obtain higher prices for their agricultural output. Marcel Fafchamps is a professor at Oxford University, United Kingdom; his email address is marcel. fafchamps@economics.ox.ac.uk. Bart Minten is a senior research fellow at the International Food Policy Research Institute; his email address is b.minten@cgiar.org. The authors thank Dilip Mookherjee, Shawn Cole, Sujata Visaria, three anonymous referees, and the editor for discussion and suggestions. This research would not have been possible without the full cooperation of Thomson Reuters India. The authors are particularly grateful to Raj Bhandari, Rantej Singh, Ranjit Pawar, and Amit Mehra for their constant support, to Alastair Sussock, Sudhansu Behera, Gaurav Puntambekar, and Paresh Kumar for their research assistance in the field, and to Grahame Dixie for his encouragements, suggestions, and comments. Funding for this research was provided by the World Bank and Thomson Reuters. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 383 –414 doi:10.1093/wber/lhr056 Advance Access Publication February 27, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 383 384 THE WORLD BANK ECONOMIC REVIEW We also investigate the channels through which the price improvement comes: better arbitrage across space and time; ability to bargain with traders; and increased awareness about quality premium leading to better agricultural practices and postharvest handling. We present simple models of the first two channels. Both models make testable predictions about farmer behavior in re- sponse to market price information. We test for the presence of these necessary channels. Given that RML circulates weather and crop advisory information, we also examine whether profits increase thanks to better crop management, reduced losses, or improved quality. There is a large interest and takeoff of similar private and public programs Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 in India (Mittal, Gandhi, and Tripathi 2010) and elsewhere (Staatz, Kizito, Weber, and Dembe ´ 2011), indicating that it is important to understand the ´ le impact of these interventions.1 Based on this, we expected to find a large and significant effect of RML on treated farmers. These expectations find less support than anticipated in our results. Many treated farmers state that they use the RML information, and they are less likely to sell at the farm-gate and more likely to change the market at which they sell their output. These findings are consistent with the idea that treated farmers seek arbitrage gains from RML information. We do not, however, find any significant difference in price between beneficiaries of the RML service and non-beneficiaries. This result is robust to the choice of estimator and method- ology. There is some evidence of heterogeneous effects: Treated farmers who are young—and presumably less experienced—receive a higher price in some regressions, but the effect is not particularly robust. Why we do not find a stronger effect of RML on the prices received by farmers is unclear. Over the study period there were important changes in prices, which are part of a general phenomenon of food price inflation in India (see, for example, World Bank 2010). Although rapidly changing prices should in principle make market information more valuable, it is conceivable that the magnitude of the change blunted our capacity to identify a significant price dif- ference, given how variable prices are. Point estimates of the treatment effect on price received are, however, extremely small and sometimes negative. We also find no significant evidence of an effect on transaction costs, net price, rev- enues, and value-added. In other channels by which RML may affect farmers, we find a statistically significant but small increase of the likelihood of grading or sorting the crop. This is especially true for younger farmers, possibly explaining why they achieve better prices. RML shows prices by grade to farmers and might have helped to inform them about the benefit from grading. With respect to other 1. In India, apart from RML, other initiatives include the Indian Farmers Fertilisers Cooperative Limited (IFFCO) Kisan Sanchar Limited (IKSL), a partnership between Bharti Airtel and IFFCO, as well as the fisher friend program by Qualcomm and Tata Teleservices, in partnership with the MS Swaminathan Research Foundation. In western Africa, Manobi and Esoko, private ICT providers, have developed a number of SMS applications to facilitate agricultural marketing there. Fafchamps and Minten 385 RML information, we find no systematic change in behavior due to weather in- formation and no change on the crop varieties grown or cultivation practices. Farmers who changed variety or cultivation practice do, however, list RML as a significant source of information in their decision. To summarize, RML bene- fits appear minimal for most farmers, even though RML is associated with sig- nificant changes in where farmers sell their crops. There is some evidence that young farmers benefit more from the service, but the effect is not robust. I. DESCRIPTION OF THE INTERVENTION AND EXP ERIMENTAL DES IG N Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 The Indian branch of Thomson-Reuters distributes agricultural information to farmers through a service called Reuters Market Light (RML). Subscribers are provided with SMS messages in English or the local language of their choice— in total 75 to 100 SMS per month. Subscribers are offered a menu of three agricultural markets and three crops to choose from. This menu is based on market research of farmers’ information needs and a pilot marketing program carried out in 2006-07. Prior to the experiment, Reuters gathered encouraging anecdotal information from farmers regarding the usefulness of RML. At the time of the experiment, Reuters had about 25,000 RML subscribers in Maharashtra. The RML content included market information, weather fore- cast, crop advisory tips, and commodity news. At the time of the study the price for the service was about $1.50 per month. Thomson Reuters is a trans- national corporation specializing in market information services, its core busi- ness. The corporation has a long experience collecting and selling market information and a reputation for accuracy. Although Thomson Reuters did not disclose how it obtains market price information, we have no reason to suspect that the provided information is inaccurate. Doubts about data accuracy would be detrimental to the firm’s reputation and would hurt its effort to market RML across India. A randomized controlled trial (RCT) was organized by the authors to test the effect of RML on the price received by farmers. The RCT was conducted in close collaboration with Thomson Reuters. Five crops were selected as the focus for the study: tomatoes, pomegranates, onions, wheat, and soybeans. In Maharashtra these crops are all grown by smallholders primarily for sale. Whereas wheat and soybeans are storable, the other three crops are not and should be subject to greater price uncertainty due to short-term fluctuations in supply and demand (Aker and Fafchamps 2011). We expect market informa- tion to be more relevant for perishable crops. Soybeans have long been grown commercially in Maharashtra, but tomatoes, onions, and pomegranates are more recent commercial crops. We expect farmers to be more knowledgeable about well-established market crops—such as wheat and soybeans—than about crops whose commercial exploitation is more recent—such as tomatoes, onions, and pomegranates. Finally, pomegranate is a tree crop that is sensitive to unusual weather and requires pesticide application. We expect the benefit 386 THE WORLD BANK ECONOMIC REVIEW from weather information and crop advisories to be more beneficial to pom- egranate farmers. For each of the five crops, we selected one district where the crop is widely grown by small farmers for sale: Pune for tomatoes, Nashik for pomegranates, Ahmadnagar for onions, Dhule for wheat, and Latur for soya. All five districts are located in the central region of Maharashtra; we avoided the eastern part of the state where sporadic Maoist activity has been reported, and the western part of the state which is less suitable for commercial agriculture. A total of 100 villages were selected for the study, 20 in each of the 5 dis- tricts. The villages were chosen in consultation with Thomson Reuters to Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ensure they were located in areas not previously targeted by RML marketing campaigns. Ten farmers were then selected from each village, yielding a total intended sample size of 1,000 farmers.2 Two treatment regimes were implemented. In the first regime—treatment 1—all 10 farmers in the selected village were offered free RML. In the second regime—treatment 2—three farmers randomly selected among the village sample were offered RML. The purpose of treatment 2 is to test whether the treatment of some farmers benefit others as well. Farmers who are not signed up are used to evaluate the externalities generated if farmers share RML information. Bruhn and McKenzie (2009) show that, in RCTs, stratification improves effi- ciency. Randomization of treatment across villages was thus implemented by constructing, in each district, triplets of villages that are as similar as possible along a number of dimensions that are likely to affect the impact of the treat- ment.3 A description of the process is provided in the supplemental appendix 1, available at http://wber.oxfordjournals.org/. We also offered RML to randomly selected extension agents covering half of the treatment 1 and treatment 2 villages. In principle, extension agents could disseminate the relevant information they receive, making it unnecessary to dis- tribute it to individual farmers. Whether they do so in practice is unclear, given that extension agents visit villages infrequently. The object of the RCT is to estimate the impact of RML on farmers who voluntarily sign up for the service because they benefit from it. In villages tar- geted by RML, only a small proportion of farmers sign up. They tend to be larger farmers with a strong commercial orientation in RML crops. There is no 2. This sample size was determined as follows. The primary channel through which we expect SMS information to affect welfare is the price received by producers. We therefore want a sample size large enough to test whether SMS information raises the price received by farmers. Goyal (2010) presents results suggesting that price information raises the price received by Indian farmers by 1.6 percent on average. Based on this estimate and its standard error, a simple power calculation indicates that a total sample size of 500 farmers should be sufficient to identify a 1.6 percent effect at a 5 percent significance level. To protect against loss of power due to clustering, we double the sample size to 1,000. We did not have sufficient information to do a proper correction of our power calculations for clustering. 3. More precisely, 6 triplets and one pair. Fafchamps and Minten 387 point estimating the effect of RML on people unlikely to benefit from it. For this reason, participants are limited to farmers growing the district-specific selected crop for sale. These farmers have a large enough marketed surplus to amortize the cost of information gathering and enough experience with the crop to benefit from agricultural information. It remains that the farmers to whom we offered RML did not express an initial interest in it and may have dismissed the usefulness of something they did not pay for. Since a mobile phone is required to receive RML messages, we limit partici- pants to farmers with a mobile phone at the time of the baseline survey. We omit farmers who already were RML customers prior to baseline because they Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 could not be used for impact analysis. Because we carefully avoided areas in which RML had already been actively promoted, few farmers were eliminated due to this condition. Further details on the sample and survey design are found in supplemental appendix 2. II. CONCEPTUAL FRAMEWORK Our primary objective is to test whether farmers benefit from the SMS-based market information and, if so, how. Information gathered by Thomson Reuters and conversations with farmers and RML customers in villages of Maharashtra suggest that farmers benefit in several ways. According to this information, timely access to market price information at harvest helps farmers decide where to sell, as in Jensen (2007). It also enables them to negotiate a better price with traders. To illustrate how informed farmers may obtain a higher price through better arbitrage, consider a farmer selecting where to sell his output. There are two possible markets, i and j, with transport costs ti and tj. We assume tj . ti, that is, market j is the more distant market. To focus on arbitrage, we assume that the distribution of producer prices F( p) is the same in both markets. In particu- lar, E[ pi] ¼ E[ pj ] ¼ m. Given this, it is optimal for an uninformed farmer to always ship his output to market i since EU ( pi - tj ) , EU ( pj – ti) for any utility function U(.).4 The average price received by farmers is thus m. A rele- vant special case is when i is selling at the farm-gate to an itinerant buyer and j is selling at the nearest market. In that case the farmer incurs no transaction cost, receives price pi, and does not learn the market price pj. Now suppose that the farmer is given information on prices in i and j. Shipping to i remains optimal if pi – ti ! pj – tj; otherwise, the farmer ships to j. The average farmer price now is:   à À Á E pi  pi À ti ! pj À tj Pr pi À ti ! pj À tj ð1Þ Â  à À Á þ E pj  pi À ti , pj À tj Pr pi À ti , pj À tj ! m 4. Since p — ti . p — tj point wise. There is no role for risk aversion in this model. 388 THE WORLD BANK ECONOMIC REVIEW This equation holds with equality only if pi is always larger than pj — (tj — ti). This arises only if tj – ti is large relative to the variance of prices or if prices in the two markets are strongly correlated. It follows that if price infor- mation allows farmers to arbitrage better across markets, the average farmer price should rise and we should observe farmers now selling in different, more distant markets.5 In the case of farm-gate sales, obtaining information about pj induces farmers to sell at the market if pj . pi þ tj where pi is the price offered by itinerant buyers. There remains the issue of why pi = pj in the first place: If farmers can arbi- trage, so can traders. Let u be the trader shipment cost between i and j. With Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 perfect information, trader arbitrage yields: ð2Þ pj þ u pi pj À u and thus j pi — pj j u. Farmer arbitrage therefore arises whenever tj — ti , u, that is, when farmers have access to cheap transport to markets.6 However, if traders have a comparative advantage in transport—for instance, because they ship larger quantities and benefit from returns to scale—then it is possible that tj — ti . u for most if not all farmers. In this case, Pr( pi - ti , pj – tj ) ¼ 0, which implies that farmers always sell at the closest market or location i, and the average farmer price is m, as in the case without information. Even when price information does not trigger farmer arbitrage, it may facili- tate arbitrage by traders, thereby ensuring that jpi — pjj u (Aker and Fafchamps 2011). If farmers are risk averse, they would benefit from the reduc- tion in the variance of prices,7 irrespective of whether they receive market in- formation pi and pj or not prior to deciding where to ship their output. The second way farmers can benefit from price information is when they sell to traders who are better informed about market prices—for example, when selling at the farm-gate. To illustrate, consider price negotiation between an informed trader who knows the market price realization pi, and an uninformed farmer who only knows the price distribution F( pi). To demonstrate how infor- mation can benefit the farmer, imagine that the farmer mimics an auction system and calls a decreasing sequence of selling prices until the trader accepts it. In a competitive market with many buyers—which makes collusion difficult—the selling price will be pi. In a one-on-one negotiation, as would 5. A similar reasoning applies to intertemporal arbitrage: Uninformed farmers may prefer to sell immediately after harvest, whereas better informed farmers may choose to sell at a later date if the anticipated price is higher. Since RML does not disseminate information about future prices, however, we do not expect an intertemporal arbitrage effect, except perhaps in the immediate vicinity of harvest. Several of the studied crops are perishable; this further limits opportunities for intertemporal arbitrage. 6. For instance, if transport costs per kilometer are identical for farmers and traders, condition tj — ti , u holds generically on a plane, except when the farmer and the two markets are exactly in a straight line. With many farmers distributed randomly on the plane, this has Lebesgue measure zero. 7. Except when they consume much of their output, something that is ruled out here since the empirical analysis focuses on commercial crops Fafchamps and Minten 389 take place at the farm-gate, the buyer correctly anticipates that the farmer will continue calling lower prices below pi. He can thus wait for the farmer to reach his reservation price, which is the value of the farmer’s next best alterna- tive, namely, selling at the nearest market. The expected payoff to an uninformed farmer of selling at the market is EU( pi — ti). Let p ~l ; pi À ti be the market price net of transport cost, and let m ~ ; E½p ~i Š. The farm-gate reservation price of a risk-averse farmer is the price pr ~ À p that solves: i ¼ m ð3Þ Uðm ~l Þ ~ À pÞ ¼ EU ðp Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Using a standard Arrow-Pratt Taylor expansion, we get: ! 1 0 ð4Þ 0 ~ Þ À U ðm ~ Þp % E Uðm Uðm ~ Þ þ U ðm 0 ~ Àp ~ Þðm ~l Þ þ U 0 ðm ~ Àp ~ Þðm ~ l Þ2 2 which we can solve for p: 00 ~Þ 2 1 1 U ðm ð5Þ p% À 0 s ¼ R CV 2 2 U ðm ~Þ 2 where R is the farmer’s coefficient of relative risk aversion, s2 is the variance of the market price, and CV ; s=m ~ is the coefficient of variation of price. It follows that a buyer can always buy from an uninformed farmer at price pr i ¼ m ~ À p. Only if the realized market price pi , pri is the farmer unable to find farm-gate buyers—in which case he must travel to the market and sell at pi , m~ þ ti À p but incur transport cost ti. The average price received by an uninformed farmer is: ðm À ti À pÞPrð pi ! m À ti À pÞ þ E½ pi j pi , m À ti À pŠ ð6Þ Prð pi , m À ti À pÞ m À ti It follows that the larger R—and thus p—is, the lower the average farmer price is. If risk aversion is negatively correlated with wealth, the above predicts that poor uninformed farmers receive a lower average price than non-poor unin- formed farmers. Similarly, the larger CV is—for instance, because the farmer is inexperienced and unsure about the price distribution—the lower the average farm-gate price is. Once we introduce price information, the farmer’s farm-gate reservation price becomes pi - ti and buyers are no longer able to exploit farmers’ risk aver- sion to buy below the market price. The expected price received by farmers is m — ti if they sell at the farm-gate, or m if they sell in the market. Hence, the average price received by informed farmers is unambiguously higher than that of uninformed farmers. The difference is largest when uninformed farmers 390 THE WORLD BANK ECONOMIC REVIEW often sell at the farm-gate. If farmers do not sell at the farm-gate at all, infor- mation has no effect on the average price farmers receive.8 The two models above do not exhaust the possible channels by which price information may affect farmer prices. In India one important possibility is ex- cessive fees collected by commission agents (Minten, Vandeplas, and Swinnen 2012). However, since RML circulates no information about market fees, it is unclear why it should lead to their reduction. Model predictions regarding prices can be summarized as follows. If price information enables farmers to arbitrage across markets, treated farmers should receive a higher price than control farmers, but only if treated farmers Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 start selling in distant markets. Otherwise, we expect no difference between control and treated farmers. The introduction of RML, however, could reduce the variance of prices for everyone through trader arbitrage, as in Aker and Fafchamps (2011). If price information helps farmers negotiate better prices with traders, treated farmers should receive a higher average price only if they were selling at the farm-gate prior to treatment. For these farmers, we expect a stronger treat- ment effect for poor and inexperienced farmers. For farmers who were selling primarily if not exclusively through wholesale markets, we expect no effect of the treatment on price. But treatment may nevertheless induce farmers to sell at the farm-gate for convenience reasons, or if traders have a comparative advan- tage in transporting produce from the farm-gate. RML may benefit farmers in other ways, which we do not model since they are more straightforward. Better knowledge of quality-driven price differentials may induce farmers to upgrade output quality, for instance by grading or treat- ing their crops. Weather information helps with farm operations. In particular, information about the probability of rainfall enables farmers to either delay ( pesticide application) or speed up (harvest) certain farm operations. Information about air moisture is a good predictor of pest infestation and hence of the need to apply pesticide. Crop advisories assist farmers to choose a more appropriate technology (choice of variety, pesticide, and fertilizer). I I I . T E S T I N G S T R AT E GY We now describe how we test the above predictions. Since the data are balanced, we ascertain the effect of RML on outcome indicators by comparing control and treatment in the ex post survey. Formally, let Yi represent an 8. Uninformed farmers would benefit if they could commit to sell at the market. If such a commitment mechanism is unavailable, however, farmers can always be tempted to sell at the farm-gate if offered a price above their reservation price. Of course, a sophisticated but uninformed farmer should infer that if a trader is willing to buy from him at the farm-gate, the market price must be above his reservation price, in which case he should sell at the market. If farmers are sophisticated, we should therefore observe few if any farm-gate sales by uninformed farmers. In this case, providing market information to farmers should make farm-gate sales more common. Fafchamps and Minten 391 outcome indicator—for example, price received—for farmer i. Let Wi ¼ 1 if farmer i was offered a free subscription to the SMS-based market service, and Si ¼ 1 if farmer i signed up for the service. All treated farmers are in treated vil- lages but the converse is not true: only some farmers in treatment 2 villages were offered RML. It is possible for treated farmers not to sign up for the service—that is, for Si ¼ 0 even though Wi ¼ 1. Although control villages were not targeted by RML marketing campaigns, it is also possible for non- targeted farmers to independently sign up—that is, for Si ¼ 1 even though Wi ¼ 0. We are interested in estimating the direct effect of RML on customers, that Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 is, those with Si ¼ 1. Since Si is subject to self-selection and Wi is not, we begin by reporting intent-to-treat estimates that compare control farmers to those who were offered the free subscription. The estimating equation is: ð7Þ Yi ¼ u þ bWi þ ei Next we investigate the effect of receiving the RML subscription. As we will see, the likelihood of signing up is much higher among farmers who received the offer of a free subscription. This means that Wi satisfies the inclusion re- striction and can be used as instrument for Si. We thus estimate an instrumental variable (IV) model of the form: ð8Þ Yi ¼ u þ aSi þ ei where ei is an error term possibly correlated with Si (self-selection effect) but uncorrelated, by design, with the instrument Wi. Provided that there are no defiers, we can interpret IV estimates from equa- tion (8) as local average treatment effects (LATE). Assuming no defiers means that farmers who did not sign up for RML even though they were offered a free subscription would not have signed up for it if they had not been offered a free service. In our setup, this assumption is unproblematic. We can therefore inter- pret a in equation (8) as the effect of RML for a farmer who would be induced to sign up if offered the service for free. This is the IV-LATE approach. Equation (7) can be generalized to investigate heterogeneous effects. Let Xi be a vector of characteristics of farmer i thought to influence the effect of the treatment. We expect larger RML benefits for commercial farmers who empha- sise crops for which RML information is useful. We also expect less experi- enced farmers to more benefit. The estimated model becomes: ð9Þ  Þ þ ei Yit ¼ u þ aWi þ gXi þ hWi ðXi À X where X  denotes the sample mean of Xi. The average treatment effect is given by a, whereas the heterogeneous effects of treatment on a farmer with charac- teristic Xi is a þ hðXi À X Þ. 392 THE WORLD BANK ECONOMIC REVIEW When interpreting models (7), (8), and (9), one must remember that identify- ing the value of information is difficult because the value of information changes with circumstances. In particular, information is useful only when it can be acted upon. Up-to-date price information is most useful around harvest time. Crop advisory and input cost information are most useful at planting time. For information to be useful it must be provided in a timely manner. How valuable information is depends on the context: because information is not useful in one year does not imply that it is never useful. Second, information circulates through channels other than RML — farmers visit markets and talk to each other and to commission agents. For models (7), Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 (8), and (9) to identify the impact of RML information, the circulation of in- formation among farmers must not be so rapid and widespread that control farmers benefit from it as well. For this reason we regard the village as the most appropriate treatment unit, because information exchange is more likely among neighbors. We cannot, however, entirely rule out the possibility of spil- lovers across villages. Third, price information may benefit farmers by improving their bargaining power with traders and commission agents. Since the latter cannot easily distin- guish between RML and non-RML farmers, it is possible that they adapt their behavior toward all farmers, for instance, by making better price offers. If this is the case, control farmers may benefit as much as treated farmers from the RML service. There is little we can do to protect against this form of contamin- ation, except check informally how agricultural wholesale prices change over time as farmers become better informed.9 I V. T H E C O N T E X T AND D ATA Take-up of RML by Maharashtra farmers is a revealed preference measure of the benefits from the service. We report in Table 1 the number of agricultural holdings in each study districts (2000/01 agricultural census) and the number of RML subscribers over the study period. RML take-up has varied over time. Take-up increased rapidly in all five districts between 2007, the time at which RML was introduced, and 2009, the time at which our experiment started. Take-up never exceeded 0.5 percent of the total farmer population, however. The table also shows that subscription levels have stabilized in recent years and have even come down in some districts in 2010. It is only in Nashik that we see a large increase in the number of subscribers between 2009 and 2010. This may be explained by Nashik having a nascent grape-growing and wine-making 9. It is also conceivable (albeit unlikely) that RML clients indirectly create a negative externality for nonclients, for instance, because the selling behavior of RML clients indirectly lowers the price received by nonclients, or because it raises the price for local consumers. If this were the case, we would overestimate the effect of RML by comparing RML and non-RML farmers within the same village. This is why we focus our analysis on comparisons across treatment and control villages. Fafchamps and Minten 393 T A B L E 1 . Number of agricultural holdings and RML subscribers in the five districts studied in Maharashtra Number of RML subscribers** Crop followed Number of agricultural District: in survey holdings* 2007 2008 2009 2010 Ahmadnagar onion 916,724 711 1,377 3,763 1,637 Dhule wheat 230,216 108 1,296 1,028 840 Latur soya 305,706 163 914 1,048 826 Nashik pomegranate 591,763 2,176 1,561 3,934 6,514 Pune tomato 667,365 392 653 3,495 781 Total 2,711,774 3,550 5,801 13,268 10,598 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 *: Government of India, Agricultural Census, 2000/01. **: Thompson-Reuters. industry that has been rapidly growing in recent years. Since grapes are grown primarily by large farmers, they are not included in our study. Next we report on contamination and noncompliance. Extensive contamin- ation could indicate that many farmers find RML beneficial and sought it out even though it was not marketed locally. In contrast, extensive noncompliance could suggest that treated farmers did not find the service useful. In Table 2 we compare the experimental design, or intent to treat, in the two upper panels to actual RML usage in the lower panel. The uppermost panel describes the ori- ginal experimental design. This design assumes that 10 farmers would be found in each of the 100 villages selected for the study. The middle panel of Table 2 describes how the experiment was implemented in practice. This represents what in the rest of the paper we call intent to treat. All farmers in treatment 1 villages were offered RML free of charge for one year. In control villages, no farmer was offered RML and no marketing of RML was done by Thompson-Reuters. In treatment 2 villages, a randomly selected subset of farmers (3 out of 10) were offered RML and the others were not. There was some attrition between the baseline and follow-up surveys: of the 1,000 farmers interviewed in the baseline, 933 were revisited in the follow-up survey. There is some difference in attrition between the control and treatment groups, that is, 91 percent versus 96 percent versus 93 percent. To investigate whether there is anything systematic about attrition, we regressed an attrition dummy on household characteristics.10 We find that onion producers (Ahmadnagar district) are more likely to drop out of the experiment, but none of the other variables is statistically significant. Triplet dummies are included as regressors throughout the analysis; they indirectly control for district/target crop. 10. That is, household size, age of household head, education of household head, land owned, total land cultivated of the selected crop in 2009, and target crop/district dummies. 394 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Compliance and contamination All villages Treatment 1 Treatment 2 Control Number RML RML RML RML of villages yes no yes no yes no yes no Intended experimental design Number of households All 100 455 545 350 0 105 245 0 300 Tomato growers 20 91 109 70 0 21 49 0 60 Pomegranate growers 20 91 109 70 0 21 49 0 60 Onion growers 20 91 109 70 0 21 49 0 60 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Wheat growers 20 91 109 70 0 21 49 0 60 Soya growers 20 91 109 70 0 21 49 0 60 Realized design or intent to treat All 100 422 511 325 0 97 239 0 272 Tomato growers 20 84 107 64 0 20 49 0 58 Pomegranate growers 20 89 107 68 0 21 52 0 55 Onion growers 20 88 105 68 0 20 49 0 56 Wheat growers 20 86 102 67 0 19 47 0 55 Soya growers 20 75 90 58 0 17 42 0 48 RML usage (from 2010 survey and phone interview) All households 100 247 686 181 144 56 280 10 262 Tomato growers 20 44 147 35 29 9 60 0 58 Pomegranate growers 20 65 131 42 26 19 54 4 51 Onion growers 20 44 149 36 32 8 61 0 56 Wheat growers 20 48 140 33 34 11 55 4 51 Soya growers 20 46 119 35 23 9 50 2 46 Extension agents: Intended design 30 70 15 20 15 20 0 30 Realized design 20 80 10 25 10 25 0 30 In the next panel we report actual RML usage, as depicted by the 2010 survey and by the ex post phone interview. We note a significant proportion of noncompliers: only 59 percent of those farmers who were offered RML actual- ly used it. Non-usage has various proximate causes. Some subscribers simply refused the service. In the ex post phone interview, respondents were asked the reason for refusal. Some indicated that they believed they would be charged for service later on; others were illiterate households who could not read SMS mes- sages and thus could not use the service anyway. Another reason for non-usage was that subscribers never activated the RML service. To activate it, the sub- scriber had to select three crops and markets; some subscribers never com- pleted the activation sequence. Non-usage was also partly due to changes in phone number or to migration—for example, a household member leaving the farm and taking the phone number with them. The RML service is tied to a specific phone number, so if this phone number is no longer used by the house- hold, the service no longer reaches its intended target. Finally, a number of Fafchamps and Minten 395 Chinese-made phones could not display the Marathi script and households with such phones could not read the RML messages. Although there is vari- ation between them, all these proximate causes indicate a certain lack of inter- est in the service: if RML had been valuable, recipients would have made more effort to secure it—for example, by keeping the SIM card and getting another phone. There is variation in noncompliance across districts: Noncompliance is lowest in Nashik among pomegranate farmers (27 percent). This finding is con- sistent with the high take-up reported in Table 1, and indicates more interest in RML in that district. In contrast, the proportion of noncompliers is close to Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 half among onion, tomato, and wheat growers. While noncompliance is high, contamination is low everywhere: only 3.7 percent of control farmers—10 out of 272 farmers—signed up for RML. This confirms that interest in the service among study farmers is limited. At the bottom of Table 2 we report variation between the intended experimen- tal design and the realized treatment for extension agents. The intent was to offer one year of RML service free of charge to the extension agents serving a random- ly selected subsample of 30 of the 70 treated villages. In practice, we only managed to locate and offer RML to extension agents serving 20 of the treated villages. In order not to introduce contamination, RML was not offered to exten- sion agents serving control villages. This means that we can only measure the additional effect that an informed extension agent may have over and above an individual RML contract (treatment 1 villages) or in addition to treatment of other farmers in the same village (control farmers in treatment 2 villages). In Table 3 we compare control and intent-to-treat farmers in terms of balance. Columns 4 and 5 report the mean value of each variable for the control group and their standard deviation, respectively. Columns 6 and 7 report the coefficient of an intent-to-treat dummy in a regression of each vari- able on triplet fixed effects.11 Reported coefficients suggest good balance on all variables, including area planted to the target crop, marketing, transaction costs, past weather, and past technological innovation. We follow Deaton’s (2009) suggestion not to include unnecessary control variables in the analysis of randomized controlled trials (RCTs), as it may artificially inflate t-values. V. E M P I R I C A L A N A L Y S I S We now turn to the econometric analysis. Unless otherwise stated, all analysis is conducted in terms of intent to treat, that is, the treated are those who were offered a free one-year subscription to the RML, whether or not they used it. We also report local average treatment effect (LATE) results in which we in- strument actual RML usage with random assignment to treatment. We refer to 11. Bruhn and McKenzie (2009) indicate that fixed effects for each stratification cell should be included in all regressions. 396 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Balancedness of treatment versus control in the 2009 baseline data Number Control group Treatment* of Unit observation Mean St. Dev. Coeff. t-value Household characteristics Education level head of household years 911 8.19 4.36 0.243 0.68 Household size number 933 6.43 2.70 0.131 0.61 Share of children in household share 933 0.26 - 2 0.008 2 0.72 Share of elderly in household share 933 0.08 - 0.011 1.38 Age head of household years 922 49.51 12.93 0.409 0.42 Farm experience years 930 26.86 13.92 2 0.379 2 0.39 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Land ownership and cultivation Land owned acres 933 9.62 8.62 0.677 0.64 Land cultivated of tomato in Pune acres 191 1.79 3.01 2 0.154 2 0.40 Land cultivated of pomegranate in Nashik acres 196 3.64 3.50 0.508 0.62 Land cultivated of onions in Ahmadnagar acres 193 2.06 1.87 0.258 0.65 Land cultivated of wheat in Dhule acres 188 5.76 3.95 2 1.073 2 1.65 Land cultivated of soya in Latur acres 165 5.76 4.09 0.841 0.39 Total crop area cultivated acres 933 14.78 13.04 1.340 0.88 Marketing characteristics studied crop Know market price of studied crop: 2 the day before he sold it share 910 0.78 - 0.037 1.13 2 the week before he sold it share 912 0.38 - 0.070 1.58 2 a month before he sold it share 912 0.08 - 0.029 1.33 2 when he planted it share 912 0.06 - 0.020 1.67 For each transaction: 2 Prices obtained in each transaction Rs/kg 1563 13.22 10.20 2 0.149 2 0.44 2 Quantities sold per transaction log(kgs) 1563 7.11 1.57 2 0.067 2 0.75 2 Produce is sold in the village share 1554 0.15 - 2 0.016 2 0.57 2 Head of household made sale share 1561 0.85 - 2 0.021 2 0.58 2 Crop was graded/sorted before sale share 1561 0.70 - 0.036 0.84 2 Produce is sold through commission share 1555 0.40 - 0.032 0.83 agent Number of sale transactions per farmer number 894 1.74 1.19 2 0.001 2 0.01 Transaction costs last transaction Paid for transport of produce share 908 0.88 - 0.013 0.45 Paid for personal transport share 797 0.11 - 0.022 0.90 Sold through commission agent share 905 0.57 - 2 0.036 2 0.80 Weather in 12 months prior to survey Did not incur storm/heavy rainfall share 933 0.53 - 2 0.021 2 0.63 Technology changes in 12 months prior to survey Changed crop varieties share 933 0.34 - 2 0.020 2 0.57 Changed cultivation practices share 933 0.28 - 2 0.004 2 0.11 All variables refer to 2009 data. *village triplet code dummies and intercept included but not reported. Fafchamps and Minten 397 these results as IV or LATE estimates interchangeably. For much of the analysis we use both treatment 1 and treatment 2 farmers to improve efficiency. When using treatment 2 farmers, the intent-to-treat variable is set to 1 if a surveyed farmer in a treatment 2 village was randomly assigned to treatment, and 0 otherwise. All reported standard errors are clustered by village triplet (see ex- perimental design section). RML Usage We begin with RML usage as reported by farmers. In the baseline survey all Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 respondents were asked to list their main sources of information for agricultur- al prices, weather forecast, and advice on agricultural practices. Answers are tabulated in Table A.1 in the supplemental appendix. Own observation/experi- mentation is the main source of information reported by all respondents, fol- lowed by conversations with other farmers. Radio and television are mentioned as a common source of information on the weather, less so for crop prices. RML is not mentioned by anyone. In the top panel of Table 4 we report the average difference in the propor- tion of respondents who mention RML as a source of information in the ex post survey. The average treatment effect on the treated (ATT) is calculated using the nearest neighbor-matching methodology described in Abadie at al. (2004), where matching is performed by triplet dummy. Reassuringly, treated farmers are significantly more likely to mention RML in all six categories. The difference is largest in magnitude for prices and weather, which are the primary focus of the RML service: 24 percent and 23 percent more treated respondents mention RML as a source of information on crop prices and weather forecasts, respectively. LATE-IV estimates are reported in the next panel of Table 4. These estimates are obtained using regression analysis. Dummies are included for village selec- tion triplets. Since contamination is low (3.7 percent) but noncompliance is high (41 percent), we expected instrumented treatment effects to be larger than the intent-to-treat effect reported at the top of the table. This is indeed what happens: We now find that farmers who were induced to use RML as a result of treatment are 46 percent more likely to mention RML as a source of infor- mation on crop prices. The corresponding figures for weather prediction and for input use are 44 percent and 39 percent, respectively. This suggests that RML is seen as a source of information by a large proportion of participating farmers. Yet, the effect is not 100 percent, which means that, since non-users do not list RML, a sizable portion of treated respondents do not list RML as a source of information. In the second part of the table we look for evidence of heterogeneous effects by farmer age and farm size. We estimate regression (3) with triplet fixed effects as suggested by Bruhn and McKenzie (2009). Farmers cultivating a larger area are significantly more likely to mention RML as a source of 398 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Use of RML Use RML as one of the sources of information for: Crop Weather Crops to Cultivation input use post-harvest Whole sample prices prediction plant practices (d) practices Number of 925 931 925 925 918 924 observations Nearest neighbor matching (a) ATT Coeff 0.243 0.231 0.106 0.086 0.200 0.054 z-value 10.600 10.530 6.550 5.390 9.360 3.970 Regression results (b) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 1. IV-LATE Treatment Coeff 0.463 0.439 0.206 0.172 0.386 0.113 t-value 10.460 10.530 5.230 4.710 8.940 3.220 Intercept Coeff 0.007 2 0.021 2 0.008 2 0.001 2 0.044 0.044 t-value 0.840 2 2.530 2 1.000 2 0.150 2 5.080 6.300 2. Heterogeneous effects (c) Treatment Coeff 0.239 0.225 0.107 0.089 0.198 0.057 t-value 8.760 9.330 5.580 4.880 8.140 3.160 Intercept Coeff 2 0.034 2 0.470 2 0.034 2 0.026 2 0.066 0.027 t-value 2 1.950 2 2.770 2 2.450 2 1.870 2 4.320 2.000 Dummy young Coeff 0.024 0.022 0.007 0.005 0.014 2 0.003 head of t-value 1.270 1.670 0.650 0.450 1.520 2 0.300 household Total crop area Coeff 0.001 2 0.000 0.001 0.001 0.000 0.001 cultivated t-value 1.200 2 0.020 1.700 1.890 0.430 1.680 Interaction with treatment Dummy young Coeff 0.045 0.018 0.057 0.019 0.007 0.034 head of t-value 1.040 0.390 1.580 0.530 0.150 1.210 household Total crop area Coeff 0.002 0.004 2 0.000 2 0.001 0.002 2 0.000 cultivated t-value 2.340 4.300 2 0.060 2 0.700 2.110 2 0.380 (a) Matching based on village triplet code dummies (b) Village triplet code dummies included but not reported (c) Mean value substracted from those control variables interacted with treatment (d) fertilizers, pesticides, and herbicides t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. information. This effect is limited to treated large farmers for crop prices, weather predictions, and input use. Farmer age is never significant. Next we examine whether treated farmers appear more knowledgeable about crop prices. In the first four columns of results in Table 5, we present ATT estimates for knowing the sale price of the target crop before the day of the sale. Results show that treated farmers consider themselves more knowl- edgeable about crop prices in general. The difference is significant in all four cases, that is, one day before sale as well as several months before sale. In the second panel of Table 5 we report IV-LATE estimates that, as for Table 4, are T A B L E 5 . Knowledge and information sharing Know price before sale: Collect price info Share Whole sample One day One week One month at planting Information farming No of people consulted Collect price at planting Number of observations 722 723 722 722 922 929 925 Nearest neighbor matching (a) ATT Coeff 0.084 0.095 0.097 0.078 0.063 0.035 0.005 z-value 3.100 2.830 3.090 2.540 4.050 0.580 0.150 Regression results (b) 1. IV-LATE Treatment Coeff 0.130 0.158 0.181 0.146 0.119 0.068 0.011 t-value 2.180 1.980 3.280 1.960 4.080 0.500 0.160 Intercept Coeff 0.717 0.304 -0.034 0.010 0.676 1.520 0.564 t-value 64.780 20.640 -3.280 0.720 116.330 55.390 39.010 2. Heterogeneous effects (c) Treatment Coeff 0.065 0.073 0.084 0.068 -0.063 0.014 0.007 t-value 2.110 1.780 2.930 1.780 -4.080 0.200 0.190 Intercept Coeff 0.702 0.265 -0.104 -0.061 0.665 1.593 0.561 t-value 23.330 7.140 -3.510 -1.540 42.340 24.530 14.020 Dummy young head of hh Coeff -0.002 0.030 0.038 0.035 -0.004 -0.229 -0.041 t-value -1.220 0.690 1.090 0.890 -0.160 -2.400 -0.970 Total crop area cultivated Coeff 0.001 0.002 0.004 0.004 0.001 0.007 0.002 t-value 1.220 1.430 2.850 1.910 0.780 3.150 1.240 Interaction with treatment Dummy young head of hh Coeff -0.010 -0.059 -0.048 0.013 0.037 0.021 0.089 t-value -0.220 -1.080 -0.830 0.220 1.180 0.150 1.500 Total crop area cultivated Coeff -0.001 -0.001 -0.000 -0.003 -0.001 -0.002 -0.002 t-value -0.960 -0.450 -0.450 -1.420 -0.680 -0.590 -1.130 (a) Matching based on village triplet code dummies. (b) Village triplet code dummies included but not reported. (c) Mean value substracted from those control variables interacted with treatment. Fafchamps and Minten t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. 399 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 400 THE WORLD BANK ECONOMIC REVIEW larger in magnitude than the intent-to-treat ATT. There is no evidence of het- erogeneous effect along those two dimensions. In the next column of Table 5 we investigate whether treated farmers report sharing more information about farming with other farmers. If RML informa- tion is valuable, we expect treated farmers to be more likely to share it with others. Results reported in Table 5 suggest that this is indeed the case, but the effect is not large in magnitude: the intent-to-treat estimated ATT is a 6 percent increase; the IV-LATE estimate is larger at 12 percent, but still relative- ly small. Both effects are statistically significant, however. There is no evidence of heterogeneous effects by farm size or farmer age. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 In the last two columns of Table 5 we check whether treated farmers econo- mize on search costs because of RML. To this effect, we examine whether treated farmers make less effort gathering price information, either by consult- ing with others or by collecting price information in person at the time of planting. Contrary to expectations, results do not suggest this to be the case. The heterogeneous effect regression results reported at the bottom of the table indicate that large farmers consult with more people and are more likely to collect price information at planting time. For these farmers, the gain from making a better informed decision are larger, hence more effort is made to gather relevant price information. But we find no significant evidence that RML helps large farmers economize on these costs. This may only be tempor- ary, however: once farmers learn they can trust RML information they may decide to rely on it more. Young farmers consult fewer people about prices, but there is no evidence of heterogeneous treatment effects by farmer age. Price Received There is considerable price variation within villages. Different crops have dif- ferent coefficients of variation: lower for nonperishable crops such as wheat (CV ¼ 0.07) and soya (CV ¼ 0.14), and higher for perishable crops such as to- matoes (CV ¼ 0.22), onions (0.44), and pomegranates (0.45). We thus expect RML to be particularly beneficial for more perishable crops since their prices are more volatile and information is potentially more valuable. This is what we investigate in Table 6. The dependent variable is the log of the unit price received by the respondent on average over all the sales transac- tions of the target crop during the 12 months preceding the survey. Similar results are obtained if we use the price level instead of the log. The unit of ob- servation is the sales transaction. Most farmers report a single sale but some report more than one, which explains why the number of observations exceeds the number of participating farmers. The first column of Table 6 reports the ATT obtained using nearest neighbor matching. Contrary to expectations, we find no beneficial effect of the treat- ment on price received: the treatment effect is negative and statistically signifi- cant. We worry that this may be due to the inclusion of treatment 2 villages in the comparison. Indeed, in these villages, the small number of randomly Fafchamps and Minten 401 T A B L E 6 . Prices obtained (expressed in log(Rs/kg)) Heterogeneous effects (d) ATT (a) ATT (b) IV-LATE OLS long-model(c) OLS IV For whole No obs. 1480 688 1480 1425 1464 1457 sample(e) Treatment Coeff 2 0.031 2 0.043 2 0.062 2 0.028 2 0.034 2 0.026 t-value 2 2.000 2 0.520 2 1.670 2 1.510 2 1.860 2 1.430 Intercept Coeff 2.260 2.159 2.248 2.249 t-value 309.620 21.250 93.64 99.080 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Dummy young Coeff 0.021 2 0.013 2 0.013 head of hh t-value 0.990 2 0.500 2 0.530 Total crop area Coeff 0.005 0.001 0.001 cultivated t-value 5.720 1.900 1.550 Dummy if sold Coeff 2 0.011 2 0.006 2 0.008 to a trader t-value 2 0.250 2 0.190 2 0.290 Treatment Coeff 2 0.013 extension t-value 2 0.500 agent Interaction with treatment Dummy young Coeff 0.057 0.059 head of hh t-value 1.750 1.850 Total crop area Coeff 2 0.001 2 0.000 cultivated t-value 2 0.590 2 0.240 Dummy if sold Coeff 0.085 0.091 to a trader t-value 1.750 1.830 For control/ No obs. 947 443 947 909 938 931 treatment 1 village Treatment Coeff 2 0.015 0.031 2 0.079 2 0.017 2 0.046 2 0.032 t-value 2 0.600 0.630 2 1.600 2 0.510 2 1.740 2 1.170 Intercept Coeff 2.211 2.071 2.218 2.209 t-value 147.890 15.100 61.610 59.530 Dummy young Coeff 0.016 2 0.014 2 0.013 head of hh t-value 0.490 2 0.410 2 0.360 Total crop area Coeff 0.005 0.001 0.001 cultivated t-value 2.940 0.950 0.800 Dummy if sold Coeff 2 0.053 2 0.020 2 0.016 to a trader t-value 2 0.970 2 0.770 2 0.510 Treatment Coeff 2 0.048 extension t-value 2 1.000 agent (Continued ) 402 THE WORLD BANK ECONOMIC REVIEW TABLE 6. Continued Heterogeneous effects (d) ATT (a) ATT (b) IV-LATE OLS long-model(c) OLS IV Interaction with treatment Dummy young Coeff 0.041 0.038 head of hh t-value 1.040 0.910 Total crop area Coeff 0.000 0.000 cultivated t-value 0.050 0.180 Dummy if sold Coeff 0.101 0.093 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 to a trader t-value 2.210 1.820 (a) impact survey only; using nearest neighborhood matching; the reported coefficient on treat- ment is the ATT. (b) diff-in-diff, nearest neighborhood matching; using average unweighted prices in baseline and impact survey. (c) including but not reported dummies for graded, sold through commission agent, sold to trader, immediate payments, and quantity sold, years of education head of household, social network in village, land owned, years of farm experience, area cultivated of studied crop. (d) Mean value substracted from those control variables interacted with treatment. (e) village triplet code dummies included but not reported. t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. treated farmers may circulate the RML information to untreated farmers, who would then also benefit from it. This may blur the comparison between control and treated farmers due to a confounding externality between control and treated farmers. To investigate whether this explains our result, we re-estimate the ATT using only treatment 1 and control villages. The results are reported in the second panel of Table 6. We again find a negative treatment effect on farmer price, but it is not statistically significant. We also checked (results not reported here to save space) whether farmers in treatment 2 villages received higher prices than in control areas—unsurprisingly, given the lack of result for stronger treatment 1, they do not. The next column reports dif-in-dif ATT esti- mates, using nearest neighbor matching. Point estimates are now slightly posi- tive, but nowhere near conventional levels of significance. Next we examine whether the lack of effect is due to non-compliance. To in- vestigate this possibility, we instrument actual RML usage with the intent-to-treat dummy and report the results in the IV-LATE column of Table 6. The estimated coefficient of receiving the RML service is still negative, but remains non-significant for the entire sample as well as for the sample without treatment 2 villages. In Table A2 in the supplemental appendix, we repeat the ATT nearest neigh- bor matching and IV-LATE analysis for each crop separately. For the whole sample, ATT point estimates are negative for all crops, significantly so for Fafchamps and Minten 403 onions. IV-LATE point estimates remain negative, but are not statistically sig- nificant. When we restrict the analysis to control and treatment 1 villages, we find negative ATT and IV point estimates for four out of five crops; except for one (IV for tomatoes), they are not significant. We then examine whether intent-to-treat results may be affected by omitted variable bias. This is unlikely because treatment is randomly assigned, but we check it anyway. To this effect, we add controls for farmer age and farm size, as well as dummies for type of sale (that is, whether sold in the village or to a trader, as opposed to sold in the local wholesale market or mandi). We also include a dummy equal to one if the extension agent serving the village Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 received the free RML service. Results are reported in Table 6 under the “OLS long model” column. Other controls are included as well, as detailed at the bottom of the table, but their coefficients are not reported to save space. Again we find no evidence of a significant treatment effect. The coefficient of the ex- tension agent treatment is similarly non-significant. In the last two columns of Table 6 we investigate the possible existence of heterogeneous effects. The OLS columns report the heterogeneous intent-to-treat effect, equation (9), with controls. We also estimate an heteroge- neous effect version of equation (8): ð10Þ  Þ þ ei Yi ¼ u þ aSi þ gXi þ hSi ðXi À X Wooldridge (2002) recommends estimating IV models of this kind as follows. Let ~Sl be the predicted value of Si from the instrumenting equation. We con- struct a variable ~  Þ and we estimate (10) using ~ Sl ðXi À X Sl and ~  Þ as Sl ðXi À X instruments. In the OLS (intent-to-treat) results we now find a negative average treatment effect but a positive heterogeneous effect on young farmers. Treated young farmers received a price that is about 6 percent higher on average. In the IV results, the average treatment effect is non-significant, but the heterogeneous age effect remains. This suggests that less experienced farmers gain something from RML. These findings, however, are not robust to dropping treatment 2 villages, as seen in the second panel of Table 6. As robustness check, we correct for the possibility of non-random attribu- tion by adding an inverse Mills ratio as additional regressor in the IV-LATE re- gression. This Mills ratio is obtained from the attrition selection regression mentioned in Section 3. Results, not shown here, are similar to those reported in Table 6, and the Mills ratio is not statistically significant from 0 in the full sample or when using treatment 1 only, suggesting that non-random attrition is unlikely to have affected our findings. We also find that, in the OLS regression, farmers that grow more of the target crop get a significantly higher price on average. One possible explanation is that, for small crop sales, farmers make less effort to obtain price informa- tion and, hence, sell at a lower price. This effect, however, is not robust—it 404 THE WORLD BANK ECONOMIC REVIEW disappears in the IV regression or if we drop treatment 2 villages. Finally, con- sistent with expectations, treated households receive a price that is 8 –9 percent higher than control households when they sell to a trader as opposed to a com- mission agent. This is in line with the idea that better informed farmers can ne- gotiate better deals from buyers when they sell outside the relative safety of the mandi. We also examined whether treatment reduced the coefficient of variation of the price received by farmers in the same village. We expect price variation across farmers to be less if they are better informed. Aker (2008), for instance, reports that the introduction of mobile phones in Niger facilitated price inte- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 gration and reduced price dispersion. We do not find a similar effect for RML: the coefficient of variation of prices in treatment 1 villages is 0.320; in control villages it is 0.228, that is, smaller than in treated villages. The difference, however, is not statistically significant: the t-value ¼ 1.52, with a p-value of 0.135. Costs and Revenues RML may affect farmers in ways other than prices. Transaction costs per trans- action average 0.84 Rs/Kg. This compares to standard deviations for prices of 2.2, 17.1, 4.6, 0.9, and 3.1 Rs/Kg for tomatoes, pomegranates, onions, wheat, and soya, respectively. There is therefore room for farmers to increase revenues by reducing transaction costs. In the first column of Table 7, we report ATT and IV estimates for total transaction costs on the farmer’s last crop sale. Transaction costs include trans- port, loading and off-loading, payment at checkpoints, personal transport, pro- cessing, and commissions. Point estimates are positive for the whole sample— suggesting that RML raises costs—but they become negative when we only use treatment 1 villages. In both cases, however, point estimates are not significant. In the next column we investigate whether farmers received a higher net price (defined as the gross price minus the variable transaction costs in the last transaction). Mattoo, Mishra, and Narain (2007) estimate that transport costs per truck in India are between 0.09 to 0.13 Rs/kg/100 kilometers, which is small relative to total transaction costs. It thus seems that, in transport cost at least, arbitraging over space is not prohibitively expensive relative to other transaction costs. If farmers use RML information to arbitrage across space, they may ship their crop to a more distant market and incur a higher transport cost, but obtain a higher price net of costs, as in Jensen (2007). This is not what we find: results remain resolutely non-significant whether we include treatment 2 villages or not. Fafchamps and Minten 405 T A B L E 7 . Profitability measures Value Transaction Net price Sale added cost (c) (d) revenues (e) For whole sample Number of 713 713 713 713 observations Nearest neighbor matching (a) ATT Coeff 0.078 2 0.760 48,247 46,352 z-value 1.420 2 1.480 0.580 0.580 Regression results (b) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 IV-LATE Treatment Coeff 0.146 2 1.450 87,074 84,530 t-value 1.050 2 1.730 0.880 0.910 Intercept Coeff 1.576 8.906 66,545 59,235 t-value 59.060 55.350 3.500 3.320 For control/treatment 1 villages Number of 458 458 458 458 observations Nearest neighbor matching (a) ATT Coeff 2 0.150 0.735 143,852 138,311 z-value 2 1.700 1.060 1.190 1.220 Regression results (b) IV-LATE Treatment Coeff 0.159 2 0.074 267,588 260,249 t-value 0.439 2 1.000 1.370 1.410 Intercept Coeff 1.602 1.977 2 22,749 2 27,914 t-value 25.230 84.880 2 0.370 2 0.480 (a) Matching based on village triplet code dummies. (b) Village triplet code dummies included but not reported. (c) last transaction only; includes costs for transport, loading, off-loading, payments at check- point/toll or road-block, personal transport, processing, commission expressed in Rs/kg. (d) last transaction only; gross price minus transaction costs expressed in Rs/kg. (e) sales minus monetary input costs (fertilizer, pesticides, spray, purchased seeds, manure). t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. Farmers may gain not on the unit price but on total revenue. This is investi- gated in column 3. We find large positive point estimates, but no significant effect.12 If we use logs instead to limit the influence of outliers, we again find no significant effect. The last column reports similar results for value-added, that is, revenues minus monetary input costs such as fertilizer and pesticides. If weather information and crop advisories raise farmers’ technical and allocative efficiency, we would expect value-added to rise. Results are similar to those for 12. Sale values are large because quantities sold are large. This is especially true for onions and pomegranates where the average size of a single transaction is 10 metric tons. There is, however, a lot of variation around this average. 406 THE WORLD BANK ECONOMIC REVIEW sale revenues: large positive point estimates, but nothing statistically significant. Similar findings obtain if we use logs instead. Marketing In the conceptual section we argued that if RML is used by farmers to increase the price they receive, we should observe differences in marketing practices. If price information makes enables farmers to arbitrage across markets, we should observe systematic changes in where farmers sell. We first note that most sales take place in a market, nearly always a whole- sale market or mandi. The only exceptions are pomegranates for which, at Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 baseline, 44 percent of sales are conducted at the farm-gate and, to a lesser extent, wheat, with 7 percent of farm-gate sales. For the other crops, farm-gate sales represent less than 2 percent of recorded sales. Second, market diversifica- tion varies from crop to crop. Sales of perishable crops are geographically con- centrated: 98 percent and 81 percent of all market sales of tomatoes and pomegranates, respectively, occur at a single district market. Concentration is less for other crops: for onions, 51 percent of sales go to one district market. Corresponding figures for wheat and soya are 54 percent and 57 percent, respectively. To investigate whether treatment changed where farmers sell their crops, we construct an overlap index that captures the extent to which a farmer sold to the same location in the baseline and follow-up surveys. There are 39 whole- sale markets listed in the data, with farm-gate sales treated as a separate loca- tion. The index is weighted by quantity. An index value of 1 means the farm sold in the same location in the two survey rounds; a value of 0 means that nothing was sold at the same place. We also construct an added market dummy, which takes value 1 if the farmer sold in a new market or location in the follow-up survey, and a dropped market dummy equal to 1 if the farmer stopped selling in a specific location in the second round. Average treatment effects for the market overlap index and for the added and dropped market dummies are reported in Table 8. In the top panel we use the entire sample; in the second panel we only use the treatment 1 and control samples. With the entire sample treatment has a significant effect: treated farmers are 10 percentage points more likely to add a new sales location (market or farm-gate) and 9 percentage points more likely to drop one sales lo- cation. Treatment also reduces the overlap index by 10 percent on average. When we instrument RML usage with assignment to treatment, point estimates double and remain significant. These results are consistent with the predictions of the arbitrage model although, as we have seen in the previous two subsec- tions, changing sales location does not appear to have resulted in a higher price on average. Point estimates are also slightly smaller when we limit the sample to treatment 1 and control villages (second panel of Table 8), but they are no longer statistically significant at the 10 percent level, perhaps because of the re- duction in sample size. Fafchamps and Minten 407 T A B L E 8 . Spatial arbitrage and market changes Number of markets Added Dropped Overlap index (c) For whole sample Number of observations 691 691 691 Nearest neighbor matching (a) ATT Coeff 0.099 0.087 2 0.095 z-value 2.980 2.680 2 3.030 Regression results (b) IV-LATE Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Treatment Coeff 0.208 0.194 2 0.197 t-value 2.120 2.080 2 2.090 Intercept Coeff 0.575 0.463 0.493 t-value 30.430 25.850 27.290 For control/treatment 1 villages Number of observations 445 445 445 Nearest neighbor matching (a) ATT Coeff 0.045 0.074 2 0.077 z-value 0.880 1.560 2 1.650 Regression results (b) IV-LATE Treatment Coeff 0.187 0.189 2 0.198 t-value 1.260 1.320 2 1.400 Intercept Coeff 0.629 0.503 0.489 t-value 13.540 11.230 11.110 (a) Matching based on village triplet code dummies. (b) Village triplet code dummies included but not reported. (c) overlap index of sales location between years, weighted by quantity – see text for details t-values based on standard errors clustered by village triplet code. We continue our investigation of crop marketing in Table 9. The unit of analysis is an individual sale transaction. We first examine whether farmers sell at a wholesale market or mandi. As we have discussed earlier, farmers may choose to sell at the mandi because it is the only way to obtain accurate price information, even though doing so raises transaction costs relative to farm-gate sales. If this is the case, the RML service may give farmers the confidence not to sell at the wholesale market, for instance, because they can better negotiate with a farm-gate buyer. To investigate this possibility, we test whether treated farmers are less likely to sell at the mandi. Results, reported in the first column of Table 9, indicate that this is not the case: The intent-to-treat ATT, reported at the top of column 1, raises the likelihood of selling at the mandi. In the rest of column 1 we examine whether the results are different when we use IV-LATE instead, or when we allow for heterogeneous effect by firm size and farmer age. Results are qualitatively similar. The magnitude of the effect, however, is small, 408 THE WORLD BANK ECONOMIC REVIEW T A B L E 9 . Other marketing characteristics, all transactions Sold in if whole-sale Sold through a Crop was wholesale market, chosen commission Sold to graded/sorted market because closest agent trader before sale For whole sample Number of 1477 1352 1482 1470 1478 observations Nearest neighbor matching (a) ATT Coeff 0.030 2 0.078 0.006 0.046 0.033 z-value 2.540 2 3.220 0.230 1.740 2.260 Regression results (b) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 1. IV-LATE Treatment Coeff 0.063 2 0.131 0.539 0.084 0.055 t-value 1.750 2 0.940 0.844 1.050 1.120 Intercept Coeff 0.923 0.199 0.933 0.450 0.925 t-value 132.800 6.940 89.080 28.350 98.140 2. Heterogeneous effects (c) Treatment Coeff 0.032 2 0.064 0.054 0.039 2 0.029 t-value 1.820 2 1.010 0.620 1.010 2 1.120 Intercept Coeff 0.955 0.277 0.898 0.355 0.045 t-value 57.990 4.760 19.650 6.900 2.830 Dummy young Coeff 2 0.013 2 0.091 0.049 0.140 2 0.024 head of hh t-value 2 1.080 2 2.100 0.690 2.950 2 1.180 Total crop area Coeff 2 0.002 2 0.001 2 0.001 0.001 2 0.002 cultivated t-value 2 1.620 2 0.400 2 0.580 0.570 2 1.410 Interaction with treatment Dummy young Coeff 0.011 0.094 2 0.015 2 0.159 0.052 head of hh t-value 0.590 0.460 2 0.120 2 2.330 1.950 Total crop area Coeff 0.003 0.002 2 0.008 2 0.002 0.002 cultivated t-value 1.790 0.460 2 1.330 2 1.130 1.380 (a) Matching based on village triplet code dummies. (b) Village triplet code dummies included but not reported. (c) Mean value substracted from those control variables interacted with treatment. t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. probably because most farmers already sell at the mandi. When we differentiate the effect by crop, it is significant for pomegranates ( point estimate 0.157 with z-value of 2.35) and—less so—for soya ( point estimate 0.079 with z-value of 1.73); it is not significant for the other three crops. That pomegranates are most affected is hardly surprising given that pomegranates are the only crop with a sizable proportion of farm-gate sales at baseline. Thus, if anything, RML makes farmers more likely to sell at the mandi. Among farmers who sell at the market, however, Table 9 has shown a change in crop destination. To verify this further, we asked farmers who sell at a particular wholesale market whether they do so because it is the closest market. We see from the second column of Table 9 that treated farmers are Fafchamps and Minten 409 less likely to say they sell at a market because it is closest. Taken together, the evidence therefore suggests that treated farmers are more likely to sell farther away from home—either by switching from farm-gate to market sale or by switching to a more distant mandi. To investigate this further, we test whether treated farmers are more likely to sell directly to a trader (typically at the farm-gate) or without the help of a commission agent. If RML improves price information, farmers may be less re- luctant to sell to a trader, knowing they can insist on a price commensurate to the price at the nearest mandi. By a similar reasoning, they may rely less on commission agents who are contractually obliged to help farmers get the best Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 price but to whom a fee must be paid. Table 9 shows this is not the case for commission agents—the ATT is not significantly different from 0 in any of the three methods we report. For selling to a trader, we find a weakly significant ATT, but significance disappears when we use IV-LATE or allow for heteroge- neous effects. In the heterogeneous effect regression reported in the last panel of Table 9, we see that young farmers are more likely to report selling to a trader, but this relationship disappears with treatment, suggesting that young farmers learn not to sell to traders. Taking columns 1 to 4 together, the evidence suggests that RML helped some farmers realize that they could obtain a higher price by going to a more distant mandi rather than selling at a closer market or at the farm-gate. It is possible that some farmers choose to sell locally because of uncertainty regard- ing the return from traveling to the more distant mandi. Providing information about the mandi price reduces the risk of traveling to the mandi, and the reduc- tion in uncertainty may be what induced some farmers to incur the additional cost of traveling. In contrast, the evidence does not support the hypothesis that better informed farmers are emboldened to sell in local markets or at the farm- gate because they can insist on receiving a price more in line with the regional wholesale price. Finally, RML provides information on the price spread due to crop quality, that is, it shows prices by grade. Consequently, we expect treated farmers to pay more attention to quality, for instance, by grading or sorting their output into separate categories to obtain a better price on the top quality. This is what we find (see the last column of Table 9) for the ATT where the effect is statis- tically significant. The magnitude of the effect, however, is small: treatment raises the proportion of farmers grading or sorting their output by 3 percentage points. The effect also disappears in the IV-LATE regression; it resurfaces in the heterogeneous effect regression, but only when interacted with farmer age—that is, young farmers are slightly more likely to grade or sort their output as a consequence of treatment. Weather Information RML provides weather forecasts that are spatially disaggregated—and hence presumably more accurate than those publicized on the radio. Do RML 410 THE WORLD BANK ECONOMIC REVIEW forecasts help farmers improve yields, for instance, because farmers can take better anticipative action? We investigate this question in Table A.3 in the supplemental appendix. Farmers were asked whether or not they incurred unusually high rainfall events, such as a storm or heaving downpour. Some 58 percent said they did. The likelihood of reporting a storm is correlated with treatment in the IV and heterogeneous effect regressions: treated farmers are more likely to report in- curring a storm. Since we have no reason to believe that the weather is corre- lated with treatment, this is most likely due to response bias: farmers who receive regular weather information become more aware of unusual rainfall Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 events, and are more likely to report them to enumerators. We test whether treated farmers were able to reduce output loss or increase output following a storm. We find no evidence that this is the case. We also find little evidence of beneficial heterogeneous effects: young farmers report more output loss at harvest following a storm, not less. Agricultural Technology and Practices In addition to price and weather information, RML provides crop advisory messages relaying information on crop varieties, pesticide use, and cultivation practices. This information may be particularly valuable for sample farmers because some of our target crops are relatively new to them. In Table 10 we examine whether farmers changed the variety of the target crop that they grow. Some 31 percent of respondents stated they did change variety between the two survey years, but this proportion is the same irrespect- ive of treatment. Of those who changed variety, 65 percent stated they did so to improve profitability. Again we find no statistical relationship with treat- ment. Farmers who stated they changed crops to improve profitability were asked whether they did so because of RML. Here we find a statistically positive treatment effect: depending on the estimation method, treated farmers are 14 – 20 percent more likely to list RML as inspiration for the change. This is re- assuring, but not necessarily conclusive given that treatment is found to have no effect on the propensity to change variety or on the reason for changing variety. In the last two columns of Table 10 we turn to cultivation practices. In 2010 farmers were asked whether they changed anything about their cultiva- tion practices since the previous year; 22 percent of respondents stated they did so. We find no evidence that treated farmers were more likely to change culti- vation practices. Those who did change were asked what made them change practices. Of those farmers who report a change, a large proportion mentions RML as the reason for the change. The effect is statistically significant and large in magnitude—a 20–41 percent higher likelihood of listing RML, depending on the estimator. As for crop variety, this evidence is reassuring but not conclusive Fafchamps and Minten 411 T A B L E 1 0 . Crop varieties grown and cultivation practices If yes, Change of crop reason is If Change in If change, For whole variety since profita- profitability, cultivation because of sample last year bility because of RML practices last year RML Number of 895 240 156 911 203 observations Nearest neighbor matching (a) ATT Coeff 0.029 0.020 0.155 2 0.027 0.211 z-value 1.100 0.460 2.830 2 1.110 3.990 Regression results (b) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 1. IV-LATE Treatment Coeff 0.043 0.006 0.200 2 0.045 0.410 t-value 0.970 0.090 2.060 2 1.240 5.170 Intercept Coeff 0.525 0.374 2 0.033 0.476 2 0.016 t-value 59.350 30.950 2 2.060 65.780 2 0.950 2. Heterogeneous effects (c) Treatment Coeff 0.021 2 0.003 0.140 2 0.025 0.199 t-value 0.910 2 0.080 2.220 2 1.270 3.580 Intercept Coeff 0.518 0.408 2 0.112 0.432 2 0.051 t-value 17.210 9.280 2 5.680 14.940 2 1.270 Dummy young Coeff 2 0.001 2 0.079 0.103 0.042 0.036 head of hh t-value 2 0.030 2 1.280 2.350 0.890 0.850 Total crop area Coeff 0.001 0.002 0.001 0.001 0.000 cultivated t-value 0.770 1.640 0.980 1.310 2 0.030 Interaction with treatment Dummy young Coeff 2 0.091 0.053 2 0.001 2 0.009 0.033 head of hh t-value 2 1.760 0.680 2 0.010 2 0.150 0.220 Total crop area Coeff 2 0.002 0.000 0.000 2 0.001 0.001 cultivated t-value 2 1.240 2 0.120 0.100 2 0.910 0.210 (a) Matching based on village triplet code dummies. (b) Village triplet code dummies included but not reported. (c) Mean value substracted from those control variables interacted with treatment. t-values based on standard errors clustered by village triplet code; t-values in bold significant at the 10% level or better. given that treatment has no noticeable effect on changing crop practices themselves. VI. CONCLUSION We have reported the results of a randomized controlled trial of the impact of an SMS-based agricultural information service in Maharashtra. This informa- tion service, called Reuters Market Light (RML), sends SMS to farmers with information on prices, weather forecasts, crop advice, and news items. The price information is expected to improve farmers’ ability to negotiate with 412 THE WORLD BANK ECONOMIC REVIEW buyers and to enable them to arbitrage better across sales outlets. Weather in- formation should help farmers reduce crop losses due to storms. Crop advisory information should induce farmers to adopt new crop varieties and improve cultivation practices. The trial was conducted in collaboration with Thomson-Reuters, the pro- vider of RML. The experiment involved 933 farmers in 100 villages of central Maharashtra. Treatment was randomized across villages and, in some cases, across farmers as well. Participating farmers were surveyed twice in face-to-face interviews. We also conducted a follow-up telephone survey to gather informa- tion on reasons for non-compliance. Randomization appears to have worked Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 well in the sense that the control and treatment samples are balanced on most relevant variables. Although contamination is limited, non-compliance is common, which is why we reported intent-to-treat estimates throughout. We also reported instrumental variable (IV) estimates in which selection into treat- ment is used to instrument RML usage. We find no statistically significant average treatment effect on the price received by farmers, crop losses resulting from rainstorms, or the likelihood of changing crop varieties and cultivation practices. Treated farmers appear to make use of the RML service and they associate RML information with a number of decisions they have made. But, based on the available evidence, on average they would have obtained a similar price or revenue, with or without RML. Although disappointing, our results are in line with the RML take-up rate in the study districts. After a rapid expansion following the introduction of the service in 2007–09, take-up shows a relative stagnation in 2009–10, suggest- ing a possible loss of interest. We cannot, however, rule out that supply-side factors played a role. We also suspect that some farmers do not know how to renew the service.13 Although the absence of positive effect on price may surprise and disap- point, we find evidence of an RML information effect on where farmers sell their crop: they are less likely to sell at the farm-gate—especially young farmers—and more likely to sell at a different, more distant wholesale market. These results contradict the idea that RML information enables farmers to ne- gotiate better prices with itinerant traders, but are consistent with using RML information to arbitrage across sales outlets. Why arbitrage does not translate into higher prices is unclear, but some possible explanations arise from the data. First, few farmers sold at the farm-gate at baseline—except for pomegra- nates—thereby limiting the number of farmers who could realize that selling at the market was more beneficial than selling at the farm-gate, as a few did. Second, before treatment crop sales were concentrated on a single wholesale market in each district. Spatial concentration probably limits the range of 13. The provider has indeed encountered difficulties in setting up a reliable system for enabling clients to easily and reliably make repeat purchases of the RML service. Fafchamps and Minten 413 alternative market destinations nearby—and thus opportunities for arbitraging by farmers. We find little evidence of other RML effects. If RML information helps farmers improve crop quality, we should observe treated farmers changing agri- cultural practices, especially for crop varieties and grading. We do not, except for grading but the magnitude of the effect is small. We also find no significant effect on transaction costs, revenues, and value-added. Taken together, the evidence is consistent and compelling. Surveyed farmers sell almost exclusively to a wholesale agricultural market nearby. If traders have a comparative advantage in transport, for example because farmers do Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 not know anyone they can trust in other markets (Gabre-Madhin 2001), trader arbitrage across markets should ensure that farmers cannot fetch a more remu- nerative price by selling elsewhere. Hence it is optimal for farmers to sell to the nearest market. Similarly, if farmers fear being cheated when they sell at the farm-gate, it is optimal for them not to do so. Given this, it is perhaps not so surprising that better price information did not translate into higher farmer prices.14 If the above interpretation is correct, it has a number of implications for the external validity of our findings. Price information could help if spatial arbi- trage across agricultural markets does not hold, for example because markets are disorganized, segmented, or too thin to attract a steady flow of buyers—or because producers have a comparative advantage in transport, as in Jensen (2007). Even in such a case, however, price information is likely to be used first by traders, as documented by Aker (2008). Price information could also help farmers who sell at the farm-gate, such as the coffee growers studied by Fafchamps and Hill (2008). A stronger effect on crop quality may be obtained if price information is detailed by variety and grade and if farmers are provided with complementary information on how to produce high-price varieties and grades. These suggestions should help steer policy intervention toward regions and markets where the effect of price information may be beneficial, and avoid wasting resources on markets where it is unlikely to matter. REFERENCES Abadie, A., D. Drukker, J. L. Herr, and G. W. Imbens. 2004. “Implementing Matching Estimators for Average Treatment Effects in Stata.” Stata Journal 4 (3): 290– 311. Aker, J.C. 2008. Does Digital Divide or Provide? The Impact of Mobile phones on Grain Markets in Niger. Berkeley: University of California Press. 14. Nothing in this argument relies on buyer competition within each market: Even if buyers act in a monopsonistic fashion, as documented by Banerji and Meenakski (2004) for Delhi wheat markets, giving farmers information about market prices would not improve farmer prices—unless buyers price discriminate across markets and spatial arbitrage does not hold. 414 THE WORLD BANK ECONOMIC REVIEW Aker, J. C., and M. Fafchamps. 2011. Mobile Phones and Farmers’ Welfare in Niger. Berkeley: University of California Press. Banerji, A., and J. V. Meenakski. 2004. “Buyer Collusion and Efficiency of Government Intervention in Wheat Markets in Northern India: An Asymmetrical Structural Auction Analysis.” American Journal of Agricultural Economics 86 (1): 236–253. Bruhn, M., and D. McKenzie. 2009. “In Pursuit of Balance: Randomization in Practice in Development Field Experiments.” American Economic Journal: Applied Economics 1 (4): 200–232. Deaton, A. 2009. Instruments of Development: Randomization in the Tropics, and the Search for the Elusive Keys to Economic Development. NBER Working Paper #14690. Fafchamps, M., and R. V. Hill. 2008. “Price Transmission and Trader Entry in Domestic Commodity Markets.” Economic Development and Cultural Change 56 (4): 729– 766. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Gabre-Madhin, E. 2001. “The Role of Intermediaries in Enhancing Market Efficiency in the Ethiopian Grain Market.” Agricultural Economics 25 (2 –3): 311–320. Goyal, A. 2010. “Information, Direct Access to Farmers, and Rural Market Performance in Central India.” American Economic Journal: Applied Economics 2 (July): 22– 45. Jensen, R. 2007. “The Digital Provide: Information (Technology), Market Performance, and Welfare in the South Indian Fisheries Sector.” Quarterly Journal of Economics 127 (3): 879 –924. Mattoo, A., D. Mishra, and A. Narain. 2007. From Competition at Home to Competing Abroad. Washington, DC: World Bank. Minten, B., A. Vandeplas, and J. Swinnen. 2012. “Regulations, Brokers, and Interlinkages: The Institutional Organization of Wholesale Markets in India.” Journal of Development Studies, forthcoming. Mittal, S., S. Gandhi, and G. Tripathi. 2010. Socio-Economic Impact of Mobile Phone on Indian Agriculture. ICRIER Working Paper no. 246. New Delhi: International Council for Research on International Economic Relations. Staatz, J.M., A.M. Kizito, M.T. Weber, and N.N. Dembe ´ 2011. Evaluating the Impact on Market ´ le Performance of Investments in Market Information Systems: Methodological Challenges. MSU Staff Paper 2011-08. East Lansing: Department of Agricultural, Food and Resource Economics, Michigan State University. Wooldridge, J. M. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press. World Bank. 2010. India Economic Update, June 23 2010. Crises, Food Prices, and the Income Elasticity of Micronutrients:Estimates from Indonesia Emmanuel Skoufias, Sailesh Tiwari, and Hassan Zaman The 2008 global food price crisis and more recent food price spikes have led to a greater focus on policies and programs to cushion the effects of such shocks on poverty and malnutrition. Analysis of the income elasticities of micronutrients and Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 their changes during food price crises can shed light on the potential effectiveness of cash transfer and nutrition supplement programs. This article examines these issues using data from two cross-sectional household surveys in Indonesia, taken before (1996) and soon after (1999) the 1997–98 economic crisis, which led to a sharp in- crease in food prices. First, using nonparametric and regression methods, the article examines how the income elasticity of calories from starchy staples as a share of total calories differs between the two survey rounds. Second, the article estimates income elasticities of important nutrients in Indonesia. The analysis finds that, although summary measures such as the income elasticity of the starchy staple ratio might not change during crises, this stability masks important differences across individual nutri- ents. In particular, income elasticities of some key micronutrients, such as iron, calcium, and vitamin B1, are significantly higher in a crisis year than in a normal year, yet the income elasticities of others—such as vitamin C—remain close to zero. These results suggest that cash transfer programs might be even more effective during crises to ensure the consumption of essential micronutrients. But to ensure that all key micronutrients are consumed, nutrition supplement programs are also likely required. JEL classification codes: I12, O12, D12, E31 keywords: Income elasticity, crisis, prices, micronutrients, vitamin C, starchy staple ratio Global food prices rose sharply in 2008 and, in real terms, reached levels not seen since the early 1970s. The Food and Agriculture Organization (FAO 2008) Food Price Index grew by 73 percent between September 2006 and mid-2008, driven by unprecedented increases across all food categories. During the same period meat and fish prices increased 25 percent, eggs and milk 91 Emmanuel Skoufias (corresponding author; eskoufias@worldbank.org) is lead economist in the Poverty Reduction and Equity Group of the World Bank. Sailesh Tiwari (stiwari@worldbank.org) is in the Young Professionals Program of the World Bank and a Ph.D. candidate in the Economics Department of Brown University. Hassan Zaman (hzaman@worldbank.org) is lead economist in the Poverty Reduction and Equity Group of the World Bank. The authors are grateful to Harold Alderman for useful comments on an earlier version of this article. A supplemental appendix to this article is available at http://wber.oxfordjournals.org. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 415 –442 doi:10.1093/wber/lhr054 Advance Access Publication November 22, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 415 416 THE WORLD BANK ECONOMIC REVIEW percent, oils and fats 149 percent, and cereal grains 123 percent. After a down- ward trend in 2009, global food prices rose again in late 2010, with prices in December 2010 close to their 2008 peak. Such food price volatility, particularly sharp spikes, needs to be monitored closely for its impact on the poor. A review of the literature on the effects of the 2008 food price increases suggests that they are likely to have had a significant impact on the incidence of poverty (Ivanic and Martin 2008) and undernourishment (Tiwari and Zaman 2010) throughout the developing world. Soaring food prices and their adverse effects have not only heightened con- cerns about food security and malnutrition in parts of the developing world, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 but have also sparked a renewed interest in the design of effective policy responses. From the household point of view, such price increases have two main consequences: they reduce the purchasing power of household income, especially among poorer households, which spend a larger share of their incomes on food, and they result in a relative price effect that induces house- holds to substitute away from more expensive foods.1 Government interventions during rising food prices have almost always been motivated by the need to compensate poor households for their lost purchasing power. These interventions—commonly known as social safety net programs— are aimed at smoothing consumption and protecting the caloric availability of households to prevent increases in poverty and hunger. A review of the safety net programs used during the 2008 food price crisis shows that they typically took the form of income support, cash transfers, price subsidies, and supple- mentary feeding programs or in-kind transfers of staple foods (Wodon and Zaman 2010). Yet even if income support or in-kind staple food distribution is successful at preventing available calories from reaching dangerously low levels, there are valid concerns about dietary diversity and the consequent risk of malnutrition. When household income drops, households may keep calorie levels more or less constant through substitutions within and between food groups, while the consumption of essential micronutrients might decrease significantly as house- holds consume less meat, vegetables, eggs, and milk (Behrman 1995). The extent to which the consumption of micronutrients responds to decreases in income among poor households is of particular concern given the long-run consequences that a diet poor in micronutrients can have on child development before and after birth. Though there is ample evidence on the income elasticity of calories (Strauss and Thomas 1995; Subramanian and Deaton 1996; Skoufias 2003), empirical evidence on the income elasticity of micronutrients is sparse. The evidence that does exist suggests substantial differences in the income elasticities of 1. Purchasing power improves for net sellers of agricultural commodities whose prices increase. In addition, for households close to subsistence and already consuming the cheapest sources of calories, substitution possibilities are much more limited. Skoufias, Tiwari, and Zaman 417 micronutrients (Behrman and Deolalikar 1987; Bouis 1991). Pitt and Rosenzweig (1985), focusing on Indonesia and using data from farm house- holds, report very low income elasticities (less than 0.03) for a set of nutrients that included calories, proteins, fats, carbohydrates, calcium, phosphorus, iron, and vitamins A (carotene) and C (ascorbic acid). Chernichovsky and Meesook (1984), using data from rural and urban areas, report much higher income elas- ticities for nutrients—for example, from 0.7 to 1.2 for the poorest 40 percent (by expenditure) of the population on Java. Similarly diverse estimates have been reported for other countries.2 Most of the empirical evidence sheds light on whether the price sensitivity Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 of the demand for food and nutrients varies with income level (Timmer and Alderman 1979; Timmer 1981; Pitt 1983) or whether the income elasticity of calories varies with income level (Behrman and Deolalikar 1987; Ravallion 1990; Strauss and Thomas 1995; Subramanian and Deaton 1996). But evi- dence is lacking on whether the income elasticity of nutrients varies significantly depending on changes in the relative prices faced by households, such as changes experienced during food price spikes. This is an important gap in the literature and perhaps a large contributor to the significant divergence in esti- mates of the income elasticity of nutrients. When income elasticities are esti- mated using cross-sectional data, price variations necessarily come from the cross-section as well. But despite being pointed out by Deaton and Muellbauer (1980), the possibility that these estimated elasticities could be sensitive to the degree of variability of prevailing relative prices is something that has neither been tested explicitly nor been used to qualify the policy recommendations that emerge from existing studies. In 1998, during the financial crisis in Indonesia, the value of the rupiah depreciated dramatically, falling from around 2,400 per US dollar in June 1997 to just under 15,000 per dollar in June 1998 and finally settling at 8,000-9,000 per dollar in December 1998. These fluctuations in the exchange rate led to large increases in the prices of tradable commodities in domestic markets. Indonesia’s consumer price index rose 107 percent between February 1996 and February 1999. During the same period the food price index jumped 188 percent. In addition, subsidies for consumer goods such as rice, oil, and fuel were removed in 1998. So it is questionable whether estimates of the income elasticity of nutrients obtained from a sample of households observed during 2. Behrman and Deolalikar (1987), using data from villages covered by the International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), report income elasticity estimates of 0.06 to 0.19 for proteins (depending on whether level estimates or differences over time are used), 0.30 to – 0.22 for calcium, –0.11 to 0.30 for iron, 0.19 to 2.01 for vitamin A, – .08 to 0.18 for thiamine, 0.69 to 0.01 for riboflavin, – 0.15 to 0.21 for niacin, and 0.15 to 1.25 for vitamin C. A Nicaraguan study by Behrman and Wolfe (1984) reports significant income elasticity estimates in the range of 0.04 to 0.11 for calories, proteins, iron, and vitamin A (with statistically significant, but quantitatively small, nonlinearities). A Philippine study by Bouis (1991) reports an income elasticity of 0.44 for iron, an income elasticity of 0.16 for calories, and insignificant income elasticities for vitamin A and vitamin C. 418 THE WORLD BANK ECONOMIC REVIEW precrisis years can provide any guidance on how caloric and micronutrient availability might respond to additional income (other things being equal) during a period with a different set of relative prices. From a policy perspec- tive, the sensitivity of the income elasticity of nutrients to relative prices implies that policies aimed at increasing household income—such as employment and cash transfer programs—might be less effective in protecting nutritional out- comes under some economic conditions. This article uses shocks to food prices in Indonesia to examine the relation- ship between nutrient consumption and prices, with the analysis conducted at two levels. First, using the starchy staple ratio (the calories from starchy staple Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 foods such as cereals and tubers as a share of total calories) as a summary measure of household nutritional welfare, it assesses the impact of dramatic changes in food prices on household dietary composition. The SSR is defined as the share of calories consumed obtained from starchy staple foods such as cereals and tubers. According to Bennett’s Law this ratio is inversely related to the importance of starches relative to higher-quality, more expensive, micronutrient-rich foods such as meat, fish, fruits, and vegetables. The focus is on how the income elasticity of the starchy staple ratio differs between survey rounds in 1996 and 1999, when relative prices were very different for cereals and the other major food groups.3 The analysis is conducted for Indonesia’s entire population and for the poor in urban and rural Central Java (one of the country’s poorest provinces) in 1996 and 1999. Results are reported using both nonparametric and regression methods. This analysis is supplemented by updated estimates of the income elasticities of important nutrients in Indonesia, such as calories, proteins, fats, carbohy- drates, calcium, phosphorous, iron, and vitamins A, B1 (thiamin), and C. In times of crisis, cash transfers may be the fastest and least costly way of reaching households most likely to be adversely affected, if the delivery infrastructure is in place and leakage is low. Reliable elasticity estimates can help policymakers determine beforehand whether cash transfers are likely to increase nutrient availability among poor households or if other interventions may be needed. If estimated elasticities are high and significantly different from zero, policies such as cash transfers that aim to compensate for price increases are likely to be effective. But if estimated elasticities are indistinguishable from zero, alter- native interventions may be needed. Particular emphasis is placed on the sensi- tivity of the elasticity estimates to biases due to measurement errors in consumption and nutrient availability at the household level. The article also presents a test of whether the income elasticity of nutrients varies with the economic conditions facing households. Other things being 3. As a measure of dietary composition, the starchy staple ratio is useful because it summarizes changes in nutritional welfare over time. There is also evidence from Indonesia that a closely related variable—the share of total food expenditure going to nongrains, such as animal and plant sources—is positively associated with lower incidence of stunting among children younger than five (Sari and others 2010). Skoufias, Tiwari, and Zaman 419 equal, changes in the relative prices of staple foods may result in unexpected responses of how the demand for nutrients responds to cash transfers. For example, if total calorie availability is already low and the price of a staple increases during a crisis, households receiving cash transfers might spend the additional income on that same staple if it continues to be the cheapest source of calories (Behrman 1988; Behrman and Deolalikar 1989).4 This article is structured as follows. Section I describes the data used for the analysis and the construction of key variables and presents background infor- mation on the changes in prices and nutrient availability between 1996 and 1999 in Indonesia. Section II discusses the empirical strategy and the results of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 estimation using both nonparametric and regression methods. Section III sum- marizes the findings and presents some policy implications. I. BACKGROUND AND D ATA The analysis in this article is based on the detailed consumption module of the National Socio-Economic Survey (SUSENAS) conducted every three years by the Central Statistical Agency of the Government of Indonesia. The consump- tion module is nationally representative of urban and rural areas in each of the country’s 26 provinces and the Jakarta metropolitan area.5 The survey included 60,678 households in 1996 and 62,217 households in 1999. In addition to the detailed nature of the survey, one of the main advantages of comparing the income elasticity of calories in these two years is the oppor- tunity to examine economic behavior under dramatically different relative price regimes. In February 1999, when the 1999 SUSENAS was conducted, inflation in Indonesia had reached its peak since the start of the financial crisis in late 1997 and its intensification in mid-1998. An additional benefit is that the same questionnaire was used at the same point in time in both survey years. Thus the possible influence of seasonal factors in the relationship of income to calories—as emphasized by Behrman, Foster, and Rosenzweig (1997)—can be controlled for.6 A detailed discussion of the SUSENAS consumption module and the construction of key variables used in the analysis is presented in appendix A. To make meaningful comparisons across two cross-sectional surveys that are three years apart, the nominal income of households in 1999 must be expressed 4. This statement is not intended to compare the effectiveness of cash transfers relative to other possible alternatives of increasing nutrient availability within households. Alternatives include in-kind food transfers and employment creation programs. 5. The analysis also uses some variables from the larger SUSENAS (core survey) containing observations for about 205, 000 households. 6. The Idul Fitri-Lebaran holiday following the fasting month (Ramadan) is a moving holiday—in 1999 it fell in late January. Central Statistical Agency officials confirmed that the survey was conducted two weeks after the holiday, so the value of household food consumption has little chance of appearing unusually high due to the holiday. 420 THE WORLD BANK ECONOMIC REVIEW in 1996 rupiah. A critical point for the construction of real income in 1999 is the fact that changes in food prices affect households differently depending on the share of their budgets spent on food. Typically, poorer households spend a much larger share of their incomes on food—nearly 60 percent for poor rural households in Indonesia, compared with 40 percent for urban households at the top of the expenditure scale. The SUSENAS consumption module includes data on the value and quantity of food items consumed, allowing calculations of unit values at the household level. A deflator was constructed combining the unit values calculated from the consumption module with province-specific prices reported for nonfood items Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 by the Central Statistical Agency.7 First, because expenditures are only col- lected for nonfood items, a deflator for nonfood items was constructed using the mean shares of major groups of nonfood items in the February 1999 survey as weights and the province-specific price indexes for these groups.8 Second, a household-specific food deflator was calculated from a weighted average of the 52 food items used in the calculation of the poverty line in Indonesia. Specifically, the household-specific food deflator was calculated using the formula X52  ! Pi ðR; 96Þ À1 Ph F ð99Þ ¼ h Si ð99Þ ð1Þ i¼1 Pi ðR; 99Þ which is the standard formula for calculating a Paasche price index (Deaton and Zaidi 1999). Sh i denotes the share for food item i of the total amount expended on the 52 food items, and the superscript h indicates that the share varies by household. The second term is the ratio of the median unit value of food item i in region R in 1996 to the corresponding unit value in 1999. Household-specific unit values of food items are replaced by median unit values for each of the urban and rural areas of the 26 provinces and the Jakarta metropolitan area (a total of 53 regions) to minimize the influence of measurement errors and differences in the quality of food consumed by wealth- ier households (Deaton 1988). With these price deflators for food and nonfood, the overall price deflator for household h in 1999, Ph(99), can be expressed as   Ph ð99Þ ¼ W ^ h ð99ÞPh ðR; 99Þ : ^ h ð99ÞPh ð99Þ þ 1 À W ð2Þ F F F NF Note that the weights applied to food and nonfood also vary across house- holds. The weight for each household is calculated from the predicted value of 7. More details on the construction of the price indexes can be found in Skoufias (2003). Suryahadi and others (2000), and Levinsohn, Berry, and Friedman (1999), all of whom take a similar approach to constructing household-specific price indexes for Indonesia. 8. The province-specific price indexes for food and nonfood groups reported by the Central Statistical Agency are based on prices for 27 cities in 1996 and 44 cities in 1999. Skoufias, Tiwari, and Zaman 421 F I G U R E 1. Changes in Price of 1,000 Calories by Food Group Relative to Cereals and Tubers in Rural and Urban Central Java, 1996 and 1999 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules using expenditure quartiles. the regression of household food share in 1999, W ^ h ð99Þ, on the logarithm of F per capita consumption and the logarithm of household size. This approach eliminates the influence of household-specific, unobserved components (such as taste preferences) on the share of food. To provide more concrete evidence about the relative price regimes prevail- ing in the two survey years, figures 1a and 1b show the changes in mean prices per 1,000 calories (1 kilocalorie) paid by rural and urban households between 1996 and 1999 in Central Java, a densely populated province with a high con- centration of poor people. Prices per kilocalorie are calculated by dividing the nominal value of household consumption for each food group by the 422 THE WORLD BANK ECONOMIC REVIEW kilocalories provided by all the items in the group.9 Poorer households may consume food items of lower quality and, as a result, the prices per kilocalorie paid by these households may be lower than those paid by richer households. To investigate for this possibility, prices per kilocalorie are calculated separate- ly for households at the bottom and top 25 percent of the distribution of total consumption per capita in each year.10 Figures 1a and 1b confirm that the relative prices faced by households changed considerably between 1996 and 1999.11 In general, the absolute prices of calories from all food groups seems to have increased dramatically between the two years in both urban and rural areas. Relative to cereals, food groups Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 such as meat and fish, fruits and vegetables, and eggs and milk became more expensive in both rural and urban areas. There is also considerable heterogen- eity between the rich and the poor in the magnitude of the relative price changes. Price increases appear more pronounced for the poor, particularly in urban areas. For example, in urban areas the price of eggs and milk relative to cereals increased 23 percent for the poorest consumers but only 4 percent for the richest consumers. Price changes of this magnitude undoubtedly had a large impact on calorie and nutrient availability at the household level. Figures 2a and 2b present the change in average daily calories per capita between 1996 and 1999 for rural and urban Central Java. There was a significant reduction in both total and proportional calorie availability. Calorie availability at the household level in Central Java declined by 8 percent in rural areas and 6 percent in urban areas. There was a much larger reduction in calories derived from food groups richer in micronutrients—meat and fish, fruits and vegetables, and eggs and milk. The figure also shows considerable heterogeneity in the reduction between the rich and the poor, with larger reductions for the poor in almost all cases in both rural and urban Central Java. The change in calories sourced from cereals and tubers is interesting. In rural Central Java calories obtained from cereals and tubers fell 12 percent among the poorest households, but only 1 percent among the richest house- holds. By contrast, roughly the same level of decline was experienced by the poorest and richest households in urban areas. This discrepancy is largely because rich rural consumers are likely to be landowners, possibly engaged in the production of some of these staples. Thus price increases for these staples 9. The calorie prices reported are derived by dividing expenditures by total calories in the food group in each year. Thus the price of calories in 1999 may be biased downward depending on the extent to which households substituted away from more expensive food items within and between groups. 10. In 1999 the percentiles of total consumption per capita are estimated after dividing consumption by the deflator discussed earlier. 11. For a related analysis of the impact of the Indonesian crisis on budget shares, with repeated observations on sampled households, see Thomas and others (1999). Skoufias, Tiwari, and Zaman 423 F I G U R E 2. Changes in Per Capita Calorie Consumption by Food Group in Rural and Urban Central Java, 1996 and 1999 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules using expenditure quartiles. may have made those rich consumers better off and not affected their con- sumption of cereals and tubers by much. II. E M PI R I CA L FR A ME WO R K , A N A LY S I S , AND R E S U LT S Economic theory provides little guidance on how the income elasticity of a given commodity might change with changes in prices. Given a Marshallian demand function for any food item xi, summarized by the function xi ¼ xðp; M; ZÞ where p is a vector of relative prices, M is real income, and Z is a vector of preference shifters (such as household demographic characteristics), it follows that, in general, the response of demand to income changes depends 424 THE WORLD BANK ECONOMIC REVIEW on the same set of variables: q xi ; M; ZÞ: ðp @M Without strong (even arbitrary) assumptions about the separability of prefer- ences between and within specific food groups, little can be said (at least from a theoretical perspective) about how changes in prevailing relative prices might affect the demand for a commodity in response to income changes. Under these circumstances this issue can be addressed only empirically. This is precisely the gap that this article aims to fill. The question addressed empirically is Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 whether income elasticity differs significantly between a noncrisis year (1996) and a crisis year (1999), which have very different relative price vectors, p 96 and p99 : q xi q xi 96 ; M; ZÞ = ðp  ; M; ZÞ ðp ð3Þ @M @ M 99 The analysis uses both nonparametric and regression methods that take into account the role of measurement error in total outlay. The “almost ideal demand system” proposed by Deaton and Muellbauer (1980) provides the empirical framework for the analysis because it makes transparent the dependence of income elasticity on relative prices.12 Starting with the Working-Leser formulation that relates value shares to the logarithm of total expenditures wi ¼ ai þ bi logðxÞ ð4Þ Deaton and Muellbauer propose a way of making the coefficient of total expenditures,bi, a function of prices. Given a cost function logcðu; pÞ ¼ að pÞ þ ubð pÞ ð5Þ and choosing the functions a( p) and b( p) to be of the form X 1X X að pÞ ¼ a0 þ a logpk þ k k g Ãlogpk logpl l kl ð6Þ 2 k b bð pÞ ¼ b0 Pk pk k ð7Þ the Engle curve can be expressed as X wi ¼ ai þ g logpij þ bi logðx=pÞ j ij ð8Þ 12. The quadratic Engel curve proposed by Banks, Blundell, and Lewbel (1997) is an extension of the almost ideal demand system specification that also explicitly recognizes the dependence of the income elasticity to prices (see equation 11 in their paper). Skoufias, Tiwari, and Zaman 425 where P is the price index defined by X 1X X log P ¼ a0 þ ak log pk þ g logpk logpl l kl ð9Þ 2 k and the parameters gS ij are defined as 1  gij ¼ gij à þ g ji à ¼ g ji : ð10Þ 2 Based on this framework, the question addressed in this article translates to 0 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 whether the b s of equation (7), estimated using cross-sectional data from 1996—the normal year—as the coefficients of the logarithm of total expendi- tures in equation (8), are statistically and economically similar to those esti- mated in 1999—the crisis year. Nonparametric Regression and the Starchy Staple Ratio The analysis begins with an investigation of how the starchy staple ratio (SSR) and its income sensitivity vary in a cross-section of households between 1996 and 1999. SSR, calculated as the share of total calories obtained from cereals and tubers, is considered a more useful aggregate measure of household welfare than is total caloric availability per capita because SSR captures diver- sity in diets. According to Bennett’s Law, SSR declines with household income.13 This analysis adopts a flexible approach that examines Indonesian households by aggregating the caloric contents of the more than 200 food items included in the SUSENAS survey. Given that the income elasticity of cal- ories in Indonesia is known to be nonlinear (Skoufias 2003), nonparametric methods are used that also allow the income elasticity of the SSR to vary with income level. Using y to denote the logarithm of SSR and x the logarithm of per capita total household consumption, the regression function is: mðxÞ ¼ Eðy=xÞ: ð11Þ Following Subramanian and Deaton (1996) and Deaton (1997), m(x) is esti- mated using a smooth local regression technique proposed by Fan (1993).14 At any given point x, a weighted linear regression is run of the logarithm of SSR on the logarithm of per capita consumption. The weights are chosen to be the largest for sample points close to x and to decrease with distance from x. The distribution of the logarithm of per capita consumption is divided into 100 evenly spaced grids, and local regressions are estimated for each grid instead of 13. Bennett’s law involves the relation between household diet and income, while Engel’s law relates the share of food expenditures in a household budget to household income. Timmer, Falcon, and Pearson (1983) provide a detailed discussion of Bennett’s law. 14. Fan (1993) demonstrates the superiority of the smooth local regression technique over kernel and other methods. 426 THE WORLD BANK ECONOMIC REVIEW for each point x in the sample. For the local regression at x, observation i has the quartic kernel weight x À x 2 !2 15 i wi ðxÞ ¼ 1À ð12Þ 16 h if 2 h x 2 xi h and zero otherwise. The quantity h is a bandwidth set to trade off bias and variance, which tends to zero with increasing sample size. This analysis uses a bandwidth of 0.8. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Figures 3a and 3b compare logarithms of SSR of dietary content in 1996 and 1999 with logarithms of per capita expenditure (lnPCE), using 1996 prices, for rural and urban Central Java. With the income level in 1999 made comparable to that in 1996, these figures make it possible to examine the effects of changes in food prices between 1996 and 1999 on household dietary composition (as summarized by SSR), assuming that family size did not change significantly during that period. An advantage of these nonparametric figures is that potential biases in the measurement of calorie availability among higher-income households do not affect estimates of nutrient income elasticity among poorer ones. For example, figures 3a and 3b make it possible to obtain a sense of the extent to which the exclusion of nutrients obtained from prepared foods (consumed more by wealthier households) affects the nutrient income elasticity of wealthier households. In rural Central Java in 1999 SSR was just below the median level of per capita expenditure. SSR was slightly below its level in 1996 for households in the lower half of the distribution and above it for households in the upper half (see figure 3a). This suggests that in the crisis year the availability of calories from cheaper food sources (such as cereals and tubers) generally increased for higher-income groups in rural areas.15 This pattern is even more obvious in urban areas, where the 1999 SSR line lies almost uniformly above the 1996 SSR line. Thus for the urban poor there is a rather clear shift to the right in 1999, indicating a larger reliance on starchy staples for calories in times of crisis. Among rural households the pattern is less clear. The differences in the slopes of the SSR lines in 1999 and 1996, in both rural and urban areas, suggest that the responsiveness of SSR to increases in income varies in crisis and noncrisis years (figures 3c and 3d). In rural areas the elasticity of SSR relative to income is higher during the crisis year and invariant to household income at about –0.27 percent. Thus it appears that an increase in income during a crisis year, such as cash transfers to poor rural households, is likely to be more effective in increasing dietary diversity (that is, lowering SSR) than in a noncrisis year. 15. Studdert, Frongillo, and Valois (2001) corroborate the increase in household food insecurity and compromised diet during the crisis in three Java provinces, including Central Java. Skoufias, Tiwari, and Zaman 427 F I G U R E 3. Features of the Starchy Staple Ratio in Urban and Rural Central Java, 1996 and 1999 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ calculations based on 1996 and 1999 SUSENAS consumption modules. In contrast, in urban areas, where there was a larger reliance on starchy staples for calories in the crisis year, the income elasticity of SSR was about the same in 1996 and 1999 for poorer households, becoming smaller than the elas- ticity in the noncrisis year.16 Note, however, that as bivariate plots of calorie 16. Supplemental appendix S3, available at http://wber.oxfordjournals.org, examines whether the elasticity estimates are significantly different at different levels of outlay by checking whether the standard error bands for the 1996 estimate overlap those for the 1999 estimate. The figures do not reveal any significant differences in the income elasticity estimates for the crisis and noncrisis years, a result confirmed by the regression methods used in the latter part of this article. 428 THE WORLD BANK ECONOMIC REVIEW shares at the household level these are descriptive only. The next section per- forms a more detailed analysis of SSR and its elasticity. Regression Analysis The analysis in the previous section focused on the bivariate relationship between SSR and per capita expenditure. Though being informative about the general shape of that relationship and how it changed between 1996 and 1999, that analysis cannot account for a number of critical factors—primarily the dif- ferences in the age and gender composition of households, as well as the Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 problem of correlated errors between household consumption and nutrients. This section examines the elasticity of SSR relative to household consumption, controlling for a set of household characteristics. The focus on the effects of relative price vectors on the income elasticity of demand faces some constraints.17 A typical cross-sectional household survey collects data within a short timeframe. As a result, most of the price variation for any household commodity comes from differences in transport costs, market segmentation, the quality of the commodity, and other transaction costs that can prevent the equalization of prices paid by consumers for the same commodity. To the extent that households in different locations are surveyed at different times of year, a survey might also capture seasonal variability in prices. Even so, it is doubtful that seasonal variation in prices provides an adequate repre- sentation of the change in relative prices that consumers face during an eco- nomic crisis. Household panel data provide an opportunity to circumvent some of these shortcomings. Behrman and Deolalikar (1987), for example, analyze the relationship of calories to income using data from village surveys conducted by the International Crops Research Institute for the Semi-Arid Tropics (ICRISAT). But even these data shed little light on the relationship between household food consumption and spending during a crisis, because the eco- nomic environment was relatively stable during the period of that study. To compensate for the fact that these are cross-sectional data from surveys taken in different years, not longitudinal data, this analysis uses a flexible speci- fication that provides an explicit test of the difference between the elasticity coefficients in the precrisis and crisis years. To implement this, the two sets of cross-sectional data were pooled, then the following regression was run: À 0 Á ln sSRjkt ¼ a96 þ b96 ln PCEjkt þ g96 Xjkt þ m96 À 0 Á þ D99 à a99 þ b99 ln PCEjkt þ g96 Xjkt þ m99 þ 1jkt ð13Þ 17. In much of the literature on the nutrient-income relationship (see Strauss and Thomas 1995), prices are typically left out of the specification of the Engel curve estimated using cross-sectional data. This is based on the ad hoc assumption that all households face the same prices. Notable exceptions are Subramanian and Deaton (1996), Behrman and Deolalikar (1987), Bouis and Haddad (1992), and Banks, Blundell, and Lewbel (1997). Skoufias, Tiwari, and Zaman 429 where lnSSRjkt is the natural logarithm of SSR for household j that lives in cluster k in year t, m96 and m99 are vectors of binary variables identifying the cluster fixed effects in the 1996 and 1999 rounds,18 and D99 is a binary dummy variable equal to one for observations in 1999 and zero otherwise. The variable denotes the natural logarithm of real per capita consumption expendi- tures for household j that lives in cluster k in year t. 1jkt is an error term sum- marizing the influence of random disturbances. The set of control variables Xjkt is a vector of household characteristics indicating the logarithm of house- hold size and variables characterizing the age and gender composition of the household, all expressed as ratios of the total family size (the number of chil- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 dren ages 0 –5, the number of children ages 6–12, the number of male and female household members ages 13 –19 and 20 –54, and the number of men older than 55). Additional binary variables include whether the household head is a woman, dummy variables on the education levels of the household head and his or her spouse (completed primary, junior high, or senior high school), and the sectors of employment of the household head and his or her spouse (self-employed, unemployed, or a wage worker). With this specification, cluster-level fixed effects are allowed to differ across the two survey years, providing control over differing relative prices between years.19 In addition, the coefficients of all the control variables and lnPCEjkt are allowed to vary across the two years, providing an explicit test of the differ- ence in the income elasticity of the various dependent variables between 1996 and 1999. The dummy variable D99 is also able to absorb any other aggregate effects (aside from relative prices) that might have changed between 1996 and 1999. As first pointed out by Bouis and Haddad (1992), a food expenditure survey can overstate the nutrient availability in wealthier households, since it is common for these households to provide meals to employees and domestic servants.20 In addition, following the 1997–98 crisis, it is plausible that there was an increase in the frequency of this practice. To minimize potential pro- blems introduced by the fact that the level (and thus the elasticity) of SSR and nutrients may be less accurately measured for wealthier households, the estima- tion limits the sample to the lower half of the distribution of consumption per capita in rural and urban Central Java. Robust standard errors are estimated to control for unknown forms of heteroskedasticity. 18. The SUSENAS survey is a clustered survey with at most 16 households surveyed per cluster each year. Clusters in 1996 and 1999 have the same code, but it is unclear whether they represent the same clusters across the two years. Because of this limitation, clusters with the same code in different years are treated as different clusters. 19. Subramanian and Deaton (1996) use the same approach with cross-sectional data. Cluster-level fixed effects also take into account other time-invariant local characteristics that determine dietary intakes and preferences. 20. In the SUSENAS survey, domestic servants are counted as household members. 430 THE WORLD BANK ECONOMIC REVIEW In defense of the ordinary least squares (OLS) estimates, it is important to bear in mind that the SUSENAS survey is a seven-day food intake and consumption survey that carefully collects information about food consumed outside the house- hold as well as food received in kind from outside sources. The survey also cap- tures food received as payment for services, food gifts, and food transfers. Since the same questionnaire was applied at the same point in time each survey year, there is no reason to believe that biases exist due to these factors. Yet the possibility remains that correlated measurement errors in total food consumption (and thus calorie and nutrient availability) are potential sources of bias in estimates of income elasticities of nutrients. As first noted by Bouis and Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Haddad (1992) in a linear version of equation (13), the likelihood that measure- ment errors in nutrient availability are positively correlated with measurement errors in household consumption implies that this type of measurement error is not the standard errors-in-variables problem, where coefficients are likely to be biased toward zero. In the context of correlated measurement errors in the de- pendent and independent variables of a regression, it is unclear whether the upward bias from the correlated errors outweighs the standard downward attenu- ation bias from the measurement error in total consumption. So the direction of net bias in income elasticity estimates obtained using OLS methods will generally depend on the relative size of the correlation between the measurement errors and the variance of the measurement error in household consumption. In the case of a log-linear equation such as that of equation (13), Deaton (1997) notes that elasticity estimates using the log of nonfood expenditures as the sole instrumental variable (IV) are likely to be biased downward, implying that the elasticities estimated by OLS and IV methods provide upper and lower bounds, respectively, for their true value. To address these considerations, the analysis uses an index of household assets, constructed using the principal com- ponents method, as an instrumental variable for lnPCEjkt. Specifically, in each survey year the index of household assets is estimated using variables for the number of cows and buffaloes, sheep and goats, chickens and ducks, and pigs owned by the household. In addition, a series of dummy variables summarizes the household resi- dence and its environment, such as whether the roof is concrete or tile, the walls are brick or wood, the floors are tile or cement, the toilet is private or shared, drinking water is accessed through a public network or purchased, and energy for cooking, lighting, and heating is obtained through the public electric or gas utility. Because the consumption variable interacting with the 1999 dummy variable enters into the estimated equation, the instrument used for this interaction term is the asset index based on 1999 data interacting with the 1999 year dummy variable. The results of the first-stage regressions are in appendix B.21 To examine whether the estimated elasticities vary significantly 21. In all cases, a Hausman-type test (Hausman 1978; Holly 1982) for the absence of measurement error in the consumption variable rejected the null hypothesis. Skoufias, Tiwari, and Zaman 431 T A B L E 1 . Elasticities of the Starchy Staple Ratio for Rural and Urban Households in Central Java, 1999 All households Poorer householdsa Ordinary least Instrumental Ordinary least Instrumental Area and variable squares variable squares variable Rural lnPCE – 0.25*** – 0.24*** –0.27*** – 0.31*** (0.01) (0.01) (0.02) (0.04) Marginal effect 0.01 0.00 0.04 -0.01 (lnPCE*D99) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 (0.02) (0.02) (0.03) (0.07) Urban lnPCE – 0.26*** – 0.22*** –0.31*** – 0.21*** (0.02) (0.02) (0.02) (0.07) Marginal effect – 0.02 0.00 0.05 0.10 (lnPCE*D99) (0.02) (0.04) (0.04) (0.11) *** Significant at the 1 percent level. Note: Numbers in parentheses are robust standard errors, corresponding to the elasticity esti- mates. Each column represents a separate regression using a wide range of household-level eco- nomic and demographic control variables. Instrumental variable estimates were obtained by instrumenting the natural logarithms of per capita expenditure with household-specific asset indexes. a. Defined as the lower half of the distribution based on consumption per capita. Source: Authors’ analysis based on SUSENAS 1996 and 1999 data. for the poor, separate analyses are performed for the entire sample and for the sample of households with per capita expenditure below the median. Table 1 presents the elasticity estimates for SSR obtained from these regres- sions, fitted separately for the entire sample and for the subset of poorer house- holds. As expected, the negative sign on the point estimates for SSR income elasticity is consistent with the nonparametric observation that the share of cal- ories derived from starchy staples declines with income. Further, the compari- son of the elasticity estimates of the IV estimates between urban and rural areas suggests that SSR elasticity is marginally higher in rural Central Java. This implies that the rate at which households, as their incomes rise, switch from cheap sources of calories to food groups that are better sources of micro- nutrients is higher in rural than in urban areas. The estimated coefficient for b99 in the regression is of particular interest because it contains information on how different the elasticities were in 1999 (the crisis year) relative to 1996 (the reference year). The estimates for b99 in the IV specification reported in table 1 suggest that SSR elasticity in 1999 was practically identical to the elasticity in 1996. The size of the marginal effect is 432 THE WORLD BANK ECONOMIC REVIEW small and not statistically significant—implying that the income elasticity of SSR was invariant to the different relative prices prevailing in 1999. Comparing these results with those for the subset sample of the poor in Central Java shows a generally higher elasticity among the poor. This indicates that cash transfers could be more effective in reducing the SSR for the poor relative to the entire sample population. But the fact that the coefficients of the interaction term are not significantly different from zero implies that cash transfers may have been ineffective during the crisis year. Even though the invariance of SSR elasticity to price increases suggests little substitution toward or away from cereals and tubers, the complex pattern of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 price changes in Indonesia during the financial crisis of the late 1990s could have induced households to substitute within and among other food groups. To probe deeper into the consequences of such substitutions for the elasticity of micronutrients, the estimation in equation (13) is repeated using the natural logarithm of the consumption of specific nutrients as dependent variables. The elasticity coefficients obtained from these regressions for the entire sample are reported in table 2. All the OLS estimates of the nutrient elasticities for both rural and urban Central Java are positive and statistically significant for 1996. For example, in rural areas the estimates for the elasticity range from 0.18 for carbohydrates to 0.63 for fat. In urban areas the spread is narrower, ranging from 0.13 for car- bohydrates to 0.44 for fats. The IV estimates, on the other hand, while general- ly statistically significant, appear to be smaller than the corresponding OLS estimates for both urban and rural areas. This suggests that the upward bias from the correlated errors may outweigh the standard downward attenuation bias in the OLS estimates. It is also noteworthy that vitamin C is the only nu- trient for which the IV estimates of elasticity are not significant in both urban and rural areas—suggesting that vitamin C consumption may not be responsive to income in Central Java, whether in a crisis year or not. The estimates for b99, the coefficient on the interaction between the loga- rithm of per capita expenditure and the year dummy for 1999, are statistically significant for calories, proteins, fats, carbohydrates, phosphorous, iron, and vitamin B for rural areas and for calories, proteins, fats, calcium, phosphorous, and iron for urban areas. This is evidence of a difference in the elasticity of these nutrients in the crisis year relative to the reference year. Moreover, these differences are quite large, particularly for urban areas. The income elasticities of calories and iron appear to have doubled in urban areas in 1999. The elasti- cities for proteins, calcium, and phosphorous also nearly doubled. The income elasticity of fats more than doubled in 1999 for urban households. But there are some nutrients for which income elasticity did not change significantly between 1996 and 1999: calcium and vitamin A in rural areas and carbohydrates, vitamin A, and vitamin B in urban areas. An increase in income elasticity of any specific nutrient in the crisis year—particularly of the magni- tudes seen in urban areas—may be considered an indicator of deteriorating T A B L E 2 . Income Elasticities of Nutrients for Rural and Urban Households in Central Java, 1999 Area, estimator, and variable Calories Proteins Fats Carbohydrates Calcium Phosphorus Iron Vitamin A Vitamin B Vitamin C Rural Ordinary least squares lnPCE 0.25*** 0.31*** 0.63*** 0.18*** 0.37*** 0.22*** 0.28*** 0.27*** 0.19*** 0.20*** (0.02) (0.02) (0.04) (0.01) (0.03) (0.02) (0.03) (0.06) (0.02) (0.06) Marginal effect (lnPCE*D99) 0.05** 0.09*** 0.03 0.07*** 0.10** 0.10*** 0.16*** – 0.08 0.11*** – 0.07 (0.03) (0.03) (0.06) (0.02) (0.04) (0.03) (0.04) (0.08) (0.04) (0.09) Instrumental variable lnPCE 0.14*** 0.19*** 0.50*** 0.07*** 0.22*** 0.10*** 0.16*** 0.13*** 0.07*** – 0.03 (0.02) (0.02) (0.04) (0.02) (0.03) (0.02) (0.03) (0.05) (0.02) (0.05) Marginal effect (lnPCE*D99) 0.07** 0.10** 0.19*** 0.06** 0.07 0.10*** 0.14*** – 0.15 0.13** – 0.10 (0.03) (0.04) (0.07) (0.03) (0.05) (0.04) (0.04) (0.09) (0.05) (0.09) Urban Ordinary least squares lnPCE 0.19*** 0.26*** 0.44*** 0.13*** 0.36*** 0.22*** 0.25*** 0.30*** 0.19*** 0.31*** (0.01) (0.02) (0.04) (0.01) (0.03) (0.02) (0.03) (0.05) (0.02) (0.05) Marginal effect (lnPCE*D99) 0.06* 0.06 0.23*** 0.04 0.11** 0.05 0.11** 0.06 0.07 0.05 (0.03) (0.04) (0.06) (0.03) (0.05) (0.04) (0.05) (0.08) (0.05) (0.08) Instrumental variable lnPCE 0.12*** 0.18*** 0.34*** 0.07*** 0.26*** 0.15*** 0.17*** 0.22*** 0.13*** 0.09 (0.02) (0.02) (0.04) (0.02) (0.03) (0.02) (0.03) (0.06) (0.03) (0.06) Marginal effect (lnPCE*D99) 0.12** 0.16** 0.38*** 0.08 0.23*** 0.14** 0.17** 0.07 0.08 0.10 (0.05) (0.07) (0.10) (0.05) (0.08) (0.07) (0.07) (0.12) (0.08) (0.11) *** Significant at the 1 percent level; ** significant at the 5 percent level; * significant at the 10 percent level. Note: Numbers in parentheses are robust standard errors, corresponding to the elasticity estimates. Each column represents a separate regression using a wide range of household-level economic and demographic control variables. All specifications control for cluster fixed effects and year dummy variables. Skoufias, Tiwari, and Zaman Instrumental variable estimates were obtained by instrumenting the natural logarithms of per capita expenditure with household-specific asset indexes. Source: Authors’ analysis based on SUSENAS 1996 and 1999 data. 433 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 434 THE WORLD BANK ECONOMIC REVIEW nutritional quality and availability. Consider the income elasticity of fats, which have the highest elasticity of all nutrients for both rural and urban con- sumers. Among households in Central Java, this reflects the “luxury good” status of foods such as meat and fish. The price increases in 1999 also appear to have severely curtailed households’ ability to obtain other basic nutrients, such as protein, calcium, phosphorous, and iron, as well as overall calories. Table 3 shows elasticity estimates for a restricted sample of the poor (per capita expenditure below the median), which are uniformly higher than for the entire sample population. The IV estimates reveal four broad types of nutrients: those for which income elasticities are statistically indistinguishable from zero Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 in 1996 and 1999, those for which elasticities are statistically indistinguishable from zero in 1996 but significant in 1999, those for which elasticities are statis- tically different from zero but not statistically different between 1996 and 1999, and those for which elasticities are statistically different from zero and statistically different between the two years. Table 4 categorizes nutrients based on the IV results for Central Java. It shows that, unlike for the entire sample, the elasticity estimates for poorer households suggest considerable heterogeneity in the elasticities of nutrients. For the urban poor in Central Java, for example, elasticity in the normal year (1996) is significantly different from zero only for fats. But in the crisis year (1999), elasticity is significantly larger for fats as well as for calories, proteins, carbohydrates, phosphorous, and vitamin B. Elasticities for calcium, iron, and vitamins A and C are never significant, indicating that cash transfers may not be effective vehicles for protecting these nutrients in poorer households.22 Robustness Check: Is It Prices or Other Factors? The empirical specification of equation (13) relies on cross-sectional variation in relative prices to identify income elasticities. Pooling the cross-sectional data from the two survey years allowed the construction of a basic test for whether the estimated elasticities were different in 1996 and 1999. Although a dummy variable for year is included in the specification, which would presumably absorb everything else that changed between the two survey years, there could be linger- ing questions about whether the findings are due to changes in food prices specifically or perhaps driven by other, unrelated effects of the economic crisis. If food prices could be observed at the household level, then the issue could be tested directly. But prices are not observed at the household level. Rather, unit values are observed, and these are known to be contaminated by variations in quality. To confirm that changes in food prices are indeed driving these results, 22. A simple calculation based on IV estimates for all households in urban Central Java shows that a cash transfer of 25 percent of income would increase the consumption of nutrients considered in this article by from 6– 18 percent in the 1999 crisis year—twice the impact range of 3–9 percent for the same set of nutrients in the 1996 noncrisis year. T A B L E 3 . Income Elasticities of Nutrients for Poorer Households in Rural and Urban Central Java, 1999 Area, estimator, and variable Calories Proteins Fats Carbohydrates Calcium Phosphorus Iron Vitamin A Vitamin B Vitamin C Rural Ordinary least square lnPCE 0.29*** 0.36*** 0.74*** 0.22*** 0.45*** 0.24*** 0.35*** 0.28*** 0.19*** 0.23** (0.02) (0.02) (0.06) (0.02) (0.04) (0.03) (0.04) (0.09) (0.03) (0.10) Marginal effect (lnPCE*D99) 0.12*** 0.14*** 0.08 0.14*** 0.11* 0.16*** 0.23*** – 0.04 0.22*** – 0.15 (0.03) (0.04) (0.09) (0.03) (0.06) (0.05) (0.05) (0.13) (0.05) (0.15) Instrumental variable lnPCE 0.03 0.12*** 0.41*** – 0.03 0.11* – 0.14 0.05 – 0.05 – 0.12** 0.37*** (0.03) (0.04) (0.07) (0.04) (0.06) (0.08) (0.05) (0.11) (0.05) (0.12) Marginal effect (lnPCE*D99) 0.04 0.06 0.36** 0.01 0.17 0.03 0.19* – 0.05 0.18 – 0.16 (0.08) (0.09) (0.16) (0.08) (0.11) (0.17) (0.10) (0.24) (0.13) (0.24) Urban Ordinary least squares lnPCE 0.22*** 0.31*** 0.79*** 0.13*** 0.49*** 0.24*** 0.36*** 0.41*** 0.19*** 0.44*** (0.03) (0.03) (0.07) (0.03) (0.06) (0.03) (0.05) (0.11) (0.04) (0.13) Marginal effect (lnPCE*D99) 0.19*** 0.14*** 0.23** 0.20*** 0.10 0.17*** 0.16** – 0.04 0.21*** – 0.10 (0.05) (0.05) (0.10) (0.05) (0.09) (0.05) (0.07) (0.17) (0.06) (0.19) Instrumental variable lnPCE – 0.02 0.05 0.47*** – 0.11 0.22 – 0.00 0.06 0.06 – 0.08 – 0.17 (0.07) (0.08) (0.17) (0.08) (0.15) (0.08) (0.13) (0.31) (0.10) (0.33) Marginal effect (lnPCE*D99) 0.38*** 0.36** 0.76*** 0.35** 0.15 0.35** 0.18 0.18 0.35* – 0.05 (0.13) (0.15) (0.27) (0.15) (0.24) (0.15) (0.21) (0.51) (0.19) (0.52) *** Significant at the 1 percent level; ** significant at the 5 percent level; * significant at the 10 percent level. Note: Numbers in parentheses are robust standard errors, corresponding to the elasticity estimates. Each column represents a separate regression using a wide range of household-level economic and demographic control variables. All specifications control for cluster fixed effects and year dummies. Instrumental variable estimates were obtained by instrumenting natural logarithms of per capita expenditure with household-specific asset indexes. Poorer Skoufias, Tiwari, and Zaman household are defined as the lower half of the distribution based on per capita expenditure. Source: Authors’ analysis based on SUSENAS 1996 and 1999 data. 435 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 436 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Nutrient Classifications for Poorer Households in Central Java Based on the Effectiveness of Cash Transfers Category 2: cash transfers have no Category 3: cash Category 4: cash Category 1: cash effect in normal years, transfers are effective, transfers are generally transfers have no but useful in crisis but no more or less so effective, but no more Area effect years in crisis years so in crisis years Rural Calories, Iron Proteins, Calcium, Fats Carbohydrates, Vitamin B, Vitamin Phosphorous, C Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Vitamin A Urban Calcium, Iron, Calories, Proteins, Fats Vitamin A, Carbohydrates, Vitamin C Phosphorous, Vitamin B Source: Authors’ analysis based on SUSENAS 1996 and 1999 data. Poorer household are defined as the lower half of the distribution based on per capita expenditure. regressions of the following form are analyzed for each nutrient: X &  ' X &  ' Pl Pl ln Yj ¼ b0 þ b1 ln PCEj þ b ln i i þ i gi ln  ln PCEj PCT PCT þ d 0 Xj þ 1j ð14Þ where i indexes four food groups: meat and fish, fruits and vegetables, eggs and milk, and others, PCT is the average of the unit values of cereals and tubers at the village (desa) level, and Pi is the average unit value of each of the four food groups such that the ratios in equation (14) represent the price of each of the four food groups relative to the price of cereals and tubers. As before, Xj represents the set of controls used in previous regressions. Using village-level averages of unit values avoids the problem of quality differences in household-level purchases of food.23 These regressions are estimated for 1996 with the primary goal of assessing 0 whether or not the coefficients on the interaction terms (the gis) are significantly different from zero. These estimations are made for each of the micronutrients and separately for urban and rural central Java. These estimations reveal that at least one interaction term is always statistically different from zero. In add- ition, an F-test of the joint significance of these interactions rejects the null hy- pothesis in almost all instances (see Appendix S4, a supplemental appendix available at http://wber.oxfordjournals.org). This makes an explicit connection between the estimated income elasticity of any one micronutrient and the 23. Deaton (1988) notes that using cluster means of unit values in regressions of this form is essentially the same as using cluster dummy variables to instrument for individual unit values in a regression of shares and prices, where prices are expressed as unit values. Skoufias, Tiwari, and Zaman 437 prevailing relative prices of food in the economy, corroborating the finding that elasticities estimated under one set of relative prices do not necessarily remain valid when relative prices change dramatically. In particular, this also provides additional confidence in the attribution of differences in the income elasticities of micronutrients between the normal year and the crisis year to food prices. I I I . I M P L I CAT I O N S There is considerable heterogeneity in the income elasticity of demand for nutrients over time based on analysis of household data from the 1996 and Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 1999 consumption modules of the SUSENAS in Indonesia. A comparison of OLS and IV estimates of the demand for nutrients suggests that OLS estimates are likely to be misleading due to bias from correlated errors in consumption and nutrient content. In particular, the finding that IV estimates are generally lower than OLS estimates suggests that the upward bias due to correlated measurement errors in nutrient intake and household-level consumption may outweigh the possible attenuation bias. The analysis also shows that for most nutrients, including micronutrients such as phosphorous, iron, and calcium, income elasticity estimates are signifi- cantly higher in a crisis year. Moreover, the magnitude of the increase appears to be larger in urban than in rural areas. On the other hand, for some micronu- trients, such as vitamin C, the income elasticity estimates obtained from the IV specification are statistically indistinguishable from zero. This suggests that income may have limited leverage to increase or protect the consumption of vitamin C, whether in a crisis year or not. The separate analysis of nutrient elasticity for poor households reinforces the message that, some nutrients could be effectively protected using cash transfers during crisis years (such as iron and fats in rural Central Java) while others are unlikely to be responsive to income supplements (such as vitamin A in rural Central Java). These results have two specific policy implications. First, given the signifi- cant increases in the income elasticities of both micronutrients and macronutri- ents during economic crises suggests that cash transfer programs can help households protect their consumption of essential nutrients, with important differences between the urban and rural poor. To the extent that delivery infra- structure is already in place exists and leakage is low, cash transfers are widely accepted as the quickest and cheapest intervention to scale up to reach house- holds most likely to be adversely affected. This research shows that they can also be more effective in protecting the consumption of some key nutrients during economic crises than in normal economic conditions. Second, a complete reliance on cash transfers may be insufficient if the policy goal of policy in response to economic crises is to protect all important nutrients. For example, consumption of vitamin C, an important micronutrient, was unre- sponsive to income in both rural and urban Central Java. This suggests that tar- geted micronutrient supplementation programs, designed with careful attention to 438 THE WORLD BANK ECONOMIC REVIEW differences between the urban and rural poor, might have to accompany cash transfers to ensure that key micronutrients are not sacrificed during crises. Future research could be directed at understanding and identifying specific nutrients that households are likely to sacrifice during a crisis in different settings. A P P E N D I X A . D ATA D E SC R I P T I O N Data from the consumption module of SUSENAS, collected every three years, included 216 food items in 1996 and 214 food items in 1999.24 The seven-day Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 food intake survey makes a very good effort to capture the total value of the food consumed by households. In both years, households were asked to report the quantity and value of food purchased, given to them as gifts, or consumed out of their own production during the previous week.25 Items are valued by local inter- viewers using the prevailing market prices in the villages where households reside. The micronutrient content of each food item is calculated using conversion factors published by the Nutrition Directorate in the Ministry of Health of Indonesia (Direktorat Gizi, Departemen Kesehatan 1988). That publication contains the micronutrient content of a comprehensive list of foods; each was matched with one or more of the food items captured in the SUSENAS con- sumption module. Because both the quantities and calories of each food item are available in the SUSENAS dataset, either may be used to derive the micro- nutrient content. Additional investigation determined that it is preferable to rely on the calorie data rather than the quantity data. In 1996, for example, the quantity of various food items was recorded in kilograms rather than in grams as the questionnaire specified.26 In addition, for a number of food items, quantity was coded in pieces, such as number of eggs, rather than in weight, but calories were provided per unit of weight.27 Similar problems were noted with the coding of quantities in 1999. Because of these issues, the analysis in this article used the calorie informa- tion provided by the BPS to derive a more reliable measure of the quantity of each food item and micronutrient consumed. First, the standard calories-to-quantity conversion formula (also applied by the Central Statistical Agency) was used to derive a new quantity for each food item. Second, the quantity-to-micronutrient formula obtained from the Ministry of Health was applied to derive the quantity of micronutrients for each food item. This approach implicitly assumes that the calorie data are more reliable than the 24. The difference arises from the fact that “high quality” and “imported” rice were treated as separate food items in the cereals category in 1996, but not in 1999. 25. Van de Walle (1988) provides a guide to the SUSENAS consumption module that is still very useful despite some changes in the questionnaire. 26. In 1996 this was the case for food items with codes 45 –52, 95, 102, 110– 114, 126–127, 158–166, 171– 180, 184–185, 187– 194, 208, 212, and 215. 27. In 1996 this was the case for food items with codes 75, 76, 81, 82, 84, 157, 167, and 203. Skoufias, Tiwari, and Zaman 439 original quantity data—a reasonable assumption because the quantity data must have been processed in some way to apply the standard conversion factors to calculate calories. The value of food consumption is the sum of spending on grains, meat and fish, eggs and milk, vegetables, pulses, fruits, seasonings, fats and oils, soft drinks, prepared foods and other food items, and alcohol.28 The reference period for consumption of these items is the seven days preceding the day of the survey interview. Weekly consumption was transformed into monthly con- sumption transformed into monthly consumption by multiplying by (30/7) For nonfood expenditures the survey collects two measures, one for the prior Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 month and one for the previous 12 months. To avoid exclusion errors, average expenditures per month were calculated from the reported expenditures of the last 12 months. Expenditures on nonfood items include tobacco, housing, clothing, health and personal care, education and recreation, transportation and communication, taxes and insurance, and ceremonial expenses. Expenditures on durables such as household furniture, electric appliances, and audiovisual equipment are excluded for aggregate household consumption. A household’s income is measured by monthly per capita consumption, denoted by and calculated by dividing the monthly total of food and nonfood consump- tion in survey period t by the size of the household in the period.29 A P P E N D I X B . F I R S T - S TA G E R E G R E S S I O N S T A B L E B - 1 . First-stage Regressions (1) (2) (1) (2) Variable lnPCE lnPCE*D99 Variable lnPCE lnPCE*D99 Log of household 0.24*** – 0.02*** Household head 0.20*** 0.01 asset index self-employed with permanent assistance (0.01) (0.00) (0.04) (0.02) Log of household – 0.12*** 0.14*** Household head 0.00 0.00 asset index*D99 working without pay (0.01) (0.01) (0.02) (0.01) Number of boys 0.03 0.04 Household head 0.00 0.04* under age 5 literate (0.08) (0.03) (0.07) (0.02) (Continued ) 28. Unlike SUSENAS, this article does not include tobacco expenditures in the food consumption total. 29. It is implicitly assumed that there are no economies of scale at the household level. For comparing income elasticity over time, this assumption is not overly limiting. In any case, the regression analysis controls for the gender and age composition of families in each survey year. 440 THE WORLD BANK ECONOMIC REVIEW TABLE B-1. Continued (1) (2) (1) (2) Variable lnPCE lnPCE*D99 Variable lnPCE lnPCE*D99 Number of girls – 0.11 0.02 Spouse has no – 0.04 0.01 under age 5 education (0.08) (0.03) (0.03) (0.01) Number of boys ages 0.07 0.02 Spouse has not – 0.09*** 0.03*** 6 – 12 completed primary school (0.08) (0.03) (0.03) (0.01) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Number of girls ages 0.06 0.02 Spouse has – 0.10*** 0.02** 6 – 12 completed primary school (0.08) (0.03) (0.03) (0.01) Number of boys ages 0.26*** 0.01 Spouse has – 0.02 0.01 13– 19 completed junior or senior high school (0.08) (0.03) (0.03) (0.01) Number of girls ages 0.27*** 0.03 Spouse 0.00 0.01 13– 19 self-employed without assistance (0.08) (0.03) (0.02) (0.01) Number of men ages 0.28*** – 0.01 Spouse 0.04 0.02* 20– 54 self-employed with nonpermanent assistance (0.07) (0.02) (0.03) (0.01) Number of women 0.14** 0.01 Spouse – 0.07 – 0.09** ages 20– 54 self-employed with permanent assistance (0.06) (0.02) (0.06) (0.05) Household head has – 0.20*** 0.05** Spouse working 0.08** 0.03* no education without pay (0.07) (0.02) (0.04) (0.01) Household head has – 0.14*** 0.00 Spouse with wage 0.01 0.00 not completed employment primary school (0.03) (0.01) (0.02) (0.01) Household head has – 0.09*** 0.01 Spouse literate 0.07** – 0.01 completed primary school (0.02) (0.01) (0.03) (0.01) Household head has – 0.06** – 0.01 Household size – 0.47*** 0.00 completed junior or senior high school (0.02) (0.01) (0.02) (0.01) (Continued ) Skoufias, Tiwari, and Zaman 441 TABLE B-1. Continued Household head – 0.01 0.01 D99 ( ¼ 1 if year is 0.58*** 12.52*** self-employed 1999) without assistance (0.02) (0.01) (0.14) (0.10) Household head 0.04* 0.01 R-squared 0.705 0.999 self-employed with nonpermanent assistance (0.02) (0.01) F-statistic 148.3 24566 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 *** Significant at the 1 percent level; ** significant at the 5 percent level; * significant at the 10 percent level. Note: These regressions also include the control variables that interact with the 1999 year dummy variable. Source: Authors’ analysis based on equation (13). REFERENCES Banks, J., R. Blundell, and A. Lewbel. 1997. “Quadratic Engel Curves and Consumer Demand.” The Review of Economics and Statistics 79 (4): 527–39. Behrman, J. 1988. “Nutrition and Incomes: Tightly Wedded or Loosely Meshed?” Pew/Cornell Lecture Series on Food and Nutrition Policy. Cornell University, Ithaca, NY. ———. 1995. “Micronutrients and Household Behavior: What We Know and What We Don’t Know.” Working Paper 2 on Agricultural Strategies for Micronutrients. International Food Policy Research Institute, Washington DC. Behrman, J., and A. Deolalikar. 1987. “Will Developing Country Nutrition Improve With Income? A Case Study for Rural South India.” Journal of Political Economy 95 (3): 492–507. ———. 1989. “Is Variety the Spice of Life? Implications for Calorie Intake.” The Review of Economics and Statistics 71 (4): 666– 72. Behrman, J., and B. L. Wolfe. 1984. “More Evidence on Nutrition Demand: Income Seems Overrated and Women’s Schooling Underemphasized.” Journal of Development Economics 14 (1– 2): 105– 26. Behrman, J., A. Foster, and M. R. Rosenzweig. 1997. “The Dynamics of Agricultural Production and the Calorie-Income Relationship: Evidence from Pakistan.” Journal of Econometrics 77 (1): 187–207. Bouis, H. 1991. “The Determinants of Household-level Demand for Micronutrients: An Analysis for Philippine Farm Households.” International Food Policy Research Institute, Washington, DC. Bouis, H., and L. Haddad. 1992. “Are Estimates of Calorie-Income Elasticities Too High? A Recalibration of the Plausible Range.” Journal of Development Economics 39 (2): 333 –64. Chernichovsky, D., and O. Meesook. 1984. “Urban-Rural Food and Nutrition Consumption Patterns in Indonesia.” World Bank Staff Working Paper 670. World Bank, Washington DC. Deaton, A. 1988. “Quantity, Quality, and the Spatial Variation of Price.” American Economic Review 78 (3): 418–30. ———. 1997. The Analysis of Household Survey Data. Baltimore, Md.: The Johns Hopkins University Press. Deaton, A., and J. Muellbauer. 1980. “An Almost Ideal Demand System.” American Economic Review 70 (3): 312–26. Deaton, A., and S. Zaidi. 1999. “Guidelines for Constructing Consumption Aggregates for Welfare Analysis.” Princeton University, Princeton, NJ. 442 THE WORLD BANK ECONOMIC REVIEW Gizi, Direktorat, and Departemen Kesehatan. 1988. Daftar Komposisi Bahan Makanan. Jakarta: Penerbit Bhratava karya Aksra. Fan, J. 1993. “Local Linear Regression Smoothers and Their Minimax Efficiencies.” Annals of Statistics 21 (March): 998–1004. FAO (Food and Agriculture Organization). 2008. “Food Outlook.” Rome. http://www.fao.org/docrep/ 011/ai474e16htm. Hausman, J. A. 1978. “Specification Tests in Econometrics.” Econometrica 46 (6): 1251–71. Holly, A. 1982. A Remark on Hausman’s Specification Test.” Econometrica 50: 749– 59. Ivanic, M., and W. Martin. 2008. “Implications of Higher Global Food Prices for Poverty in Low-income Countries.” Agricultural Economics 39 (s1): 405– 16. Levinsohn, J., S. Berry, and J. Friedman. 1999. Impacts of the Indonesian Economic Crisis: Price Changes Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 and the Poor. NBER Working Paper 7194. Cambridge, MA: National Bureau of Economic Research . Pitt, M. M. 1983. “Food Preferences and Nutrition in Rural Bangladesh.” The Review of Economics and Statistics 65 (1): 105–14. Pitt, M. M., and M. R. Rosenzweig. 1985. “Health and Nutrient Consumption across and within Farm Households.” The Review of Economics and Statistics 67 (2): 212– 23. Ravallion, M. 1990. “Income Effects on Undernutrition.” Economic Development and Cultural Change 38 (3): 489 –515. Sari, M., S. de Pee, M.W. Bloem, K. Sun, A. L. Thorne-Lyman, R. Moench-Pfanner, N. Akhter, K. Kraemer, and R. D. Semba. 2010. “Higher Household Expenditure on Animal-Source and Nongrain Foods Lowers the Risk of Stunting among Children 0-59 Months Old in Indonesia: Implications of Rising Food Prices.” Journal of Nutrition 140 (1): 195–200. ——. 2003. “Is the Calorie-Income Elasticity Sensitive to Price Changes? Evidence from Indonesia.” World Development 31 (7): 1291–1307. Strauss, J., and D. Thomas. 1995. “Human Resources: Empirical Modeling of Household and Family Decisions.” In Handbook of Development Economics vol. 3, ed. J. Behrman, and T. N. Srinivasan. Amsterdam: North-Holland. Studdert, L., E. Frongillo, and P. Valois. 2001. “Household Food Insecurity Was Prevalent in Java during Indonesia’s Economic Crisis.” Journal of Nutrition 131 (10): 2685–91. Subramanian, S., and A. Deaton. 1996. “The Demand for Food and Calories.” Journal of Political Economy 104 (1): 133–62. Suryahadi, A., S. Sumarto., Y. Suharso, and L. Pritchett. 2000. “The Evolution of Poverty during the Crisis in Indonesia, 1996 to 1999 (Using Full SUSENAS Sample).” SMERU Working Paper. Social Monitoring and Early Response Unit, Jakarta. Thomas, D., E. Frankenberg, K. Beegle, and G. Teruel. 1999. “Household Budgets, Household Composition, and the Crisis in Indonesia: Evidence from Longitudinal Household Survey Data.” Paper presented at the 1999 Population Association of America Meetings, RAND, Santa Monica, CA. Timmer, P. 1981. “Is There Curvature in the Slutsky Matrix?” The Review of Economics and Statistics 62 (3): 395 –402. Timmer, P., and H. Alderman. 1979. “Estimating Consumption Parameters for Food Policy Analysis.” American Journal of Agricultural Economics 61 (S): 982–94. Timmer, P., W. P. Falcon, and S. Pearson. 1983. Food Policy Analysis. Published for the World Bank. Baltimore, Md.: The Johns Hopkins University Press. Tiwari, S., and H. Zaman. 2010. “The Impact of Economic Shocks on Global Undernourishment.” Policy Research Working Paper 5215. World Bank, Washington, DC. Van de Walle, and Dominique. 1988. “On the Use of the SUSENAS for Modeling Consumer Behavior.” Bulletin of Indonesian Economic Studies 24 (2): 107–22. Wodon, Q., and H. Zaman. 2010. “Higher Food Prices in Sub-Saharan Africa: Poverty Impact and Policy Responses.” The World Bank Research Observer 25 (1): 157– 76. Economic Geography and Economic Development in Sub-Saharan Africa Maarten Bosker and Harry Garretsen Sub-Saharan Africa’s (SSA) physical geography is often blamed for its poor economic per- formance. A country’s geographical location does, however, not only determine its agri- cultural conditions or disease environment. It also pins down a country’s relative position Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 `-vis other countries, affecting its ease of access to foreign markets. This paper assesses vis-a the importance of market access for manufactures in explaining the observed income dif- ferences between SSA countries over the period 1993–2009. We construct yearly, theory- based measures of each SSA country’s market access using the information contained in bilateral manufacturing trade flows. Using these measures, we find a robust positive effect of market access on economic development that has increased in importance during the last decade. Interestingly, when further unraveling this finding, access to other SSA markets in particular turns out to be important. JEL codes: O10, O19, O55, F1 Sub-Saharan Africa (SSA) is home to the world’s poorest countries. Alongside factors such as poor institutional quality, low (labour) productivity and low levels of human capital, the region’s geographical disadvantages are often viewed as an important determinant of its dismal economic performance. A country’s geography directly affects economic development through its effect on disease burden, agricultural productivity, the availability of natural resources, or its accessibility (see Gallup and others 1999; Collier and Gunning, 1999; Ndulu, 2007; Nunn and Puga, 2011). Recently, the new economic geography (NEG) literature (see Krugman, 1991; Fujita and others 1999; World Bank 2008) has highlighed another mechanism through which geography affects a country’s prosperity. A country’s Maarten Bosker (bosker@ese.eur.nl; corresponding author) is assistant professor at the Erasmus University Rotterdam. He is also affiliated with the CEPR, the Tinbergen Institute and Utrecht University, The Netherlands. Harry Garretsen ( j.h.garretsen@rug.nl) is professor at the University of Groningen, The Netherlands. He is also affiliated with Cambridge University and CESifo. The authors thank Rob Alessie, Bernard Fingleton, Henri Overman, Giacomo Pasini, Joppe de Ree, Steve Redding, Marc Schramm and seminar participants in Cambridge, Glasgow, Milan, Oxford, Rome, Rotterdam, Savannah, and Utrecht for useful comments and discussions on an earlier version of this paper. In particular we thank the editor, Elisabeth Sadoulet, and three anonymous referees for very helpful remarks that have significantly improved our paper. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 443 –485 doi:10.1093/wber/lhs001 Advance Access Publication February 14, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 443 444 THE WORLD BANK ECONOMIC REVIEW location not only determines its physical geography; it also pins down its pos- ition on the globe vis-a ` -vis all other countries (its relative geography). The NEG literature in particular emphasizes the important role of relative geog- raphy in determining a country’s access to international markets. It predicts that the better this market access, the higher a country’s level of income. Redding and Venables (2004) were the first to establish empirically that market access indeed matters for economic development. Based on the estima- tion results for a worldwide sample of 101 countries, they find for example that if Zimbabwe were located in central Europe, the resulting improvement in its market access would ceteris paribus increase its per capita income by almost Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 80 percent. Subsequently, several other studies confirmed the positive effect of market access on economic development (e.g. Knaap 2006; Breinlich 2006; or Mayer, 2008). More recently, it has also been found to hold for developing countries (see Deichmann, Lall, Redding and Venables, 2008 for a good over- view). Amiti and Cameron (2007) show that wages are higher in Indonesian districts that enjoy better market access, Hering and Poncet (2010) and Bosker and others (2010) find similar evidence in case of Chinese cities, and Fally and others (2010) do so for Brazilian states.1 The importance of relative geography in shaping global and regional patterns of economic development has also not gone unnoticed in policy circles: it was the main topic of the World Bank’s 2009 World Development Report (World Bank, 2008). Despite the attention given to the role of economic geography in shaping patterns of economic development in both the developing and developed world, we are unaware of a study that empirically establishes whether, and if so, to what extent, it can help to explain the observed differences in economic development between SSA countries.2 SSA is only a marginal player in the world’s export and import markets. Since 1970, the region’s share in global trade has declined from about 4 percent to a mere 2 percent in 2005 (IMF, 2007). Through their detrimental effect on market access, high trade costs are generally viewed as one of the main causes for SSA’s poor trade performance (see Freund and Rocha, 2011; Collier, 2002; Foroutan and Pritchett, 1993; Coe and Hoffmaister, 1999; Limao and Venables, 2001; Amjadi and Yeats, 1995; Portugal-Perez and Wilson, 2008). Increasing SSA participation in world markets, as well as 1. Moreover, Amiti and Javorcik (2008) find that market access positively affects the amount of FDI in Chinese provinces and Lall, Shalizi and Deichmann (2004) show that market access is an important determinant of firm level productivity in India. 2. The only paper we know of focusing on the role of market access in SSA is Elbadawi, Mengistae and Zeufack (2006) that shows that differences in terms of export performance between firms in 10 SSA countries and firms in other developing countries (e.g. India, China, Malaysia or Peru) can partly be explained by SSA’s poor market access. Their paper does not link export performance — or market access — to income per capita. Another paper that is similar in spirit to ours is that by Arora and Vamvakidis (2005), which looks at how South Africa’s economy influences development in the rest of SSA. Bosker and Garretsen 445 stimulating trade relations between SSA countries, is viewed as very important to its future economic success (IMF, 2007; World Bank 2007; or UNCTAD 2010). Especially, expanding the (exporting) manufacturing sector is seen as crucial to the region’s chances on future economic success (Collier and Venables, 2007; IMF, 2007; World Bank, 2007). It has been one of the key ingredients of the sustained growth witnessed in the rapidly developing Asian countries (see e.g. Johnson et al., 2006; or Jones and Olken, 2008). Developing an exporting manufacturing sector will not only help to diversify SSA coun- tries’ export portfolio, making them less vulnerable to price fluctuations on world commodity markets, it is also expected to increase overall productivity Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 through increased knowledge spillovers and learning by doing (Van Biesebroeck, 2005; Bigsten and So ¨ derbom, 2006). As a result, improving the region’s market access by investing in infrastruc- ture, stimulating regional integration, or providing preferential access to European and U.S. markets are all seen as a vital ingredients for improving SSA’s trade potential and its overall economic performance (Buys et al., 2010; UNCTAD, 2009, 2010; Frazer and van Biesebroeck, 2010; Freund and Rocha, 2011). Against this background the main contribution of our paper is to empirically establish the importance of SSA market access, and market access for manufac- tures in particular, for its economic development over the last two decades.3 To do this, we follow the empirical strategy introduced by Redding and Venables (2004) that is firmly based in the theoretical new economic geography (NEG) literature. We first construct yearly measure(s) of each SSA country’s market access over the period 1993–2009, making use of bilateral manufactur- ing export data involving at least one SSA country. Next, and by using our con- structed measure(s) of market access, we estimate the impact of market access on GDP per worker. We do adapt the Redding and Venables (2004) strategy in three different ways. First, when constructing our market access measure(s) using the information contained in bilateral trade flows, we take explicit account of the fact that most SSA countries only trade with a fraction of their possible trade partners. Second, our 17-year sample period allows us to use panel data methods and control for time-invariant unobserved heterogeneity in all our estimations (hereby most notably capturing all possible differences in SSA countries’ physical geography). Finally, we distinguish explicitly between the importance of access to other SSA markets and to markets in the rest of the world (ROW). Overall, our main finding is that economic geography is an important deter- minant of SSA’s economic development. Even after controlling for many other posited explanations of SSA’s poor economic performance such as its physical geography, education levels, or institutional quality, market access for 3. Throughout the paper, unless explicitly noted otherwise, market access refers to a country’s market access for manufactures. 446 THE WORLD BANK ECONOMIC REVIEW manufactures has a significant positive effect on GDP per worker. The effect of market access that we find for SSA countries is, however, significantly lower than that found in comparable studies looking at Brazil (Fally and others 2010), Indonesia (Amiti and Cameron, 2007), or China (Hering and Poncet, 2010). But, although lower than those found in other parts of the world, our results do show that the effect of market access has increased markedly in SSA over the last two decades: the positive relationship between market access and economic development is strongest in the second half of our sample period. Interestingly, when further unraveling this finding by distinguishing between the importance of access to other SSA markets and to markets in the rest of the Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 world (ROW), we find that it is the variation in access to other SSA markets in particular that drives our findings. ROW market access loses its significance after controlling for other (more standard) explanations for SSA’s poor eco- nomic performance. It confirms the (increased) importance of SSA markets for SSA’s own economic development (see also Easterly and Reshef, 2010, foot- note 2; UNCTAD, 2009, 2010). Finally, based on our estimation results, we tentatively ‘decompose’ the con- tribution of policy-relevant variables to overall market access, and look at the predicted effect of several ( policy induced) changes aimed at improving SSA market access. This shows that improving SSA infrastructure (see also Buys and others, 2010), alleviating the burden of landlocked countries, and increasing re- gional economic integration, all positively affect a country’s market access in varying degrees, carrying important benefits for its economic development. I. ECO N O M I C D E V E LO P M E N T A N D MA R K E T AC C E S S: T H E O R E T I CA L FR AM EWOR K A ND EMPIRICAL STRATEGY At the heart of our analysis lies the theoretical relationship between market access in manufactures and income levels that follows from standard economic geography theory. Referring to Appendix C for a more formal exposition of the NEG model that underlies our analyses,4 this relationship is shown in log- linear form in equation (1) below (corresponding to equation (C5) in Appendix C, with an added subscript t to denote years): MAijt R zfflfflfflfflfflffl}|fflfflfflfflfflffl{ X ð1ÀsÞ ln wit ¼ gt þ x1 ln cit þ x2 ln m jt Tijt ¼ gt þ x1 ln cit þ x2 ln MAit ð1Þ j |fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl} MAit 4. See Fujita, Krugman, and Venables (1999), Puga (1999), Head and Mayer (2004) for more detailed expositions of various basic NEG models. See also Head and Mayer (2010), who show that the relationship between market access and economic development not only follows from NEG models but can be derived from a more general class of models. Bosker and Garretsen 447 Equation (1) is the so-called wage equation that lies at the heart of virtually all empirical NEG studies (see e.g. Hanson, 2005; Redding and Venables, 2004; Knaap, 2006 and Amiti and Cameron, 2007; Hering and Poncet, 2010). It pre- dicts that wages in country i in year t, wit, are higher the better a country’s pro- duction efficiency, cit, and, most importantly for our present purposes, the better its so-called real market access MAit.5 This market access is a trade cost (Tij ) weighted sum of all countries’ market capacities (mj ). Each country j’s contribution to country i’s market access con- sists of country j’s market capacity, a reflection of its real spending power, weighted by the level of trade costs incurred when shipping goods from Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 21 country i to country j, i.e. MAij ¼ mj /Ts ij . The closer (or better connected) a country is to world markets, the better its market access. It is equation (1) that constitutes the backbone of our empirical analysis into the relevance of market access for SSA economic development (see section III). Estimating the Wage Equation: The Redding and Venables (2004) Approach Estimating equation (1) is not as straightforward as it may seem. The difficulty comes from the fact that the market access term, MAit, is not directly observed. Two estimation strategies have been proposed to deal with this issue. The first strategy follows Hanson (2005) and estimates equation (1) directly. A drawback of this method is that it requires additional assumptions on how to proxy each country’s market capacity, mj. Particularly problematic in this respect is the fact that a country’s price index of manufacturing varieties, one of the two main components of mj (see Appendix C) is not directly observed.6 This is why we base our empirical analysis on the second proposed strategy to estimate (1). This method does not face the problem of having to make ad-hoc assumptions on how to proxy a country’s market capacity. It was first introduced by Redding and Venables (2004)7 and involves a two-step 5. In Redding and Venables (2004) and Knaap (2006), each firm also uses a composite intermediate input (made up of all manufacturing varieties) in production, allowing them to also look at the relevance of so-called supplier access for income levels. Since our goal is to establish the relevance of market access we, in line with Breinlich (2006), skip intermediate inputs and thereby ignore supplier access [this also has the advantage that we avoid the multicollinearity problems when including both market and supplier access in the estimations, see Redding and Venables (2004) and Knaap (2006)]. In this respect our derivation and application of the wage equation is similar to Hanson (2005), see also Head and Mayer (2004, pp. 2622–2624), or Head and Mayer (2010). As a robustness check, Table 4c shows results when also including constructed measures of supplier access to (1). 6. Moreover, this direct estimation strategy jointly identifies the relative importance of the different components making up a country’s overall market access and the overall effect of market access on income levels. It does so solely from the spatial distribution of GDP ( per capita) across countries. The nonlinear nature of (1) makes this an impossible task without putting a priori restrictions on (some) of the parameters (see e.g. Amiti and Cameron, 2007). Econometrically, the parameter on market access, x2, and the parameters within the market access term (e.g. s) are not separately identified when directly estimating (1). 7. Other papers using this strategy include inter alia Knaap (2006), Breinlich (2006), Mayer (2008), Head and Mayer (2006, 2010), Hering and Poncet (2010) or Bosker and Garretsen (2010). 448 THE WORLD BANK ECONOMIC REVIEW procedure. In a first step, additionally collected information contained in (bilat- eral) trade data is used to provide estimates of the (relative) role of trade costs, Tij, and market capacity, mj, in determining a country’s market access. The way this is done is firmly based on NEG theory. As derived in Appendix C, equation (C6) shows that the connection between bilateral exports and market access follows directly from a standard NEG model (where we have again added a subscript t to denote years). ð1ÀsÞ EXijt ¼ sit ½m jt Tijt Š ð2Þ |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 MAijt Exports EXij from country i to country j depend on the ‘supplier capacity’ of the exporting country, si (see (C6) for its definition), the market capacity of the importing country, mj, and the magnitude of bilateral trade costs Tij between the two countries. Comparing MAij in (1) and (2) immediately shows that we can estimate (2), and use the resulting predicted values of MAijt to construct yearly measures of each country’s market access. More formally: A. Estimate the bilateral export equation (2) in log-linear form using informa- tion on bilateral export flows and trade costs, capturing each country’s supplier and market capacity by a full set of importer-year and exporter-year dummies: ln EXijt ¼ rit þ m jt þ b ln Tijt þ 1ijt ð3Þ B. Use m ^ jt and b ^ , the estimated parameters on the included importer-year dummies and on trade costs respectively, to construct each country’s market access based on the direct relationship between bilateral exports and market access [again compare (1) to (2)]: X ^b ^ it ¼ MA expðm ^ jt ÞTijt ð4Þ j |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} ^ ijt MA The constructed measure of each country’s market access shown in (4) is then used in the second step to get an estimate of the impact of market access on income levels: ^ it þ h ln wit ¼ gt þ x2 ln MA ð5Þ it where the error term hit in (5) captures a country’s level of technological efficiency [cit in (1)]. The estimated value of x2, together with its standard deviation, is the most important parameter for the purpose of our paper. It provides us with an in- dication of the size, sign, and significance of the effect of market access on income levels. In the next two sections we implement the above-described two-step method in order to verify the importance of market access for SSA economic Bosker and Garretsen 449 development over the last two decades (1993–2009). In section II, we focus on estimating the trade equation and constructing our measure(s) of market access. Next, in section III, we estimate the wage equation making use of our constructed measures of market access and show that market access is of in- creasing importance in understanding the observed differences in economic de- velopment between SSA countries. Moreover, we decompose each country’s market access into access to other SSA markets and to markets in the ROW, and show that having good access to other SSA markets has become particular- ly important over the last two decades. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 I I . E S T I M AT I N G THE T R A D E E Q U AT I O N AND CONSTRUCTING M A R K E T AC C E S S The starting point of our empirical analysis is the trade equation (3) capturing each country’s supplier and market capacity by an exporter-year and importer- year dummy respectively. In order to estimate (3), we collected information on yearly bilateral manufacturing exports to and from SSA countries over the period 1993–2009.8 We take this data from the UN COMTRADE database, focusing on manufacturing goods as defined by the Standard International Trade Classification (SITC Rev.3). It contains information on bilateral manu- facturing export flows from each SSA country to and from 47 other SSA coun- tries and 153 countries in the rest of the world. We think a particular focus on manufacturing exports is warranted for two important reasons. First, it most closely follows the NEG theory that underlies our analysis. NEG theory only predicts a relationship between market access in the manufacturing sector and income levels (see Appendix C). It is not evident from theory that a similar relationship should hold for primary goods’ market access (trade patterns of which are more likely to be dominated by more stand- ard comparative advantage or Heckscher-Ohlin type forces). Taking total export flows when estimating (3) is likely to give biased estimates of the para- meters needed to build our measures of market access, particularly in SSA where overall exports are dominated by exports of natural resources and/or agriculture (although this dominance varies substantially between SSA coun- tries, see Figure B2 and Table B2 in Appendix B).9 8. See Appendix A for a full list of variables (including data sources) that we use in our analysis. 9. This problem is much less present when looking at different samples of countries (e.g. Europe, North America and even South-East Asia and parts of Latin America) where trade is dominated by manufacturing goods. We have also done our analysis using total bilateral SSA exports as the basis for constructing our market access measures. When using these measures we do not find a significant effect of market access on economic development (results are available upon request). This could be an indication that relative location to markets for a country’s natural resources (which dominate SSA exports to the rest of the world) does not matter. However, given that (as stressed in the main text) a theoretical underpinning of a relationship between market access for primary products and economic development is lacking, we decided not to particularly stress this finding. We do add controls for a country’s economy’s dependence on natural resources when estimating (5). 450 THE WORLD BANK ECONOMIC REVIEW Second, developing the (exporting) manufacturing sector is viewed by many as crucial to the region’s chances on future economic success (Collier and Venables, 2007; IMF, 2007; World Bank, 2007). Previous spells of sustained growth (mostly experienced by Asian countries) were all accompanied by a rapid expansion of international trade, and trade in manufacturing goods in particular (see e.g. Johnson and others 2006 or Jones and Olken, 2008). In this respect it is interesting to note that manufacturing goods already dominate SSA exports to the rest of Africa (see UNCTAD, 2009; 2010 [see Annex 4.5 for a country-by-country overview]). They constitute an average 40 percent of total intra-SSA exports. When disregarding primary exports (fuel, ores, minerals, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 etc.), that account for roughly 75 percent of total exports to the ROW, the same is true of SSA exports to the ROW. Next, we need to decide on how to measure trade costs, Tijt. The NEG-model does not specify trade costs in any way (except that they are of the iceberg type). In the absence of actual trade cost data and following the modern empirical trade and economic geography literature (see e.g. Anderson and van Wincoop, 2004; Limao and Venables, 2001; Redding and Venables, 2004; Bosker and Garretsen, 2010), we specify Tijt to be a multiplicative function10 of the follow- ing observable variables that are commonly used in the literature11: bilateral dis- tance (Dij ), sharing a common border (Bij ), a common language (CLij ), or a common colonial heritage [distinguishing between sharing a common colonizer (CCij ) and having had a colony-colonizer relationship (CRij )], and finally a dummy variable indicating membership of the same African regional or free trade agreement (RFTAijt) in year t (see Appendix A for a full list of the RFTAs included in this definition). In loglinear form this amounts to substituting the following trade costs specification for b ln Tijt in (3): b ln Tijt ¼ d1 ln Dij þ d2 ln Bij þ d3 ln CLij þ d4 ln CCij þ d5 ln CRij þ d6 RFTAijt ð6Þ Overall, this results in the following bilateral export equation that we estimate for each year of our sample period separately (which explains the added sub- script t to all coefficients): ln EXijt ¼ rit þ m jt þ d1t ln Dij þ d2t ln Bij þ d3t ln CLij þ d4t ð7Þ ln CCij þ d5t ln CRij þ d6t RFTAijt þ 1ijt 10. This is the usual choice in the gravity literature (see e.g. Limao and Venables, 2001; Subramanian and Tamarisa, 2003). See Hummels (2001) for a critique on this, arguing in favor of an additive specification instead. 11. Tariffs are also an important component of trade costs. However, using the available SSA tariff data available in UN TRAINS, reduces our sample from 16560 to 546, 1251 or 4859 for 1993, 2000 and 2009 respectively. For this reason we excluded tariffs from our trade cost specification. Bosker and Garretsen 451 Equation (7) forms the basis for constructing our yearly measures of each SSA country’s market access over the 1993–2009 period. Estimating the Trade Equation. Dealing with the ‘Zeroes’ in Bilateral SSA Trade The actual estimation of (7) raises a number of issues of its own. In particular, the presence of zero trade flows complicates matters. The average SSA country exports manufacturing goods to only 20 percent of possible partner countries, so that about 80 percent of bilateral SSA manufacturing export flows are Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 zeroes. And, although the number of zeroes drops over our sample period (from 89 percent in 1993 to 77 percent in 2003), it does complicate matters when estimating the parameters of (7) that we need to construct our measures of market access. Failing to adequately take account of these zeroes results in inconsistent estimates of these parameters, and thus in wrongly constructed market access measure(s). To deal with these zero observations, several estimation strategies have been proposed that each have their (dis)advantages. We follow Helpman and others (2008) and use a Heckman 2-step estimation strategy to estimate the para- meters of (7). This method has the virtue of not having to impose exogenous sample selection, that is, assuming that there is no unobserved variable related to both the probability to trade and the amount of trade [as e.g. discarding the zero observations and applying OLS on the non-zeroes only, or applying zero-inflated Poisson or negative binomial methods do]. Nor do we have to assume a priori that the exact same model explains both the zero and the non-zero bilateral trade flows [as e.g. using Tobit, NLS or pseudo-Poisson (PPML) techniques would imply].12 The Heckman 2-step procedure amounts to first estimating, using probit, how each of the included explanatory variables affects the probability to trade. Next, in the second stage, the effect of each variable on the amount of trade is estimated, including the inverse Mills ratio (constructed using the results from the first step) to control for the endogenous selection bias that would plague the results when simply discarding the non-zero observations (see for instance ch. 17 in Wooldridge, 2003). However, using the Heckman 2-step procedure is also not free of assumptions: results are only convincing when one can rely on a valid exclusion restriction: a valid exclusion restriction: that is, having at least 12. Note that, due to the assumed CES utility function, the NEG model set out in Appendix C in principle implies that each country trades at least something with every other country. This implies that using the NEG trade equation in explaining both the zero and the non-zero trade flows ascribes the zero observations to the error term only (relying on arguments of measurement error or reporting errors, see also Santos Silva and Tenreyro, 2006, p. 643). Although maybe defendable when looking at samples with a limited amount of ‘zeroes’, we think this is very unlikely in our SSA case, where about 80 percent of the observations are zeroes. 452 THE WORLD BANK ECONOMIC REVIEW one variable that determines the probability to trade but not the amount of trade (see Wooldridge, 2003, p. 589).13 The choice of such a variable is generally quite difficult. However, in our case we can build on a recent paper by Helpman and others (2008), and use their suggested measure of the religious similarity of two countries as the vari- able explaining the probability to trade but not the amount of trade condition- al upon trading. The economic rationale behind the use of this variable is firmly based on recent trade models that show that in order to trade at all, exporters have to be able to cover the fixed costs of exporting. The higher these costs between two countries, the higher the probability of not observing Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 any bilateral trade between them. Helpman and others (2008, p. 466) show that religious (dis)similarity serves as a useful proxy of these fixed costs,14 and they moreover show convincing evidence that, econometrically, it can not be rejected as a valid ‘instrument’.15 Using Helpman and others’s (2008) religious similarity variable to fulfill the (necessary) exclusion restriction, we estimate (7) for each year in our sample period separately.16 To explicitly allow intra-SSA trade to be differently influ- enced by our included variables, we interact all variables, including all importer- and exporter-dummies, with an intra-SSA trade dummy-variable. Table 1 shows the results. In Table 1, the postfix “- SSA” denotes that a vari- able is interacted with this intra-SSA trade dummy. Significance of an “SSA”-variable indicates a significantly different effect of that particular 13. Another disadvantage of the Heckman two-step method is that it does not adequately take account of the heteroscedasticity inherently present in bilateral trade data (see Santos Silva and Tenreyro, 2006). However, we think that the disadvantages of the current methods available that do do this (see the discussion in the text and also footnote 12), that is, either assuming exogenous sample selection (OLS on the non-zeroes or zero-inflated Poisson) or imposing that the zero trade flows are the result of measurement or reporting errors (PPML), do not outweigh the ability of the Heckman two-step procedure to take account of endogenous sample selection. 14. We use religious similarity in all main results presented in this paper. We have also looked into the possibility of using the other two ‘instruments’ proposed by Helpman and others (2008): “the number of days and procedures needed to start a business” and “the costs incurred when starting a business.” We constructed the same two variables using the available data in the World Bank’s “Doing Business Survey”. However, using these two alternative variables in the first stage reduces our sample size significantly (data are missing for the period up to 2003, and for the 2003– 2009 period using them reduces the average yearly sample size from 16560 to 12781). Moreover, we find that, in case of our sample, these two variables are poor predictors of the probability to trade in the first stage of our Heckman-estimation strategy. Results are available upon request. 15. Helpman and others (2008) also show the importance of taking explicit account of firm heterogeneity when estimating the trade equation. We decided not to do this in this paper because it is not clear what the consequences are of introducing firm heterogeneity in the NEG model that we use, and for the wage equation (5) in particular. It lies beyond the current scope of the paper to develop a fully-fledged NEG model incorporating firm heterogeneity. As such, we decided to stick to the more standard NEG model used in all previous empirical work looking at the relationship between market access and economic development, and refrain from explicitly incorporating firm heterogeneity into our analysis. This is certainly not to deny that it would be a very interesting avenue for future research. 16. Although religious similarity itself does not change over time, we hereby do allow its effect on the probability to trade to differ between each year in our sample. Bosker and Garretsen 453 T A B L E 1 . Trade Equation with Importer and Exporter Dummies 1993– 2009 Dep: ln Manuf Exports Coefficients eq. (7) 0/1 trade ln dist 2 1.96*** 2 0.97*** [0.00] [0.00] ln dist - SSA 2 0.44* 2 0.38 [0.24] [0.14] Contiguity 0.92 1.16* [0.38] [0.18] Contiguity - SSA 0.33 2 0.56 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 [0.57] [0.36] Com. lang. 0.81*** 0.45*** [0.00] [0.00] Com. lang. - SSA 0.23 0.02 [0.42] [0.50] Com. col. 0.76*** 0.30** [0.003] [0.03] Com. col. – SSA 2 0.18 0.42 [0.51] [0.11] Colonizer 1.55*** 0.80 [0.00] [0.19] RFTA 0.73* 2 0.25 [0.32] [0.43] RFTA - SSA 0.62 0.44 [0.32] [0.60] Religion HMR 2 0.613** 2 [0.034] P-value Mills’ ratio [0.000]*** 2 nr. obs. 16560 Censored 13344 (80%) Uncensored 3216 (20%) Notes: We report estimated coefficients and not marginal effects. Marginal effects are avail- able upon request. The coefficients are used as input in the construction of our market access measures. These coefficients actually differ by year, the numbers reported in the Table are the mean coefficients, mean p-values (in brackets), and mean number of observations over the period specified. *, **, *** denotes significant at the 5 percent level in at least 50, 80, or 100 percent of the years. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. variable on intra-SSA trade than on SSA trade with the ROW. The coefficients give the overall effects of each of the included variables on the amount of trade (after taking the first stage into account) and the results for 0/1 trade refer to the estimated coefficients in the first stage probit estimations. First of all, the final rows of Table 1 show that also in case of our trade sample restricted to SSA bilateral exports, the usefulness of the Helpman and others (2008) approach can not be rejected:17 religious similarity has a 17. Note that, as in case of two stage least squares, one can never fully test the validity of religious similarity as our ‘instrument’. This ultimately hinges upon believing the arguments put forward by Helpman and others (2008) in favour of using this variable to satisfy the needed exclusion restriction. 454 THE WORLD BANK ECONOMIC REVIEW significantly positive effect on the probability to trade18 and, moreover, the inverse Mills’ ratio is significant in the second stage (hereby not rejecting the need to take account of endogenous sample selection). Turning to the results on the importance of our included trade costs vari- ables,19 we confirm the standard result that distance negatively affects the amount of trade between countries. We do not find convincing evidence that the penalty on distance is significantly higher for intra-SSA trade (see also Foroutan and Pritchett, 1993).20 We only find a significantly higher distance penalty for intra-SSA trade in 50 percent of our sample years. Interestingly, we find this significantly higher distance penalty during the later years in our Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 sample in particular, suggesting that the improvements in SSA trade costs that have been made in recent years have been biased towards improving trade costs with the ROW instead of better connecting the sub-continent. Second, we do not find evidence of a border effect in SSA trade. For SSA trade with the ROW this may not be that surprising. The only SSA countries that border non-SSA countries are those bordering North African countries, and SSA countries trade less with these countries than with other non-African countries (see e.g. IMF, 2007). The lack of a “border-effect” is arguably more surprising for intra-SSA trade given that studies looking at other parts of the world (e.g. Europe or the United States) usually find strong evidence that neigh- bors trade disproportionately more with each other. By contrast, we do find strong effects of language and colonial history on trade volumes of SSA countries. Sharing a colonial history has a strong positive effect on the amount of trade. Especially SSA trade with its former colonizer(s) is much higher than trade with other countries in the world. Having a common colonizer also boosts bilateral trade, and this effect is not significantly different for intra-SSA trade compared to trade with other former colonies in the ROW. Sharing a common language also stimulates both intra-SSA and SSA-ROW trade in largely the same way (see also e.g. Foroutan and Pritchett, 1993; or Coe and Hoffmaister, 1999). Finally, we do not find very convincing evidence that SSA trade in manufac- tures benefits significantly from the many RFTAs that are in existence on the sub-continent. We only find a significant positive effect of having an RFTA on export volumes in 50 percent of the years in our sample period. It is interesting 18. Religious similarity is significant at the 1 percent level in all years, except for 1993 ( p-value ¼ 0.43), 1994 and 1995 (in both years it is significant at the 10 percent level). 19. Given our main interest in the estimated coefficient of the trade equation in the second stage for our main purpose to construct various market access measure(s), the results of the first stage probit estimation are not explicitly discussed. 20. Note that comparing our findings to other studies looking at SSA-trade (e.g. Foroutan and Pritchett, 1993; Coe and Hoffmaister, 1999; Subramanian and Tamirisa, 2003 or Limao and Venables, 2001) is difficult due to the difference in estimation strategy used. These other studies use, for example, Tobit or NLS techniques to estimate the trade equation. Moreover, they usually do not include importer-year and exporter-year dummies in their regressions. Bosker and Garretsen 455 to note however, that we find these significant positive effects of RFTAs on export volumes for the latest years in the sample (2006–2009) in particular. A tentative indication that African RFTAs, many of whom often only exist(ed) on paper, could be becoming more effective in coordinating policies favorable to trade (see also UNCTAD, 2009). Constructing Market Access, Distinguishing Between Access to SSA and to the ROW Using the yearly-estimated coefficients of (7) and the relationship between the trade equation in (2) and market access in (1), the next step is to construct Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 market access using (4). Besides calculating overall market access, we also look at three different subcomponents of market access. Following among others Redding and Venables (2004), Breinlich (2006) or Head and Mayer (2010), we distinguish explicitly between the respective contribution of domestic market access (DMA) and foreign market access (FMA). Furthermore, in order to be able to distinguish between the relevance of access to other SSA markets and to markets in the rest of the world (ROW) respectively, we in turn split foreign market access (FMA) into access to other SSA markets and access to ROW markets: MAit ¼ DMAit þ MASSA it þ MAit ROW ; ð8Þ ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflffl FMAit P R P R where MASSA it ¼ MAijt , MAROW it ¼ MAijt and DMAit ¼ MAiit . We j[SSA;j=i jÓSSA construct the different components of a country’s total market access for each year separately according to (4), adapted to take into account the estimated parameters of (7): Xh ^1t ^ ROW ¼ MA ^ jt ÞDd expðm ^ ^ ^ it ij expðd2t Bij þ d3t CLij þ d4t CCij jÓSSA i ð9Þ ^5t CRij þ d þd ^6t RFTAijt Þ X h ^ SSA ^ SSA ¼ d ^ SSA CLij þ d ^ SSA Bij þ d ^ SSA CCij MA it ^ SSA expðm jt ÞDij 1t expðd 2t 3t 4t j[SSA;j=i i ^ SSA RFTAijt Þ ^ SSA CRij þ d þd 5t 6t ^ it ¼ expðm SSA ðd^ SSA =2Þ ^ SSA þ d ^ SSA þ d ^ SSA Þ; ^ SSA þ d DMA ^ it ÞDii 1t expðd 2t 3t 4t 6t 456 THE WORLD BANK ECONOMIC REVIEW where coefficients with superscript “SSA” denote the estimated effect of a vari- able on intra-SSA trade (i.e. the coefficient on a variable plus the coefficient on that variable interacted with the intra-SSA dummy). We construct domestic market access (DMA) in Redding and Venables’ (2004) preferred way. That is, we use a country’s internal distance (see Appendix A for its exact definition) as input in the trade cost function, assuming that the speed at which internal trade decays with distance is half as strong as for trade with other countries ^ SSA by two). Moreover, we take each country as sharing a (i.e. we divide d1t common border, a common language, a common colonial history, and having an RFTA with itself. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Having constructed the three different components of a country’s total market access using (9), it is straightforward to obtain both total market access (MA) and access to foreign markets (FMA) according to equation (8). III. EVIDENCE ON THE ROLE OF ECONOMIC GEOGRAPHY IN SSA E CO N O M I C DE V E LO PM E N T Having constructed yearly market access measures for all 48 SSA countries in our sample, we are finally in a position to assess the effect of market access on economic development. Market Access and Economic Development We start by visualizing the relationship between market access and income levels. Figure 1a plots market access (MA) against GDP per worker (our pre- ferred proxy of wages, see below for more on this choice) over our entire sample period. It shows an overall positive relationship between GDP per worker over our sample period. Furthermore, when looking for possible differences over our 17 year sample period, Figure 1b suggests that the relationship between income levels and market access has strengthened in the most recent years of our sample. F I G U R E 1a. Market Access and GDP Per Worker in SSA Notes: The raw correlation between ln MA and ln GDP per worker is 0.22 [ p-value: 0.00]. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Bosker and Garretsen 457 F I G U R E 1b. Market Access and GDP Per Worker in SSA, Changes Over Sample Period Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Notes: The raw correlation between ln MA and ln GDP per worker is 0.05 [ p-value: 0.35] for the 1993– 2000 period and 0.26 [ p-value: 0.00] for the 2001– 2009 period. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. To look at this in a more rigorous way, we turn to the estimation of the NEG wage equation, (5). In the absence of reliable wage data for all SSA coun- tries in all years of our sample, we need to proxy wages. Since many SSA face high unemployment rates, we decided not to use GDP per capita (as e.g. Redding and Venables, 2004; Breinlich, 2006; or Head and Mayer, 2010 do), but take GDP per worker as a more appropriate measure.21 The error term hit in (5) captures cit, a country’s level of production efficiency. Again following Redding and Venables (2004), we start by assuming that these cross-country differences in technology are captured by an idiosyncratic error term and esti- mate (5) using pooled OLS (implicitly only allowing for other variables deter- mining technological efficiency that are uncorrelated with our market access measure). The result is shown in the first column of Table 2 below.22 We find that the estimated market access coefficient is positive and significant. A 1 percent increase in a country’s market access increases GDP per worker by 0.07 percent. This conclusion is, however, somewhat premature. It is only valid under the earlier-mentioned assumption of idiosyncratic differences in countries’ produc- tion efficiency, cit, that are uncorrelated with market access. As this assumption is likely to be violated, we subsequently make use of the panel data nature of our data set. We include country fixed effects to capture all time-invariant country-specific variables that affect a country’s production efficiency. Most 21. All results in this paper also hold when using GDP per capita instead. They are available upon request. 22. We only show bootstrapped standard errors for all our estimation results (they are based on 200 replications). The bootstrapped standard errors take explicit account of the fact that our measures of market access are all generated regressors. See Redding and Venables (2004, p. 64) for more details. Results only become stronger when using robust standard errors instead. 458 T A B L E 2 . Market Access and Economic Development in SSA Dep: ln GDP Per Worker 1 2 3 4 5 ln MA 0.067*** 0.031*** 0.021*** 0.011 0.031** [0.000] [0.003] [0.006] [0.129] [0.039] Polity IV 2 2 2 0.002 0.002 0.0004 2 2 [0.559] [0.633] [0.964] Urbanization rate 2 2 2 0.009 0.01 2 0.005 2 2 [0.333] [0.473] [0.812] Gr. prim. enrollment 2 2 2 0.002 2 0.001 0.0004 2 2 [0.179] [0.659] [0.815] % Oil in GDP 2 2 0.018*** 0.026** 0.001 2 2 THE WORLD BANK ECONOMIC REVIEW [0.000] [0.018] [0.887] Civil war 2 2 2 0.159** 2 0.218** 0.015 2 2 [0.015] [0.044] [0.809] Civil conflict 2 2 2 0.055 2 0.03 2 0.053 2 2 [0.112] [0.374] [0.163] % Agriculture in gdp 2 2 2 0.014*** 2 0.012** 2 0.02*** 2 2 [0.001] [0.012] [0.001] ln working pop. dens. 2 2 2 0.022 2 0.275 1.043 2 2 [0.943] [0.623] [0.114] Nr. obs 775 775 583 268 315 Time-period 1993– 2009 1993– 2009 1993– 2009 1993– 2000 2001– 2009 P-value country FE 2 [0.000] [0.000] [0.000] [0.000] P-value year FE 2 [0.000] [0.000] [0.000] [0.001] Notes: P-values in brackets. Bootstrapped p-values on the basis of 200 replications. ***, **, * denotes significance at 1, 5, or 10 percent respectively. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bosker and Garretsen 459 notably, we hereby control for differences in physical geography that is often blamed for Africa’s poor development (climate, primary resource endowments, soil quality, etc). By including time (year) fixed effects as well, we also take into account any shocks that are affecting all countries similarly. Examples are the introduction of new technological innovations made in developed countries (a prime example here are mobile phones, which have rapidly spread all over SSA) or worldwide economic shocks such as changes in the world price of agri- cultural products or natural resources. The second column of Table 2 shows that the inclusion of fixed effects is quite important (corroborating findings by Head and Mayer, 2010): the effect of total market access on GDP per worker Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 is still positive and significant, but it is about half that found in column 1: a 1 percent increase in a country’s total market access increases GDP per worker by 0.03 percent. However, the inclusion of country- and year-fixed effects may still not provide us with accurate estimates of the effect of market access. They only control for time-invariant country-specific or country-invariant time-specific variables. It is not unlikely that a country’s production efficiency is also deter- mined by time- a ` nd country-varying variables that are correlated with market access. If this is the case, we would still obtain biased estimates of the coeffi- cient on market access, even when allowing for country- and year fixed effects. We therefore include several additional control variables to our regressions (see also Breinlich, 2006; Redding and Venables, 2004; Hering and Poncet, 2010; Fally and others, 2010). They are related to a country’s quality of institutions ( polity IV), its human capital (gross primary enrolment), its scope for econom- ics of density (working population per km2 urbanization rate), and whether or not it is in a state of civil war or conflict. Moreover, given that our market access measures all focus on the import- ance of manufactures, whereas SSA countries vary widely in the importance of this sector in their overall exports (see Figure B2 and Table B2 in Appendix B), we control for a country’s economy’s dependence on natural resources by con- trolling for the importance of oil, and of agriculture respectively in its overall economy (note that, given its time-invariant nature, the presence of oil, or any other natural resource, is already controlled for by the included country fixed effects). Column 3 of Table 2 shows the corresponding estimation results. Adding these additional controls further lowers the effect of market access23 but it remains significantly related to GDP per worker: a 1 percent increase in a country’s market access raises its income level by 0.02 percent. As to the control variables, we find that three of them are significant.24 SSA countries 23. This is also partly driven by the significant reduction in sample size resulting from the fact that not all controls are available for all countries in all years (using the reduced sample in column three without including any of the controls but only country- and year-FE gives an estimated coefficient of 0.23 [ p-value: 0.036]). 24. Note that the non-significance of some of our included controls may also be the result of them having very little within-variation, leaving us with a danger of making type II errors on these variables. 460 THE WORLD BANK ECONOMIC REVIEW that are more dependent on agriculture, and those plagued by civil war tend to have lower levels of GDP per worker. Also, we find that the more oil- dependent SSA countries have higher levels of income per worker. Finally, column 4 and 5 show the results of estimating the relationship between market access and GDP per worker for the first and second half of the sample respectively (always including fixed effects and the eight above- mentioned control variables). This confirms the preliminary evidence shown in Figure 1b: the positive relationship between market access and economic devel- opment is strongest in the second half of our sample period (2001–2009). Although positive, we do not find a significant effect of market access on Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 income levels for the early years in our sample.25 Columns 3–5 constitute our baseline results. They show that market access is a significantly positive determinant of a SSA country’s economic develop- ment. Moreover, the importance of market access has increased over the last two decades. Based on the results shown in column 5, a 1 percent increase in total market access increases GDP per worker by 0.031 percent. When com- paring this result to similar studies using samples encompassing both developed and developing countries (e.g. Redding and Venables, 2004; or Head and Mayer, 2010), but also to other studies looking at developing economies like Brazil (Fally and others, 2010), China (Hering and Poncet, 2010) or Indonesia (Amiti and Cameron, 2007), we find a substantially lower effect of market access on economic development.26 Given the fact that the manufacturing sector is still relatively undeveloped in SSA compared to many countries con- sidered in these other studies, this finding may not be that surprising. Nevertheless, our results show that economic geography matters, also in SSA. Moreover, its importance has increased in recent years. Decomposing the Importance of Access to SSA Markets and to Markets in the ROW Having established the importance of market access for SSA economic develop- ment, we further decompose this finding in this section. In particular, we look at the relative importance of access to other SSA markets and markets in the ROW respectively. To do so, we use the decomposition of a country’s market access in domestic market access (DMA) and foreign market access (FMA) [see (9)], further decomposing the latter into access to other SSA markets (MASSA it ) and access to markets in the ROW (MAROW it ). 25. But we note that it is also not significantly different from the effect we find during the later years. 26. The closest estimate we found in these studies is the one reported on China by Hering and Poncet (2010). In their specification that includes most possible other controls related to wages, they find a positive effect of 0.05 percent in response to an increase in market access of 1 percent. Bosker and Garretsen 461 F I G U R E 2a. Decomposing Market Access – variation in access to SSA and access to ROW Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Notes: The overall mean share of SSA in FMA is 56% (s.d. 0.31). Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. F I G U R E 2b. Market Access and Distance to Major Markets Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. SSA Market Access and ROW Market Access Compared In Figures 2a and 2b, we start by showing some stylized facts about the two most interesting components of market access, ROW and SSA market access. The left-hand panel of Figure 2a shows that having good access to SSA markets is virtually uncorrelated with good access to markets in the ROW. The right- hand panel adds to this by plotting the mean relative importance of SSA in a country’s FMA against mean GDP per worker over our sample period. This shows that ROW market access dominates FMA for SSA’s island nations, countries located in SSA’s north east (e.g. Sudan Djibouti, or Eritrea), and South Africa (SSA’s economic powerhouse27). By contrast, for countries close to South Africa or Nigeria (e.g. Botswana, Swaziland; Togo or Benin), SSA 27. We note that all results presented in the paper are robust to the exclusion of South Africa from the sample. 462 THE WORLD BANK ECONOMIC REVIEW market access dominates their overall FMA. But, despite the substantial differ- ence between countries in the degree to which the ROW dominates their FMA, Figure 2a shows no clear relationship between the relative importance of SSA or the ROW in FMA, and income levels. Figure 2b shows part of the reason why SSA and ROW market access are little correlated. Given the strong distance decay effects we found when esti- mating the trade equation (see Table 1), countries located closest to Europe (Africa’s largest export market) have the best ROW market access.28 On the contrary, countries closest to South Africa, but also those in West Africa (see Figure B1 in Appendix B), tend to have the best SSA market access.29 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 SSA market access, ROW market access and SSA economic development Next, Figure 3 depicts the relationship between the various (sub)components of total market access and economic development. We separately plot DMA, ROW market access and SSA market access against GDP per worker. Given our finding of an increased importance of market access over the years (see Table 2), Figure 3 distinguishes between the first and second half of our sample (1993–2000 and 2001–2009). These simple scatterplots show that the rela- tionship of all three different (sub)components of a country’s market access with GDP per worker appears strongest during the 2nd half of our sample period.30 To more formally assess the importance of market access’s different compo- nents, we re-estimate equation (5) replacing total market access by DMA, access to SSA markets (denoted as FMA-SSA in Table 3) and ROW market access (denoted as FMA-ROW in Table 3). The estimation results are shown in Table 3 below, always distinguishing between the first and second half of our sample period. All regressions include the same control variables as in columns 3–5 of Table 2 and a full set of country- and year fixed effects.31 First of all, the results confirm our earlier finding that market access has become increasingly important for SSA countries. In the first half of our sample period we only find some evidence that, of the various market access categories, domestic market access is weakly significant (at 10 percent). FMA, also when further split between SSA and ROW market access, is not 28. We find a very similar picture when plotting ROW market access against distance to the USA. See Figure B1 in Appendix B. 29. Congo (COG) and the Democratic Republic of Congo (ZAR), are two exceptions here. These countries’ SSA market access is the best in our sample, an artifact of the fact that the two main cities in these two countries (used to calculate the distance between them) are located only 10.5 km apart (the next smallest distance is that between Nigeria and Benin: 105km). Leaving these two countries out of our sample does not change any of the results presented in our paper. 30. Note that Equatorial Guinea (GNQ) shows as somewhat of an outlier in these scatterplots. This is due to its rapid economic growth over our sample period following the discovery of large oil and gas reserves. All results in our paper our fully robust to leaving this country out of the sample. 31. The results on these control variables are very similar to those in Table 2. They are available upon request, as are the results when doing the estimations using our entire sample period. F I G U R E 3. Market Access’ Different Components and Economic Development Notes: All figures plot demeaned ln GDP per worker against a demeaned subcomponent of market access. Demeaned meaning that we substract a country’s mean ln GDP per worker or mean subcomponent of market access from each observation. Doing this already removes the influence of any Bosker and Garretsen unobserved time-invariant from the data, making our plots more directly related to our findings in Table 2 that always include a full set of country fixed effects. It is also done to show that there exists substantial within country-variation in both each respective subcomponent of market access as well as in ln GDP per worker. Source: Authors’ analysis based on data sources discussed in the text. 463 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 464 T A B L E 3 . Foreign Market Access and Economic Development in SSA Dep: ln GDP Per Worker 1 2 3 4 5 6 7 8 ln DMA 0.004* 0.004* 0.02 0.02 2 2 2 2 [0.087] [0.087] [0.151] [0.195] 2 2 2 2 ln FMA 0.009 2 0.048** 2 0.011 2 0.053** 2 [0.679] 2 [0.036] [0.606] 2 [0.019] 2 ln FMA - ROW 2 2 0.087 2 2 0.039 2 2 0.075 2 2 0.039 2 [0.396] 2 [0.655] 2 [0.435] 2 [0.665] ln FMA - SSA 2 0.014 2 0.049** 2 0.016 2 0.054*** THE WORLD BANK ECONOMIC REVIEW 2 [0.370] 2 [0.023] 2 [0.311] 2 [0.004] Controls: see Table 2 nr. obs 268 268 315 315 268 268 315 315 time-period 1993– 2000 1993– 2000 2001– 2009 2001–2009 1993– 2000 1993– 2000 2001– 2009 2001– 2009 p-value country FE [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] year FE [0.000] [0.000] [0.001] [0.001] [0.000] [0.000] [0.001] [0.001] Notes: P-values in brackets. Bootstrapped p-values on the basis of 200 replications. ***, **, * denotes significance at 1, 5, or 10 percent, respectively. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bosker and Garretsen 465 significantly related to income levels in these years (see column 1 and 2). This changes markedly during the latest years in our sample (2001–2009). In par- ticular, we find that access to foreign markets has become much more import- ant in explaining the differences in economic development between SSA countries (see column 3). Even more interesting, when further decomposing this effect of FMA in column 4, we find that it is access to SSA markets in par- ticular that drives these findings: A 1% increase in a country’s access to other SSA markets is associated with a 0.05% increase in GDP per worker. The effect of ROW market access instead is insignificant. Some care has to be taken in interpreting these results. Our findings do go Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 against those proclaiming that intra-SSA economic linkages are too weak and under-developed to be of importance to SSA countries. They support the view that stimulating intra-SSA manufacturing export markets is very important for the future viability of SSA countries that want to become less dependent on natural resource revenue (Collier and Venables, 2007). Indeed, manufacturing already dominates intra-SSA exports (UNCTAD, 2009). Moreover, intra SSA-exports have increased faster than SSA-exports to the rest of the world in recent years (Easterly and Reshef, 2010; ARIA IV). However, given that we focus on market access for manufacturing goods (see also our discussion at the beginning of section II), our results should not be taken as saying that access to markets in the ROW does not matter for SSA. SSA exports to the ROW are dominated by natural resources (for more than 75 percent). The fact that market access to the ROW for SSA manufacturing products is not significant, does not say much about the importance of, for example, lowering tariffs on SSA agricultural products for SSA countries’ economic development. It can also be taken as an indication that, to date, most SSA manufactures are not yet finding their way to markets outside the (sub)continent.32 Finally, also note that DMA loses its (weak) significance in the second half of our sample period. It confirms the idea that for most SSA countries their own domestic market size is too small to be of significant importance.33; pos- sibly even posing constraints on firms’ prospects (see Collier and Venables, 2007). However, endogeneity problems are inherently present when including DMA. One basically regresses a measure dominated by a country’s own GDP (DMA) on its GDP per worker (see Head and Mayer (2010) for a (critical) dis- cussion on this issue). Therefore, columns 5–8 in Table 3 show that our results on FMA, and its two subcomponents (FMA-SSA and FMA-ROW), also come through when totally abstracting from DMA. To summarize our main findings: 32. One could argue that our ROW market access measure suffer from too little cross-sectional variance to find any effect when controlling for country- and year-specific fixed effects. However, the scatterplots in Figure 3 suggest otherwise. Also, when including SSA and ROW market access separately we find the same results. 33. Note that this was also already borne out by the estimated effect of DMA in the early years of our sample. Although significantly positive at the 10 percent level, this effect is very small: a 1 percent increase in DMA increasing GDP per worker by only 0.004 percent. 466 THE WORLD BANK ECONOMIC REVIEW access to foreign markets is increasingly important for SSA countries. Moreover, it is access to other SSA countries in particular that is positively associated with income levels. Additional Robustness Checks Several issues could still invalidate our main findings. First, even when abstract- ing from domestic market access (DMA), there is still the issue of endogeneity. The assumption under which our baseline results are valid is that, after control- ling for fixed effects and the included control variables, the remaining error term is uncorrelated with our measures of foreign market access. One way in Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 which this may be violated is when the error term still contains other variables influencing a country’s GDP per worker that are correlated with market access. Another way is reverse causality: if market access not only influences GDP per worker, but GDP per worker in turn also influences market access, the error term is by construction correlated with market access. To control for both possible sources of endogeneity, we employ an instru- mental variable approach,34 using the distance to SSA’s most important export markets in SSA and in the ROW as instruments for our measures of foreign market access (i.e. the EU, South Africa and Nigeria; see Figures 2b and B1). The relevance of this approach relies on the arguments put forward to justify the usefulness of these ‘distance instruments’ (see among others Redding and Venables, 2004 and Hanson, 2006 for more on this). Table 4a shows the results. First of all, Table 4a shows that we cannot reject the validity of our instru- ments (see the statistics at the bottom of the Table). The F-statistic for their joint significance in the first stage is larger than 10 (Staiger and Stock, 1997), except in column 3 where it is 7.44. Moreover, they always pass the Hansen J test for overidentification. The results confirm our baseline findings. Foreign market access significantly positively affects income per worker in the latest years of our sample. When further subdividing this into ROW- and SSA market access, we again find that this positive effect holds for access to SSA markets only. A drawback of our IV results is that the distance instruments used are time- invariant. This precludes the use of country-fixed effects. Columns 1 –4 of Table 4b therefore also show results when including each market access measure lagged one period (all control variables are also lagged one period). This to some extent controls for reverse causality, while still allowing for the inclusion of country-fixed effects.35 Reassuringly, all our baseline results again come through. 34. This also controls for the third way by which endogeneity issues may be raised, that is, measurement error. 35. Note that this argument breaks down in case of autocorrelation in the residuals. Also including lagged variables does not solve possible endogeneity resulting from omitted variables or measurement error. Bosker and Garretsen 467 T A B L E 4 a . IV-Results dep: ln GDP per worker 1 2 3 4 ln FMA 0.091 2 0.313*** 2 [0.316] 2 [0.008] 2 ln FMA - ROW 2 2 0.059 2 2 0.273* 2 [0.533] 2 [0.065] ln FMA - SSA 2 0.064 2 0.176*** 2 [0.249] 2 [0.001] Controls see Table 2 (no country FE) nr. obs 268 268 315 315 time-period 1993– 2000 1993– 2000 2001– 2009 2001– 2009 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 F-stat. instrument 14.02 2 7.44 2 FMA - ROW 2 198.1 2 143.94 FMA - SSA 2 17.62 2 18.41 p-value over ID-test [0.415] [0.269] [0.501] [0.822] Notes: P-values, based on robust standard errors, in brackets. ***, **, * denotes significance at 1, 5, or 10 percent, respectively. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. T A B L E 4 b . Lagged Market Access Dep: ln GDP Per Worker 1 2 3 4 ln FMA 2 0.006 2 0.037* 2 [0.695] 2 [0.072] 2 ln FMA - ROW 2 0.005 2 2 0.01 2 [0.951] 2 [0.908] ln FMA - SSA 2 2 0.007 2 0.042** 2 [0.540] 2 [0.024] Controls see Table 2 (also lagged) nr. obs 203 203 299 299 time-period 1993– 2000 1993– 2000 2001– 2009 2001– 2009 Notes: P-values in brackets. Bootstrapped p-values on the basis of 200 replications. ***, **, * denotes significance at 1, 5, or 10 percent, respectively. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Finally, our last robustness check extends our empirical model.36 Footnote 5 already hinted at the possibility of extending the NEG model that we use (see Appendix C) to also include an intermediate goods sector. This would (see Redding and Venables 2004 for the details) add an additional term to equation (5). Besides a country’s market access (i.e. its ease of access of final goods markets), a country’s supplier access (i.e. its ease of access to markets for 36. We also estimated (5) in first differences as an alternative way to deal with unobserved country-specific variables that are correlated with market access. Again, we find that SSA market access is the only component of market access for which we systematically find a significant positive effect on GDP per worker (results available upon request). 468 T A B L E 4 c . Supplier Access dep: ln GDP Per Worker 1 2 3 4 5 6 ln MA 0.011* 2 2 0.031** 2 2 [0.096] 2 2 [0.029] 2 2 ln FMA 2 0.010 2 2 0.050** 2 2 [0.646] 2 2 [0.018] 2 ln FMA - ROW 2 2 2 0.056 2 2 2 0.042 2 2 [0.597] 2 2 [0.581] ln FMA - SSA 2 2 0.031* 2 2 0.050** 2 2 [0.080] 2 2 [0.016] ln DMA 2 2 2 2 2 2 2 2 2 2 2 2 THE WORLD BANK ECONOMIC REVIEW ln SA 0.001 2 2 0.002 2 2 [0.806] 2 2 [0.729] 2 2 ln FSA 2 2 0.003 2 2 0.007 2 2 [0.851] 2 2 [0.536] 2 ln FSA - ROW 2 2 0.091 2 2 0.042 2 2 [0.247] 2 2 [0.451] ln FSA - SSA 2 2 2 0.023** 2 2 0.034*** 2 2 [0.044] 2 2 [0.006] Controls see Table 2 Nr. obs 268 268 268 315 315 315 Time-period 1993– 2000 1993– 2000 1993– 2000 2001– 2009 2001– 2009 2001– 2009 Notes: P-values in brackets. Bootstrapped p-values on the basis of 200 replications. ***, **, * denotes significance at 1, 5, or 10 percent, respectively. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bosker and Garretsen 469 intermediate goods needed in final goods production) would be an important determinant of its income level. A yearly-measure of each SSA country’s sup- plier access can be constructed in a very similar way as its market access (basic- ally replacing m ^ jt in (4), but again see Redding and Venables 2004 for ^ jt with r more details). Table 4c shows the results when also taking account of supplier access (SA). Similar to market access, it is very straightforward to decompose overall supplier access into domestic supplier access (DSA), and access to sup- pliers in SSA and in the ROW respectively. For supplier access too, we find that it increased in importance over the last two decades, and that access to SSA suppliers in particular is positively asso- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ciated with higher GDP per worker. However, when focusing on foreign sup- plier access, or supplier access as a whole, results are much weaker than those for market access. Most importantly for our purposes, all our baseline findings regarding the importance of market access hold up to also considering coun- tries’ supplier access. I V. T H E E F F E C T OF DIFFERENT POLICIES AIMED AT IMPROVING M A R K E T AC C E S S Our main findings show that improving a country’s market access, and in par- ticular access to other SSA countries, will have significant positive effects on its economic development. In this section we use this positive relationship between market access and income levels to gain insight into the relative effect of different policies aimed at improving a country’s market access.37 For example, we look at the effect of improving a country’s infrastructure, increas- ing SSA regional integration, or alleviating the burden of landlocked countries. To be able to do so, we change our baseline estimation strategy in one import- ant way (see also Elbadawi and others, 2004 or Redding and Venables 2004). Our baseline strategy includes importer-year and exporter-year dummies when estimating the trade equation (3). This does, however, not allow one “to quantify the effects . . . of particular country characteristics (for example, land- locked or infrastructure), since all such effects are contained in the dummies” (Redding and Venables 2004, p. 75). As such, it becomes impossible to look at the effect of country-specific policies aimed at lowering trade costs. To over- come this problem we follow Redding and Venables (2004, section 7), and esti- mate the following trade equation instead of equation (3), proxying each 37. We note at this point that our policy experiments do not quantify the full general equilibrium effects on income levels. We are confining ourselves to the “short-run” effects of improving market access on income levels. We abstract from any subsequent changes in economic geography induced by e.g. firms or consumers changing their location decistion as a results of the changes in income levels induced by the change in market access resulting from one of our policy experiments. 470 THE WORLD BANK ECONOMIC REVIEW r m country’s market and supplier access by its GDP (sit % GDPitt ; mit % GDPit t ): ln EXijt ¼ rt ln GDPit þ mt ln GDP jt þ b ln Tijt þ 1ijt ð10Þ In order to conduct our various ‘policy experiments’, we augment our specifi- cation of trade costs (6) by three different exporter- and importer-specific determinants of trade costs: being landlocked, being an island, and the state of a country’s infrastructure, that is, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 b ln Tijt ¼ d1 ln Dij þ d2 ln Bij þ d3 ln CLij þ d4 ln CCij þ d5 ln CRij þ d6 RFTAijt þ d7 llit þ d8 llit þ d9 islit þ d10 islit þ d11 infrait þ d12 infra jt ð11Þ Furthermore we take note of Martin and others (2008) and control for whether or not a country is experiencing civil war or civil conflict (see Appendix A for the full details on all these additional variables included to the trade equation). We estimate (10) with (11) substituted for trade costs Tijt for the latest pos- sible year in our sample (which is 2008 instead of 2009 because of missing in- frastructure and conflict data for 2009). We take the latest possible year in order to make our prediction of the effect of each of our ‘policy experiments’ as up-to-date as possible. Table B1 in Appendix B shows the resulting esti- mated coefficients. Again, we allow for a different effect of each trade-cost related variable on intra-SSA trade. Confining our discussion to the newly added country-specific trade cost variables,38 we find a large burden of being landlocked, and to a lesser extent of being an island (landlocked countries export and import significantly less (84 percent and 125 percent respectively) than coastal nations; and these numbers are 80 percent and 34 percent for island countries). Moreover, we find that bad infrastructure is an important de- terrent to trade, and exports in particular. Based on the estimates shown in Table B1, we calculate each SSA country’s market access and its components [in a similar way as in (9)]. Next, we recal- culate these measures taking into account one of eight different policy experi- ments that we set out in more detail below, and calculate the resulting change in foreign market access (and its two subcomponents). We do not look at do- mestic market access as most of our policy experiments are not that interesting to look at for domestic market access (e.g. no longer being landlocked only affects a country’s trade costs with other countries, and it is hard to think about a country establishing an RFTA with itself ). 38. Given that we no longer include importer- and exporter dummies the results on the magnitude of the effect of bilateral trade costs on bilateral exports differs from those reported in Table 1. However, the direction of their effect is never different. Bosker and Garretsen 471 T A B L E 5 . Policy Experiments – Increase in Market Access and GDP Per worker 1 2 3 4 5 6 SSA - RFTA Wide No All with Free No Longer Longer Infrastructure distances South Trade Country Landlocked Island þ 1 s.d. Halved Africa Zone Cape Verde FMA 2 71.89 20.67 63.72 0.34 2.68 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 FMA - SSA 2 51.10 76.57 112.09 2.94 21.11 FMA - ROW 2 74.31 10.39 55.26 2 2 GDP per worker 2 2.76 4.14 6.05 0.16 1.14 Botswana FMA 85.89 2 34.61 75.43 2 0.41 FMA - SSA 90.29 2 76.57 112.09 2 1.41 FMA – ROW 84.02 2 10.39 55.26 2 2 GDP per worker 4.88 2 4.14 6.05 2 0.08 Central African Republic FMA 84.60 2 18.55 61.96 0.06 0.35 FMA - SSA 90.29 2 76.57 112.09 0.69 3.83 FMA – ROW 84.02 2 10.39 55.26 2 2 GDP per worker 4.88 2 4.14 6.05 0.04 0.21 Ethiopia FMA 85.01 2 23.81 66.34 2 0.27 FMA - SSA 90.29 2 76.57 112.09 2 1.76 FMA – ROW 84.02 2 10.39 55.26 2 2 GDP per worker 4.88 2 4.14 6.05 2 0.10 Sudan FMA 2 2 15.84 59.73 2 0.18 FMA - SSA 2 2 76.57 112.09 2 2.99 FMA – ROW 2 2 10.39 55.26 2 2 GDP per worker 2 2 4.14 6.05 0.16 Notes: All numbers are percentages. The effect on GDP per worker is calculated by multiply- ing the change in SSA MA by the coefficient on SSA MA reported in column 8 in Table 3. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Finally, the effect of the resulting improvement in market access on GDP per worker easily follows from the estimated coefficient(s) on foreign market access in Table 3. In particular, we take our finding on the relative import- ance of access to other SSA markets seriously and use the coefficient on SSA market access, reported in column 8, in combination with the change in SSA market access to get at the overall impact of each policy experiment on economic development.39 Table 5 shows the results of doing this for the 39. It would be straightforward to redo these calculations using any other reported coefficient in Table 3 or Table 2 for that matter. This would change the absolute effect of each of the different policy measures on GDP per worker, but it leaves the relative magnitude of each of the different policy experiments unchanged. 472 THE WORLD BANK ECONOMIC REVIEW first six of our eight different “policy experiments,” focusing on five differ- ent countries. Given the way we modelled trade costs, see (11), the first four policy experi- ments affect all countries’ SSA market access and ROW market access similar- ly. This is not true for overall FMA. Since we allowed all trade costs variables to have a different effect on intra-SSA trade and SSA trade with the ROW re- spectively, changes in FMA depend on the relative importance of SSA market access and ROW market access in overall FMA (explaining why the overall change in FMA resembles that in ROW market access for Cape Verde and Sudan, and that in SSA market access for the other three countries). Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Halving distances to all trade partners (a rough proxy for improving SSA countries’ connectivity through, e.g., cross-border infrastructure projects, or more effective border procedures) results in the largest improvement in GDP per worker, raising it by about 6 percent. Next comes alleviating a landlocked country’s burden of having no direct access to the coast (raising incomes by almost 5 percent), followed by a 4 percent increase in GDP per worker as a result of a one standard deviation improvement in a country’s infrastructure (e.g. corresponding to upgrading Ethiopia’s infrastructure to resemble that in Botswana). With a resulting increase of 2.8 percent, alleviating the remoteness of an island country has the smallest effect on GDP per worker. Finally, columns 5 and 6 show the effects of a newly established RFTA. These are also positive but much smaller compared to the other policy experiments. Not surprisingly, the effect on GDP per worker is larger, the larger the number of new partner countries in the new RFTA.40 (compare columns 5 to 6, or the impact of the SSA-wide free trade zone on Cape Verde to that on Botswana [a SSA-free trade zone would more than triple the number of RFTA partners for Cape Verde whereas ‘only’ doubling it for Botswana]). This much smaller effect of the establishment of an SSA-wide free trade zone compared to those of our other ‘policy experiments’ should in our view be taken with a pinch of salt. Due to the plethora of RFTAs officially in existence in SSA in 2008, the average SSA country already shared official RFTA membership with 46 percent of its SSA trade partners. However, the effectiveness of SSA RFTAs in actually implement- ing policies favourable to intra-SSA trade varies widely (compare, e.g., SADC to CENSAD). Our simple dummy variable for the existence of an RFTA is unable to take the varying degrees of effectiveness of each RFTA into account, so that our findings in Table 5 are most likely an underestimate of the effect on econom- ic development were SSA countries able to establish an SSA-wide trade zone op- erating at the same level effectiveness as e.g. ASEAN or MERCOSUR, let alone NAFTA or the EU (see also UNCTAD, 2009). Our last two experiments do not so much concern policy. They are aimed at giving an idea of the magnitude of spatial spillovers across SSA countries. How 40. Moreover, the impact also depends on the relative importance of a country’s newly added RFTA partners for its market access compared to that of the countries with which it already shares an RFTA. Bosker and Garretsen 473 F I G U R E 4. Spillovers to Neighbors of Positive 10 Percent GDP Shock in Nigeria or South Africa Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. large are the benefits of growth in one particular country for its neighbors as a result of the increased market access that these neighboring countries enjoy? Figure 4 gives some idea of this. It plots the increase in GDP per worker in all SSA countries resulting from a 10 percent increase in the GDP of one of SSA’s economic powerhouses, South Africa and Nigeria, against distance to these countries. Given the magnitude of the estimated distance penalty in SSA (see Table 1), we find that the effects of such shocks quickly peters out with distance. Countries located closest to South Africa and Nigeria respectively experience the largest spillovers. The overall spillover effect is small compared to some of our earlier “trade cost experiments” (the nearest neighbors experiencing “spillover-growth” of about 0.2 percent).41 This is due to the fact that for each country, South Africa and Nigeria constitute only one of many trading part- ners, that determine a country’s market access. V. C O N C L U S I O N S The role of geography in explaining sub-Saharan Africa’s poor economic per- formance is often confined to its physical geography, focusing on, for example, its hostile disease environment or poor climate. This paper focuses on a 41. This effect is much smaller that the spillover-effects of South African growth on its neighbors found by Arora and Vamvakidis (2005). Their analysis is a reduced form exercise, making it hard to compare to our theory-based approach (although our findings could be reconciled with theirs by arguing that we only capture the trade-induced spillover effect, whereas they capture a composite spillover effect of South African growth including also other non-trade related spillovers). Moreover, given their chosen empirical strategy they are unable to include year fixed effects in their panel estimation so that it is impossible to exclude the possibility that (part of ) their findings are driven by an omitted variable affecting both South African economic growth as well as that of other SSA countries. 474 THE WORLD BANK ECONOMIC REVIEW different role of geography and establishes the importance of relative or eco- nomic geography for economic development in sub-Saharan Africa (SSA). Using an empirical strategy that is firmly based upon a new economic geog- raphy model, our paper is among the first to test for the importance of market access and thereby of economic geography in explaining the observed differ- ences in economic development between SSA countries. Building on the framework introduced by Redding and Venables (2004), we first construct theory-based measures of market access for manufactures for each SSA country, relying on bilateral manufacturing trade data to reveal the relative importance of trade costs and market size in determining each coun- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 try’s market access. In doing so, we explicitly allow for a different impact of trade costs on intra-SSA trade and SSA trade with the rest of the world (ROW), and subsequently decompose each country’s total market access into market access to other SSA countries and into market access to the ROW respectively. Using these constructed measures of market access, we estimate the impact of market access for manufactures on GDP per worker. We find that market access positively affects income levels. Economic geography matters for eco- nomic development, also in SSA. Moreover, it has increased in importance over the last two decades. The relationship between market access and econom- ic development is strongest and most robust during the 2001–2009 period: a 1 percent increase in a country’s market access is associated with a 0.03 percent increase in its GDP per worker. This finding is robust to controlling for other variables affecting economic development (most notably human capital, insti- tutions and natural resource dependence), to controlling for unobserved hetero- geneity by allowing for country (and year) specific fixed effects, and to instrumenting market access by distance to major markets. Arguably even more interesting is our finding that, when decomposing our overall market access effect into the respective effects of domestic, SSA-, and ROW-market access, access to other SSA markets has the most significant (and also the most robust) impact on a country’s economic development. This finding becomes less surprising when considering the fact that most SSA coun- tries sell the bulk of their manufacturing exports to other SSA countries (UNCTAD 2009; 2010). Moreover, SSA export growth has been regional in recent years. Intra SSA-exports have increased faster than SSA-exports to the ROW (Easterly and Reshef 2010; UNCTAD 2010). By contrast, SSA exports to the ROW are to date still dominated by natural resources and agricultural pro- ducts, with only little SSA manufactured goods finding their way to European, US, or Asian markets. Our findings stress the importance of the rest of SSA for most SSA countries’ prospects on developing a more diversified economy with a profitable, exporting, manufacturing sector (one of the backbones of Asia’s sustained growth over the last decades). Improving market access among SSA countries alleviates the constraint of small domestic market size faced by most Bosker and Garretsen 475 SSA countries (Collier and Venables 2007), carrying positive effects for eco- nomic development. Based on our findings, we also show tentative evidence on the impact of several policies specifically aimed at improving SSA countries’ market access. Overall, this lends support to the view that current efforts to improve SSA market access by, for example, investing in infrastructure (Sub-Saharan African Transport Policy Program or The Infrastructure Consortium for Africa), allevi- ating the burden of landlocked countries (the Almaty Programme), or by aiming to increase effective intra-SSA integration (African Union), are indeed important, although in varying degrees, in further stimulating SSA economic Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 development. Above all, see also Henderson, Shalizi and Venables (2001), our results are a reminder that distance or relative geography matters for economic develop- ment. Despite room for ( policy-induced) improvements in market access, the (economic) remoteness of many SSA countries remains an important burden on their economic development prospects. APPENDIX A. D ATA D E F I N I T I O N S AND SOURCES GDP (also per capita and per worker): Gross Domestic Product (also per capita and per worker) in current US dollars. From World Bank Development Indicators, 2011, or World Bank Africa Database, 2010. Distance: Great circle distance between main cities, from CEPII. Internal distance: This often-used specification of Dii reflects the average dis- tance from the centre of a circular disk with areai to any point on the disk (as- suming these points are uniformly distributed on the disk). It is calculated on the basis of a country’s area: Dii ¼ 2=3ðareai =pÞ1=2 . Contiguity: Dummy variable indicating if two countries share a common border, from CEPII. Common official language: Dummy variable indicating if two countries share a common official language, from CEPII Common colonizer: Dummy variable indicating if two countries have been colonized by the same colonizer, from CEPII. Colony – Colonizer relationship: Dummy variable indicating if two coun- tries have ever had a colony-colonizer relationship, from CEPII. Landlocked: Dummy variable indicating if a country has no direct access to the sea. Island: Dummy variable indicating if a country is an island. Infrastructure index: Following Limao and Venables (2001), the index is constructed as the unweighted average of four variables (each normalized to have a mean of 0 and standard deviation 1 over the whole sample period as well as in each year). As Limao and Venables (2001), we ignore missing values, 476 THE WORLD BANK ECONOMIC REVIEW making the implicit assumption that the four variables are perfect substitutes to a transport services production function. The four components are: - Roads: Km road per km2. - Paved roads: Km paved road per km2. - Railways: Km railways per km2. - Telephone main lines: Telephone main lines per 1000 inhabitants. All four are taken from the World Bank Development Indicators 2011, or the World Bank Africa Database, 2010. African regional or free trade agreement: Dummy variable indicating if two countries in a particular year are both a member of one of the following Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 African regional or free trade agreements: ECOWAS, ECCAS, COMESA, SADC, UEMOA, CEMAC (or UDEAC), EAC, IGAD, CENSAD, or the recent- ly established AFTZ. Civil conflict: Dummy variables indicating if a country experienced the use of armed force between two parties, of which at least one is the government of a state that resulted in at least 25 and at most 999 battle-related deaths, from the International Peace Research Institute, Oslo. Source: World Development Indicators, 2011. Civil war: Dummy variables indicating if a country experienced the use of armed force between two parties, of which at least one is the government of a state that resulted in at least 1000 battle-related deaths, from the International Peace Research Institute, Oslo. Source: World Development Indicators, 2011. Urbanization rate: Share of the population living in urban areas, from the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. Gross primary enrollment: Gross enrollment ratio is the ratio of total enroll- ment, regardless of age, to the population of the age group that officially corre- sponds to the level of education shown. Primary education provides children with basic reading, writing and mathematics skills along with an elementary understanding of such subjects as history, geography, natural science, social science, art and music. From the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. Percentage oil rents in total GDP: Oil rents are the difference between the value of crude oil production at world prices and total costs of production. From the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. Percentage agriculture in total GDP: Agriculture includes forestry, hunting and fishing as well as cultivation of crops and livestock production. From the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. Bosker and Garretsen 477 Working population per km2: Data on the working population and a coun- try’s overall area are separately taken from the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. Polity IV: The "Polity Score" captures a regime’s authority spectrum on a 21-point scale ranging from -10 (hereditary monarchy) to þ 10 (consolidated democracy). From the Polity Project. Religious similarity: Fraction measuring the probability of two people from different countries adhering to the same religion. In particular we follow Helpman and others (2008) and construct this variable as: (% Protestants in country i . % Protestants in country j ) þ (% Catholics in country i . % Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Catholics in country j ) þ (% Muslims in country i . % Muslims in country j ). Information on each country’s religious composition is taken from Robert Barro’s website at Harvard. Percentage manufacturing exports in merchandise exports: Manufactures comprise commodities in SITC sections 5 (chemicals), 6 (basic manufactures), 7 (machinery and transport equipment), and 8 (miscellaneous manufactured goods), excluding division 68 (non-ferrous metals). From the World Bank Development Indicators, 2011, or the World Bank Africa Database, 2010. APPENDIX B F I G U R E B1. ROW and SSA Market Access, and Distance to the USA and Nigeria Resp. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. 478 THE WORLD BANK ECONOMIC REVIEW T A B L E B 1 . Trade Equation Without Importer and Exporter Dummies. Year ¼ 2008 Coefficients Coefficients Dep: ln Manuf Exports Variable ROW Extra SSA Variable ROW Extra SSA ln dist 2 0.797*** 2 0.82*** ln GDP exp 1.405*** 0.087 [0.00] [0.003] [0.00] [0.472] contiguity 2 1.944*** 4.182*** ln GDP imp 0.782*** 2 0.326*** [0.055] [0.00] [0.00] [0.001] com. lang. 0.528*** 2 0.378 civil war exp 0.255 2 [0.00] [0.272] [0.207] 2 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 com. col. 1.20*** 2 0.258 civil war imp 2 0.03 2 [0.00] [0.49] [0.888] 2 colonizer 1.611*** 2 civil conflict exp 0.234 2 [0.00] 2 [0.205] 2 RFTA 1.151*** 2 0.165 civil conflict imp 0.113 2 [0.00] [0.717] [0.498] 2 infra exp 0.416*** 2.647*** [0.00] [0.00] p-value Mills’ ratio [0.066] infra imp 0.192* 0.733 p-value religion [0.000] [0.098] [0.233] ll exp 2 0.84*** 2 0.063 nr. obs 14946 [0.00] [0.86] censored 11184 ll imp 2 1.252*** 0.189 uncensored 3762 [0.00] [0.56] isl exp 2 0.797*** 0.284 [0.00] [0.616] isl imp 2 0.342** 2 0.41 [0.044] [0.443] Notes: We report estimated coefficients and not marginal effects. Marginal effects, and the results for the 1st stage probit are available upon request. The coefficients are used as input in the construction of our market access measures. *, **, *** denotes significant at the 5 percent level in at least 50, 80, or 100 percent of the years. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Bosker and Garretsen 479 F I G U R E B2. Share of Manufacturing in Total Merchandise Exports 1993– 2009 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Notes: The figure shows box-plots for each year in our sample. Each box ranges from the 25th to the 75th percentile of the distribution of SSA countries’ percentage of manufacturing exports in total merchandise exports. The horizontal line within each box denotes the median percentage of manufacturing exports in total merchandise exports. This median increases from less than 10 % to more than 20 % over our sample period. The lines extending from each box denote the upper and lower adjacent values respectively. These are calculated as 1 . 5 times the interquartile range (IQR), the difference between the 25th and 75th percentile of the overall SSA distribution of manufacturing shares in total merchandise trade in each particular year. The countries explicitly shown in the figure are those for which manufacturing exports constitute a share in overall merchandise exports that is even higher that the upper adjacent value of the overall SSA distribution in a particular year. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. APPENDIX C In this Appendix, we briefly set out the new economic geography (NEG) model that underlies our empirical framework.42 Assume the world consists of i ¼ 1, . . . ,R countries, each being home to an agricultural43 and a manufactur- ing sector. As in virtually all NEG models, we focus on the manufacturing sector. Moreover, and in line with e.g. Redding and Venables (2004), Breinlich (2006), Knaap (2006) and Head and Mayer (2006), we restrict our attention to the ‘short-run’ version of the model. This amounts to, as Redding and Venables 2004, p.59) put it, “taking the location of expenditure and produc- tion as given and asking the question what wages can manufacturing firms in each location afford to pay its workers.” 42. See Fujita, Krugman, and Venables (1999), Puga (1999), Head and Mayer (2004) for more detailed expositions of various NEG models and the derivation of market access and the equilibrium wage equation in particular. 43. The agricultural sector uses labor and land to produce a freely tradable good under perfect ´ raire good. competition that acts as the nume 480 THE WORLD BANK ECONOMIC REVIEW T A B L E B 2 . Average Share of Manufacturing in Merchandise Exports Avg. % Manufacturing in Total Merchandise Exports Country 1993–2000 2001– 2009 1993– 2009 Sub-Saharan Africa 29.4 32.4 31.1 Djibouti 2 90.7 90.7 Lesotho 94.9 88.8 90.0 Botswana 89.6 82.9 83.6 Cape Verde 81.3 68.2 73.8 Mauritius 72.2 66.0 68.9 Swaziland 54.4 65.2 63.9 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 South Africa 49.6 54.7 52.3 Namibia 55.8 47.1 48.0 Central African Republic 49.0 37.4 44.5 Senegal 43.1 41.7 42.3 Madagascar 28.4 46.5 38.0 Eritrea 28.0 40.5 36.3 Zimbabwe 31.4 33.5 32.5 Togo 12.8 53.1 31.4 Kenya 25.4 30.6 28.1 Gambia, The 19.0 21.9 20.7 Guinea 22.8 18.4 20.4 Ghana 14.7 22.3 18.9 Tanzania 16.6 19.5 18.6 Cote d’Ivoire 14.1 17.2 16.0 Comoros 23.6 6.3 14.3 Burkina Faso 12.7 11.5 12.1 Uganda 6.9 14.5 11.0 Malawi 8.4 11.9 10.4 Zambia 10.7 10.1 10.4 Mali 4.1 13.6 9.9 Mozambique 11.6 6.9 8.8 Ethiopia 7.5 9.4 8.6 Sierra Leone 9.7 7.5 8.6 Benin 5.8 8.7 7.0 Niger 2.8 9.3 6.6 Burundi 2.2 10.2 6.4 Rwanda 5.8 5.9 5.9 Cameroon 6.8 4.0 5.1 Gabon 2.7 7.8 5.0 Sao Tome and Principe 2.0 4.3 3.9 Congo, Rep. 2.4 2 2.4 Seychelles 1.1 3.7 2.4 Nigeria 1.6 2.9 2.3 Sudan 3.7 0.6 2.1 Guinea-Bissau 0.2 1.2 1.0 Mauritania 0.2 0.0 0.1 Notes: All numbers denote percentages. Angola, Chad, the Democratic Republic of Congo, Equatorial Guinea, Liberia and Somalia are not shown in the Table due to lack of information. Source: Authors’ analysis based on data sources discussed in the main text or Appendix A. Bosker and Garretsen 481 In the manufacturing sector, firms operate under internal increasing returns to scale, represented by a fixed input requirement ciF and a marginal input re- quirement ci. Each firm produces a different variety of the same good under monopolistic competition using the same Cobb-Douglas technology combining two different inputs. The first is an internationally immobile factor (e.g. labor), with price wi and input share b, the second is an internationally mobile factor with price vi and input share g, where g þ b ¼ 1.44 Manufacturing firms sell their products to all countries. This involves ship- ping them to foreign markets incurring trade costs in the process. These trade costs are assumed to be of the iceberg-kind and the same for each variety pro- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 duced. In order to deliver a quantity xij (z) of variety z produced in country i to country j, xij (z)Tij has to be shipped from country i. A proportion (Tij-1) of output ‘is paid’ as trade costs (Tij ¼ 1 if trade is costless). Taking these trade costs into account gives the following profit function for each firm in country i, X R X R pi ¼ pij ðzÞxij ðzÞ=Tij À wb g i vi ci ½F þ xij ðzފ ðC1Þ j j where pij (z) is the price of a variety produced in country i. Turning to the demand side, consumers combine each firm’s manufacturing variety in a CES-type utility function, with s being the elasticity of substitution between each pair of product varieties. Given this CES-assumption, it follows directly that in equilibrium all manufacturing varieties produced in country i are demanded by country j in the same quantity (for this reason varieties are no longer explicitly indexed by (z)). Denoting country j’s expenditure on manu- facturing goods as Ej, country j’s demand for each product variety produced in country i can be shown to be ðsÀ1Þ: xij ¼ pÀ s ij Ej Gj ðC2Þ where Gj is the price index for manufacturing varieties that follows from the assumed CES-structure of consumer demand for manufacturing varieties. It is defined over the prices, pij, of all goods produced in country i and sold in country j, " #1=ð1ÀsÞ X R Gj ¼ ni p1 ij Às ðC3Þ i Maximization of profits (C1) combined with demand as specified in (C2) gives 44. Since our main aim is to establish the relevance of market access we, in line with for instance Breinlich (2006), skip intermediate inputs and thereby ignore supplier access for most of our analysis, do however see Table 4c for estimation results for supplier access. 482 THE WORLD BANK ECONOMIC REVIEW the well-known result in the NEG literature that firms in a particular country set the same f.o.b. price, pi, depending only on the cost of production in loca- tion i, i.e. pi is a constant markup over marginal costs: pi ¼ wb g i vi ci s=ðs À 1Þ ðC4Þ As a result, price differences between countries in a good produced in country i can only arise from differences in trade costs, i.e. pij ¼ piTij. Next, free entry and exit drive (maximized) profits to zero, pinpointing equi- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 librium output per firm at x  ¼ ðs À 1ÞF. Combining equilibrium output with equilibrium price (C4) and equilibrium demand (C2), and noting that in equi- librium the price of the internationally ( perfectly) mobile primary factor of production will be the same across countries (vi ¼ v for all i), gives the equilib- rium manufacturing wage: 0 11 MAij bs BXR zfflfflfflfflffl}|fflfflfflfflffl{ C À1=b B B ð1ÀsÞ C C wi ¼ Aci B mj Tij C ðC5Þ @ j A |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} MAi where A is a constant that contains inter alia the substitution elasticity, s, and the fixed costs of production, F), and mj denotes country j’s market capacity that is a combination of its expenditure on manufacturing goods (Ej ) and the price index for manufacturing varieties that it faces (Gj ), i.e. mj ¼ Ej Gj(s21). In log-linear form (C5) equates to equation (1) that underlies all our estimation results in the main text of the paper. Finally, aggregating demand from consumers in country j for a good pro- duced in country i (see (C2)) over all firms, ni, producing in country i, gives the following aggregate export equation describing the total amount country i exports to county j. ðsÀ1Þ ð1ÀsÞ ð1ÀsÞ EXij ¼ ni p1 i Às Ej Gj Tij ¼ si ½mj Tij Š; ðC6Þ |fflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflffl{zfflfflfflfflfflffl} MAij MAij where we make use of the fact that pij ¼ piTij and redefine si ¼ ni p1 i 2s (what Redding and Venables 2004 refer to as a country’s supplier capacity). Equation (C6) forms the basis of the two-step estimation procedure that we use in our paper. Bosker and Garretsen 483 REFERENCES Amiti, M., and B.S. Javorcik. 2008. Trade Costs and Location of Foreign Firms in China. Journal of Development Economics 85(1–2): 129 –149. Amiti, M., and L. Cameron. 2007. Economic Geography and Wages. Review of Economics and Statistics 89(1): 15– 29. Amjadi, A., and A.J. Yeats. 1995. Have Transport Costs Contributed to the Relative Decline of Sub-Saharan African Exports? Some Preliminary Evidence. World Bank Policy Research Paper no. 1559. Anderson, J.E., and E. van Wincoop. 2004. Trade Costs. Journal of Economics Literature vol. XLII: 691–751. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Arora, V., and A. Vamvakidis. 2005. The implications of South African economic growth for the rest of Africa. IMF Working Paper WP/05/58. Bigsten, A., and M. So ¨ derbom. 2006. “What have we learned from a decade of manufacturing enter- prise surveys in Africa?” The World Bank Research Observer 21: 241–265. Bosker, E.M., and H. Garretsen. 2010. Trade Costs, Market Access and Economic Geography. Papers in Regional Science 89 (1): 485–513. Bosker, E.M., S. Brakman, H. Garretsen, and M. Schramm. 2010. Relaxing Hukou-Increased Labor Mobility and China’s Economic Geography. CEPR discussion paper 8106. Breinlich, H. 2006. The spatial income structure in the European Union – what role for Economic Geography? Journal of Economic Geography 6: 593– 617. Buys, P., U. Deichmann, and D. Wheeler. 2010. Road Network Upgrading and Overland Trade Expansion in Sub-Saharan Africa. Journal of African Economies 19: 399– 402. Coe, D.T., and A.W. Hoffmaister. 1999. North-South Trade: Is Africa Unusual? Journal of African Economics 8: 229–256. Collier, P. 2002. Primary Commodity Dependence and Africa’s Future. Keynote Speech, World Bank’s Annual Conference on Development Economics. Collier, P., and J.W. Gunning. 1999. Explaining African Economic Performance. Journal of Economic Literature 37: 64 –111. Collier, P., and A.J. Venables. 2007. Rethinking trade preferences: how Africa can diversify its exports. The World Economy 30(8): 1326–1345. Deichmann, U., S.V. Lall, S.J. Redding, and A.J. Venables. 2008. “Industrial location in developing countries.” World Bank Research Observer 23(2): 219–246. Easterly, W., and A. Reshef. 2010. African export successes: surprises, stylized facts, and explanations. NBER Working Paper no. 16597. Elbadawi, I., T. Menigstae, and A. Zeufack. 2006. Market access, supplier access, and Africa’s manufac- tured exports: a firm level analysis. Journal of International Trade & Economic Development 15(4): 493–523. Fally, T., R. Paillacar, and C. Terra. 2010. Economic geography and wages in Brazil: evidence from micro-data. Journal of Development Economics 91: 155– 168. Foroutan, F., and L. Pritchett. 1993. Intra-Sub-Saharan African Trade: Is It Too Little?. Journal of African Economics 2: 74–105. Frazer, G., and J. van Biesebroeck. 2010. Trade growth under the African growth and opportunity pact. Review of Economics and Statistics 92(1): 128– 147. Freund, C., and N. Rocha. 2011. What constrains Africa’s exports?. World Bank Economic Review 25(3): 361–386. Fujita, M., P. Krugman, and A.J. Venables. 1999. The Spatial Economy MIT Press. Gallup, J.L., J.D. Sachs, and A.D. Mellinger. 1999. Geography and Economic Development, International Regional Science Review. 22: 179– 232. 484 THE WORLD BANK ECONOMIC REVIEW Hanson, G. 2005. Market Potential, Increasing Returns and Geographic Concentration. Journal of International Economics 67(1): 1 –24. Head, K., and Th. Mayer. 2004. The Empirics of Agglomeration and Trade. in V. Henderson, and J-F. Thisse (eds.) The Handbook of Regional and Urban Economics, volume IV North Holland 2609– 2665. ———. 2006. Regional wage and employment responses to market potential in the EU. Regional Science and Urban Economics 36(5): 573– 594. ———. 2010. Gravity, Market Potential and Economic Development U British Columbia/Sciences-Po. Processed. Helpman, E., M. Melitz, and Y Rubinstein. 2008. Estimating trade flows: trading partners and trading volumes. Quarterly Journal of Economics Vol. 123, 441– 487. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Henderson, J.V., Z. Shalizi, and A.J. Venables. 2001. Geography and Development. Journal of Economic Geography 1: 81– 105. Hering, L., and S. Poncet. 2010. Market Access and individual wages: evidence from China. Review of Economics and Statistics 92: 145– 159. Hummels, D. 2001. Toward a Geography of Trade Costs. mimeo, Purdue University IMF, 2007. Regional Economic Outlook: Sub-Saharan Africa, Washington, D.C. Johnson, S., J. Ostry, and A. Subramanian. 2006. Levers for growth. Finance and Development 43(1): 28 –31. Jones, B.F, and B.A. Olken. 2008. The anatomy of start-stop growth. Review of Economics and Statistics 90(3): 582 –587. Knaap, T. 2006. Trade, location, and wages in the United States. Regional Science and Urban Economics 36(5): 595– 612. Krugman, P. 1991. Increasing Returns and Economic Geography. Journal of Political Economy nr. 3, 483 –499 Lall, S.V., Z. Shalizi, and U. Deichmann. 2004. Agglomeration economies and productivity in Indian in- dustry. Journal of Development Economics 73: 643–673. Limao, N., and Mayer, Th. 2008. Market Potential and Development. CEPR Discussion Paper no. 6798, CEPR, London. Ndulu, B. 2007. Challenges to African Growth World Bank. Nunn, N., and D. Puga. 2011. Ruggedness: the blessing of bad geography in Africa. Review of Economics and Statistics forthcoming. Portugal-Perez, A., and J.S. Wilson. 2008. Trade costs in Africa: barriers and opportunities for reform. Policy research working paper, no.4619. World Bank. Puga, D. 1999. The rise and fall of regional inequalities. European Economic Review Vol. 43, 303– 334. Redding, S., and A.J. Venables. 2004. Economic Geography and International Inequality. Journal of International Economics 62(1): 53 –82. Subramanian, A., and N.T. Tamirisa. 2003. Is Africa Integrated in the Global Economy?. IMF Staff Papers 50(3): Washington, D.C. Staiger, D., and J.H. Stock. 1997. Instrumental variables regression with weak instruments. Econometrica 65: 557– 586. ¨ ssmann. 2006. Dualism and Cross-Country Growth Regressions. Journal of Temple, J., and L. Wo Economic Growth 11: 187–228. United Nations Economic Commission for Africa, 2010. Assessing Regional Integration in Africa IV (ARIA IV) – Enhancing Intra-African Trade. Addis Ababa, Ethiopia. United Nations Conference on Trade and Development. 2009. Economic Development in Africa Report 2009 – Strengthening Regional Economic Integration for Africa’s Development. New York and Geneva. Bosker and Garretsen 485 Van Biesebroeck, J. 2005. Exporting raises productivity in Sub-Saharan African manufacturing firms. Journal of International Economics 67: 373–391. Wooldridge, J.M. 2003. Introductory Econometrics-A Modern Approach Thomson, USA. World Bank. 2007. Accelerating Development Outcomes in Africa-Progress and Change in the Africa Action Plan Africa Region, The World Bank, Washington, D.C. ———. 2008. World Development Report 2009 The World Bank, Washington, D.C. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 The Decision to Import Capital Goods in India: Firms’ Financial Factors Matter Maria Bas and Antoine Berthou Are financial constraints preventing firms from importing capital goods? Sourcing capital goods from foreign countries is costly and requires internal or external finan- cial resources. A simple model of foreign technology adoption shows that credit con- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 straints act as a barrier to importing capital goods under imperfect financial markets. In our study, we investigate this prediction using detailed balance-sheet data from Indian manufacturing firms having reported information on financial statements and imports by type of good over the period 1997– 2006. Our empirical findings shed new light on the micro determinants of firms’ choices to import capital goods. Baseline es- timation results show that firms with a lower leverage and higher liquidity are more likely to source their capital goods from foreign countries. Quantitatively, a 10 per- centage point improvement of the leverage or liquidity ratio increases the probability of importing capital goods by 11 percent to 13 percent respectively. Different robust- ness tests demonstrate that these results are not driven by omitted variable bias related to changes in firm observable characteristics as well as ownership status. These find- ings are also robust to alternative specifications dealing with the potential reverse causality issues. JEL codes: F10, F14, D92 Globalization is characterized by a significant increase in world imports of capital goods and intermediate inputs. In developing countries, a number of firms rely on capital goods and inputs from abroad since they are more advanced in terms of technology relative to the domestic goods. While the lit- erature on endogenous growth provides theoretical grounds for the role of foreign technology to enhance economic growth, recent firm-level studies confirm that firm performance depends critically on the access to inputs used Maria Bas is an economist at the Centre d’Etudes Prospectives et d’Informations Internationales (CEPII); her email is maria.bas@cepii.fr. Antoine Berthou (corresponding author) is an economist at the Banque de France and is associate researcher at CEPII; his email is antoine.berthou@banque-france.fr. The authors thank Ame ´ lie Maingault for her excellent research assistance. We have benefited from discussions with Jens Arnold, Agne ´ nassy-Que ` s Be ´ , Matthieu Crozet, Joze Damijan, Joep Konings, ´ re Tibor Besedes, Ben Li, Kalina Manova, Bruno Merlevede, Sandra Poncet, Romain Ranciere and Elisabeth Sadoulet. We also thank three anonymous referees for their valuable comments to the manuscript and their constructive suggestions. This research was carried out while Antoine Berthou was an economist at CEPII, and does not reflect the views of the Banque de France. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 486– 513 doi:10.1093/wber/lhs002 Advance Access Publication February 8, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 486 Bas and Berthou 487 in the production of final goods.1 Amiti and Konings (2007) find that input- trade liberalization in Indonesia boosts firm productivity up-to 12 percent, while Khandelwal and Topalova (2010) show that it improved firm productiv- ity by 4.8 percent in India. Goldberg and others (2010) demonstrate that input- tariff cuts in India account on average for 31 percent of the new products intro- duced by domestic firms. Using firm-level data from Argentina, Bas (2011) finds that input tariff reductions are associated with 8 percent increase in the probability of exporting.2 While the use of foreign technology is expected to increase firm efficiency, foreign technology adoption is conditioned by the access to financial resources. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Importing capital goods implies incurring fixed costs associated with gathering information on foreign markets, establishing linkages with foreign suppliers, learning the new technology or adapting the production process, which requires external financing.3 In our study, we argue that financial constraints represent an important barrier to firms’ imports of capital goods, thereby limit- ing their opportunities to benefit from technological spillovers of foreign countries. First, we present a simple theoretical framework to rationalize the main mechanisms through which financial access affects firms’ foreign technology choice. In this framework, using foreign capital goods increases the efficiency to produce final goods, but requires paying an additional fixed cost. In the presence of financial constraints wealthier firms have a better access to external finance and are more likely to use the foreign technology by importing capital goods. Second, we test this relationship between firms’ financial statements and their decision to import capital goods using a detailed Indian firm-level dataset, Prowess. This data was collected by the Centre for Monitoring the Indian Economy (CMIE) for the period 1997–2006.4 During this period, about 75 percent of imports of capital goods in India are originated from high income OECD countries.5 The Prowess data provides information on financial charac- teristics of firms as well as imports distinguished by type of goods (capital equipment, intermediate goods, or final goods). This information allows us to compute the liquidity and leverage ratios that are used throughout the paper to measure firms’ financial factors. These balance sheet statements are expected to 1. Ethier (1982), Markusen (1989), Grossman and Helpman (1991), Rivera-Batiz and Romer (1991) develop theoretical models where foreign technology acts as a driver of economic growth. 2. Kasahara and Rodrigue (2008), Halpern and others (2009), Schor (2004), Kugler and Verhoogen (2009), Bas and Strauss-Kahn (2011) find empirical evidence that the use of foreign inputs enhances firms’ total factor productivity, the quality of final goods, and the number of products exported by firms. 3. See Eaton and Kortum (2001), who quantify that about 25 percent of cross-country productivity differences can be explained by the relative price of equipment, half of it being due to barriers to trade in equipment. 4. We focus on the period 1997–2006 in order to maximize the number of firms each year. 5. This number is obtained by using the HS6 product-level bilateral trade BACI dataset from CEPII, combined with the Broad Economic Product Classification provided by the United Nations. 488 THE WORLD BANK ECONOMIC REVIEW be positively related to the borrowing capacity in the presence of financial con- straints. Our empirical strategy disentangles the impact of the liquidity and le- verage ratios of the firm on the decision to invest in foreign capital goods. Our empirical findings confirm the theoretical prediction that those firms that are ex-ante more liquid and less leveraged are more likely to import capital goods. In our baseline estimations, a 10 percentage point decrease in the leverage ratio or an equivalent increase in the liquidity ratio for the average firm increases the likelihood of importing capital goods by 11 percent and by more than 13 percent, respectively. These results are robust to changes in firm observable characteristics such as firm size, capital and skill-intensity. We carry Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 out different tests that demonstrate that our results are not influenced by omitted variable bias related to India’s trade liberalization. Our results remain also robust to the exclusion of multinational firms, state-owned firms, and local business groups. We provide robustness tests to account for the possibility that using foreign capital goods may improve financial factors of firms ex-post. First, we focus on the sample of firms that have started importing foreign capital goods, by con- sidering in the empirical analysis only those firms that did not import capital goods in the previous two to four years. As an additional related test, we include in the baseline specification the past importer status to take into account previous import experience. Second, we use the measure of external dependence proposed by Rajan and Zingales (1998) to test whether financial factors are more important in industries where firms rely more on external finance. These results confirm that the leverage (liquidity) of the firm has a strong negative ( positive) effect on the probability of importing foreign equipments. These results complete the existing evidence regarding the determinants of firm performance in the case of the Indian economy. Many of these works have used the Prowess data over a comparable period of time. Alfaro and Chari (2009) show that although the importance of private firms in the Indian economy has been growing after the economic reforms in the early 1990s, state-owned firms still represent an important share of total production and assets in some sectors. Khandelwal and Topalova (2010) and Goldberg and others (2010, 2009) study the micro-economic effects of trade liberalization in India. Input-tariffs cuts have contributed significantly to firm productivity growth (Khandelwal and Topalova 2010), and also to the ability of firms to introduce new products (Goldberg and others 2010, 2009). On the same line, Arnold and others (2010) find that product market reforms in services sectors have an important effect on firm productivity in the manufacturing sector in India. Evidence regarding the importance of financial factors in explaining Indian firm performance is more scarce: Topalova (2004) shows that although Indian firms improved their financial statements during the period of economic reforms, some firms still face problems servicing their debt obligations. Bas and Berthou 489 This work also contribute to the literature in finance that connects financial factors and firms’ investment decision. In the presence of information asymmet- ries, uncollateralized external financing becomes more costly than internal fi- nancing, thus introducing a positive relation between a firm’s net worth and its investment decision. This link has been empirically observed for a number of countries and surveyed by Hubbard (1998). These studies (Fazzari and others, 1988, Whited 1992, Bond and Meghir 1994, Bond and others 2003) use firms’ financial indicators such as the cash flow, the debt to assets ratio, or the liquid- ity ratio as proxies for firms’ net worth or collateral. Most of these papers rely on data for OECD economies and show that wealthier firms invest more. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Similar evidence is found for Ecuador (Haramillo and others 1996) and Cote d’Ivoire (Harrison and McMillan 2003). In a different setting, Gorodnichenko and Schnitzer (2010) use a survey of firms in Eastern European countries and show that financial constraints decrease investment in innovation by domestic firms. Aghion and others (2008) alternatively use measures of firms’ payment incidents for France to analyze the relation between credit constraints and re- search and development along the business cycle. We build on this literature and provide new evidence that financial constraints are preventing firms located in India to invest in foreign capital equipment. Previous empirical studies have investigated whether financial constraints in- fluence the export decisions of firms in the United Kingdom (Greenaway and others 2007) and in several developing economies (Berman and He ´ ricourt 2010). The negative effect of financial constraints on export decisions is observed for the sample of developing countries, but not in the case of the UK. These studies, however, elude the question of financial constraints as a deter- minant of foreign technology adoption through the imports of foreign capital goods. This is the focus of our study. In the next section, we present a simple theoretical framework of import de- cision and credit constraints. Section II describes the data and introduces the estimation strategy. In Section III we present the baseline empirical results. Section IV presents several robustness checks. In the last section, we present our conclusion. I . T H E O R E T I C A L M O T I VA T I O N The aim of this section is to motivate our empirical analysis by introducing a simple model of endogenous adoption of foreign technology. The theory ratio- nalizes the mechanisms through which credit constraints affect firms’ decision to upgrade foreign technology. The model is based on firm heterogeneity in terms of productivity a` la Melitz (2003). Firms are also characterized by their initial wealth as in Chaney (2005).6 They use this wealth as a collateral to get 6. Previous models of heterogeneous firms and credit constraints have also used this framework to ˆ ls (2008). explain the determinants of export decision. See Manova (2008) and Muu 490 THE WORLD BANK ECONOMIC REVIEW external finance in the presence of financial constraints. The representative household allocates consumption from among the range of domestic goods ( j ) produced using domestic-low technology (Vd) and those produced using foreign-high technology (Vf ).7 Production There is a continuum of firms, which are all different in terms of their initial productivity (w). This productivity draw is derived from a common distribution Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 density g(w), after firms decide to enter the market. Each firm produces its own variety in a monopolistic competition market structure. In order to produce the final good ( y) firms must pay a fixed production cost (F) and they need to combine two inputs: labor (l ) and physical capital (k). There are two types of capital equipment goods: domestic (z) and imported (m).8 However, only those firms that are productive enough to adopt the foreign technology are able to produce with imported capital goods. Heterogeneous firms in terms of different productivity levels (w) are introduced. Technology is represented by the follow- ing Cobb-Douglas production function that combines labor (l ) and capital goods (k) to produce output with factor shares h and 1 2 h:  h  1Àh ki li ð1Þ yi ¼ wgi i ¼ fd; f g h 1Àh The subscript d corresponds to firms producing with domestic technology and f to those producing with foreign technology embodied in imported capital goods. The coefficient g represents the efficiency of imported capital goods relative to domestic ones. Firms using only domestic capital goods (i ¼ d) have g ¼ 1 and kd ¼ z. Firms producing with foreign technology (i ¼ f ) Àcombine ÀzÁ a m Á1Àa both types of capital goods by a Cobb-Douglas function: kf ¼ a 1Àa . Firms that decide to adopt foreign technology increase their productivity level by a factor g . 1. To access imported capital goods firms must pay a fixed foreign technology acquisition cost (FT). The fixed technology costs are asso- ciated with gathering information on foreign markets, learning about the foreign technology and establishing linkages with foreign suppliers of this tech- nology. To keep the model simple, we assume that the fixed cost for domestic capital goods is included in the fixed production cost. These assumptions reflect the fact that for a developing country like India, foreign capital goods 7. The standard CES utility function (C) represents the consumer preferences wÀ1 wÀ1 wÀ1 Ð Ð C w ¼ Cdjw dj þ w The elasticity of substitution between both types of goods is given j[Vf C fj dj. j[Vd  w P by f . 1. The optimal relative demand functions are: Ci ¼ p C, where P represents the price index, i C the global consumption and pi the price set by a firm. 8. To keep the model simple, we assume that one unit of domestic capital good is produced using one unit of labor, which is elastically supplied and the wage is normalized to one. Bas and Berthou 491 are more advanced in terms of technology relative to domestic goods, but they are also associated with ahigher initial investment.9 The first-order condition of monopolistic firms is such that prices reflect a   1 constant mark-up, r ¼ wÀ ci w , over marginal costs: pi ¼ rw. ci represents the per unit cost of production, which is different among firms depending on whether or not they have adopted the foreign imported technology: cd ¼ ph z and ah ð1ÀaÞh cf ¼ ð pz Þ ðtm g pz Þ : The price of domestic capital good is pz and the price of imported capital takes into account transport costs and tariffs (tm): pm ¼ tmpz . hð1ÀaÞ c ¼ tm g . We assume that the efficiency Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 f The relative per unit cost is equal to cd parameter of imported capital goods ( g) is higher than its additional variable cost (tm) relative to domestic ones.10  w Combining the demand faced by each firm, qi ðwÞ ¼ piP ðwÞ C; and the price  wÀ1 function, pi ðwÞ ¼ rc i Ai , revenues are given by ri ðwÞ ¼ qi ðw Þ pi ð wÞ ¼ P pi R; where R ¼ PC is the aggregate revenue of the industry exogenous to the firm. Firm profit is then pi ¼ r w À F; where F is the fixed production cost. i Firm’s Decision under Perfect Financial Market Conditions Only those firms with enough profits to afford the fixed production (F) cost will be able to survive and produce. Profits of the marginal firm are equal to r ð w ÃÞ zero. The zero cutoff condition is given by: d wd ¼ F. The value wà d represents the productivity cutoff to produce in the domestic market. Once a firm has received its productivity draw, it may also decide to adopt a foreign technology to reduce its marginal costs on the basis of its profitability. Only a subset of the most productive firms will switch to foreign technology since the fixed importing cost is higher than the fixed production cost. The con- dition to acquire the foreign technology is given by: pf ðwf ÃÞ ¼ 0. The value wÃf rf ðwf ÃÞ represents the productivity cutoff to import foreign goods: w ¼ F þ FT . Firm’s Decision under Imperfect Financial Market Conditions Importing technology embodied in foreign capital goods implies a sunk cost of investment (FT). In the presence of financial constraints, firms cannot use their future expected revenues rf(w) to get external finance ex-ante. In this context, 9. Using product-level imports for India (from the BACI data), we find that about 75 percent of imports of capital goods in India during the 1997-2006 period are sourced from high income OECD economies. This confirms that capital goods are mostly imported from countries that are more advanced in terms of technology. 10. Note that the relative per unit cost is a function of tariffs on capital goods and the efficiency parameter. A reduction of import tariffs on capital goods reduces the relative per unit costs of foreign technology. Similar results hold in the case of an increase in the efficiency parameter of foreign technology ( g). 492 THE WORLD BANK ECONOMIC REVIEW firms can make use of two sources of cash to finance the extra fixed cost FT. First, firms are able to borrow up to rd(w), which corresponds to the sales of the final good for firms using the domestic technology. Financial intermediaries have perfect information about firms’ profitability in the case where they produce with the domestic technology, and will be willing to provide cash in advance up to rd(w). Second, firms can use their exogenous wealth A as a col- lateral to borrow additional liquidity lA, where l corresponds to the credit multiplier and is inversely related to the extent of credit constraints in the economy, as in Aghion and others (1999). We assume that the productivity and the exogenous collateral distributions Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 are independent. The total liquidity that is available to the firm is equal to pd(w) þ lA. Importing foreign capital goods relates to the liquidity constraint condition (LCC) given by ð2Þ pd ðwÞ þ lA ! FT We can define the lowest productivity level below which firms with an exogen- ous wealth A, wðAÞ, are liquidity constrained. wðAÞ is given by pd ðwðAÞÞ þ lA ¼ FT . Firms that face liquidity constraints have a productivity level below wðAÞ: They are not able to import capital goods due to financial constraints. Following Chaney (2005), we set wà d ¼ gðFÞ and use the zero cutoff profit conditions and the liquidity constraint condition, equation (2), to define two productivity cutoffs11:   1 !   1 hð1ÀaÞ F þ FT wÀ1 tm FT þ F À lA wÀ1 à wà f ¼ à wd ; wðAÞ ¼ wd F g F All the firms with a productivity level between maxfwà à f ; wðAÞg . w . wd produce with domestic technology. Only those firms with a productivity w . maxfwà f ; wðAÞg are able to finance the fixed technological cost of import- ing and thus they use both types of capital goods. Which are the firms that face credit constraints to import capital goods? There is a subset of firms that are profitable enough to be viable importers, but prevented from accessing foreign capital goods because of liquidity constraints. Firms that have a productivity level w below wðAÞ are liquidity constrained, and are not able to source imported inputs from abroad no matter how profit- able they could be by importing more efficient foreign capital goods. All firms with a productivity level above wÃf could profitably import, if they had sufficient liquidity. Hence, there is a subset of liquidity constrained firms with a 11. For tractability purposes we assume, as in Chaney (2005), that the price index only depends on local firms’ prices. In the Appendix we define the price index approximation. Bas and Berthou 493 productivity level above wf Ã, but below wðAÞ: In the appendix we demonstrate the existence of liquidity constrained importers. Testable Prediction Firms’ import decision is determined by domestic revenues and by the exogen- ous collateral. These two sources of finance allow firms to afford the fixed tech- nology cost of importing. Using equation (2) we can define the probability that a firm i imports capital goods at time t:   1 r fÀ1 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ð3Þ Prðpd þ lA À FT . 0Þ ¼ PrðwfÀ1 RPwÀ1 þ lA À F À FT Þ . 0 w cd The probability of importing is directly determined by the two sources of finance. On the one hand, in this monopolistic competition framework with heterogeneous firms, the most productive firms set lower prices and have larger domestic revenues to finance the fixed importing cost. On the other hand the higher the exogenous collateral, the greater the financial resources of the firm to afford the fixed foreign technology cost. Testable prediction: In the presence of financial constraints, wealthier firms are more likely to import foreign equipment and upgrade foreign technology. I I . D ATA AND EM P I R I CA L ME T H O D O LO GY In the empirical part of the paper, we present a test of the prediction that is derived from the theoretical model. The empirical strategy is based on the esti- mation of an equation where the import decision of a firm is explained by its fi- nancial factors such as the liquidity or leverage ratios. Estimations are performed using information for a sample of 3,500 Indian listed companies (Prowess data) over the period 1997–2006. Data The Indian firm-level dataset is compiled from the Prowess database by the Centre for Monitoring the Indian Economy (CMIE). This database contains in- formation from the income statements and balance sheets of listed companies comprising more than 70 percent of the economic activity in the organized in- dustrial sector of India. Collectively, the companies covered in Prowess account for 75 percent of all corporate taxes collected by the Government of India. The database is thus representative of large and medium-sized Indian firms.12 The dataset covers the period 1997–2006 and the information varies by year. It provides quantitative information on sales, capital stock, income from 12. Since firms are under no legal obligation to report to the data collecting agency, the Prowess data do not allow properly identifying entry and exit of firms. 494 THE WORLD BANK ECONOMIC REVIEW financial and non financial sources, consumption of raw material and energy, compensation to employees and ownership group. The Prowess database provides detailed information on imports by category of goods: finished goods, intermediate goods and capital goods. In our main empirical specification, we use imports of capital goods (machinery and equip- ment) as a proxy of foreign technology. Although we are not able to test direct- ly for the impact of imported capital goods depending on the countryof origin (e.g developed vs. developing countries), one realistic assumption for the case of a developing country like India is that most imports of capital goods are sourced from more advanced economies. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 The dataset contains also comprehensive information about the financial statements of firms such as total assets, current assets, total debt and liabilities. We construct two financial variables: (1) the leverage ratio and (2) the liquidity ratio. Leverage is the ratio of borrowing over total assets and liquidity ratio is measured by the ratio of current assets over total liabilities of the firm. Summary statistics are provided in the Appendix Table. Our sample contains information for 3,500 firms on average each year in organized industrial activ- ities from manufacturing sector for the period 1997–2006. The total number of observations firm-year pairs is 34,735. In order to keep a constant sample throughout the paper and to establish the stability of the point estimates, we keep firms that report information on all the firm and industry level control variables. On average 32 percent of firms import capital goods in a year and 62 percent of firms import intermediate goods. Firms are categorized by indus- try according to the 4-digit 1998 NIC code (116 industries). Most of the firms in our sample are private-owned firms (81 percent). 39 percent of firms are largest firms belonging to local business groups and only 7 percent are multinational firms. Although our panel of firms is unbalanced, there is no statistical difference in the average firm characteristics presented in the Appendix Table between the initial year (1997) and the final year (2006) of our sample. Two industry-level controls are included in the empirical specifications to control for competitive pressures. Since the period under analysis covers trade liberalization process started in the early 1990s, we introduce effectively applied output tariffs (collected rates) at the 4-digit NIC code level obtained from the World Bank (WITS).13 In order to capture domestic competition we use an Herfindhal index computed at the 2-digit NIC industry level. The Herfindahl index measures the concentration of sales for each industry within 2-digit industry categories. 13. Tariffs data provided by WITS are at the industry level ISIC rev 2 4-digit level. We use correspondence tables to convert tariffs into ISIC rev 3.1. that match almost perfectly with NIC 4-digit classification. Bas and Berthou 495 Empirical Methodology A unique feature of our database is that firms report imports by type of pro- ducts: finished goods, capital goods and intermediate goods. Keeping in line with our theoretical framework, the baseline econometric analysis is therefore performed on capital goods. The rationale for this is that importing capital goods implies incurring fixed costs associated with gathering information on foreign markets, establishing linkages with foreign suppliers, and learning about the new foreign technology. In the case of a developing economy like India, firms’ importing capital goods decision can be interpreted as foreign Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 technology adoption. We estimate a linear probability model, where the decision of a firm i to import capital goods from abroad in year t is explained by its financial factors and additional control variables. Our preferred specification estimates the fol- lowing equation using the following model: ðIÞ ImporterðisÞðtÞ ¼ b0 þ b1 FinanceðiÞðtÀ1Þ þ b2 ZðiÞðtÀ1Þ þ b3 XðsÞðtÞ þ yt þ mi þ nit where Importer(is)(t) is a dummy variable equal to one if the firm i producing in 4-digit NIC code industry s, has positive imports of capital goods in year t and zero otherwise. Finance measures firms’ financial statements. The financial vari- ables of interest that we use to proxy the financial factors (the empirical coun- terpart of the exogenous collateral in the model) are the liquidity ratio and the leverage ratio. The liquidity ratio is the share of firms’ current assets over total liabilities. The liquidity ratio is related to the firm’s ability to pay off its short- terms debts obligations. The leverage ratio indicates the proportion of borrow- ing over total assets of the firm. A higher level of leverage decreases, everything else equals, the net worth of the firm. According to the model’s predictions, an improvement of the firm’s wealth (measured by a higher liquidity ratio or a lower leverage), increases the access to external finance. Since the access to ex- ternal finance determines the decision to source capital goods from abroad, we expect a positive coefficient for the liquidity ratio and a negative coefficient for the leverage ratio. Unobserved firm characteristics could lead to inconsistent estimates. For this reason, all estimations include firm-level fixed effects (mi). The introduction of firm fixed effects is important to control for unobservable firm characteristics that do not vary over time. Our specification shows how improvements in firms’ financial factors over time affects firms’ decisions to import. Estimates also include controls for firm and industry characteristics that vary over time. First, we introduce a set of firm level variables (Z(i)(t21)) expressed in logarithm in year (t-1) that control for observable firm characteristics that 496 THE WORLD BANK ECONOMIC REVIEW mightaffect firms’ import choices. We use the value added to measure firms’ size (the number of employees is not available in the Prowess data). In alterna- tive specifications, we use firm total factor productivity (TFP) computed using Levinsohn and Petrin (2003) methodology, by relying on wage bill rather than labor.14 Since larger firms tend to be more skill intensive and to pay higher wages, we also control for the wage-bill. As we focus on the import decision of capital equipment goods, we also include the past capital intensity of the firm measured as total capital stock over the wage-bill. We expect a positive coeffi- cient of capital intensity. The more firms rely on capital goods in the produc- tion process, the more likely they are to import capital goods from abroad. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Second, we introduce a set of industry level variables X(st) that control for ob- servable industry characteristics that might affect firms’ import choices of capital goods. Several studies show that competition might enhance firm efficiency and create incentives for firms to invest in R&D activities and in foreign technology (Aghion and others 2005). We construct a Herfindahl index at the 2-digit NIC in- dustry level to control for competition in the domestic market. We also control for foreign competition pressures associated with the trade liberalization process experienced by India at the beginning of the 1990s, by including the average ef- fective applied import tariffs for final goods at the 4-digit NIC industry level. All explanatory variables are expressed in logarithm and they are lagged by one period. We also introduce year fixed effects to control for macroeconomic shocks (y t). This is an important control since India was affected by the Asian financial crisis in 1997-1998. The introduction of year fixed effects allows us to control for the effects of this crisis on both financial statements of firms and their import deci- sions. In the last section we deal explicitly with the potential reversecausality between financial factors and firms’ investment decision in imported capital goods. I I I . E S T I M AT I O N R E S U LT S The estimation results of the import decision equation are presented in this third section of the article. All estimations are performed using the above men- tioned firm-level data for India (Prowess). The testable prediction from the model states that “in the presence of financial constraints, wealthier firms are more likely to import foreign equipment and upgrade foreign technology.” Baseline Results Are Financial Factors Related to Firms’ Decision of Sourcing Foreign Capital Goods? Estimation results of the linear probability model (equation I) are pro- vided in Table (1). The estimation includes firm and year fixed effects. The effect of the leverage ratio lagged of one period on firms’ decision to import 14. Because our dataset does not contain the number of employees, we can not rely on the extension of Olley and Pakes (1996) and Levinsohn and Petrin (2003) developed by Ackerberg and others (2007) to estimate total factor productivity. T A B L E 1 . Access to External Finance and Import of Capital Goods Decision (1997–2006) Dummy equal one if firm(i) imports capital goods in t Dependent variable (1) (2) (3) (4) (5) (6) (7) (8) Leverage(i)(t 2 1) 2 0.156*** 2 0.132*** 2 0.114*** 2 0.114*** (0.016) (0.016) (0.016) (0.016) Liquidity ratio (i)(t 2 1) 0.166*** 0.109*** 0.126*** 0.126*** (0.026) (0.026) (0.026) (0.026) Log value added(i)(t 2 1) 0.029*** 0.016*** 0.016*** 0.030*** 0.015*** 0.015*** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Capital intensity(i)(t 2 1) 0.036*** 0.036*** 0.044*** 0.044*** (0.005) (0.005) (0.005) (0.005) Log wage(i)(t 2 1) 0.055*** 0.055*** 0.060*** 0.060*** (0.007) (0.007) (0.006) (0.006) Output tariffs(s)(t 2 1) 0.026 0.020 (0.036) (0.036) Herfindahl index(s)(t 2 1) 0.001 0.000 (0.003) (0.003) Firm fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Observations 34,735 34,735 34,735 34,735 34,735 34,735 34,735 34,735 R2 0.016 0.021 0.025 0.025 0.013 0.018 0.024 0.024 Notes: The table reports estimates from linear probability estimations of Equation (I). The dependent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatory variables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill. The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is at the 2-digit NIC indus- try level. In parentheses we report heteroskedasticity-robust standards errors. ***,**, and * indicate significance at the 1, 5 and 10 percent levels respectively. Bas and Berthou Source: Authors’ estimations using Prowess data. 497 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 498 THE WORLD BANK ECONOMIC REVIEW capital goods is negative and significant at the 1 percent level: firms having a higher ratio of borrowing over total assets are less likely to import their capital goods from the foreign market (column 1). Next, we include firm level variables to control for firm characteristics that vary over time and that could be picking up the effect of firms’ financial factors. As expected, bigger firms are more likely to import capital goods from abroad (column 2). We next introduce two additional firm-level controls: capital intensity and wage-bill in column (3). More capital and skill-intensive firms ( paying higher wages) have a higher probability of upgrading foreign technology. The coefficient of interest remains robust and also stable when we control for firm observable characteristics. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Moreover, firms producing in industries growing faster might be less credit constrained. If this is the case, changes in firms’ financial statements might be capturing the effects of industry characteristics. We address this issue by intro- ducing in column (4) additional controls at the industry level. Concerning the time variant industry characteristics, both coefficients of output tariff and Herfindahl index are not significant. More importantly, the negative effect of leverage on firms’ foreign technology decision remains unchanged and robust to the inclusion of this set of industry controls. The estimated coefficients imply that a 10 percentage point fall in the leverage ratio leads to an 11 percent to 15 percent increase in the likelihood of importing capital goods for the average firm. We test how a firm’s liquidity ratio affects its probability to upgrade foreign technology embodied in imported capital goods in columns (5) to (8). The lagged liquidity ratio is subsequently introduced. The coefficient of the liquidity ratio is positive and significant at the 1 percent level, indicating that more liquid firms are more likely to import their capital goods from abroad (column 5). Results on the liquidity ratio remain robust to the inclusion of firm size, capital and skill-intensity in column (6) and (7) and also to the inclusion of industry-level characteristics in column (8). Moreover, the point estimates of the liquidity ratio are stable under different specifications. The estimated coeffi- cients imply that a 10 percentage point increase in the liquidity ratio leads to a 13 percent to 17 percent increase in the likelihood of importing capital goods for the average firm. Based on these estimations, we use the standard deviation of the leverage and liquidity ratios within firms to have a quantification of their economic impact on the import decision of capital goods by the average firm. We find that a one standard deviation reduction of the leverage ratio, corresponding to a decrease of leverage of the average firm by 32 percent, increases the probabil- ity of sourcing capital goods from abroad by 3.6 percent. A one standard devi- ation increase of the liquidity ratio, corresponding to an increase of the liquidity of the average firm by 17 percent, improves the probability of Bas and Berthou 499 importing capital goods by 2.1 percent. These results confirm that firms finan- cial factors are important determinants of the decision to import capital goods.15 Are Imports of Intermediate Inputs also Affected by Financial Constraints ? In order to disentangle the mechanisms through which financial access affect firms’ foreign technology upgrading, we consider separately the special case of intermediate goods imports. This test allows us to determine whether financial factors affect differently imports of intermediate goods relative to capital goods. First, we estimate equation (I) for the subsample of firms using foreign inter- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 mediate goods. Results are reported in Table 2. Once we control for firm and industry characteristics, the leverage ratio has a negative but not significant effect on firms’ import decision to use foreign intermediate goods (column (1)). Next, we investigate whether the decision to import intermediate goods is asso- ciated to the decision of upgrading foreign technology embodied in capital goods. If such a complementarity exists, the effect of financial factors on imports of intermediates may arise because of its effect through capital goods. To isolate the effect of capital goods decision from foreign input decision, we restrict our sample to firms that have never imported capital goods in the period (columns 2). The results are similar to the previous ones. These findings indicate that credit constraints are not crucial for importing foreign inputs. In the next columns, we reproduce the same specifications using the liquidity ratio. The effect is positive and significant in both cases for the full sample (column 3) and the sample of firms that have never imported capital goods in the period (column 4). The higher the current assets over total liabilities ratio of the firm, the more likely firms are to import their inputs from abroad. Given that firms tend to import intermediates on a regular basis, the positive effect of the liquidity of the firm may not be specifically related to the decision to start importing intermediates. Indeed, 62 percent of firms in our sample report importing intermediates, whereas 32 percent import capital goods (Appendix Table). This evidence suggests that firms find it more difficult to import capital goods than intermediates due to a larger fixed cost. In order to explore the effect of financial health of firms on their decision to start sourcing inputs from abroad, we carry out an additional test focusing on firms that have not imported intermediate goods in the previous two years. These results are reported in columns (5) and (6). As can be seen when we focus on the decision to start importing intermediate inputs, not only the leverage ratio is not signifi- cant but now the liquidity ratio is no longer significant. Firm size, capital and 15. These results are also robust to alternative econometric specifications, available upon request, such as Conditional Logit estimations with firm fixed effects. We only report the linear probability model estimations since the parameters are easier to interpret and their stability easier to establish. 500 T A B L E 2 . Decision to import intermediates Dummy equal one if firm(i) imports intermediates in t (5) (6) Firms that have not imported Dependent variable (1) (2) (3) (4) inputs in the past 2 years Leverage(i)(t-1) 2 0.025 2 0.024 2 0.008 (0.017) (0.025) (0.021) Liquidity ratio (i)(t 2 1) 0.069*** 0.124*** 0.003 (0.026) (0.034) (0.034) Log value added(i)(t 2 1) 0.021*** 0.018*** 0.020*** 0.016*** 0.015*** 0.015*** (0.003) (0.004) (0.003) (0.004) (0.003) (0.003) Capital intensity(i)(t 2 1) 0.042*** 0.037*** 0.045*** 0.039*** 0.041*** 0.041*** (0.005) (0.007) (0.005) (0.007) (0.007) (0.007) THE WORLD BANK ECONOMIC REVIEW Log wage(i)(t 2 1) 0.086*** 0.073*** 0.087*** 0.073*** 0.085*** 0.086*** (0.007) (0.010) (0.007) (0.009) (0.009) (0.009) Output tariffs(s)(t 2 1) 0.039 0.095* 0.038 0.092* 2 0.012 2 0.012 (0.042) (0.051) (0.042) (0.051) (0.048) (0.048) Herfindhal index(s)(t 2 1) 0.003 0.005 0.003 0.005 0.002 0.002 (0.002) (0.004) (0.002) (0.004) (0.004) (0.004) Firm fixed effects Yes Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Yes Observations 34,735 18,477 34,735 18,477 16,190 16,190 R2 0.033 0.027 0.034 0.028 0.036 0.036 Notes: The table reports estimates from linear probability estimations of Equation (I). The dependent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatory variables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill. The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is at the 2-digit NIC indus- try level. In parentheses we report heteroskedasticity-robust standards errors. *** , ** and * indicate significance at the 1, 5 and 10 percent levels respectively. Source: Authors’ estimations using Prowess data. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bas and Berthou 501 skill intensity are positive, significant and more over their coefficients remain stable relative to the previous estimations.16 These results suggest that credit constraints are more binding for importing capital goods than intermediate goods. One of the main differences between the decision of importing capital goods and the decision of using foreign inter- mediate inputsis related to the nature of this choice and the way these produc- tion factors enter into the production process. While intermediate goods are variable inputs that firms have to buy on a regular basis, capital goods are fixed investments (machines and capital equipment) used in the production process that are not renewed every year. As discussed above, the fact that less Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 firms import capital goods than intermediates suggests that the barriers to start importing are larger in the case of capital goods. Moreover, the evidence that most foreign capital goods in India are sourced from developed economies rein- forces this idea that importing capital goods is associated with a higher fixed cost. This may be due to a larger up-front cost in the case of foreign technology upgrading, and also to additional sunk costs related to learning about foreign technologies, finding foreign suppliers, and the adaptation period of the pro- duction process. In the theoretical part of the paper, the model shows that the larger thefixed cost, the more binding are credit constraints. This reasonably explains why firms’ financial factors play a higher role for starting importing capital goods than intermediates. Financial Constraints Versus Tariffs on Capital Goods. In a context of trade liberalization, firms could upgrade foreign technology easily thanks to the removal of import barriers on capital equipment goods. Thereby, the effect of better financial access on foreign technology adoption might just be picking up the effects of lower tariffs on capital equipment goods. In the previous specifications, we include tariffs on final goods at the 4 digit industry level to capture the impact of India’s trade liberalization that took place at the beginning of the nineties. We now explore the robustness of our results when we take into account tariff reductions on capital goods over the period. The average yearly reduction of import tariffs on machinery and equip- ment goods is 2.3 percent during the period. Since trade liberalization in India in the early 90s consisted in a unilateral trade reform, we use effectively applied tariff rates at the HS6 product level set by India to the rest of the world for the period 1997 to 2006. This is possible thanks to the match of the firm level data with the average import tariff data of products corresponding to HS6 codes between 840000 and 859999 (machinery and mechanical appli- ances) from the World Bank (WITS). The results including the variation of the average import tariffs on capital equipment goods are presented in Table 3. As expected, a reduction of import barriers on capital goods increases the likelihood of firms to upgrade foreign technology. Most importantly, our 16. The results are similar when we use as an alternative specification, available upon request, the previous import status of intermediate goods as a control variable relying on the full sample of firms. 502 T A B L E 3 . Trade liberalization and imports of capital goods Dummy equal one if firm(i) imports of capital goods Dependent variable (1) (2) (3) (4) (5) D tariffs capital goods 2 0.321*** 2 0.295*** 2 0.309*** 2 0.297*** 2 0.311*** (0.052) (0.051) (0.050) (0.052) (0.051) Leverage(i)(t 2 1) 2 0.133*** 2 0.133*** (0.016) (0.016) Liquidity ratio (i)(t 2 1) 0.162*** 0.161*** (0.026) (0.026) Log value added(i)(t 2 1) 0.019*** 0.019*** 0.018*** 0.018*** (0.003) (0.003) (0.003) (0.003) Capital intensity(i)(t 2 1) 0.042*** 0.042*** 0.052*** 0.052*** (0.005) (0.005) (0.005) (0.005) THE WORLD BANK ECONOMIC REVIEW Log wage(i)(t 2 1) 0.046*** 0.047*** 0.052*** 0.053*** (0.007) (0.007) (0.007) (0.007) Output tariffs(s)(t 2 1) 0.027 0.027 (0.029) (0.029) Herfindahl index(s)(t 2 1) 0.001 0.001 (0.003) (0.003) Firm fixed effects Yes Yes Yes Yes Yes Observations 34,735 34,735 34,735 34,735 34,735 R2 0.001 0.017 0.017 0.015 0.015 Notes: The table reports estimates from linear probability estimations of Equation (I). The dependent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatory variables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill. The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is at the 2-digit NIC indus- try level. In parentheses we report heteroskedasticity-robust standards errors. *** , ** and * indicate significance at the 1, 5 and 10 percent levels respectively. Source: Authors’ estimations using Prowess data. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bas and Berthou 503 results remain unaffected by the introduction of import tariffs on capital goods. Once we take into account directly the effects of trade reform, a reduc- tion of the leverage ratio and an increase in the liquidity of the firm have both a positive impact on the probability of adopting a foreign technology. Comparing the point estimates of the coefficients of the leverage and liquidity ratios with those reported in baseline estimations (Table 1) reveals that the coefficients on the variables of interest remain stable. Estimation results in Table 3 show that a 10 percentage point reduction of the leverage ratio increases the probability of upgrading foreign technology by 13 percent. Similarly, a 10 percentage point increase in the liquidity ratio is associated Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 with an increase in the likelihood of importing capital goods of 16 percent in this specification. These effects are very comparable to the baseline specification. Additional firm characteristics. This section presents alternative sensitivity tests related to other firm-characteristics that might be driving our results. In the previous estimations, firm controls include the wage bill, value added and capital intensity. Although value added is positively correlated with firm productivity, it is a raw measure for efficiency gains. As an additional test we use firm total factor productivity measured by Levinsohn and Petrin (2003) methodology, using wage bill as a proxy for firms’ labor utilization. Our find- ings show that most productive firms have a higher probability of upgrading foreign technology (columns (1) and (2) of Table (4)). The inclusion of firms’ total factor productivity does not modify the results as compared with the baseline specification. The point estimates suggest that a reduction of leverage ratio of 10 percentage points or an analogous increase in the liquidity ratio increases the probability of sourcing capital goods from abroad by 14 and 15 percent, respectively. Next we address whether firms’ ownership is driving our results. Previous studies on multinational firms show that foreign firms in developing countries tend to use more advanced technologies and be more productive relative to do- mestic firms (Javorcik 2004). One reason may be that foreign multinationals have a better access to finance, and are more likely to source capital goods from abroad. Javorcik and Spatareanu (2009) also show that the suppliers of multinationals are less credit constrained.17 In general, the fact that foreign companies are wealthier firms and use more advanced technology could poten- tially explain our results. In order to address this issue, we exclude from our sample multinational firms in columns (3) and (4) of Table (4). Our coefficients of interest on financial variables remain robust and stable when we restrict the sample to domestic firms, implying that financial factors matter when consider- ing the sample of domestic firms.18 17. Manova and others (2009) also show that in the case of China, multinationals have a better propensity to export in sectors where firms are typically more financially vulnerable. 18. We thank anonymous referees for having pointed out this channel. T A B L E 4 . Additional firm characteristics 504 dummy ¼ 1 if firm imports capital goodsit = 1 Subsample Non-Business Controlling for TFP Subsample Non-MNF Subsample Private firms groups Dependent variable (1) (2) (3) (4) (5) (6) (7) (8) Leverage(i)(t-1) 2 0.141*** 2 0.119*** 2 0.118*** 2 0.120*** (0.018) (0.017) (0.017) (0.020) Liquidity ratio (i)(t 2 1) 0.148*** 0.132*** 0.119*** 0.094*** (0.028) (0.027) (0.029) (0.030) TFP(i)(t 2 1) 0.026*** 0.028*** (0.009) (0.009) Log value added(i)(t 2 1) 0.016*** 0.015*** 0.015*** 0.015*** 0.011*** 0.011*** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) THE WORLD BANK ECONOMIC REVIEW Capital intensity(i)(t 2 1) 0.038*** 0.048*** 0.035*** 0.043*** 0.034*** 0.042*** 0.032*** 0.040*** (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) (0.006) (0.006) Log wage(i)(t 2 1) 0.082*** 0.087*** 0.052*** 0.057*** 0.051*** 0.056*** 0.058*** 0.063*** (0.007) (0.007) (0.007) (0.007) (0.007) (0.007) (0.008) (0.008) Output tariffs(s)(t 2 1) 0.025 0.018 0.042 0.037 0.012 0.006 2 0.006 2 0.011 (0.038) (0.038) (0.037) (0.038) (0.040) (0.040) (0.044) (0.044) Herfindahl index(s)(t 2 1) 0.002 0.002 0.001 0.001 0.002 0.001 0.002 0.001 (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Firm fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Observations 32,936 32,936 32,275 32,275 28,301 28,301 21,321 21,321 R2 0.026 0.024 0.025 0.023 0.023 0.021 0.024 0.021 Notes: The table reports estimates from linear probability estimations of Equation (I). The dependent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatory variables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill. The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is at the 2-digit NIC indus- try level. In parentheses we report heteroskedasticity-robust standards errors. *** ,** , and * indicate significance at the 1, 5 and 10 percent levels respectively. Source: Authors’ estimations using Prowess data. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bas and Berthou 505 Besides, previous works using the same firm-level dataset have emphasized the role of state-owned firms in India (Topalova 2004, Alfaro and Chari 2009). State-owned firms might benefit from special access to credit from State-owed banks. In order to address this issue, we restrict the sample to private firms (columns 5 and 6). The point estimates of leverage and liquidity ratio remain robust and stable. Finally, the largest domestic firms belonging to the Indian business groups in India could also benefit from a better access to finance. Our results are robust to the exclusion of these groups from the estimation sample (columns 7 and 8). These tests confirm that different firm ownership character- istics are not picking up our results. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Robustness Analysis One of the challenges when investigating the relationship between the access to external finance and firms’ technology adoption decisions is the potential reverse causality. In the medium or long run, importing foreign capital goods is expected to increase the profitability of the firm and therefore its financial state- ments (reduce the leverage or increase the liquidity ratio). This mechanism would result in a positive bias in the relation between imports and financial factors of the firm. In the short run, the cost associated with the imports of a new technology is expected to increase the leverage of the firm, or decrease its liquidity. This mechanism would result in a negative bias. We perform two ro- bustness checks to address this potential reverse causality issue.19 Decision to Start Importing Capital Goods. We explore the robustness of our baseline specification when we restrict our sample to firms that have not imported capital goods in the previous years. We investigate whether an in- crease in the access to external finance is associated with the decision to start sourcing capital goods from abroad. By focusing on firms that have not imported capital goods in the previous period, this specification deals with the possible endogeneity issues between financial access and foreign technology adoption that the previous specifications might suffer. The estimates from linear probability estimations of equation (I) with firm and year fixed effects for the restricted sample of firms that have not imported capital goods in the last two years are reported in columns (1) to (2) of Table 5. In this case, the coefficients of the financial variables are smaller com- pared to the baseline specification due to the reduction of the sample size from 34,735 observations to almost 21,000. The point estimates indicate that a 10 percentage point reduction of the leverage ratio increases the probability to 19. In estimations available upon request, we also carry out a two-stage least square (2SLS) linear probability model where the liquidity ratio, the leverage ratio, and the capital intensity are instrumented with lagged values (three to four years) and the mean capital intensity of the industry. The results remain robust for the leverage ratio under the instrumental variable specification, while the liquidity ratio is no longer significant. However, this last result should be interpretedwith caution, given the reduction of the sample size due to the use of lagged instruments. This restriction leaves us with half of the sample under the IV relative to the baseline estimations. 506 THE WORLD BANK ECONOMIC REVIEW T A B L E 5 . Decision to start importing capital goods dummy ¼ 1 if firm imports capital goodsit ¼ 1 Firms that do not import capital goods in the previous Two years Four years Dependent variable (1) (2) (3) (4) Leverage(i)(t 2 1) 2 0.058*** 2 0.052*** (0.012) (0.012) Liquidity ratio (i)(t 2 1) 0.050** 0.022 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 (0.021) (0.020) Log value added(i)(t 2 1) 0.004* 0.004* 0.000 0.001 (0.002) (0.002) (0.002) (0.002) Capital intensity(i)(t 2 1) 0.012*** 0.016*** 0.013*** 0.017*** (0.004) (0.004) (0.004) (0.003) Log wage(i)(t 2 1) 0.029*** 0.032*** 0.038*** 0.040*** (0.006) (0.005) (0.005) (0.005) Output tariffs(s)(t 2 1) 2 0.055* 2 0.058** 2 0.069*** 2 0.072*** (0.028) (0.029) (0.027) (0.027) Herfindahl index(s)(t 2 1) 2 0.002 2 0.003 2 0.001 2 0.001 (0.002) (0.002) (0.002) (0.002) Firm fixed effects Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Observations 20,993 20,993 18,600 18,600 R2 0.014 0.013 0.024 0.022 Notes: The dependent variable is a dummy equal to one if the firm i imports capital goods in t and have not imported in the previous two years (columns 1 and 2) or four years (columns 3 and 4). We use the same control variables as in Table 1. ***, ** , and * indicate significance at the 1, 5 and 10 percent levels respectively. Source: Authors’ estimations using Prowess data. start importing capital goods by 6 percent (column (1)). Similarly, a 10 per- centage point increase in the amount of liquidity increases the probability to upgrade foreign technology for the first time by 5 percent (column (2)). When we restrict our sample to firms that have not imported capital goods in the last four years, the effect of the leverage ratio is still negative, significant and stable, while the liquidity ratio is no longer significant (columns (3) and (4)). As an alternative test we include the past import experience in the baseline estimations. In this case, we keep the full sample of firms and include the lagged importer status of the firm measured by a dummy variable that is equal to one if the firm has been an importer of capital goods in the previous years. This specification allows us to take into account the past experience of import- ing capital goods that can reduce the fixed costs in the present.20 These results are reported in Table 6. As expected the previous import status has a positive effect on the decision of importing capital goods in year t. The point estimates 20. We thank an anonymous referee for having pointed out this channel. Bas and Berthou 507 T A B L E 6 . Controlling for past import experience dummy ¼ 1 if firm imports capital goodsit ¼ 1 Dependent variable (1) (2) (3) (4) Importer status(i)(t 2 1) 0.121*** 0.110*** 0.124*** 0.111*** (0.009) (0.009) (0.009) (0.009) Leverage(i)(t 2 1) 2 0.143*** 2 0.109*** (0.015) (0.015) Liquidity ratio (i)(t 2 1) 0.167*** 0.132*** (0.024) (0.024) Log value added(i)(t 2 1) 0.014*** 0.013*** Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 (0.003) (0.003) Capital intensity(i)(t 2 1) 0.029*** 0.036*** (0.005) (0.005) Log wage(i)(t 2 1) 0.044*** 0.049*** (0.006) (0.006) Output tariffs(s)(t 2 1) 0.020 0.014 (0.034) (0.035) Herfindahl index(s)(t 2 1) 0.000 0.000 (0.003) (0.003) Firm fixed effects Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Observations 34,735 34,735 34,735 34,735 R2 0.031 0.037 0.029 0.036 Notes: The table reports estimates from linear probability estimations of Equation (I). The de- pendent variable is a dummy equal to one if the firm i imports capital goods in t. All explanatory variables are lag of one period. Firms’ capital intensity is the ratio of capital over the wage-bill. The financial variables that we use are leverage(i) and liquidity ratio(i). Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. The output tariffs are at the 4-digit NIC industry level and the Herfindahl index is at the 2-digit NIC industry level. In parentheses we report heteroskedasticity-robust standards errors. ***, ** and * indicate significance at the 1, 5 and 10 percent levels respectively. Source: Authors’ estimations using Prowess data. of leverage and liquidity ratio remain almost unchanged relative to the ones presented in the baseline specifications in Table 1. These findings confirm the importance of financial access to start sourcing capital goods from abroad. Dependence with Respect to External Finance. As a final exercise, we use the measure of firms’ dependence on external finance (“external dependence”), proposed by Rajan and Zingales (1998) and updated by Braun (2002) and Braun and Larrain (2005), to identify an exogenous effect of financial con- straints on capital goods imports across different industries. In the presence of financial constraints, the borrowing capacity of a firm is closely related to its fi- nancial statement. Financial constraints are therefore expected to affect more the investment decision in sectors where firms rely more on the use of external finance. 508 T A B L E 7 . Imports of capital goods - dependence on external finance in the industry dummy ¼ 1 if firm imports capital goodsit ¼ 1 Dependent variable (1) (2) (3) (4) Leverage(i)(t 2 1) 2 0.077*** 2 0.098* (0.024) (0.051) Leverage(i)(t 2 1)  Ext.Dep.(s) 2 0.124* 2 0.121* (0.069) (0.070) Leverage(i)(t 2 1)  Cap.Int(s) 0.258 (0.537) Liquidity ratio (i)(t 2 1) 0.048 2 0.033 (0.036) (0.066) Liquidity(i)(t 2 1)  Ext.Dep.(s) 0.241*** 0.255*** (0.087) (0.089) Liquidity(i)(t 2 1)  Cap.Int.(s) 1.016 (0.650) Log value added(i)(t 2 1) 0.016*** 0.016*** 0.015*** 0.015*** THE WORLD BANK ECONOMIC REVIEW (0.003) (0.003) (0.003) (0.003) Capital intensity(i)(t 2 1) 0.036*** 0.036*** 0.044*** 0.044*** (0.005) (0.005) (0.005) (0.005) Log wage(i)(t 2 1) 0.055*** 0.055*** 0.060*** 0.060*** (0.007) (0.007) (0.007) (0.007) Herfindahl index(s)(t 2 1) 0.001 0.001 0.001 0.001 (0.003) (0.003) (0.003) (0.003) Output tariffs(s)(t 2 1) 0.018 0.019 0.014 0.017 (0.037) (0.037) (0.037) (0.037) Firm fixed effects Yes Yes yes Yes Year fixed effects Yes Yes Yes Yes Observations 33,773 33,773 33,773 33,773 R2 0.026 0.026 0.024 0.025 Notes: ***, ** and * indicate significance at the 1, 5 and 10 percent levels respectively. Leverage(i)(t-1)  Ext.Dep. and Liquidity(i)(t-1)  Ext.Dep. are interaction variable between the Leverage ratio and the variable external dependence provided by Braun (2002) and Braun and Larrain (2005). Leverage(i)(t-1)  Cap.Int. is the interaction of the leverage ratio with the capital intensity of the industry, also provided by Braun (2002). External de- pendence and capital intensity are sector-specific, with ISIC rev. 2 3-digits classification. Source: authors’ estimations using Prowess data. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bas and Berthou 509 The empirical strategy proposed by Rajan and Zingales (1998) is adapted to the context of our study.21 The measure of external dependence at the 2-digit industry level updated by Braun (2002) and Braun and Larrain (2005) is inter- acted with our measures of firms’ financial statements. The baseline empirical specification is then augmented with the Leverage(i)(t-1)  Ext. Dep.(s) and Liquidity ratio (i)(t-1)  Ext. Dep.(s) variables. The coefficient on the inter- action variable between the leverage of the firm and the external dependence of the industry is expected to be negative, and the coefficient on the interaction between the liquidity ratio of the firm and the degree of external dependence is expected to be positive: in the presence of financial constraints, the liquidity Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ratio and leverage of the firm are expected to be more closely related to the imports of foreign capital goods for firms that rely more on the use of external finance. Estimation results are reported in Table 7. The leverage ratio is interacted with the external dependence variable in column (1). The coefficient on the interaction variable reports a negative sign, confirming that the negative impact of the leverage of the firm, on its probability to import foreign capital goods, is higher in sectors where firms require more external finance. This es- timation is replicated in column (2), including as well an interaction between the leverage of the firm and the capital intensity of the industry. The capital intensity of the industry is provided by Braun (2002), and is sector-specific. This new variable allows to control for the possibility that importing capital goods is more likely to affect firms’ financial factors in sectors where firms are typically more capital intensive. Since the external dependence of the firm and the capital intensity are positively correlated, reverse causality would bias the coefficient on the Leverage(i)(t-1)  Ext. Dep.(s) variable. Estimation results in column (2), though, confirm that the point estimate of the inter- action term, Leverage(i)(t-1)  Ext. Dep.(s), is robust and stable under this specification. A similar analysis where the liquidity ratio is interacted with the sectoral ex- ternal dependence and the sectoral capital intensity is provided in columns (3) and (4). The coefficient on the Liquidity ratio(i)(t-1)  Ext. Dep.(s) is positive and significant (column 3) and it remains robust and stable when we introduce the interaction term between the liquidity of the firm and the capital intensity of the industry in column (4). These sensitivity tests therefore provide addition- al evidence confirming that the liquidity of the firm affects the import decision of capital goods. 21. Rajan and Zingales (1998) propose to identify the effect of financial development on economic growth, using an interaction term between the country’s financial development and the industry level of external dependence. The degree of dependence on external finance is a technology parameter (measured using Compustat data for the United States), and is independent of countries’ characteristics. The coefficient on the interaction term is therefore expected to be unrelated to countries’ characteristics, and unaffected by future economic growth. 510 THE WORLD BANK ECONOMIC REVIEW I V. C O N C L U D I N G R E M A R K S Adopting foreign technology is costly and requires using internal and external financial resources. This paper investigates the influence of firms’ financial factors on their decision to source foreign capital goods. We test whether firms that experience an improvement in their financial statements have a higher probability to upgrade foreign technology embodied in imported capital goods. We find strong evidence that this is the case in India. Firms with a lower leverage and a higher liquidity have a higher probability of up- grading foreign technology. Different sensitivity tests demonstrate that these Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 results are not driven by omitted variable bias related to changes in firm ob- servable characteristics (size, capital, and skill intensity) as well as ownership status (multinational, state-owned firms and local Indian business groups). Finally, these findings are also robust to alternative specifications dealing with the potential reverse causality issues between financial factors and foreign technology adoption. Our findings suggest that financial market imperfections have a negative effect on purchases of foreign technology. This is an important issue for aggre- gate productivity growth in developing countries, like India, that rely heavily on foreign technology in their production process. One important policy impli- cation of our findings is that the success of trade reforms is closely related to the capacity of the financial intermediaries to provide funding to domestic firms. A.THEORETICAL APPENDIX A.1. Price index approximation Following Chaney (2005), we assume that the price index only depends on local firms’prices and that foreign firms do not face any liquidity constraints. The price index approximation is ð ! 1 1Àw 1Àw P% pd ðwÞ LdFw ðwÞ w!wd à We define a function g(.) in the following way:  ð  ÃfÀ1 w fÀ1 gð:Þ : w ¼ w dFw ðwÞ Â F , w à ¼ gðFÞ m w!w à Bas and Berthou 511 A.2. Credit constrained firms A sufficient condition for the existence of liquidity constraints importers is cf cd , 1: This is the assumption cf that hð1ÀaÞ we introduce concerning the relative per unit cost is then equal to cd ¼ tm g , 1: This condition implies that the effi- ciency parameter of imported capital goods is higher than its additional vari- hð1ÀaÞ able cost relative to domestic ones ðg . tm Þ: Proposition 1: Under the assumption that x , 1, there is a subset of firms (denoted F) subject to liquidity constraints with a productivity level between wà f , w , wðAÞ: Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 B.EMPIRICAL APPENDIX A p p e n d i x Ta b l e : Descriptive statistics of Indian manufacturing firms (1997–2006) Mean Std. Dev. Number of firms Average number of firms per year 3,473 Importers of capital goods (%) 32 Importers of intermediate goods (%) 62 Private firms (%) 81 Local business groups (%) 39 Foreign firms 7 percent Financial variables Liquidity ratio 0.50 0.20 Leverage ratio 0.38 0.31 Firm level characteristics Value added 50 214 Wage bill 7.65 46 Capital stock 113 534 Industry level controls Effectively applied output tariffs (NIC 4 digit) 0.30 0.13 Herfindahl index (NIC 2 digit) 0.94 0.78 Notes: Mean values and standard errors in parentheses are reported. Leverage(i) is the ratio of borrowings over total assets and liquidity ratio(i) is the ratio of current assets over total liabilities of the firm. Source: authors calculations based on Prowess data. Proof In order to prove that F is not empty we investigate whether wð0Þ . wf Ã:   1   1   F þ FT À lA wÀ1 F þ FT wÀ1 cf wd à . w à F F cd d 512 THE WORLD BANK ECONOMIC REVIEW REFERENCES Ackerberg, D., K. Caves, and G. Frazer. 2007. “Structural Identification of Production Functions.” Working paper, UCLA. Aghion, P., P. Askenazy, N. Berman, G. Cette, and L. Eymard. 2008. “Credit Constraints and the Cyclicality of R&D Investment: Evidence from France.” PSE Working Papers 2008-26. Aghion, P., A. Banerjee, and T. Piketty. 1999. “Dualism and Macroeconomic volatility.” Quarterly Journal of Economics 114: 1359–1397. Aghion, P., N. Bloom, Blundell, R. Griffith, and P. Howitt. 2005. “Competition and Innovation: An Inverted U Relationship.” Quarterly Journal of Economics 120: 701– 728. Alfaro, L., and A. Chari. 2009. “India Transformed? Insights from the Firm Level 1988-2005.” NBER Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Working Papers 15448. Amiti, M., and J. Konings. 2007. “Trade liberalization, Intermediate Inputs, and Productivity: Evidence from Indonesia.” American Economic Review 97(5): 1611– 1638. Arnold, J., B. Javorcik, M. Lipscomb, and A. Mattoo. 2010. “Services Reform and Manufacturing Performance: Evidence from India.” CEPR Discussion Paper 8011. Bas, M. 2011. “Input-trade Liberalization and Firm Export Decisions: Evidence from Argentina.” Journal of Development Economics 97 (2012) 481– 493. Bas, M., and V. Strauss-Kahn. 2011. “Does Importing more Inputs Raise Exports? Firm Level Evidence from France.” Working Paper CEPII. Berman, N., and J. He ´ ricourt. 2010. “Financial Factors and the Margins of Trade : Evidence from Cross-country Firm-level Data.” Journal of Development Economics 93 (2): 206– 217. Bond, S., J. Elston, J. Mairesse, and B. Mulkay. 2003. “Financial Factors and Investment in Belgium, France, German and the UK: A comparison using Company Panel Data.” Review of Economics and Statistics 85: 153– 165. Bond, S., and C. Meghir. 1994. “Dynamic Investment Models and the Firm’s Financial Policy.” Review of Economic Studies 61(2): 197–222. Braun, M. 2002. “Financial Contractibility and Asset Hardness.” Mimeo. Braun, M., and B. Larrain. 2005. “Finance and the Business Cycle: International, Inter-industry Evidence.” Journal of Finance 60(3): 1097–1128. Chaney, T. 2005. “Liquidity Constrained Exporters.” Chicago University. Processed. Eaton, J., and S. Kortum. 2001. “Trade in Capital Goods.” European Economic Review 45(7): 1195– 1235. Ethier, W. 1982. “National and International Returns to Scale in the Modern Theory of International Trade.” American Economic Review 72(3): 389 –405. Fazzari, S. M., R. G. Hubbard, and B. C. Petersen. 1988. “Financing Constraints and Corporate Investment.” Brookings Papers on Economic Activity 19(1988-1): 141– 206. Goldberg, P., A. Khandelwal, N. Pavcnik, and P. Topalova. 2009. “Trade Liberalization and New Imported Inputs.” American Economic Review Papers and Proceedings 99(2): 494– 500. Goldberg, P. K., A. K. Khandelwal, N. Pavcnik, and P. Topalova. 2010. “Imported intermediate inputs and Domestic Product Growth: Evidence from India.” The Quarterly Journal of Economics 125(4): 1727– 1767. Gorodnichenko, Y., and M. Schnitzer. 2010. “Financial Constraints and Innovation: Why Poor Countries Don’t Catch Up.” NBER Working Paper No. 15792. Greenaway, D., A. Guariglia, and R. Kneller. 2007. “Financial Factors and Exporting Decisions.” Journal of International Economics 73(2): 377–395. Grossman, G., and E. Helpman. 1991. “Innovation and Growth in the Global Economy.” Cambridge, MA: MIT. Halpern, , Koren, and Szeidl. 2009. “Imported Inputs and Productivity.” CEPR N. 5139. Bas and Berthou 513 Haramillo, F., F. Schiantarelli, and A. Weiss. 1996. “Capital Market Imperfections before and after Financial Liberalization: An Euler Equation Approach to Panel Data for Ecuadorian Firms.” Journal of Development Economics 51(2): 367– 386. Harrison, A. E., and M. S. McMillan. 2003. “Does Direct Foreign Investment Affect Domestic Credit Constraints?” Journal of International Economics 61(1): 73– 100. Hubbard, R. G. 1998. “Capital-market Imperfections and Investment.” Journal of Economic Literatture 36(1): 193 –225. Javorcik, B. S. 2004. “Does Foreign Direct Investment Increase the Productivity of Domestic Firms? in Search of Spillovers through Backward Linkages.” American Economic Review 94(3): 605–627. Javorcik, B. S., and M. Spatareanu. 2009. “Liquidity Constraints and Firms’ Linkages with Multinationals.” World Bank Economic Review 23(2): 323–346. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Kasahara, H., and J. Rodrigue. 2008. “Does the Use of Imported Intermediates Increase Productivity? plant-Level Evidence.” Journal of Development Economics 87: 106– 118. Khandelwal, A., and P. Topalova. 2010. “Trade Liberalization and Firm Productivity: The Case of India.” The Review of Economics and Statistics Forthcoming. Kugler, M., and E. Verhoogen. 2009. “Plants and Imported Inputs: New Facts and an Interpretation.” American Economic Review Papers and Proceedings 99(2): 501–507. Levinsohn, J., and A. Petrin. 2003. “Estimating Production Functions Using Inputs to Control for Unobservables.” Review of Economic Studies 70(2): 317 –341. Manova, K. 2008. “Credit Constraints, Heterogeneous Firms and International Trade.” NBER Working Paper 14531. Manova, K., S.-J. Wei, and Z. Zhang. 2009. “Firm Exports and Multinational Activity under Credit Constraints.” Stanford University. Processed. Markusen, J. 1989. “Trade in Producer Services and in other Specialized Intermediate Inputs.” American Economic Review 79(1): 85 –95. Melitz, M. 2003. “The Impact of Trade on Intra-industry Reallocations and Aggregate Industry Productivity.” Econometrica 71: 1695– 1725. ˆ ls, M. 2008. “Exporters and Credit Constraints. A Firm-Level Approach.” National Bank of Muu Belgium Working Paper. Olley, G., and A. Pakes. 1996. “The Dynamics of Productivity in the Telecommunications Equipment Industry.” Econometrica 64(6): 1263–1297. Rajan, R., and L. Zingales. 1998. “Financial Dependence and Growth.” American Economic Review 3: 559–586. Rivera-Batiz, L., and P. Romer. 1991. “International Trade with Endogenous Technological Change.” European Economic Review 35(4): 971– 1001. Schor, A. 2004. “Heterogeneous Productivity Response to Tariff Reduction: Evidence from Brazilian Manufacturing Firms.” Journal of Development Economics 75: 373– 396. Topalova, P. 2004. “Overview of the Indian Corporate Sector: 1989– 2002.” IMF Working Papers 04/64. Whited, T. M. 1992. “Debt, Liquidity Constraints, and Corporate Investment: Evidence from Panel Data.” Journal of Finance 47(4): 1425– 1460. Coffee Market Liberalisation and the Implications for Producers in Brazil, Guatemala and India* Bill Russell, Sushil Mohan, and Anindya Banerjee The standard approach to modelling the relationship between world and producer prices of coffee does not incorporate the effects of changing government policies and market structures. These changes have led to large structural breaks in the relationship Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 between the prices implying the standard estimates are biased. We model coffee prices in Brazil, Guatemala and India allowing for the structural breaks and show that the liberalisation of coffee markets has benefited producers substantially both in terms of a higher share of the world price of coffee and higher real prices. This suggests that calls to re-regulate coffee markets may be misplaced. JEL Classification: Q11, Q17, Q18, C32, C52, F13, F14 Before the 1990s, unilateral and multilateral interventions in coffee markets were common. The governments of most coffee-producing countries considered regulation of coffee marketing and pricing necessary because of coffee’s im- portance as a source of foreign exchange and government revenue.1 For major coffee-producing countries such as Brazil and Colombia the main objective for regulation was to raise world coffee prices. Countries also used regulation to maintain fixed producer prices so as to shield coffee producers (hereafter * Bill Russell (corresponding author), Economic Studies, School of Business, University of Dundee, Dundee DD1 4HN, United Kingdom. þ 44 1382 385165 (work phone), þ 44 1382 384691 (fax), email brussell@brolga.net. Sushil Mohan, Economic Studies, School of Business, University of Dundee, United Kingdom. Anindya Banerjee, Department of Economics, Birmingham Business School, University of Birmingham, United Kingdom. We would like to thank David Hendry and Hassan Molana for their helpful advice, Ivan Carvalho, Denis Seudieu and Martin Wattam from the International Coffee Organization for help in providing the data, and Tom Doan and Pierre Perron for graciously making available the programmes for estimating structural breaks. A Supplementary Appendix to this article is available at http://wber.oxfordjournals.org and http://billrussell.info. The data are available at http:// billrussell.info. 1. Unless specified otherwise, ‘coffee’ means green (raw or un-roasted) beans and coffee prices imply prices of green beans. Coffee producers include growers and/or semi-processors, who sell their coffee as cherries, parchment or green beans. If producers sell their coffee as cherry or parchment, the prices are converted to green beans by using a ‘green bean equivalent’. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 514– 538 doi:10.1093/wber/lhr055 Advance Access Publication January 10, 2012 # The Author 2012. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 514 Bill Russell, Sushil Mohan and Anindya Banerjee 515 referred to as producers) from price fluctuations and assured them a minimum price.2 At the multilateral level the interventions generally took the form of regula- tion of export supply and prices. The first International Coffee Agreement (ICA) was signed in 1962 by the major coffee producing (exporting) and con- suming (importing) countries. According to its regulatory provisions, basic export quotas were allocated to each of the exporting member countries for export to importing member countries and they were tightened if international coffee prices fell below a particular level and loosened when they rose above that level. The monitoring requirements of the ICA supported the domestic Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 regulation of coffee markets in coffee-producing countries to ensure compliance with quota restrictions.3 The ICA was undermined by some member countries distributing their exports at lower prices through non-member countries, the in- ability to agree on quotas and the continuing fragmentation of the geography of production. Quotas were operational between October 1963 to December 1973, October 1980 to March 1986, and November 1987 to July 1989 (ICO, 1989). The agreement was suspended in 1989. The suspension of the International Coffee Agreement (ICA) in 1989 and broader economic reforms including exchange-rate reforms in most developing countries in the late 1980s resulted in most coffee-producing countries liberalis- ing their coffee sector by replacing state-controlled marketing systems with markets run by private agents. The pace and scope of liberalisation has varied across countries but has led to a more competitive international coffee market in which producer prices reflect more accurately the domestic and international market conditions. In terms of their impact on producer welfare the interventions are generally regarded as unsuccessful. The cost of reduced volatility seemed too high given that the administered prices were usually far below the certainty equivalent that would be accepted by producers.4 Jarvis (2005) and Mehta and Chavas (2008) find that the interventions resulted in high levels of rent seeking by a range of beneficiaries including bureaucrats, intermediaries such as coffee mar- keting boards and foreign importers, but not by producers. The accruing of these economic rents to intermediaries in the coffee supply chain creates a larger margin between the international and producer prices of coffee resulting in a lower share of the international price of coffee going to producers. However, some commentators do not accept these assertions and raise con- cerns on the effects of these reforms on the real price of coffee received by pro- ducers. They feel that the principal beneficiaries of liberalisation have been 2. See Akiyama (2001). 3. For a brief description of coffee market interventions see Supplementary Appendix S1 available at http://wber.oxfordjournals.org and http://www.billrussell.info. See also Raffaeli (1995), Gilbert (1996), Bates (1997), McIntire and Varangis (1999), Jerome and Ogunkola (2000), Akiyama (2001), Varangis et al. (2002) and Winter-Nelson and Temu (2002). 4. See Akiyama et al. (2001), Krivonos (2004) and Anderson et al. (2008). 516 THE WORLD BANK ECONOMIC REVIEW coffee roasters and international traders who are able to capture all the monop- oly rents and dictate the prices they are willing to pay to producers.5 Lately there have been calls for a return to some form of coffee market interventions or regulations on the grounds that the liberalisation of coffee markets has not improved the plight of producers.6 To evaluate if producers were better or worse off due to the regulation and interventions in coffee markets requires us to compare the actual returns they received with some counter-factual return that they would have received in un- regulated markets. The terminal (or international) price of coffee has not been directly regulated or administered over the last four decades.7 Consequently, if Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 the ‘law of one price’ between the terminal and producer prices of coffee holds then we can use the terminal price to calculate the corresponding un-regulated producer price of coffee. This allows us to compare the actual returns to produ- cers with the return that producers would have received in an un-regulated market. Unfortunately, the recent extensive literature does not provide unambiguous empirical evidence in support of the ‘law of one price’. The results depend on the level of aggregation of the data and the methods employed to ‘test’ the rela- tionship between prices in different markets. For example, the work on com- modity markets by Ardeni (1989) and Baffes (1991) provide mixed evidence on accepting the ‘law’ while Mundlak and Larson (1992), Michael et al. (1994), Vataja (2000) and Batista and da Silveira (2010) provide evidence that supports the ‘law’. Moreover, this literature mostly focuses on the ‘law’ in terms of exchange rate movements, distance between markets, shipping costs and price discrimination. We argue instead that the relationship between the prices of a commodity in any two markets is also highly dependent on the prevailing international and national policies. This is especially important when long samples of data are examined and policy interventions by governments are extensive and changing as in the case of coffee markets. What is important, therefore, is how we in- corporate the effects of changing coffee market policies into our model of coffee prices. We use the term ‘policies’ to include all measures to implement the inter- national and domestic agreements such as administering the producer price and controlling the production and marketing of coffee as well as local taxes, export levies and subsidies, value added taxes, the setting of coffee grading standards, exchange rate and foreign exchange regulations and the provision of credit. Many of these policies change frequently and are difficult to document. However, they redistribute income either towards or away from producers and 5. See for example Fitter and Kaplinsky (2001), Ponte (2002), Oxfam (2002), Calfat and Flores (2002), Shepherd (2004), Talbot (2004) and Daviron and Ponte (2005). 6. For example see ActionAid (2008) and South Centre (2008). 7. The terminal price of coffee is the spot price of coffee as traded in international markets and the producer price of coffee is the cash price received at the ‘gate’ by producers. Bill Russell, Sushil Mohan and Anindya Banerjee 517 other participants in the coffee supply chain and may introduce a shift, or break, in the mean of the share of the terminal coffee price going to producers.8 The changes in policies at both the domestic and international levels and the eventual liberalisation of coffee markets raise a number of issues when model- ling coffee prices. For example, consider the terminal and producer prices for Brazilian, Guatemalan and Indian Arabica coffee measured in US cents per pound.9 There are three striking features of the data. First, the difference, or gap, between the terminal and producer prices of coffee varies considerably over time in all three countries. This gap is an indirect measure of the costs of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 transferring coffee from the producer to the terminal markets. Between January 1973 and December 1989 the average transfer costs were around 85 (141 per cent of the average producer price), 45 (54 per cent), and 58 (72 per cent) US cents per pound of coffee for Brazil, Guatemala and India respective- ly. Since January 1990, following liberalisation, there has been a marked decline in these costs to around 18 (28 per cent), 32 (46 per cent) and 30 (42 per cent) US cents per pound of coffee respectively. It is unlikely that such large reductions in transfer costs can be explained by changes in freight, hand- ling and related costs alone. It is more likely the reductions are due to changes in the economic rents received by intermediaries and governments in the trans- fer process arising from the greater degree of vertical integration in coffee markets.10 The variation in the gap between the terminal and producer coffee prices for each of these countries appears to provide prima facie evidence that the ‘law of one price’ does not hold. The ‘law’ suggests that the prices of two identical goods in two separate markets will differ in equilibrium by the cost of transfer- ring the goods between the markets. Importantly this implies that the two prices will evolve together and that the gap between the two prices in equilib- rium will be constant. This ‘law’ is predicated on an unchanging economic en- vironment which in the case of coffee markets is difficult to sustain. Changes in coffee market policies at the domestic and international levels will alter the economic environment and may lead to discrete changes in the gap between the two prices. Consequently, the changes that we observe in the gap between the two coffee prices may be due to either a change in the economic environ- ment (i.e. changes in policy) or because the ‘law’ does not hold. Therefore, to examine empirically the dynamics of coffee prices with a model that incorpo- rates the ‘law’ we need to control for the effects of changes in coffee market policies. 8. For example, if the administered producer price of coffee is not changed in line with the terminal price of coffee it will lead to a change in the share of the terminal price received by producers. 9. Prices are in nominal terms unless stated otherwise. See Supplementary Appendix S2 for details of the data and S3 for a graph of the coffee price data. 10. See Mohan (2007). 518 THE WORLD BANK ECONOMIC REVIEW The second feature follows from the first. Large changes in transfer costs associated with changing government policies, regulations and market struc- ture may cause structural breaks in the mean of the producers’ share of the terminal price of coffee. The share is the ratio of the producer to the ter- minal price of coffee and is referred to in the paper as the coffee price ratio. These breaks are at times quite sudden and persistent as demon- strated in Figure 1 of the coffee price ratios for the three countries.11 Third, the nominal price of coffee received by producers at the beginning of the 21st century is much the same as it was in the early 1970s suggest- ing a large fall in real terms. This fall in the real price of coffee is demon- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 strated in Table 1. All of these features lead us to argue that if the terminal and producer prices of coffee are closely related then any modelling of the two prices must take the structural breaks in the coffee price ratio into account. Otherwise estimating the model will result in biased and poor estimates that may lead to incorrect inferences. We therefore model the relationship between the terminal and pro- ducer prices of coffee in Brazil, Guatemala and India allowing for the breaks in the mean of the coffee price ratio. These three countries are chosen due to the variation in their coffee policies and market structures over time. We demon- strate that the modelling approach we adopt is successful in dealing with this variation. We develop a two-step model of coffee prices. In the first step we employ the Bai and Perron (1998) (Bai-Perron) technique to identify multiple breaks in the mean of the natural logarithm of the coffee price ratio. In the second step we estimate a vector autoregressive error correction model (VAR-ECM) of coffee prices conditioned on the identified breaks. A major advantage of this approach is that if the law of one price holds then the estimated error correc- tion term is equivalent to the log of the coffee price ratio and the structural breaks in the coffee price ratio identified in the first step are simultaneously the breaks in the mean of the error correction term. A further advantage of this ap- proach is that we can examine directly the empirical relevance of the ‘law of one price’ after accounting for the influence of changing policies on coffee prices. The next section sets out the standard approach to modelling coffee prices based on the ‘law of one price’ and explains the biases from not allowing for the breaks in the coffee price ratio. Section II reports the results of the esti- mated models allowing for these breaks. We find that once we account for the breaks in the coffee price ratio the ‘law of one price’ is strongly supported by the data. This allows us to show that the producers’ share of the terminal price of coffee in equilibrium has increased in all three countries since liberalisation to around 0.85 in Brazil and India and 0.79 in Guatemala. Assuming that 11. In the empirical analysis that follows including estimating the breaks in the coffee price ratio, unit root tests and the estimation of the models, the variables are in natural logarithms. Bill Russell, Sushil Mohan and Anindya Banerjee 519 F I G U R E 1. THE COFFEE PRICE RATIO Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Note: Thick line is the coffee price ratio (not logged). The horizontal thin lines are the mean of the coffee price ratio for the ‘regimes’ estimated by the Bai-Perron technique as described in Supplementary Appendix S4. liberalised markets led to these higher producer shares we demonstrate in Section III that the loss of revenue to producers from coffee market regulations and interventions over the years is substantial. Our analysis shows that produ- cers have benefited since liberalisation from an increase in real prices, output and higher share in the terminal price of coffee. This indicates that calls for a 520 T A B L E 1 . Nominal and Real Coffee Prices Brazil Nominal Values Real Value Producer Price Terminal Price Transfer Costs Producer Price Terminal Price Transfer Costs 1973 31.02 69.20 38.19 1 1 1 THE WORLD BANK ECONOMIC REVIEW 1990 55.58 82.97 29.39 0.5832 0.4048 0.2599 2007 98.23 111.66 13.44 0.9273 0.4725 0.1030 Percentage Change 1973 – 1990 72.7 (3.3) 19.9 (1.1) 2 23.0 ( 2 1.5) 2 41.7 ( 2 3.1) 2 59.5 ( 2 5.2) 2 74.0 ( 2 7.6) 1990 – 2007 83.3 (3.6) 34.6 (1.8) 2 54.3 ( 2 4.5) 59.0 (2.8) 16.7 (0.9) 2 60.7 ( 2 5.3) 1973 – 2007 216.7 (3.4) 61.4 (1.4) 2 64.8 ( 2 3.0) 2 7.3 ( 2 0.2) 2 52.8 ( 2 2.2) 2 89.7 ( 2 6.5) Guatemala Nominal Values Real Value Producer Price Terminal Price Transfer Costs Producer Price Terminal Price Transfer Costs 1973 46.03 62.30 16.27 1 1 1 1990 54.58 89.46 34.87 0.4004 0.4848 0.7235 2007 98.09 123.55 25.46 0.6240 0.5807 0.4581 Percentage Change 1973 – 1990 18.6 (1.0) 43.6 (2.2) 114.3 (4.6) 2 60.0 ( 2 5.2) 2 51.5 ( 2 4.2) 2 27.6 ( 2 1.9) 1990 – 2007 79.7 (3.5) 38.1 (1.9) 2 27.0 ( 2 1.8) 55.9 (2.6) 19.8 (1.1) 2 36.7 ( 2 2.7) 1973 – 2007 113.1 (2.3) 98.3 (2.0) 56.5 (1.3) 2 37.6 ( 2 1.4) 2 41.9 ( 2 1.6) 2 54.2 ( 2 2.3) Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 India Nominal Values Real Value Producer Price Terminal Price Transfer Costs Producer Price Terminal Price Transfer Costs 1973 45.37 62.30 16.93 1 1 1 1990 66.75 89.46 22.70 0.4968 0.4848 0.4527 2007 108.34 123.55 15.21 0.6992 0.5807 0.2630 Percentage Change 1973 – 1990 47.1 (2.3) 43.6 (2.2) 34.1 (1.7) 2 50.3 ( 2 4.0) 2 51.5 ( 2 4.2) 2 54.7 ( 2 4.6) 1990 – 2007 62.3 (2.9) 38.1 (1.9) 2 33.0 ( 2 2.3) 40.8 (2.0) 19.8 (1.1) 2 41.9 ( 2 3.1) 1973 – 2007 138.8 (2.6) 98.3 (2.0) 2 10.2 ( 2 0.3) 2 30.1 ( 2 1.0) 2 41.9 ( 2 1.6) 2 73.7 ( 2 3.9) Notes: Figures in brackets are the compounded annualised percentage change. Unit Values of Exports US CPI Real values are calculated with reference to the UN index of unit values of 1973 – 1990 196.2 (6.6) 194.1 (6.6) exports. Similar results are obtained if the real price of coffee is measured in 1990 – 2007 15.3 (0.8) 58.7 (2.8) terms of the United States consumer price index (CPI). Annual values are 1973 – 2007 241.5 (3.7) 366.6 (4.6) averages of the monthly values. Nominal values are in US cents per pound. Transfer costs are the difference between the terminal and producer price. The table on the right shows the percentage changes in the UN index of unit values of exports and the United States CPI. Bill Russell, Sushil Mohan and Anindya Banerjee 521 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 522 THE WORLD BANK ECONOMIC REVIEW return to coffee market interventions of the kind that existed prior to liberalisa- tion are not justified. I. MODELLING COFFEE PRICES The standard approach to modelling coffee prices can be motivated with refer- ence to the Enke (1951), Samuelson (1952), Takayama and Judge (1971) spatial models of prices and the ‘law of one price’.12 These models argue that when two markets attain equilibrium the prices differ only by the costs of transferring the goods between the markets. The transfer costs include the ship- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 ping and storage costs associated with moving produce between markets along with the economic rents of intermediate agents in the supply chain. In this model, demand and supply shocks are fully transmitted between the two markets in equilibrium. The standard approach incorporates Samuelson’s (1957) spatial competitive market where prices of the same good in two markets diverge subject to the constraint that: À TC12 P1 À P2 TC21 ð1Þ where P1 and P2 are prices in markets one and two respectively and TC12 are the costs of transferring the goods from market 1 to market 2 and TC21 is the reverse. Samuelson points out that with positive transport costs both equalities cannot hold simultaneously and that if both inequalities hold then there is no trade between the two markets. However, if one inequality and one equality holds then there may be a uni-directional trade from the lower priced market to the higher priced market. In our case we assume the trade flow is from the producer to the terminal market for coffee such that arbitrage in a competitive market leads to: PT ¼ PP þ TCPT ð2Þ where PT . PP and PT and PP are the terminal and producer prices of coffee re- spectively and TCPT are the transfer costs associated with transferring coffee between the producer and terminal markets. The equilibrium in coffee prices between the terminal and producer markets can then be represented in these spatial models as: PP; t ¼ PT ; t à Ue ð3Þ where U e is the constant ratio that coffee prices in the two markets attain in equilibrium and the ‘t’ subscript indicates the time period of the data. Note that even though it is assumed that the physical trade in coffee is uni- 12. The model can be thought of as ‘spatial’ in terms of the geographic locations of producer and terminal coffee markets. Bill Russell, Sushil Mohan and Anindya Banerjee 523 directional the causation between the two prices in equation (3) is conceptually bi-directional. With U e , 1 in equation (3) we can identify transfer costs, TCt, associated with the movement of coffee between the two markets such that TCt ¼ PT ; t À PP; t which are also measured in price per unit of coffee. In the short run the equilibrium relationship (3) need not apply due to the incomplete transfer of information between markets and other rigidities. For example, pro- ducer prices may not immediately respond to changes in the terminal price of coffee due to the fragmented nature of the supply chain and the need for all prices and costs in that chain to adjust. One might also assume that there are Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 small ‘menu costs’ associated with adjusting costs and prices within the supply chain implying the producer price may diverge from its equilibrium level. However, in the long run we expect all prices and costs to respond to competi- tive pressures and the equilibrium relationship of equation (3) will hold if in- formation is shared efficiently between markets and all agents make normal profits. An assumption of this model that is not often highlighted in the literature is that the transfer costs in equilibrium are a fixed ratio of both the terminal and producer prices of coffee. For example, transfer costs as a ratio to, or share of, the terminal price of coffee in equilibrium is: TCt PT ; t À PP; t ¼ ¼ 1 À Ue ð4Þ PT ; t PT ; t Importantly, this implies that if the equilibrium in coffee prices is adequately described by this model, the statistical process of the transfer costs is the same as that of the two coffee prices in equilibrium. If this were not the case, when equilibrium coffee prices are attained in the two markets, transfer costs will not have returned to its equilibrium ratio in terms of the respective coffee prices. This implication is important for our modelling of coffee prices below. The equilibrium relationship in equation (3) above has a straightforward time series interpretation in terms of a vector autoregressive-error correction model of coffee prices. Consider the following error correction representation of a two variable VAR of order k: X k À1 Dxt ¼ d þ ab0 xtÀ1 þ Pi D xtÀi þ 1t ð5Þ i¼1     pP a11 a12 where, xt ¼ , and a ¼ is a matrix of equilibrium speed of pT t a21 a22 adjustment coefficients, b is a matrix containing the equilibrium vectors   1 b1 2 , Pi is a matrix of short-run coefficients, lower case variables are in 1 b2 2 524 THE WORLD BANK ECONOMIC REVIEW natural logarithms and D represents the change in the variable such that Dxt ¼ xt À xtÀ1 . The error correction representation argues that the producer and terminal prices of coffee move together and coffee prices converge on the equilibrium re- lationship. Changes in coffee prices depend on the deviation from the equilib- rium relationship and the speed of adjustment (i.e. the error correction mechanism) along with lagged changes in the prices of coffee. The standard approach to estimating the VAR-ECM of coffee prices pro- ceeds within a cointegration framework.13 If coffee prices, pP and pT are inte- grated processes and cointegrate, the matrices a and b must be of reduced rank Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 and in our case would be equal to one. No cointegration implies a rank of zero. While it is common to model price data as integrated processes, in reality prices cannot be ‘truly’ integrated as they have a lower boundary of zero. It is more likely, therefore, that prices are trend stationary processes with breaks and it is these breaks that make the data appear integrated.14 For coffee prices the trend is very small or non-existent and the natural logarithm of coffee prices and the coffee price ratio appear to be stationary processes over the past thirty four years. This can be verified by the ADF and KPSS univariate unit root tests reported in columns 1 and 2 of Table 2. Note however that even though the trend is difficult to identify in the coffee price data the ADF test indicates that the coffee price ratio is a trend stationary process for all three countries. Consequently, the standard cointegration approach to modelling coffee prices is not appropriate and the modelling framework is a VAR-ECM of stationary price variables with possible breaks in the error correction term. The equilibrium relationship in equation (5) can be written: b0 xt ¼ pP; t þ b pT ; t ¼ zt ð6Þ where zt is a stationary process. Furthermore, if b ¼ 2 1 then the equilibrium vector can be interpreted as the equilibrium relationship of equation (3) where: pP; t À pT ; t ¼ ut ð7Þ and the error correction mechanism is now equivalent to the natural logarithm of the coffee price ratio, ut. Note that when b ¼ 2 1, the standard model in equations (3) and (6) dis- plays long-run homogeneity which is equivalent to accepting the ‘law of one price’. This means that a 1 per cent increase in one coffee price leads to a 1 per cent increase in the other price in equilibrium so that transfer costs as a propor- tion of either the terminal or the producer price are unchanged. This also 13. For example see Rapsomanikas et al. (2004), Fortenbery and Zapta (2004), Krivonos (2004) and Alizadeh and Nomikos (2005). 14. See Perron (1989). Bill Russell, Sushil Mohan and Anindya Banerjee 525 T A B L E 2 . Univariate Unit Root Tests Series Original Data De-meaned Data Brazil ADF KPSS ADF KPSS Ln Producer Price 2 3.48 0.0688*10% 2 4.33 0.043710% [0.0089] [0.0005] Ln Terminal Price 2 2.91 0.1195*5% 2 4.02 0.042710% [0.0454] [0.0014] Ln Ratio of Coffee Prices 2 3.33* # 0.1975*1% 2 7.57# 0.019710% [0.0634] [0.0000] Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Guatemala Ln Producer Price 2 3.01 0.165810% 2 4.33 0.046810% [0.0350] [0.0005] Ln Terminal Price 2 2.91 0.1252*5% 2 4.02 0.043410% [0.0454] [0.0014] Ln Ratio of Coffee Prices 2 6.10* 0.1363*5% 2 11.61 0.023610% [0.0000] [0.0000] India Ln Producer Price 2 3.00 0.108110% 2 3.84# 0.125210% [0.0361] [0.0027] Ln Terminal Price 2 2.91 0.1252*5% 2 4.62 0.123210% [0.0454] [0.0001] Ln Ratio of Coffee Prices 2 5.71* 0.1154*10% 2 10.00 0.020310% [0.0000] [0.0000] Notes: The data are in natural logarithms as indicated by ‘Ln’. ‘Ln Ratio of Coffee Prices’, ut, is measured as pP,t 2 pT,t. ADF is the augmented Dickey-Fuller t-statistics that assumes a null hy- pothesis of a unit root in the data. Associated probability values are shown as [ ]. ADF 5 per cent critical values with a constant is – 2.87 and with a trend and constant -3.43. A lag length of one in the ADF test was chosen on the basis of SIC in all cases except when zero as indicated by #. KPSS is the Kwiatkowski-Phillips-Schmidt-Shin LM test statistic that assumes a null hypothesis that the data are stationary. KPSS 1, 5 and 10 per cent critical values with a constant are 0.7390, 0.4630 and 0.3470 respectively and with a trend and constant 0.2160, 0.1460 and 0.1190. The percentage shown with the KPSS test statistic is the significance level that the null hypothesis is accepted at. De-meaned data adjusts the original data for the shifts in mean as identified by the Bai-Perron technique for each country (see Supplementary Appendix S4). * indicates a significant trend in the unit root test. In all other cases the trend is insignificant and excluded prior to inference. implies that transfer costs have increased by 1 per cent in the long run. With trending price and cost variables we need b ¼ 2 1 so that a persistent change in the level of coffee prices does not lead to a persistent change in the gap between the two prices in equation (5). In other words, b ¼ 2 1 implies that a change in the level of prices leaves the coffee price ratio unchanged in equilib- rium and transfer costs increase in line with prices. Two important assumptions concerning the equilibrium relationship (7) and the VAR-ECM analysis are that the coffee price ratio, ut, converges on a con- stant value, u e, in equilibrium and that ut is a stationary process with a con- stant mean equal to u e. Looking closely at Figure 1 we can however see discrete shifts in the mean of the coffee price ratio. Thus, the finding of 526 THE WORLD BANK ECONOMIC REVIEW stationarity of the coffee price ratio notwithstanding, it is worth investigating whether breaks are present in this series. These visual shifts in the ratio can be tested formally by applying the Bai-Perron technique, which allows for breaks in otherwise stationary processes, for identifying multiple structural breaks in the mean of the coffee price ratio, ut.15 A justification for the likely presence of such breaks is provided in the paragraph below. The Bai-Perron technique iden- tifies 11, 8 and 10 discrete shifts in the mean of the coffee price ratio for Brazil, Guatemala and India and the shifting mean coffee price ratio are shown in Figure 1 by horizontal thin lines.16 Our reasons for proceeding with our analysis with the maintained as- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 sumption that the identified breaks in the coffee price ratio are valid are two-fold. First, incorporating the breaks improves the estimated model’s de- scription of the coffee price data considerably. Second, the converse assump- tion, namely that the coffee price ratio is stationary with a constant mean, is difficult to sustain. For example, if the converse were true then all the changes in coffee market policies over the past thirty five years have had no persistent impact on the share of the terminal price of coffee that goes to producers. More importantly, the vociferous arguments surrounding the merits of either regulating coffee markets or de-regulating coffee markets are largely irrelevant as neither policy would have had any persistent impact on the coffee price ratio. Therefore, both these implications of assuming a sta- tionary process with constant mean are hard to defend empirically and at a policy level. The number of breaks that we find in the mean of the coffee price ratio may seem large relative to that usually reported in the applied structural breaks lit- erature. However, this result is consistent with producer prices being largely administered during the first seventeen years of the data and the coffee market being de-regulated in irregular steps over the next seventeen years. Consequently, over the entire thirty four years examined in the empirical ana- lysis, the coffee price ratio can be characterised as being subject to discrete and irregular shocks. Finding around ten breaks over a thirty-four year period implies that a break occurred only once every thirty six months on average which is broadly consistent with the rate of structural change and interventions in coffee markets over the period studied. We return to the issue of the validity of the identified structural breaks in Section II following the estimation of the models below. It appears, therefore, that the two important assumptions mentioned above, namely that the coffee price ratio is a stationary process with a constant mean and converges on a constant value in equilibrium, are not valid. This suggests the mean of the coffee price ratio may be discretely time-varying and the 15. See Supplementary Appendix S4 for details of the Bai-Perron estimates of the structural breaks. 16. If the breaks in the coffee price ratio are valid this implies Type 1 and Type 2 errors in the ADF and KPSS tests respectively in Table 2. Bill Russell, Sushil Mohan and Anindya Banerjee 527 equilibrium coffee price ratio could be written in a more general form that con- tains a trend, t, and n shift dummies Di: Xn pP; t þ b1 pT ; t þ b2 t þ d D ¼ vt i¼1 i i ð8Þ The inclusion of the trend is intended to capture a systematic divergence between the terminal and producer prices in equilibrium. This would occur if transfer costs evolved differently to coffee prices over the longer term. Transfer costs may be driven by factors such as the wage rate, productivity, shipping rates, insurance, technological advances incorporated in the supply chain, real Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 interest rates, energy costs and inventory control which may not affect to the same extent the prices of homogeneous agricultural products like coffee. The shift dummies represent discrete structural breaks in the coffee price ratio in response to changing domestic and international coffee market policies. If estimation of the model proceeds assuming that the mean ratio of coffee prices is time invariant as implicitly (or explicitly) assumed in the standard lit- erature then the estimates of the equilibrium coefficients, b, and the adjustment coefficients in equation (5) will be poor and biased if the coffee price ratio, ut, is in fact non-stationary (i.e as described by (8) above). The direction of the bias in b depends on whether the coffee price ratio is on average increasing or decreasing over the period. The adjustment coefficients will be biased down- wards if the shifts in the mean coffee price ratio are not accounted for. From the perspective of the estimated model it will appear that the coffee price ratio is taking a long time (i.e. the speed of adjustment is low) to return to the equi- librium coffee price ratio following a change in the mean coffee price ratio in the data. Consequently, how we model the coffee price ratio may affect our estimates in important ways. II. EST IMATIN G A VA R - E C M OF COFFEE PRICES The models are estimated in natural logarithms using the monthly average International Coffee Organization (ICO) Indicator Price for Arabica coffee (‘Brazilian Natural’ Arabica for Brazil and ‘Other Mild’ Arabica for Guatemala and India) as a measure of terminal prices and the monthly average producer price for Arabica coffee in Brazil, Guatemala and India for the period January 1973 to October 2007. All price data are in nominal terms and measure a monthly average price in US cents per pound. Further details concerning the data are provided in Supplementary Appendix S2. Modelling coffee prices without structural breaks Estimates of the standard VAR-ECM model that do not account for the struc- tural breaks in the mean of the coffee price ratio are reported in Table 3. While there is some variation in the estimates there is a strong pattern in the results. A trend is significant for all three countries suggesting that there has been a 528 T A B L E 3 . VAR Error Correction Model of Coffee Prices: Model 1 – Standard Model without Breaks BRAZIL Equilibrium Coefficients Adjustment Coefficients pP pT Trend DpP Dp T Unrestricted 1.0000 2 0.7010 ( 2 4.7) 2 0.1828 ( 2 3.1) 2 0.0630 ( 2 3.0) 2 0.0044 ( 2 0.3) Restricted 1.0000 2 1.0000 2 0.0048 ( 2 4.3) 2 0.0351 ( 2 1.6) 0.0199 (1.2) TLRR ¼ 0.2875, TT ¼ 0.1513, LM1 ¼ 0.2688, LM2 ¼ 0.4058. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0883, pT ¼ 0.0121. Exclusion: pP ¼ 0.0121, pT ¼ 0.0883, Trend ¼ 0.0830. Exogeneity: PP ¼ 0.0494, pT ¼ 0.8594. GUATEMALA Equilibrium Coefficients Adjustment Coefficients pP pT Trend DpP Dp T Unrestricted 1.0000 2 0.9571 ( 2 13.9) 2 0.0495 ( 2 2.2) 2 0.2022 ( 2 6.0) 0.0218 (1.0) Restricted 1.0000 2 1.0000 2 0.0528 ( 2 2.3) 2 0.1935 ( 2 5.9) 0.0268 (1.3) THE WORLD BANK ECONOMIC REVIEW TLRR ¼ 0.6281, TT ¼ 0.0564, LM1 ¼ 0. 1204, LM2 ¼ 0. 0683. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000, pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000, Trend ¼ 0.0381. Exogeneity: pP ¼ 0.0000, pT ¼ 0.3724. INDIA Equilibrium Coefficients Adjustment Coefficients pP pT Trend DpP Dp T Unrestricted 1.0000 2 0.7783 ( 2 12.0) 2 0.0514 ( 2 2.4) 2 0.0913 ( 2 4.0) 0.0864 (3.7) Restricted 1.0000 2 1.0000 2 0.3474 ( 2 2.4) 2 0.0508 ( 2 2.8) 0.0788 (4.3) TLRR ¼ 0.0316, TT ¼ 0.0364, LM1 ¼ 0.3603, LM2 ¼ 0.5283. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000, pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000, Trend ¼ 0.0226. Exogeneity: pP ¼ 0.0003, pT ¼ 0.0009. Notes: Reported as ( ) are t-statistics. The trend is multiplied by 100. The models and statistics are estimated with two lags of the core variables and an effective sample of 416 monthly observations for the period January 1973 to October 2007. The number of lags was chosen by a likelihood ratio test for lag reduction. TLRR and TT are the finite sample Bartlett corrected probability values of the test of the equilibrium restriction that b ¼ 2 1 and the likelihood ratio exclusion test of the estimated trend respectively. LM1 and LM2 are the probability values of the Lagrange Multiplier tests of no serial correlation in the errors of lags 1 and 2 respectively. Normality is the probability value of the Doornik-Hansen test for normal errors. Stationarity, Exclusion and Exogeneity are the probability values of the likelihood ratio tests that pP and pT (and trend if applicable) are stationary, can be excluded from the equilibrium relationship and/or weakly exogenous respectively. Estimated with CATS 2.0. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Bill Russell, Sushil Mohan and Anindya Banerjee 529 trend increase in the producer price of coffee relative to terminal prices over the period examined. The estimated models for Brazil and India appear to be relatively poor representations of the dynamics of the coffee prices with the ad- justment coefficients largely insignificant for Brazil and small for India. The results for Guatemala are the best behaved where we accept the long-run homogeneity restriction that b ¼ 2 1 with the adjustment coefficients moder- ately large and significant. Consequently, with the possible exception of Guatemala, we might conclude that a VAR-ECM is a relatively poor represen- tation of the dynamics of coffee prices. The inconclusive nature of these results is consistent with the view that structural breaks are an important feature of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 the price series being modelled and therefore need to be taken into account in order to develop a more accurate description of the system generating coffee prices. Modelling coffee prices with structural breaks Breaks in the mean of the coffee price ratio are due to breaks in the component coffee price series. However, simultaneous breaks of equal magnitude will not affect the ratio. That is, the coffee price ratio will break if either the breaks in the price series occur at different points of time or when occurring simultan- eously they are of different magnitudes. While there may be numerous events that impact simultaneously on both price series the discussion above focuses on changes to policy and market structure that impact on the coffee price ratio. Therefore, our preferred approach is to identify breaks in the mean of the coffee price ratio directly by applying the Bai and Perron (1998) algorithm. The identified breaks are those shown in Figure 1 and reported in Supplementary Appendix S4. Once we allow for these breaks in the coffee price ratio the univariate unit root tests reported in columns 3 and 4 of Table 2 lead to the unambiguous conclusion that the natural logarithm of the prices of coffee and their ratio are stationary. Estimates of the VAR-ECM incorporating the breaks leads to the results reported in Table 4. The breaks are introduced as level shifts in the equilibrium relationship which account for the level shifts in the mean coffee price ratio evident in Figure 1. The inclusion of the shift dummies now leads to the rejec- tion of the trend in the equilibrium relationship at the 5 per cent level for all three countries. This implies the shift dummies dominate the trend as a proxy for the shifts in the equilibrium coffee price ratio over the period.17 The shift dummies are highly significant which reinforces our conclusion that the Bai-Perron identified breaks are valid. The restriction that b ¼ 2 1 can now be easily accepted for all three countries at the 5 per cent level implying we 17. Over the full sample the mean coffee price ratio increases and this results in a positive estimated trend in the standard model (see Figure 1 and Table 3). However, the structural breaks in Model 2 reported in Table 4 represent the evolution of the mean coffee price ratio better, resulting in the trend becoming insignificant. 530 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . VAR Error Correction Model of Coffee Prices: Model 2–Bai-Perron Break Adjusted BRAZIL Equilibrium Coefficients Adjustment Coefficients pP pT Dp P Dp T Unrestricted 1.0000 2 0.9058 ( 2 17.2) 2 0.2652 ( 2 6.7) 2 0.0338 ( 2 1.1) Restricted 1.0000 2 1.0000 2 0.2420 ( 2 5.9) 2 0.0036 ( 2 0.1) TLRR ¼ 0.2085, LM1 ¼ 0.5148, LM2 ¼ 0.0577. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000, Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity: pP ¼ 0.0000, pT ¼ 0.3981. Dummies: December 1974, 2 0.2449 ( 2 2.8); March 1977, 0.3939 (4.8); August 1979, 0.2806 (3.3); August 1981, 2 0.1225 ( 2 1.5); December 1984, 2 0.5702 ( 2 7.6); April 1987, 0.4842 (5.7); April 1989, 2 0.5049 ( 2 6.1); December 1991, 2 0.2544 ( 2 3.8); August 1996, 2 0.0645 ( 2 1.1); November 2000, 0.1965 (2.6); and December 2002, 2 0.2014 ( 2 2.8). GUATEMALA Equilibrium Coefficients Adjustment Coefficients pP pT Dp P Dp T Unrestricted 1.0000 2 0.9189 ( 2 23.8) 2 0.4547 ( 2 10.2) 2 0.0578 ( 2 1.9) Restricted 1.0000 2 1.0000 2 0.4117 ( 2 9.4) 0.0852 (2.9) TLRR ¼ 0.1309, LM1 ¼ 0.0001, LM2 ¼ 0.6317. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000, pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity: pP ¼ 0.0000, pT ¼ 0.1319. Dummies: June 1975, 0.2924 (5.4); December 1979, 2 0.1562 ( 2 3.3); October 1983, 2 0.3316 ( 2 6.1); May 1986, 0.3101 (4.8); May 1988, 0.0607 (1.1); April 1993, 0.1043 (1.8); April 1995, 2 0.1062 ( 2 1.7); and February 1998, 2 0.1693 ( 2 2.8). INDIA Equilibrium Coefficients Adjustment Coefficients pP pT Dp P Dp T Unrestricted 1.0000 2 1.0391 ( 2 24.1) 2 0.1699 ( 2 5.8) 0.2250 (7.8) Restricted 1.0000 2 1.0000 2 0.1813 ( 2 6.0) 0.2243 (7.5) TLRR ¼ 0.4704, LM1 ¼ 0.0083, LM2 ¼ 0.0777. Normality ¼ 0.0000. Normality ¼ 0.0000. Stationarity: pP ¼ 0.0000, pT ¼ 0.0000. Exclusion: pP ¼ 0.0000, pT ¼ 0.0000. Exogeneity: pP ¼ 0.0000, pT ¼ 0.0000. Dummies: March 1976, 0.5463 (10.5); June 1978, 2 0.2231 ( 2 3.9); June 1980, 2 0.1491 ( 2 2.6); September 1982, 0.2442 (5.0); December 1986, 2 0.2019 ( 2 4.3); June 1989, 2 0.1842 ( 2 3.6); April 1992, 2 0.1102 ( 2 2.0); April 1994, 0.4594 (8.1); September 1996, 2 0.2669 ( 2 6.2); and November 2004, 2 0.2021 ( 2 5.1). Notes: The shift dummies are those estimated in the restricted model. For further details see the notes to Table 3. Bill Russell, Sushil Mohan and Anindya Banerjee 531 simultaneously accept the ‘law of one price’ in relation to terminal and produ- cer coffee prices. The error correction mechanisms are also strongly significant with large adjustment coefficients. Therefore our results for the VAR-ECM with structural breaks are in accord- ance with the theoretical spatial model of prices. For example, the terminal and producer prices of coffee move very closely together as demonstrated by accepting the restriction that b ¼ 2 1 and the existence of a powerful error cor- rection mechanism in all three countries. Consequently, these results strongly favour the ‘law of one price’ once we adjust for the breaks in the mean of the coffee price ratio which we characterise as due to changing coffee market pol- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 icies. The large adjustment coefficients reported in Table 4 imply that devia- tions from the equilibrium coffee price ratio are relatively small and transitory. Following a shock to the coffee price ratio, half of the adjustment to equilib- rium occurs in around 2 1 2 months for Brazil and 1 month for Guatemala and India.18 This can be seen in Figure 1 where the estimated equilibrium coffee price ratios are equivalent to the horizontal thin lines and deviations in, ut, from the equilibrium ratio are mostly small and short lived. Finally, it appears that modelling the discrete shifts in the coffee price ratio as structural breaks plays an important role in explaining the behaviour of coffee prices and the associated changing shares in the terminal prices going to producers. Extending the standard analysis to include the breaks improves our understanding of the behaviour of coffee prices significantly. Are the identified structural breaks valid? There are two ways to approach this question. The first is to identify particular events in the regulation of coffee markets that coincide roughly with the breaks in the coffee price ratio used in the empirical analysis. To this end Supplementary Appendix S1 provides a brief description of the regulations in the coffee markets of each of the three countries and relates the changing regu- lations with the breaks in the coffee price ratio reported in Supplementary Appendix S4. Of the twenty nine breaks identified in the data of the three countries by the Bai-Perron technique, twenty roughly coincide with historical events described in Supplementary Appendix S1. The second approach focuses on the empirical results themselves and asks are the results robust to the number of breaks? In particular, is the large number of identified breaks causing the high speed of adjustment back to the equilibrium coffee price ratio and the acceptance of the law of one price? To examine whether our results are robust to the number of breaks we repeat the analysis by imposing the number of breaks to be half those identified in the models reported in Table 4 and to allow the Bai-Perron technique to choose 18. The adjustment speeds are calculated from simulations based on the estimates in Table 4. 532 THE WORLD BANK ECONOMIC REVIEW the breaks optimally subject to this constraint.19 The estimates imposing half the number of breaks are found to be very similar qualitatively and statistically to those reported in Table 4 that incorporate the full number of breaks.20 For example, for all three countries: (i) the appropriate modelling framework is a VAR-ECM of stationary price variables with possible breaks in the error cor- rection term; (ii) the trend remains insignificant; (iii) the long-run restriction of b ¼ 2 1 can be accepted suggesting the law of one price continues to hold; (iv) the adjustment coefficients are slightly smaller but similar in magnitude and significance to those reported in the ‘full’ model reported in Table 4; (v) the shift dummies are highly significant; and (vi) the residuals are largely free Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 of serial correlation as in the full model. Therefore our estimated model is robust to halving the number of breaks. Consequently, if we reject the number of breaks that we find using statistical tests and instead impose a smaller number of breaks so as to conform with our prior views of the data then our general empirical results along with the asso- ciated policy conclusions and economic implications are largely unaffected. I I I . T H E LO S S I N R E V E N U E TO P RO D U C E R S FROM CO F F E E M A R K E T R E G U L AT I O N The equilibrium coffee price ratio is particularly useful for examining the impact of government policies and changing market structures on the coffee market as it abstracts from short-run variations and shocks to this ratio. Following the liberalisation of coffee markets the equilibrium coffee price ratio in levels has increased to around 0.85 for Brazil and India and 0.79 for Guatemala in the most recent period. Given that both the producer and termin- al prices of coffee are largely market determined in the most recent period fol- lowing liberalisation we might infer the coffee price ratios are their un-regulated market values. Furthermore, the empirical evidence presented in Section II suggests the ‘law of one price’ is strongly supported by the data. Consequently, as the terminal price of coffee has been largely market deter- mined over the entire period we can calculate the counter-factual unregulated producer price of coffee by multiplying the terminal price of coffee in each period by the un-regulated market value for the coffee price ratio in 2007. This allows us to estimate the loss of revenue to producers due to regulation by examining the extent to which the actual producer price of coffee deviates from its counterfactual unregulated value. The thin line in Figure 2 is the nominal loss in revenue  to producersà due to e e the regulation of coffee markets calculated as Qt † PT ; t † U2007 À Ut where e e Qt is the production of coffee and U2007 and Ut are the equilibrium coffee 19. The ‘half-breaks’ models imposes 6, 4 and 5 breaks for Brazil, Guatemala and India respectively. 20. The results are available in Supplementary Appendix S5. Bill Russell, Sushil Mohan and Anindya Banerjee 533 F I G U R E 2. THE NOMINAL AND REAL LOSSES OF COFFEE PRODUCERS Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Note: The thin and thick lines are the nominal and real values (in 2007 prices) of the loss to producers. Real values are in terms of the UN index of unit values of exports. price ratios in 2007 and in each year for each country respectively. This calcu- lation assumes that production and the terminal price of coffee are independent of the market based coffee price ratio. Note that the law of one price means that U e is ‘unit free’ and a real number. This allows the unit of the loss to pro- ducers to be measured in US cents. The thick line in Figure 2 shows the real loss to producers measured in 2007 prices. 534 THE WORLD BANK ECONOMIC REVIEW We observe that producers suffered from a loss of revenue in all three coun- tries during the 1970s and early 1980s. In 2007 prices, the real losses to produ- cers peak at around US$7 billion in the early 1980s for Brazil and around US$0.45 billion for Guatemala and India in the mid-1970s. The losses trail off to negligible levels in the most recent periods following liberalisation in all three countries indicating that market interventions over the years have not been in the overall interest of producers. Note that these measures of loss are in respect to the producers alone and not necessarily to the country as a whole. The loss, or a part of the loss, may simply represent a transfer from the produ- cer to either the government or intermediaries in the transfer process. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 I V. A R E C O F F E E P R O D U C E R S B E T T E R O F F D U E TO LIBERA LISATION? To answer this question we consider the impact that liberalisation has had on the revenue and input costs of producers. The impact on revenues can be thought of in terms of three interrelated issues. First, what has happened to the producer price of coffee? Second, what has happened to the quantity of coffee produced? And third, what has happened to the share of the terminal price of coffee that goes to the producer? The common expectation from coffee market interventions was that they would result in an increase in the terminal and producer prices of coffee. If this were the case then the producer price of coffee should have fared better during the period of regulation than after liberalisation. This is not supported by Table 1. In the seventeen years prior to liberalisation, producer coffee price in- flation is only half that of the ‘world’ inflation rate of 6.6 per cent as measured by the UN index of unit values of exports or the United States CPI. Producer prices increased by 3.3, 1.0 and 2.3 per cent per annum in Brazil, Guatemala and India but in real terms they fell over the period prior to liberalisation by 42, 60 and 50 per cent respectively. In contrast, during the period 1990 to 2007 when coffee markets witnessed the phase of liberalisation, producer prices increased by 3.6, 3.5, and 2.9 per cent per annum in Brazil, Guatemala and India compared with ‘world’ inflation of less than 1 per cent per annum as measured by the UN index of unit values of exports. This means that the producer price of coffee increased by around 60, 55 and 40 per cent in real terms after liberalisation.21 Coffee prices there- fore did not match the growth in ‘world’ prices during the years of coffee market regulation up to 1990 but increased by more than ‘world’ prices after the liberalisation of coffee markets. The data also shows there has been a large increase in coffee production fol- lowing liberalisation in all three countries. Average coffee production in the 21. In the period 1990 to 2007 United States CPI deviates from the UN index of unit values of exports (see Table 1). However, if we use United States CPI as a measure of world inflation we find that producer prices still increase in real terms by around 15, 13 and 2 per cent respectively. Bill Russell, Sushil Mohan and Anindya Banerjee 535 seventeen years after 1990 is 1.3, 1.3 and 1.8 times the coffee production in the seventeen years prior to 1990 in Brazil, Guatemala and India respectively (ICO, 2007). There may be many reasons for this increase in production and we cannot say that the increase is solely due to liberalisation. But to the extent that liberalisation caused an increase in producer prices we can expect its effect on production was beneficial. We can also expect some increase in production because of the removal of restrictions that were imposed on the production of coffee prior to liberalisation. Akiyama (2001) argues that the removal of con- straints of the international quota system meant that exports increased along with production. In addition, the liberalisation meant that higher quality coffee Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 is no longer mixed with lower grades prior to export and the opportunity to export and obtain a premium on higher quality coffee has expanded. There was therefore an added incentive after liberalisation for producers to supply better quality coffee, which in turn has helped the growth in the consumption of coffee worldwide. In any case, as far as producers are concerned they have gained both in terms of prices and production following liberalisation. The empirical analysis above also demonstrates that the equilibrium coffee price ratio has increased sys- tematically since liberalisation of coffee markets in all three countries. This in- crease means that liberalisation has improved the returns to producers by reducing the net transfer costs throughout the coffee supply chain. Therefore, the answer to the question of whether producers are better off in terms of the revenue that they receive since liberalisation appears to be unam- biguously yes. Producers have benefited from a higher real price of coffee, higher coffee production and a higher share of the terminal price of coffee. The impact of liberalisation on the input costs of producers is obscured by the lack of data at the producer level.22 However, there are studies that esti- mate the relative input shares for the production of coffee. For example, ICO (1996/97) reports that the annual cost shares for small holder Arabica produc- tion in India are 54, 26, 7 and 14 per cent for labour, materials (fertiliser, pesticide, bags, seedlings), processing (machinery, energy) and overheads re- spectively. Consequently, if the prices of these inputs are largely determined at the economy wide level and mostly responsive to macroeconomic policies and the exchange rate then the evolution of input costs are independent of liberal- isation. It would then be reasonable to argue that there are no input cost impli- cations to liberalisation and that the benefits to producers in terms of revenue outlined above correspond to the total benefits to producers.23 22. The lack of input cost data may be overcome in the near future. In April 2010 the Executive Director of the International Coffee Council asked all members of the organisation to provide input cost data for crop years 2000/01 to 2009/10. See ICO (2010). 23. We have not come across any studies showing that liberalisation has increased input costs. It is possible that input costs after liberalisation increased faster than revenues because of factors other than liberalisation. However, producers are still better off after liberalisation (due to the higher revenues) because without liberalisation they would have been even worse off financially. 536 THE WORLD BANK ECONOMIC REVIEW V. C O N C L U S I O N AND PO L I C Y I M P L I CAT I O N S This paper argues that to understand the dynamics of coffee prices we need to allow for structural breaks in the equilibrium coffee price ratio due to changes in government policies. To simply undertake the standard analysis ignoring the breaks is inadequate. In contrast, estimating the model allow- ing for the shifts in the coffee price ratio as reported in Table 4 provides estimates that are consistent with our understanding of coffee markets. First, terminal and producer prices move closely together in equilibrium. Second, shocks to the relationship between the two coffee prices are eradi- Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 cated very quickly. Third, liberalisation of the coffee markets has coincided with increases in the coffee price ratio which is equivalent to the producers’ share of the terminal price of coffee. In Brazil, the equilibrium share has risen from 0.6254 in the late 1980s to 0.8461 in the most recent period up to October 2007. Over the same period the equilibrium share to producers has increased from 0.6325 to 0.7896 for Guatemala and from 0.5485 to 0.8494 for India. The systematic increase in the producers’ equilibrium share of terminal coffee prices over the last seventeen years has greatly benefited the producers in these three coffee producing countries. The combination of the increase in the equilibrium producers share and the increase in the real price of coffee after liberalisation beginning in 1990 suggest that the benefit to producers in terms of revenue in 2007 is around 2.15, 0.15 and 0.22 $US billion for Brazil, Guatemala and India.24 This can be compared with the actual payments to coffee producers of 4.63, 0.47 and 0.65 $US billion in 2007. It should be stressed that the analysis does not suggest that liberalisation poses no risk for producers. Instead it gives rise to new problems by expos- ing producers to the vagaries of the market. In particular, concerns have been raised that liberalisation has exposed producers to the full volatility of coffee prices.25 The international community and policy planners could do more for producers by helping to develop missing credit and insurance markets so as to handle this increase in volatility. However, leaving aside the issue of what more could be done for producers, this paper emphasises that responding to calls for returning to the kind of interventions that  24. The total à e benefit h to PT ;2007 producers e i in 2007 17 is calculated as: Q2007 †PT ;2007 † U2007 À Q2007 † ð1 þ dL Þ17 † U1989 where ð 1 þ d L Þ is the compounded annual real increase in terminal prices during the 17 years of liberalisation. The first component of the calculation is what the producer actually received in 2007. The second component is what the producer would receive PT ;2007 if the real terminal price was lower, ð1 þ dL Þ17 ; and they only received the 1989 equilibrium producers share, Ue1989. The difference between the two components is the benefit to producers from higher real coffee prices and a higher equilibrium share of terminal prices. This measure of the benefit assumes that the quantity of coffee produced has been unaffected by the liberalisation of the coffee markets. Alternatively, if coffee output has increased since liberalisation then our estimate understates the ‘true’ benefit to producers. 25. For example see Gemech et al. (2011). Bill Russell, Sushil Mohan and Anindya Banerjee 537 existed prior to the liberalisation of coffee markets is not in the interest of producers. REFERENCES ActionAid, 2008. Commodity Dependence and Development: Suggestions to Tackle the Commodities Problem. Report prepared for the ActionAid and the South Centre, April, Johannesburg and Geneva. Akiyama, T., 2001. Coffee Market Liberalization since 1990. In: T. Akiyama, J. Baffles, D.F. Larson, and P. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Varangis Eds., Commodity Market Reforms: Lessons of Two Decades, World Bank, Washington DC. Akiyama, T., J. Baffes, D. Larson, and P. Varangis, 2001. Market Reforms: Lessons from Country and Commodity Experiences. In: T. Akiyama, J. Baffles, D.F. Larson, and P. Varangis Eds., Commodity Market Reforms: Lessons of Two Decades, World Bank, Washington DC. Alizadeh, A., and N. Nomikos, 2005. Agricultural Reforms and the Use of Market Mechanisms for Risk Management. Report prepared for the Futures and Options Association, London. Anderson, K., M. Kurzweil, W. Martin, D. Sandri, and E. Valenzuela, 2008. Measuring Distortions to Agricultural Incentives Revisited, World Trade Review, vol. 74, pp. 675–704. Ardeni, P.G., 1989. Does the Law of One Price Really Hold for Commodity Prices? American Journal of Agricultural Economics, vol. 71, pp. 303–28. Baffes, J., 1991. Some Further Evidence on the Law of One Price, American Journal of Agricultural Economics, vol. 4, pp. 21– 37. Bai, J., and P. Perron, 1998. Estimating and Testing Linear Models with Multiple Structural Changes. Econometrica, vol. 66, pp. 47– 78. Bates, R.F., 1997. Open Economy Politics: the Political Economy of the World Coffee Trade, Princeton University Press, Princeton. Batista, J.C., and G.B da Silveira, 2010. Trade Costs and Deviations from the Law of One Price, American Journal of Agricultural Economics, vol. 92, pp. 1011–23. Calfat, G., and R. Flores, 2002. Government Actions to Support Coffee Producers — an Investigation of Possible Measures from the European Union Side. Economics working paper, vol. 448. Getulio Vargas Foundations, Brazil. Daviron, B., and S. Ponte, 2005. The Coffee Paradox: Global Markets, Commodity Trade and the Elusive Promise of Development. Zed Books, London. Enke, S., 1951. Equilibrium among Spatially Separated Markets: Solution by Electrical Analogue, Econometrica, vol. 191, pp. 40 –47. Fitter, R., and R. Kaplinsky, 2001. Who Gains from Product Rents as the Coffee Market becomes more Differentiated? A Value Chain Analysis, IDS Bulletin Paper, May 2001, University of Sussex, UK. Fortenbery, R., and H. Zapta, 2004. Developed Speculation and Underdeveloped Markets – the Role of Futures Trading on Export Prices in Less Developed Countries’, European Review of Agricultural Economics, vol. 314, pp. 451– 71. Gemech, F., S. Mohan, A. Reeves, and J. Struthers, 2011. Market-based Price-Risk Management: Welfare Gains for Coffee Producers from Efficient Allocation of Resources. Oxford Development Studies, vol. 39, no. 1, pp. 49– 68. Gilbert, C., 1996. International Commodity Agreements: An Obituary Notice, World Development, vol. 241, pp. 1 – 19. ICO, 1989. Agreement: Renegotiation of the International Coffee Agreement, Outline Joint Approach for Implementing the Universal Quota Concept, ICC 51/5 E 4/6, International Coffee Organization, London. ICO, 1996/1997. Coffee Market Studies: India, International Coffee Organization, London. ———, 2007. Coffee Statistics, International Coffee Organization, London. 538 THE WORLD BANK ECONOMIC REVIEW ———, 2010. Production Costs, ICC 105-6, International Coffee Council, 105th Session 22 – 24 September 2010, International Coffee Organization, London. Jarvis, L.S., 2005, The Rise and Decline of Rent-Seeking Activity in the Brazilian Coffee Sector: Lessons from the Imposition and Removal of Coffee Export Quotas, Working Paper No. 04-004, Department of Agricultural and Resource Economics and the Giannini Foundation, University of California, California, USA. Jerome, A., and O. Ogunkola, 2000. Characteristics and Behavior of African Commodity/Product Markets and Market Institutions and their Consequences for Economic Growth. CID Working Papers 35, Center for International Development, Harvard University, USA. Krivonos, E., 2004. The Impact of Coffee Market Reforms on Producer Prices and Price Transmission, Policy Research Working Paper No. 3358, July, World Bank, Washington DC. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 McIntire, J., and P. Varangis, 1999. Reforming Cote d’Ivoire’s Cocoa Marketing and Pricing System, Policy Research Working Paper No. 2081, March, World Bank, Washington DC. Mehta, A., and J-P. Chavas, 2008. Responding to the Coffee Crisis: What can we Learn from Price Dynamics? Journal of Development Economics, vol. 85, pp. 282–311. Michael, P., A.R. Nobay, and D. Peel, 1994. Purchasing Power Parity Yet Again: Evidence from Spatially Separated Commodity Markets, Journal of International Money and Finance, vol. 136, pp. 637–57. Mohan, S., 2007. Market-based Price-risk Management for Coffee Producers, Development Policy Review, vol. 253, pp. 333– 54. Mundlak, Y., and D.F. Larson, 1992. On the Transmission of World Agricultural Prices, The World Bank Economic Review, vol. 6, pp. 399– 422. Oxfam, 2002 Mugged: Poverty in your Coffee Cup, Oxfam, Oxford. Perron, P., 1989. The Great Crash, the Oil Price Shock, and the Unit Root Hypothesis, Econometrica, vol. 57, no. 6, November, pp. 1361– 401. Ponte, S., 2002. The ‘Latte Revolution’? Regulation, Markets and Consumption in the Global Coffee Chain. World Development, vol. 30, pp. 1099–122. Raffaeli, M., 1995. Rise and Demise of Commodity Agreements: An Investigation into the Breakdown of International Commodity Agreements, Woodhead Publishing, Cambridge, UK. Rapsomanikis, G., D. Hallam, and P. Conforti, 2004. Market Integration and Price Transmission in Selected Food and Cash Crop Markets of Developing Countries: Review and Applications, Commodity Market Review 2003-2004, Food and Agricultural Organisation, Rome. Samuelson, P.A., 1952. Spatial Price Equilibrium and Linear Programming, American Economic Review, vol. 42, pp. 560 –80. ———, 1957. Intertemporal price equilibrium. Welwirtschaftliches Archiv, Band, 79, pp. 181– 221. Shepherd, B., 2004. Market Power in International Commodity Processing Chains: Preliminary Results from the Coffee Market. GEM, Sciences-Po, Paris. South Centre, 2008. From Declaration to Action on Commodities: Making the Turning Point at UNCTAD X11, Policy Brief 14, South Centre, Geneva. Takayama, T., and G.G. Judge, 1971. Spatial and Temporal Price Allocation Models, Amsterdam, North Holland. Talbot, J.M., 2004. Grounds for Agreement: the Political-Economy of the Coffee Commodity Chain. Rowman and Littlefield, Oxford. Varangis, P., D. Larson, and J.R. Anderson, 2002. Agricultural Markets and Risks - Management of the Latter, not the Former, Policy Research Paper No. 2793, World Bank, Washington DC. Vataja, J., 2000. Should the Law of One Price be Pushed Away? Evidence from International Commodity Markets, Open Economies Review, vol. 11, pp. 399– 415. Winter-Nelson, A., and A. Temu, 2002. Institutional Adjustment and Transaction Costs: Product and Input Markets in the Tanzanian Coffee System, World Development, vol. 304, pp. 561– 74. Implications of COMTRADE Compilation Practices for Trade Barrier Analyses and Negotiations Alexander J. Yeats U.N. Commodity Trade (COMTRADE) statistics have major shortcomings for many Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 analyses relating to tariffs and other trade barriers. The use of a cost-insurance-freight valuation base for these data results in an upward (sometimes severe) bias in the implied dutiable value of imports for countries that utilize free-on-board tariffs. This problem can be greatly exacerbated by the “general” trade system procedure used to compile the U.N. statistics, as opposed to the “special” trade practice used for the World Trade Organization Integrated Database. U.S. International Trade Commission statistics show that the combined effects of these biases can reach magnitudes that pre- clude the legitimate use of COMTRADE for many tariff-trade simulations or related trade negotiations. JEL codes: F13, F17 To facilitate meaningful cross-country comparisons, the United Nations Statistical Office (UNSO) justifiably requests that members report import statistics to the Commodity Trade (COMTRADE) Statistics Division on a common cost-insurance-freight (c.i.f.) basis, even in the case of countries like the United States, Canada, Australia, and New Zealand, where free-on-board (f.o.b.) valu- ation tariffs are used. As a result, COMTRADE overstates the actual dutiable value of these countries’ imports. A relevant question, which appears to have been given little consideration, is whether this valuation procedure invalidates the use of COMTRADE for analysis of these countries’ tariffs and similar trade restrictions. A second important point relates to the “reporting system” used for the compilation of COMTRADE statistics. Most countries employ the “general” system, which includes imports for direct consumption as well as imports under customs bond or into officially designated foreign trade zones (FTZs). The latter are not subject to national tariffs unless they eventually clear Alexander Yeats is a former member of the “trade team” in the World Bank’s Development Economics group. His email address is ayeats@msn.com. Major parts of this paper were written while the author was a consultant to the World Bank’s Africa Region Department. The author wishes to thank Francis Ng, members of the USITC “trade data web” support staff, and an anonymous referee for comments and suggestions. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 3, pp. 539 –555 doi:10.1093/wber/lhr053 Advance Access Publication October 28, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 539 540 THE WORLD BANK ECONOMIC REVIEW customs controls. Statistics compiled by the United States International Trade Commission (USITC) can be used to determine if inclusion of FTZ transactions in COMTRADE causes the U.N. data to significantly further misstate the duti- able value of imports. These questions are of importance because recent efforts have attempted to utilize COMTRADE for tariff change simulations and related issues in multilateral trade negotiations. Before proceeding, one important qualification should be noted. In this study, the term “COMTRADE bias” is sometimes used. This is in no way intended as a criticism of the methodology or procedures used in constructing the database which are fully consistent with the appropriate and intended applications of the Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 U.N. statistics. Rather, the term is directed at users who appear to be unaware of important basic characteristics of COMTRADE and have attempted to employ the data in ways that are incorrect and inappropriate. I. CHARACTERISTI CS OF TRADE PROJECTION MODELS During the Uruguay Round, the World Bank and UNCTAD (1987) developed a partial equilibrium projection model (named SMART, for Software for Market Analysis and Restrictions to Trade) to simulate the effects of proposed tariff cuts in the negotiations. The projections employed tariff line level import statistics contained in the WTO Integrated Data Base since it was acknowl- edged that the COMTRADE data then available were too aggregate, were sometimes tabulated in inappropriate values, and were compiled using a meth- odology not appropriate for tariff simulations. The fact that far more detailed six-digit Harmonized System (HS) statistics have now become generally avail- able has generated renewed interest in the possible use of COMTRADE for tariff analyses and simulations (see http://go.worldbank.org/IJIR5D0T80). In partial equilibrium trade models, the projected trade creation for product i (TCi) resulting from a tariff cut is normally derived from TCi ¼ Mi  ed  Dti =ð1 þ ti Þ Â ð1 À ðed =es ÞÞ ð1Þ where Mi is the initial value of imports of the product, ed and es are the import elasticities of demand and supply for the item, and ti is the import tariff (see World Bank and UNCTAD 1987 or UNCTAD 1986 for previous applications). Two important points are evident from this equation. First, if the initial value of imports is overstated by a given amount (say 20 percent), the projected value for trade creation will be upward biased by this same percentage. Second, in cases where the percentage overstatement in the trade base is greater than the applied nominal tariff, the trade creation projection error will exceed the actual value of this parameter. Similar issues arise if tariff changes are ana- lyzed in a computable general equilibrium (CGE) framework. Although different procedures have been used for estimating trade diversion, valuation biases in COMTRADE will generate similar errors for these Yeats 541 projections. For example, Baldwin and Murray (1977) simulate trade diversion using a methodology which incorporates the trade creation projection estimate as an explanatory factor. In this and other empirical approaches (see World Bank and UNCTAD 1987), the higher the COMTRADE valuation bias, the larger will be the resulting error in trade diversion projections. II. THE MAGNITUDE OF C R O S S - P R O D U C T VA L U A T I O N B I A S E S The United States International Trade Commission (USITC) provides online public access to national import statistics down to the level of individual ten-digit Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 HTS products (see http://USITC.gov). Aside from the f.o.b. value of imports for direct consumption, this system (named “trade data web”) also provides infor- mation on international transport and insurance costs incurred in bringing the product to the first U.S. port of entry. As requested by the United Nations Statistical Office, these c.i.f. import values are reported for inclusion in U.N. COMTRADE records. As a result, COMTRADE overstates the dutiable value of imports for countries like Australia, Canada, New Zealand, and the United States that utilize f.o.b. import tariffs. A key question relating to tariff analyses and projections concerns the magnitude of this bias. Since the USITC “trade data web” provides both free-on-board import values and international transport costs for all imported products it can be employed to quantify the tariff valuation bias for countries trading with the United States. Specifically, the nominal bias for imports from a given country can be derived by taking the ratio of all transport and insurance costs for a given product to its landed f.o.b. value. For an empirical assessment of the bias, the following procedure was employed. First, a selection of 45 “test” countries was chosen with an effort made to achieve as much geographic and economic diversity as possible. The selection included countries in Asia, Africa, Europe, and South America, island countries like the Maldives, Fiji, and Sri Lanka, as well as several that were landlocked (Nepal and Paraguay). Next, available statistics were drawn from the USITC trade data web to compute nominal transportation costs for all six- digit HS level products imported from these countries by the United States. These freight factors were then ranked in ascending order, that is, from the lowest nominal transport cost for each product to the highest. The percentile values for each country’s distribution of biases are given in Table 1. As an illustration, the table shows that the United States imported 132 indi- vidual 6-digit HS products from the Cote d’Ivoire with a total 2007 value of $639 million. (All dollar amounts are current U.S. dollars.) The median nominal freight factor (i.e., the valuation bias) was 12.7 percent, while three- tenths of all shipments had a freight factor of 25 percent or more. A similar share of imports from Ghana, Nepal, Togo, and the Maldives incorporated biases exceeding 20 percent, while the biases for one-tenth of all imports from Belize, Senegal, Honduras, Nepal and other countries exceeded 35 percent. Ten 542 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . The Distribution of Valuation Biases for All Six-Digit HS Products Exported to the United States by Forty-Five Partner Countries Percentile Values for the 2007 United States Imports Distribution of Biases (%) Exporter No. of 6-digit HS Products Value ($ million) Median 70th 90th 95th Maldives 18 4 9.8 30.4 95.6 119.7 Tonga 21 8 40.0 51.0 78.9 90.0 Coˆ te d’Ivoire 132 639 12.7 25.1 59.4 79.6 Ghana 264 212 13.6 22.8 49.5 78.0 Senegal 108 21 8.9 15.0 38.2 74.6 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Fiji 164 191 8.0 13.9 34.3 65.5 Guinea 61 139 7.1 24.4 56.0 64.9 Belize 114 113 6.1 11.6 47.4 62.5 Honduras 589 4,101 7.5 14.9 36.8 61.5 Uruguay 474 524 8.3* 13.5 31.0 58.5 Egypt, Arab Rep. 718 2,545 8.1 13.7 31.5 54.7 Togo 49 6 12.9 22.8 36.7 54.0 Nepal 350 97 12.1 20.0 39.1 53.4 Guyana 150 146 8.7 18.1 36.4 52.1 Ecuador 808 6,540 8.8 14.7 34.1 51.2 Peru 1,162 5,489 7.9 11.8 30.1 50.4 Bolivia 362 377 7.7 12.1 28.5 49.4 Bangladesh 435 3,635 7.6* 11.7 26.2 46.9 Argentina 1,556 4,820 7.9 12.6 28.5 44.5 Guatemala 779 3,269 8.1 12.9 29.5 44.2 Philippines 1,550 9,813 7.0 11.6 24.8 44.1 Panama 532 391 6.7 12.3 26.4 43.4 Chile 1,028 9,784 7.6 12.5 28.8 43.1 Greece 879 1,297 6.3 11.0 25.7 41.0 Sri Lanka 604 2,178 6.6* 10.2 22.1 39.4 St. Lucia 74 36 7.5 11.7 26.6 39.0 Paraguay 134 80 9.6 15.4 28.2 38.0 New Zealand 1,480 3,316 5.3 9.2 20.8 37.9 Costa Rica 973 4,209 5.8 10.0 22.8 36.9 Vietnam 1,489 11,425 7.7* 12.6 25.5 36.7 Sierra Leone 174 60 4.2 8.1 21.4 35.9 Australia 2,436 8,971 5.2 9.0 23.2 35.0 Pakistan 1,060 3,831 8.2* 11.5 23.0 34.9 Turkey 1,869 4,897 6.8 10.6 22.8 34.8 Thailand 2,355 23,793 6.1 10.6 21.0 34.2 Indonesia 1,816 15,208 6.7 11.5 22.6 33.7 Colombia 1,486 10,034 6.2 10.0 20.8 29.9 Brazil 2,738 27,193 6.1 9.8 19.2 29.5 Poland 1,780 2,350 5.4 8.5 18.6 28.9 Tunisia 453 478 4.3 6.9 18.9 28.2 Morocco 601 664 8.1 9.6 19.9 27.5 Austria 2,249 10,893 4.4 7.3 17.5 27.4 Cyprus 155 18 4.5 7.3 18.7 26.8 India 3,342 25,113 7.0 9.9 17.7 24.6 Taiwan 3,198 39,853 5.6 8.1 15.3 23.5 MEDIAN VALUE 602 2,264 7.6 11.7 26.3 42.1 * The median U.S. tariff exceeds the median nominal transport cost ratio for this country. Note: As an illustration, the table shows 30 percent of the valuation biases for the Cote d’Ivoire exceed 25.1 percent. Thirty percent of all biases for Guyana exceed 18.1 percent, while 10 percent of Senegal’s biases exceeded 38.2 percent. Yeats 543 percent of the combined U.S. imports from all countries listed in the table incurred nominal freight costs exceeding 26 percent. Overall, for the 45 countries, only five (Uruguay, Bangladesh, Sri Lanka, Vietnam, and Pakistan) faced median Unites States tariffs that exceeded their median ad valorem transport costs. As a result, the percentage error in trade creation estimates for most countries will exceed the actual percentage change in this parameter, often by very high margins. The implications are that COMTRADE statistics can generate highly inaccurate estimates of the level and composition of the trade response to tariff changes. While Table 1 examined the distribution of individual countries’ valuation Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 biases across all exports, an important related question is whether significant differences exist in the biases between product groups. If so, one would have an interest in identifying those sectors where COMTRADE most seriously overstates the dutiable value of imports. Second, if the valuation biases within specific product groups fall in a relatively narrow range this might suggest the possibility of employing a standard f.o.b.–c.i.f. correction factor to offset, or reduce, the bias. This approach would be similar to the ten percent factor employed by the IMF to account for differences in export and import statistics. For relevant information, U.S. trade statistics for the countries in Table 1 were combined into six different regional groups (i.e., Southern Cone South America, Other South America and the Caribbean, Oceania, sub-Saharan Africa, Southeast Asia, and South Asia. Next, the COMTRADE bias was com- puted for U.S. imports of several broad categories of goods from these country groups. For four of the six country groups the valuation bias for textiles and footwear products exceeds the median bias for all goods, while meat, fish and vegetable imports frequently incur higher than average nominal transportation costs. Almost one-third of all meat exports from Oceania had nominal transport costs exceeding 21 percent, while the corresponding freight rate for sub-Saharan African vegetable products was about 31 percent. One possible explanation for these results is that due to their perishable nature, food products rely more heavily on relatively expensive air transport to access U.S. markets. Other products sectors with relatively high valuation biases include hides and leather goods, as well as articles of plastic and glassware products. Conversely, nominal freight costs for vehicles and machinery, optical goods, and scientific instruments are generally among the lowest in the table. The overall diversity of the biases across groups provided little evidence that the COMTRADE valuation problem could be corrected by a standard f.o.b.–c.i.f. adjustment factor.1 1. While this investigation focused on COMTRADE valuation biases for the United States, several published transport cost studies indicate the conclusions concerning similar unacceptably high biases can be generalized to countries like Australia, Canada, and New Zealand that also employ free-on-board import tariffs. See Curtis and Chen (2003), Conlon (1982), Pomfret and Sourdin (2010), Hummels (2007), or Lloyd (1976) among others. 544 THE WORLD BANK ECONOMIC REVIEW III. EFFECTS OF G E N E R A L T R A D E R E PO R T I N G P R AC T I C E S While utilization of an inappropriate cost-insurance-freight valuation base can significantly overstate the dutiable value of U.S. imports, there is an additional problem that may produce even greater biases. Specifically, two methodologies are used for compiling import statistics; namely, the general and special record- ing systems. Special trade statistics tabulate the value of goods imported direct- ly for final consumption. This exchange encounters any existing tariffs and related trade control measures so special trade data are submitted by the U.S. to the WTO Integrated Data Base. In contrast, general trade statistics record Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 the value of merchandise imports, either for direct immediate consumption or into bonded warehouses and foreign trade zones (FTZs) under customs custody. Imports under the general trade regime destined for FTZs are exempt from tariffs unless they are redirected toward domestic markets for consump- tion. Due to these special import provisions, general trade statistics, which are employed in the COMTRADE database, have major shortcomings for analyses of trade restrictions.2 As a result of their compilation procedures, general statistics may seriously mis- state dutiable import values, and may also fail to correctly identify the goods facing trade restrictions. A hypothetical example can illustrate this point. Assume the U.S. imports $10 billion of crude petroleum (HTS 270900) into a foreign trade zone. In the zone the shipment is further processed (refined) into distillate fuel oils (HTS 271019) and this product is then transferred to the domestic market for consumption—at which point applicable U.S. tariffs are assessed. Under normal zone procedures, importers generally have the option of paying duties on the original materials imported into the zone or on the finished fabri- cated product. Since the nominal equivalent of specific tariffs on distillate fuels is relatively lower than those on crude oil, the former would be reported on customs forms. The statistical records of these transactions would be as follows: (i) COMTRADE would record statistics on the actual value ($10 billion) of crude oil imports. In contrast, the WTO Integrated Database would not report any imports of crude oil, because this specific product did not cross the customs frontier for domestic consumption. (ii) The IDB would report statistics on the distillate fuel oil imports, since these are the shipments upon which relevant United States import duties are assessed. In contrast, U.N. COMTRADE would not report import 2. Recent surveys indicate that more than 250 general purpose United States FTZs have been established. These zones are considered to be outside of U.S. Customs territory for the purpose of import duty liability. Therefore, goods destined for FTZs are not subject to customs tariffs unless they formally enter into U.S. Customs territory—at which point they will be reported to the IDB. Merchandise shipped to foreign countries from FTZs is exempt from duty payments. This provision is especially important for firms that import components to manufacture finished products for export. Various activities can be conducted in a zone, including assembly, packaging, storing, cleaning, repacking, sorting, grading, testing, labeling, repairing, combining foreign or domestic components, or further processing. See MacLeod (2000) for a useful discussion of activities in U.S. foreign trade zones. Yeats 545 information for these goods due to their significant physical transform- ation within U.S. geographic territory, that is, the foreign trade zone. Table 2 provides examples of the magnitude of the general system bias using statistics on selected six-digit HS U.S. imports from Korea. The table shows the dutiable f.o.b. value for each item (column 3), which is reported to the WTO Integrated Data Base. These numbers represent the value of goods imported for direct consumption and, as such, are subject to existing tariffs. In addition, column 5 shows the corresponding c.i.f. general import value reported to COMTRADE. Differences between these values indicate the magnitude of bias Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 associated with the use of the U.N. data for tariff analyses and/or projections. Column (4) has been added to help assess the relative size of the f.o.b.–c.i.f. valuation bias in COMTRADE data as opposed to biases originating from tabulating data using the general trade system. As an example, the c.i.f. value of dutiable imports of gear boxes (HTS 870840) is $1.5 million higher than the dutiable customs value, while the difference between the latter and the c.i.f. general import value is $319 million. These comparisons indicate that the general trade compilation practices account for almost all of the differences in import values reported in the IDB and U.N. COMTRADE. The clear impression from Table 2 is that COMTRADE biases are of a mag- nitude that invalidates use of these statistics for tariff simulations or negotia- tions. Biases incorporated in the U.N. data may seriously misdirect national priorities for a liberalization across products, and may also significantly over- state overall potential trade gains. Specifically, – COMTRADE overstates the dutiable value of the first item (HTS 252329 — Portland Cement) by about $57 million, or 56 percent. Since the customs c.i.f. and general import values are equal ($158.9 million), the bias is entirely attributable to the inappropriate (for tariff analysis) valuation base employed for the U.N. statistics. – Differences of just under 1 billion dollars (approximately 67 percent) occur in the import value for petroleum oils (HTS 271019) reported to the IDB for domestic consumption and the general import total ($2.4 billion) recorded in COMTRADE. Only about $63 million, or 7 percent of the difference, is attributable to the alternative valuation bases (see columns 3 and 4). – COMTRADE-IDB differences exceeding several thousand percent occur for rare earth imports (HTS 284610). The table indicates that about 95 percent of all imports of this product were destined for the processing zones. As such, the IDB correctly does not record most shipments of this item. Similarly, the import value reported in COMTRADE for organic solvents (HTS 381400) is more than ten times greater than the dutiable value of imports for domestic consumption. – COMTRADE overstates the dutiable value of electrical or nonelectrical tapes (HTS 391990) by about $44 million, or approximately 300 546 T A B L E 2 . Examples of the Magnitude of Trade System and Valuation Biases in COMTRADE Data on U.S. Imports from Korea 2006 Import Value ($000) COMTRADE Bias HTS Dutiable Customs Recorded Customs General Imports Percent Value No. Description (f.o.b.)* (c.i.f.) (c.i.f.)** (%) ($000) 252329 Portland cement 101,985 158,860 158,860 55.8 56,875 271019 Other petroleum oils 1,437,721 1,500,482 2,405,261 67.3 967,539 284610 Rare earth compounds 27 42 713 2,567.4 686 381400 Organic solvents 130 140 1,594 1,128.8 1,465 382490 Other cultured crystals 11,276 11,783 15,135 34.2 3,859 390319 Other polymers 13,841 14,875 20,720 49.7 6,880 391910 Self adhesive plates 4,137 4,432 6,086 47.1 1,949 391990 Electrical or non-electrical tapes 14,477 15,222 58,048 301.0 43,571 THE WORLD BANK ECONOMIC REVIEW 392630 Fittings for furniture 1,590 1,704 2,676 68.3 1,086 400912 Pipes with fittings 284 302 636 124.2 352 401039 Conveyer belts 2,528 2,652 3,238 28.1 710 540753 Woven synthetic fabric 472 511 1,632 245.8 1,160 551612 Dyed woven fabrics 105 112 305 191.0 200 620690 Women’s blouses 185 194 467 152.1 282 721633 Iron shapes 15,688 17,193 20,353 29.7 4,665 730820 Iron structures 17,963 26,032 26,032 44.9 8,069 731814 Self tapping screws 2,248 2,539 2,914 29.7 666 731824 Cotter pins 295 311 639 116.9 345 830170 Lock keys 514 528 3,679 616.1 3,165 840734 Internal combustion engines 19,625 20,173 46,173 135.3 26,548 840991 Parts for aircraft engines 83,158 86,194 159,711 92.1 76,553 842131 Oil or fuel filters 15,866 17,534 20,353 28.3 4,486 848130 Copper valves 7,374 7,604 10,855 47.2 3,481 848330 Transmission bearings 4,337 4,467 15,950 267.8 11,613 848410 Metal gaskets 1,375 1,439 2,666 93.9 1,291 850110 Electrical motors 29,940 31,274 37,521 25.3 7,581 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 850140 Other alternating current 18,719 19,570 24,446 30.6 5,727 motors 850511 Electromagnets of metal 2,102 2,145 5,334 153.8 3,232 851110 Spark plugs 2,373 2,512 4,019 69.4 1,646 851120 Ignition magnets 1,181 1,184 2,050 73.6 869 851130 Ignition coils 3,562 3,843 8,756 145.8 5,194 851140 Starter motors 25,403 26,006 33,358 31.3 7,955 851150 Other generators 14,840 15,155 28,779 93.9 13,939 851890 Parts of telephone head sets 10,406 10,636 13,138 26.3 2,732 852692 Radio remote control apparatus 3,388 3,612 7,260 114.2 3,871 852812 Monitors and projectors 266,492 272,165 640,937 140.5 374,445 853650 Electronic AC switches 40,588 42,191 53,288 31.3 12,701 853690 Other electronic switches 12,098 12,549 18,574 53.5 6,476 854430 Ignition wiring sets 6,705 7,092 12,287 83.3 5,582 870829 Other parts of automobiles 212,032 232,815 258,040 21.7 46,007 870839 Parts of brakes 175,642 184,337 215,421 22.6 39,779 870840 Gear boxes and parts 63,023 64,523 382,867 507.5 319,844 870880 Suspension systems and parts 13,526 14,258 17,705 30.9 4,179 900220 Lenses and filters 266 271 12,921 4,754.6 12,655 950639 Other gymnastic equipment 6,136 6,506 15,024 144.8 8,888 TOTAL OF ABOVE 2,665,621 2,847,972 4,776,421 79.2 2,110,800 * The United States reports the statistics in this column to the WTO IDB. ** The United States reports the statistics in this column to U.N. COMTRADE. The UNSO may reclassify some USITC data to conform with the general treatment of reimported products. Yeats 547 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 548 THE WORLD BANK ECONOMIC REVIEW percent. Compilation practices for general imports were a major cause of the overall difference between these statistics and the data reported to the IDB. – If COMTRADE were used as a trade base for simulations of the effects of import tariff changes for woven synthetic fabrics (HTS 540753) the upward bias in the projection error would be 246 percent. Similar pro- blems occur for dyed woven fabrics (HTS 551612), where COMTRADE overstates the dutiable value of imports by approximately 190 percent. General imports account for most of the difference between COMTRADE data and the relevant numbers in the IDB. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 – Similarly, COMTRADE would generate an upward bias in projections of the effects of tariff changes for lock keys (HTS 830170) exceeding 600 percent. The corresponding bias for transmission bearings (HTS 848330) would be over $11.6 million, or almost 270 percent. The general reporting system accounts for almost all of the differences between the WTO and U.N. statistics. – Differences of approximately $700 million occur between the true duti- able values for the combined imports of radio remote control apparatus (HTS 852692) plus gear boxes and parts (HTS 870840) and the general import totals ($1.0 billion) recorded for these items in COMTRADE. The U.N. statistics overstate the dutiable value of gear boxes by over 500 percent. – The value reported in COMTRADE for imports of photographic lenses and filters (HTS 900220) is almost 50 times greater than the dutiable value of imports for consumption reported in the Integrated Data Base. Other examples in the table also reflect similar, very high, COMTRADE biases. Overall, for the products listed in Table 2 the combined import values reported to COMTRADE are approximately $2.1 billion higher than the true dutiable value of these goods. These comparisons indicate the U.N. statistics overstate the dutiable value of imports (and potential trade creation gains from a tariff liberalization) by about 80 percent—see the column totals. In general, major statistical discrepancies occur between COMTRADE and IDB data for raw materials, semi-finished goods, and components imported into FTZs for further processing.3 3. There are several reasons to believe that the biases associated with general trade statistics may be even greater for some other countries. Assemble operations for parts and components are often a major activity in foreign trade zones and high wage countries like the United States typically do not have an extensive competitive advantage in these operations. Ng and Yeats (1999) construct multicountry “revealed” comparative advantage indices for the assembly of parts and components. Their results suggest that these types of operations, and potential biases from general statistics, may be much greater in low wage, relatively high skill countries like Thailand, Malaysia, Indonesia, and the Philippines. Yeats 549 Even stronger negative conclusions concerning the magnitude of COMTRADE biases emerge from statistics on U.S. imports from Austria. For example, Table 3 shows general imports of cars with cylinders under 3,000cc (HTS 870323) from Austria are reported as $3.0 billion in COMTRADE, while the dutiable value of these items imports reported to the IDB is roughly $2.1 billion (or 218 percent) lower. General imports of $333 million are reported for spark ignition engines (HTS 840732) which is almost fifty times the dutiable value of these shipments. Overall, the reported general imports of the 18 items listed in Table 4 are approximately $3 billion higher than the duti- able customs value of imports. Differences for several products, like articles of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 magnesium, actually exceed one thousand percent.4 Two key points emerge from these statistics. – First, COMTRADE data have the capacity to significantly misdirect na- tional priorities for a tariff liberalization across products. This point is reflected in the fact that rankings of general import product values may differ substantially from those based on actual dutiable values of imports. As an example, the value of U.S. general imports of spark ignition engines ($333 million) is the third highest in the table even though the actual duti- able import value ($7 million) was exceeded by 8 other items—often by very high margins. – Second, the COMTRADE general statistics may also provide very inaccur- ate and unreliable information concerning the magnitude of potential overall gains resulting from a trade liberalization. This point is reflected in the fact that the actual dutiable value of U.S. imports from Austria is approximately 40 percent lower than totals reflected in the general trade statistics (see the memo item in the table). This figure represents the potential overall error in trade creation projections that utilize COMTRADE statistics. A further major defect of COMTRADE is that it may indicate no, or relatively limited, domestic consumption of a good occurred when there was in fact sig- nificant dutiable trade. This situation is the converse of that reflected in Tables 2 and 3 where COMTRADE overstated the customs value of imports. The underreporting problem occurs when imported raw materials, semi- finished goods, or components experience significant transformation in a foreign trade zone before shipment to domestic markets. As a result of the pro- cessing activity the originally imported good may be classified under a different 4. As noted, one cause of the COMTRADE-IDB statistical discrepancies is that a product experienced significant transformation in an EPZ and cleared customs under a different HS classification than that recorded in COMTRADE. In addition, the finished product may never have entered the domestic market. For example, the automotive products listed in 3 may only have had modifications to comply with domestic standards required in their final destinations in (say) Central or South America. These products would be reported in COMTRADE because they entered United States geographic territory, but would not be recorded in the IDB because they never cleared the U.S. customs frontier in any form. 550 T A B L E 3 . Examples of the Magnitude of Trade System and Valuation Biases in U.N. COMTRADE Data on United States Imports from Austria 2007 Import Value ($000) COMTRADE Bias HTS No. Description Dutiable Customs (f.o.b.)* Recorded Customs (c.i.f.) General Imports (c.i.f.)** Percent (%) Value ($000) 870323 Car cylinders under 3000cc 955,878 963,889 3,038,858 217.9 2,082,980 870324 Car cylinders over 3000cc 93,732 94,316 645,034 588.2 551,302 840734 Spark ignition engines 7,048 7,163 332,870 4,622.9 325,823 870840 Gear boxes and parts 7,209 7,397 12,110 68.0 4,901 840991 Parts for aircraft engines 14,810 15,277 18,597 25.6 3,787 870829 Other parts of automobiles 36,757 38,441 40,259 9.5 3,502 392330 Bottles and flasks goods 22,030 23,514 24,798 12.6 2,768 848180 Copper taps, cocks and valves 18,929 19,339 20,771 9.7 1,842 392390 Other containers 13,077 14,137 14,614 11.8 1,537 THE WORLD BANK ECONOMIC REVIEW 848190 Hand operated appliance parts 13,851 14,682 15,261 10.2 1,410 220860 Vodka in containers 1,131 1,306 2,312 104.4 1,181 870899 Parts of tractors 23,630 24,152 24,387 3.2 757 810411 Articles of magnesium 51 52 738 1,358.7 687 854430 Ignition wiring sets 4,474 4,527 5,075 13.4 601 848220 Tapered roller bearings 4,676 4,906 5,200 11.2 524 40690 Other cheeses 4,968 5,171 5,449 9.7 481 711719 Other ropes, cables and chains 4,337 4,381 4,667 7.6 331 848250 Cylindrical roller bearings 3,266 3,284 3,507 7.4 241 TOTAL OF ABOVE 1,229,853 1,245,933 4,214,506 242.7 2,984,653 Memo Item Imports of All Goods 7,736,054 7,941,642 10,893,325 40.8 3,157,271 * The United States reports the statistics in this column to the WTO IDB. ** The United States reports the statistics in this column to U.N. COMTRADE. The UNSO may reclassify some USITC data to conform with the general treatment of reimported products. Source: All statistics computed from USITC Trade Data Web. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 T A B L E 4 . Examples of COMTRADE Over- and Under-reporting Dutiable United States Energy Imports 2008 Import Values ($million) COMTRADE Bias Dutiable Customs Recorded Customs General Imports Value HTS No. Description (f.o.b.)* (c.i.f.) (c.i.f.)** Percent (%) ($million) 270900 Crude petroleum oils 274,950.2 281,825.3 363,391.1 32.2 88,441.0 271112 Propane 6,609.1 6,829.0 3,574.7 2 45.9 2 3,034.4 271113 Butane 2,665.8 2,738.4 1,316.6 2 50.6 2 1,349.1 271114 Ethylene 5,385.9 5,573.8 1,064.8 2 80.2 2 4,321.1 271119 Other petroleum gas 598.1 620.3 173.0 2 71.1 2 425.1 271129 Other propane 10,189.6 10,538.7 84.1 2 99.2 2 10,105.5 271311 Petroleum coke 14,906.6 15,451.9 174.8 2 98.8 2 14,731.8 271320 Petroleum coke bitumen 4,507.2 4,648.3 752.0 2 83.3 2 3,755.2 271390 Other Petroleum residue 670.5 699.1 68.3 2 89.8 2 602.2 ABOVE REFINED PRODUCTS 45,532.7 47,099.6 7,208.2 2 84.2 2 38,324.5 * The United States reports the statistics in this column to the WTO IDB. ** The United States reports the statistics in this column to UN COMTRADE. The UNSO may reclassify some USITC data to conform with the general treatment of reimported products. Source: All statistics computed from USITC Trade Data Web. Yeats 551 Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 552 THE WORLD BANK ECONOMIC REVIEW COMTRADE HTS heading than the final product. In these cases, statistics on the end products for domestic consumption are reported to the IDB, but not the components that originally entered the FTZ. COMTRADE would not record statistics on the final product actually consumed, but would report imports of the raw materials. While it can be difficult to precisely identify the goods fabricated in FTZs, the situation is less problematic for energy imports since refinery operations generally create a range of products that are clearly petroleum or natural gas derivatives. For example, Table 4 shows the IDB is reporting total U.S. crude petroleum (HTS 270900) imports whose value is about $88 billion less than Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 COMTRADE. The overstated COMTRADE statistics could cause exporters to place a higher than warranted priority on liberalization of existing U.S. specific tariffs on crude oil imports. However, this strategy would be misdirected since much of the discrepancy can be accounted for by differences in reported IDB-COMTRADE imports of refined energy products. This transformation is the reason why reported IDB imports of petroleum coke and bitumen are ap- proximately $18 billion higher than the corresponding figures in U.N. COMTRADE.5 Aside from energy products, COMTRADE may significantly underreport dutiable imports of a broad range of products that experience transformation in FTZs. Statistics on foodstuffs, beverages and tobacco, machinery, electron- ics, and transport equipment are among the sectors often affected by this problem. In extreme cases, COMTRADE fails to report any imports of items for which the IDB shows dutiable trade exceeding millions of dollars occurred. Examples include U.S. imports of unwrought zinc (HTS 790111) and marine propulsion engines (HTS 840721) from Korea. In other cases, COMTRADE significantly under-reported dutiable imports by over $400 million in the case of motorized transport equipment (HTS 870323) from Korea, and by over $2 billion for cellular telephones (HTS 851712) from China. I V. C O N C L U S I O N S This study examined characteristics of COMTRADE statistics to assess their utility for tariff analysis and related applications in multilateral trade negotia- tions. This issue is of major importance since recent attempts have been made to use COMTRADE for tabulating the value of imports subject to tariff and nontariff restrictions and simulating the trade response to negotiated tariff changes. Accurate and reliable information on these points are key 5. An important point is that discrepancies between dutiable customs and general statistics often become sharply smaller at higher levels of aggregation. This occurs when individual six-digit HS products differentiate between unassembled components and the assembled form of a good while these items are combined in a single category at (say) the four-digit level. Statistics compiled at very high levels of aggregation (like two-digit data) may completely conceal the magnitude of the differences occurring in the underlying, more detailed statistics. Yeats 553 requirements for the formulation of national trade strategies, or to support multilateral trade negotiations. For several reasons, negative conclusions were reached regarding the utility of unadjusted U.N. statistics for such efforts. First, a serious problem exists concerning the valuation base employed for COMTRADE. These statistics overstate the dutiable value of all United States imports, often very significantly, since they are expressed in cost-insurance-freight values although the U.S. employs free-on-board import tariffs.6 As a result, the error in COMTRADE-based trade creation projections could seriously misdirect national priorities in multilateral negotiations. This projection error extends across all regional groups of countries, as well as Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 major product categories. Evidence was cited that indicates these problems also occur for other countries like Australia, Canada, and New Zealand that employ free-on-board import tariffs. Second, the general trade compilation procedure used for COMTRADE may greatly amplify the detrimental effect of the valuation bias. In some cases, general import statistics overstated the dutiable value of individual six-digit HS products by several hundred percent, or by billions of dollars. Automotive equipment, machinery, electronics and energy products were often prone to this statistical bias. Third, COMTRADE may incorrectly identify specific items which are, or are not, subject to tariffs and other trade restrictions. This is due to the fact that the U.N. records tabulate information on products entering a country’s geographic territory, but may fail to record relevant information on the nature and value of the goods actually clearing customs. This problem occurs when imports experience significant transformation in foreign trade zones and then clear customs under a different HTS code than that recorded in COMTRADE.7 Another possible cause is that the processed products were for- warded to final destinations in third countries and did not clear U.S. customs in any form. As a result, COMTRADE may both seriously overstate and under- state dutiable import values. 6. A further problem is that COMTRADE is often compiled at too high a level of aggregation to be accurately used for tariff analyses and/or projections. Some six digit HS products (the lowest level of detail available in COMTRADE) may contain multiple tariff lines having widely divergent import duties. As an example, the six-digit HS product 610439 (women’s suits) has two tariff lines with duties of 0.0 and 24.0 percent. The average of these duties (12.0 percent) would not accurately reflect the level of protection afforded either product. Similarly, HS product 640199 (waterproof footwear) contains four line items with duties ranging from 0 to 39.5 percent. These are not extreme outliers as HS code 210690 (other edible food preparations) incorporates 42 individual tariff line products. 7. The relative magnitude of tariffs on production inputs and the processed product provides a useful indicator of where the largest COMTRADE-IDB data discrepancies may occur. As noted, firms operating in FTZs have the option of declaring imports of either the production inputs or the final good on customs vouchers when the item is transferred to the domestic market for consumption. In situations where tariffs are relatively high on the inputs, an incentive would exist to declare imports of the fabricated product to customs (which would be reported to the IDB), while COMTRADE would record statistics on the unprocessed components initially imported into the zone. 554 THE WORLD BANK ECONOMIC REVIEW Each of these factors is of major importance by itself. However, taken to- gether, the combined biases can reach magnitudes that clearly preclude the le- gitimate use of unadjusted COMTRADE data for trade projections and negotiations. For example, recent USITC statistics report FTZ imports of $285 billion that were exempt from tariffs. To put this value in perspective, these imports were only slightly lower than the total combined customs value of U.S. imports from Japan, Germany, and the United Kingdom. In addition, transport and insurance charges on all U.S. imports were $65.8 billion—a figure that reflects the further overall COMTRADE bias associated with tabulation of cost-insurance-freight trade values. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Third, statutory regulations exempt certain imports from tariffs. These include all U.S. government imports, imports for the treatment of specific medical pro- blems, and imports by overseas territories—all of which were about $24 billion. Altogether, these combined biases totaled $366 billion, or almost 20 percent of the customs value of all United States imports. However, as previously noted, the relative importance of the general trade bias may be larger in other countries (particularly those in East Asia) where international production sharing is prac- ticed more extensively than in the United States. These biases would normally be incorporated in COMTRADE data given the U.N. recommendation that trade statistics be tabulated using general reporting practices: “The general trade system provides a more comprehensive recording of the external trade flows than does the special system. It is recommended, therefore, that countries use the general system for compilation of their international mer- chandise trade statistics” (U.N., Department of Economic and Social Affairs (ESA/STAT/AC.137.5, p. 30). The key point that follows is that analyses of tariffs and other trade barriers should ideally utilize tariff line level import statistics compiled on the same valu- ation base employed for assessing these duties. Furthermore, the data must accur- ately account for specific exemptions like those normally afforded foreign trade zones or government entities, as well as for country specific exemptions asso- ciated with preferences or the withholding of “most-favored-nation” trade status. Brenton and others (2009) presents a useful illustration of the nature of these required adjustments within the context of formulating national structural adjust- ment policies. The United Nations (2007) provides comprehensive information on the trade compilation practices of about 40 countries, which should also be useful for identifying required adjustments. As this study shows, a failure to prop- erly account for these factors may adversely influence a country’s strategies in multilateral negotiations or the formulation of national trade policies. REFERENCES Baldwin, Robert, and Tracy Murray. 1977. “MFN Tariff Reductions and Developing Country Benefits Under the GSP.” The Economic Journal 87(March): 30– 46. Yeats 555 Brenton, Paul, et al. 2009. “Assessing the Adjustment Implications of Trade Policy Changes Using TRIST.” World Bank. http://siteresources.worldbank.org. Curtis, John, and Shenjie Chen. 2003. “Transport Costs and Changes in Canada’s Trade Pattern.” World Economy 26(July): 975 –91. Conlon, Richard. 1982. “Transport Costs and Tariff Protection of Australian Manufacturing.” Economic Record, 58(March): 73 –81. Laird, Sam, and Alexander Yeats. 1986. “The UNCTAD Trade Policy Simulation Model.” United Nations Conference on Trade and Development. Geneva. Lloyd, Peter. 1976. “Transport Costs on Indonesia’s Exports: The Australian Case.” Bulletin of Indonesian Economic Studies 12(November): 116–20. Hummels, David. 2007. “Transportation Costs and International Trade in the Second Era of Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on January 30, 2013 Globalization.” Journal of Economic Perspectives 21(3): 131–54. MacLeod, Ian. 2000. "Foreign Trade Zones." International Trade Administration. http://ia.ita.doc.gov. Ng, Francis, and Alexander Yeats. 1999. “Production Sharing in East Asia: Who Does What for Whom and Why?” World Bank Policy Research Paper 2197. Washington, D.C. Pomfret, Richard, and Patricia Sourdin. 2010. “Why Do Trade Costs Vary?” Review of World Economics 146(4): 709– 30. United Nations Department of Economic and Social Affairs. 2007. ESA/STAT/AC.137.5. “International Merchandise Trade Statistics: Concepts and Definitions.” New York. United Nations Statistical Office. “International Merchandise Trade Statistics: National Compilation and Reporting Practices.” http://UN.org/UNSD/tradeport/introduction_mm.asp. World Bank and UNCTAD. 1987. “SMART – Software for Market Analysis and Restrictions to Trade.” Washington, D.C. Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW • The Impact of the Global Food Crisis on Self-Assessed Food Security Derek D. Headey • How Is the Liberalization of Food Markets Progressing? Market Integration and Transaction Costs in Subsistence Economies Wouter Zant • Decomposing the Labor Market Earnings Inequality: The Public and Private Sectors in Vietnam, 1993–2006 Clément Imbert • Chinese Trade Reforms, Market Access and Foreign Competition: the Patterns of French Exporters Maria Bas and Pamela Bombarda • Firms Operating under Electricity Constraints in Developing Countries Philippe Alby, Jean-Jacques Dethier & Stéphane Straub • Antidumping, Retaliation Threats, and Export Prices Veysel Avsar • Information and Participation in Social Programs David Coady, César Martinelli, and Susan W. Parker THE WORLD BANK 1818 H Street, NW Washington, DC 20433, USA World Wide Web: http://www.worldbank.org/ E-mail: wber@worldbank.org