62985 Volume 23 • Number 2 • 2009 ISSN 0258-6770 (print) ISSN 1564-698X (online) THE WORLD BANK ECONOMIC REVIEW Volume 23 • 2009 • Number 2 THE WORLD BANK ECONOMIC REVIEW Dollar a Day Revisited Martin Ravallion, Shaohua Chen, and Prem Sangraula Evidence on Changes in Aid Allocation Criteria Stijn Claessens, Danny Cassimon, and Bjorn Van Campenhout Does Education Affect HIV Status? Evidence from five African Countries Damien de Walque A Cost–Benefit Analysis of Cholera Vaccination Programs in Beira, Mozambique Marc Jeuland, Marcelino Lucas, John Clemens, and Dale Whittington Do Exporters Pay Higher Wages? Plant-level Evidence from an Export Refund Policy in Chile Ivan T. Kandilov The Determinants of Funding to Ugandan Nongovernmental Organizations Marcel Fafchamps and Trudy Owens Pages 163–346 Liquidity Constraints and Firms’ Linkages with Multinationals Beata S. Javorcik and Mariana Spatareanu www.wber.oxfordjournals.org oxford THE WORLD BANK ECONOMIC REVIEW editor Jaime de Melo, University of Geneva assistant to the editor Marja Kuiper editorial board Scott Barrett, John Hopkins University, USA Thierry Magnac, Université de Toulouse I, Kaushik Basu, Cornell University, USA France Alok Bhargava, Houston University, USA William F. Maloney, World Bank François Bourguignon, Paris School of David McKenzie, World Bank Economics, France Jonathan Morduch, New York University, USA Luc Christiaensen, United Nations Jacques Morisset, World Bank University, WIDER, Finland Juan-Pablo Nicolini, Universidad Torcuato di Stijn Claessens, International Monetary Fund Tella, Argentina Shantayanan Devarajan, World Bank Boris Pleskovic, World Bank Eliana La Ferrara, Università Bocconi, Italy Martin Rama, World Bank Augustin Kwasi Fosu, United Nations Martin Ravallion, World Bank University, WIDER, Finland Elisabeth Sadoulet, University of California, Paul Glewwe, University of Minnesota, USA Berkeley, USA Karla Hoff, World Bank Joseph Stiglitz, Columbia University, USA Hanan Jacoby, World Bank Jonathan Temple, University of Bristol, UK Emmanuel Jimenez, World Bank Romain Wacziarg, University of California, Elizabeth M. King, World Bank Los Angeles, USA Aart Kraay, World Bank Ruslan Yemtsov, World Bank Justin Yifu Lin, World Bank Yaohui Zhao, CCER, Peking University, China The World Bank Economic Review is a professional journal for the dissemination of World Bank–sponsored and other research that may inform policy analysis and choice. 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 quantitative development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but not necessarily proficient in advanced mathematical or econometric techniques. Articles illustrate how professional research can shed light on policy choices. Consistency with World Bank policy plays no role in the selection of articles. Articles are drawn from work conducted by World Bank staff and consultants and by outside researchers. Non-Bank contributors are encouraged to submit their work. Before being accepted for publication, articles are reviewed by three referees—one from the World Bank and two from outside the institution. Articles must also be endorsed by two members of the Editorial Board before final acceptance. 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. 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 23, 3 Issues, 2009): Institutions—Print edition and site-wide online access: £135/$202/E202, Print edition only: £128/$192/E192, Site-wide online access only: £128/$192/ E192; Corporate—Print edition and site-wide online access: US$273/£177/E273, Print edition only: £192/$288/E288, Site-wide online access only: £192/$288/E288; Personal—Print edition and individual online access: £41/$61/E61. US$ 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 considered 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. BACK ISSUES: Previous 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@oxfordjournals.org. Tel: þ44 (0)1865 353907. Fax: þ44 (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@oxfordjournals.org. 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@oxfordjournals.org. Tel: þ81 3 5444 5858. Fax: þ81 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. 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: þ44 (0)1865 354767; Fax: þ44(0)1865 353774; E-mail: jnlsadvertising@oxfordjournals.org. 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. PAPER USED:The World Bank Economic Review is printed on acid-free paper that meets the minimum requirements 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 # 2009 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 trans- mitted 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. THE WORLD BANK ECONOMIC REVIEW Volume 23 ·2009 . Number 2 Dollar a Day Revisited 163 Martin Ravallion, Shaohua Chen, and Prem Sangraula Evidence on Changes in Aid Allocation Criteria 185 Stijn Claessens, Danny Cassimon, and Bjorn Van Campenhout Does Education Affect HIV Status? Evidence from five African Countries 209 Damien de Walque A Cost-Benefit Analysis of Cholera Vaccination Programs in Beira, Mozambique 235 Marc Jeuland, Marcelino Lucas, John Clemens, and Dale Whittington Do Exporters Pay Higher Wages? Plant-level Evidence from an Export Refund Policy in Chile 269 ivan T. Kandilov The Determinants of Funding to Ugandan Nongovernmental Organizations 295 Marcel Fafchamps and Trudy Owens Liquidity Constraints and Firms' Linkages with Multinationals 323 Beata S. Javorcik and Mariana Spatareanu SUBSCRIPTIONS: A subscription to The World Bank Economic Review (ISSN 0258-6770) compri~ .. 3 issues. Prices include postage; for subscribers outside the Americas, issues are sent air freight. Annual Subscription Rate (Volume 23, 3 Issues, 2009): Institutions-Print edition and site-wide online access: £1351$202/€202, Print edition only: £128/$1921€192, Site-wide online access only: £1281$1921 €192; Corporate-Print edition and site-wide online access: US$273/£177/€273, Print edition only: £192/$288/€288, Site-wide online access only: £192/$288/€288; Personal-Print edition and individual online access: £41/$611€61. US$ rate applies to US & Canada, Euros€ 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.orglsubscriptions. 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 considered 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. BACK ISSUES: The current year and two previous years' issues are available from Oxford University Press. Previous 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. CONTAG INFORMKI10N: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: inls.cust.serv@oxfordjournals.org. Tel: +44 (0)1865 353907. Fax: +44 (0)1865353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oxfordjournals.org. Tel: (800) 852-7323 (toll-ftee 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.ip@oxfordjoumals.org. Tel: +81 354445858. Fax: +81334542929. POSTAL INFOR!vIATION: 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 DevelopmentfTI-iE 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. 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.orgljnlslpermissions. 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: +44 (0)1865354767; Fax: +44(0)1865353774; E-mail: jnlsadvertising@oxfordjournals.org. DISCLAL\1ER: 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 DevelopmenthHE 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. PAPER USED: The World Bank Economic Review is printed on acid-free paper that meets the minimum requirements 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 ContentslSocial and Behavioral Sciences, Journal of Economic LiteraturelEconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT c[') 2009 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 trans­ mitted 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 ('~nter, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA. ._ u \1'1.1lI Dollar a Day Revisited Martin Ravallion, Shaohua Chen, and Prem Sangraula The article presents the first major update of the international $1 a day poverty line, pro­ posed in World Development .Report 1990: Poverty for measuring absolute poverty by the standards of the world's poorest countries. In a new and more representative data set of national poverty lines, a marked economic gradient emerges only when consumption per person is above about $2.00 a day at 2005 purchasing power parity. Below this, the average poverty line is $1.25, which is proposed as the new international poverty line. The article tests the robustness of this line to alternative estimation methods and explains how it differs from the old $1 a day line. JEL codes: 132, E31, 010 The widely used $1 a day poverty line was set for World Development Report 1990: Poverty (World Bank 1990) based on research for that repott documen­ ted in Ravallion, Datt, and van de Walle (1991). The aim was to set a global poverty line that defined poverty in the developing world as a whole by the standards of what "poverty" means in the world's poorest countries, recogniz­ ing that richer countries naturally have higher standards. This (intentionally) frugal basis for measuring global poverty gives the $1 a day line a salience in focusing international attention on the world's poorest-a salience that a higher line would not have. I A consensus emerged in the international develop­ ment community on this standard for measuring extreme povetty in the world, and it became the basis of the first Millennium Development Goal, to halve the 1990s $1 a day poverty rate by 2015. This article provides the first major revision of the original $1 a day line. Understanding why this revision is necessary requires understanding how the original international poverty line was set in 1990 and what new data have become available since then. Martin Ravallion (corresponding author) is a director of the Development Research Group at the World Bank; his email address is mravallion@worldbank.org. Shaohua Chen is a senior statistician in the Development Economics Research Group at the World Bank; her email address is schen@Worldbank.org. Prem Sangraula is an economist in the Development Research Group at the World Bank; his email address is psangraula@worldbank.org. 1. For example, Pritchert (2006) proposes a poverty line of around $10 a day. Calculations using the World Bank's Povea/Net (http://econ.worIdbank.orglpovcalnet) indicate that 95 percent of people in developing countries live below this line. THE WORLD BANK ECONOMIC REVIEW, VOL 23, No.2, pp. 163-184 doi:l0.1093/wber/lhp007 Advance Access Publication June 26, 2009 15) The Author 2009. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I THE WORLD BANK. All rights reserved. For permissions, please e-mail: iournals.permissions@oxfordiournals.org 163 164 THE WORLD BANK ECONOMIC REVIEW Ravallion, Datt, and van de Walle (1991) studied how poverty lines varied with mean consumption when both were converted to a common currency at purchasing power parity (PPP, meaning a currency conversion rate that is intended to ensure a common purchasing power over commodities). They found that national poverty lines have a positive economic gradient above some critical level. The elasticity rises with average consumption, approaching unity in rich countries. It can thus be argued that absolute poverty (measured using a poverty line with a constant real value) is the more relevant concept in poor countries, while relative poverty (in which the poverty line rises with the mean) is more salient in middle- and high-income countries. The poverty lines that prevail in each country (or that would be expected given the country's mean consumption) could be used in assessing global poverty. But then the resulting aggregate poverty measures would not be treat­ ing people at the same level of real consumption the same way. And by treating absolutely poor people similarly to relatively poor people such a measure of global poverty would risk diverting the focus from what is surely the highest priority: raising the living standards of the poorest people in the world. But what absolute line should be used? Ravallion, Datt, and van de Walle (1991) proposed measuring global poverty by the standards of the poorest countries, based on a survey of national poverty lines? Drawing on 33 national poverty lines for the 1970s and 1980s (for both developed and developing economies), Ravallion, Datt, and van de Walle proposed a line of $23 a month ($0.76 a day) at 1985 consumption PPP. That value was the predicted poverty line for the poorest country in the sample, based on a regression model. A higher line of $31 a month ($1.02 a day) that was more representative of the poverty lines in low-income countries. Subsequently, the higher line became more accepted in the World Bank and internationally, and it became known as the "$1 a day" line. The PPPs used by Ravallion, Datt, and van de Walle were from the Penn World Table (Summers and Heston 1991) and were based on the price surveys for 1985 done by the International Comparison Program (ICP). New price surveys were done in 1993, and the World Bank started estimating its own PPPs, using methods that were considered more appropriate for measuring poverty. 3 The changes in ICP benchmark years create comparability problems, due to differing estimation methods for the PPPs and differences in the ICP price surveys. Recognizing these problems, Chen and Ravallion (2001) revised past estimates of poverty measures to ensure consistency with new data avail­ able when the ICP benchmark round changed from 1985 to 1993. Chen and Ravallion (2001) applied the new PPPs to the original Ravallion, Datt, and 1. Prior to Ravallion, Datt, and van de Walle (1991), the World Bank had used explicitly arbitrary lines; see Ahluwalia (1974). 3. Ackland, Dowrick, and Freyens (1006) and Deaton and Heston (2008) discuss alternative approaches to measuring PPPs and their appropriateness for different applications. MIt _. I ~$ " Ravallion, Chen, and Sangraula 165 van de Walle (1991) data set of 33 national lines in local currency units. Employing the same regression method as in Ravallion, Datt, and van de Walle, Chen and Ravallion found the predicted poverty line for the poorest country to be $31.96 a month ($1.05 a day). However, a slightly higher line was considered more representative; the new $1 a day line was set at $32.74 a month, or $1.08 a day, at 1993 PPPs. In 2004, about one in five people in the developing world (1 billion people) were deemed to be poor by this standard (Chen and Ravallion 2007). These estimates all relied on the original Ravallion, Datt, and van de Walle (1991) compilation of poverty lines. However, much new analytic work on poverty at the country level has been done since 1990, notably under the World Bank's program of country poverty assessments and the Poverty Reduction Strategy Papers prepared by national governments, often with assist­ ance from the World Bank, other governments, or international agencies. Few of these studies were available in 1990, but they have since been completed for about 100 developing economies. They provide a rich source of data on poverty at the country level, and almost all include estimates of national poverty lines. The poverty studies done since 1990 also allow us to correct for the sampling biases in the original Ravallion, Datt, and van de Walle (1991) compilation of national poverty lines-biases that could not be avoided in the original compilation given the data available at the time. Another important new source of data is the 2005 round of the ICP (World Bank 2008). As the most ambitious round of the ICP (which began in 1968), it is expected to entail substantial improvements in data quality for estimating PPPs. These new data prompt a reassessment of the international poverty line. The analysis in this article leads to a proposed new international poverty line of $1.25 a day at 2005 PPP for household consumption. Section I presents the new compilation of national poverty lines, which are shown to rise with mean consumption but with a low elasticity at low consumption. Based on these empirical results, section II discusses the proposed new international poverty line. Section III compares the proposed new line to the old $1 a day line. Section IV concludes. 1. NATIONAL POVERTY LINES ACROSS DEVELOPING ECONOMIES The two most obvious ways of updating the old $1 a day poverty line have serious drawbacks. First, one might simply apply the U.S. consumer price index (CPI). This assumes that the old $1 a day line, based on an old sample of national poverty lines and an old set of PPPs, is still valid; that assumption ignores possible biases in past data sets-biases the new data can go at least some way toward addressing. Second, one might keep the ratio of the poverty line to (say) the developing world's mean income the same. With growth, this would imply a higher real poverty line over time. Indeed, the poverty line would have an elasticity of unity with respect to mean income, implying that 166 THE WORLD BANK ECONOMIC REVIEW distribution-neutral growth (all incomes grow at the same rate) would leave poverty measures unchanged, even though the poor gained in absolute terms. An elasticity of unity would seem hard to defend, especially for poor countries. 4 The approach taken here returns to the logic of the original $1 a day line, armed with new data. The set of national poverty lines collected by Ravallion, Datt, and van de Walle (1991) covered 33 countries and drew on specialized, country-specific, mostly academic studies of poverty spanning 1971-90. Clearly, this data set is now rather old. Since then there has been considerable expansion in research and analysis on poverty in developing economies, notably through the World Bank's country-level poverty assessments, which have now been completed for many developing economies. These are core reports within the World Bank's program of analytic work at the country level; each report describes the extent of poverty and its causes in a given country. The poverty assessment is conducted in consultation with the government, and most poverty assessments claim government ownership. Most low-income countries have also prepared Poverty Reduction Strategy Papers, which are pre­ pared by the government, often with some financial support from aid donors. A large share of the work on poverty assessments and Poverty Reduction Strategy Papers typically goes into poverty measurement and both typically lay out what is known about poverty in each country, including a detailed poverty profile as well as aggregate poverty statistics and how they have changed over time. Both reports are important sources of information on the accepted national poverty lines. For the purpose of this article, a new data set of 88 national poverty lines was compiled from the most recent poverty assessments and Poverty Reduction Strategy Papers over 1988-2005. In the source documents, each poverty line is given in the prices for a specific survey year (for which the subsequent poverty measures are calculated). In most cases, the poverty line was also calculated from the same survey (though there are some exceptions, for which pre-existing national poverty lines were updated using the CPI). Sometimes the national poverty lines were old lines updated over time for inflation, and sometimes the poverty line was calculated afresh at each survey, though typically anchored to a common food bundle. Such recalculated poverty lines would not generally have the same real value over time when assessed according to a reasonable price index, since the Engel curve may shift for other reasons and thus change the real value of the nonfood component of the poverty line. When a choice had to be made, the most recent national poverty line available was selected. The new data set on national poverty lines differs from the old (Ravallion, Dart, and van de Walle 1991) data set in four main respects. First, while the old data were drawn from sources for the 1980s (with a mean year of 1984), 4. For further discussion, see Ravallion (2008b) and Ravallion and Chen (2009). $ ; • . f Ravallion, Chen, and Sangraula 167 the new data are all post-1990 (mean of 1999), such that in no case do the proximate sources overlap. Second, the new data set covers 88 developing economies (74 with complete data for the subsequent analysis), while the old data set included only 22 developing economies (plus 11 developed countries). Third, the old data set used rural poverty lines when there was a choice, whereas the new one estimates national average lines. Fourth, the old data set was unrepresentative of Sub-Saharan Africa, with only five countries from that region (Burundi, Kenya, South Africa, Tanzania, and Zambia), whereas the new data set has a good spread across regions, including 25 countries in Sub-Saharan Africa. The proportion of African countries in the old sample was about half what it should have been to be considered representative of poor countries. The sample bias in the Ravallion, Datt, and van de Walle data set was unavoidable at the time (1990), but it can now be corrected. The fact that the poverty assessments are World Bank reports raises two con­ cerns. First, it might be conjectured that these are external poverty lines, rather than poverty lines accepted by the country. However, the process of producing a poverty assessment entails (often extensive) consultation with the govern­ ment, including discussion about the most appropriate poverty line. Thus, this new set of poverty lines has a stronger claim to being national poverty lines than those used by Ravallion, Datt, and van de Walle (1991), which were based largely on academic studies. Second, it might be thought that the poverty lines used in the World Bank poverty assessments reports and in governments' Poverty Reduction Strategy Papers are biased toward the World Bank's old international poverty line. This does not appear to be a serious concern. The poverty assessments (and the Poverty Reduction Strategy Papers) typically either use a pre-existing national poverty line or derive a new line, and in both cases the line has no obvious origins in the World Bank's $1 a day poverty line. The aim is to use a poverty line appropriate to the country. Some 80 percent of these reports use a version of the cost of basic needs method in which the food component of the poverty line is the expenditure needed to purchase a food bundle specific to each country (or region) that yields a stipulated food energy requirement. s To this amount an allowance is added for nonfood spending, which is typically anchored to the nonfood spending of people whose food spending (or some­ times total spending) is near the food poverty line. There is considerable scope for discretion in setting such a poverty line. Although the stipulated food-energy requirements are similar, the food bundles that can yield a given food energy intake can vary enormously, and some will be preferable to others in any given context. The nonfood spending that is deemed adequate will also vary. The judgments made in setting the various parameters of a poverty line are likely to reflect prevailing notions of what poverty means in each country setting. 5. This method, and alternatives, are discussed in detail in Ravallion (1994, 1998, 2008a). " 168 THE WORLD BANK ECONOMIC REVIEW These poverty lines are converted to a common currency using the PPP for individual consumption expenditure hy households from the 2005 ICP, as documented in World Bank (2008).6 The 2005 ICP is clearly the most com­ plete assessment to date of how the cost of living varies across countries. The ICP collected primary data on a region-specific list of prices for 600-1,000 (depending on the region) goods and services. The prices were obtained from a large sample of outlets in each country. All regions participated, but the par­ ticipation rate was markedly lower for Latin America. The 2005 ICP introduced several improvements over previous ICP rounds. The number of countries participating rose from 117 in 1993 to 146 countries. The new countries include China, which had not previously participated in the ICP. The surveys have been implemented on a more scientific basis. New methods were used for measuring government compensation and housing. Adjustments were made for the lower average productivity of public sector workers in developing economies (lowering the imputed value of the services derived from public administration, education, and health), Ring comparisons (linking regional PPP estimates through global prices) were done for more countries (18 in all). The 2005 data were also subject to more rigorous supervi­ sion and validation methods than was the 1993 round, including stricter stan­ dards in defining internationally comparable quality standards for the goods identified in the ICP price surveys. Otherwise, the PPPs calculated from the ICP data (and in World Bank 2008) follow standard methods; as in the past, the World Bank uses a multilateral extension of the bilateral Fisher price index. 7 While these are clearly improvements, the new PPPs still have some limit­ ations. The ICP aimed to survey prices that were nationally representative. This was not the case in China, where the Iep survey was confined to 11 cities. Although the survey included some surrounding rural areas, it cannot be con­ sidered representative of rural China, where the cost of living is lower than in urban areas. The correction method described in Chen and Ravallion (200Sa) was used to derive a PPP for rural areas based on a prior estimate of the 6. The rcp started in 1968. Before 2000, the Penn World Table (Summers and Heston 1991) was the main source of the PPP rates for consumption derived from the ICP, as used in the Bank's global poverty measures. In 2000, there was a switch to the 1993 PPPs estimated by the World Bank's Development Data Group; the most recent results are reported in World Bank (2008). There are methodological differences in these two sets of PPPs. The Penn World Table used the Geary-Khamis (GK) method, while the Bank used the Elteto-Koves-Szulc (EKS) method, which is the multilateral extension of the bilateral Fisher index. On the differences between the GK and EKS methods and implications for global poverty measures, see Ackland, Dowrick, and Freyens (2006). There were also improvements in country coverage and data quality in the 1993 PPPs as compared with the Penn World Table. 7. As argued in Ravallion, Datt, and de Walle (1991), the weights attached to different commodities in the conventional PPP rate may not be appropriate for the poor. Results reported in Deaton and Dupriez (2008) do not suggest that the reweighting needed to derive a "PPP for the poor" will have much impact on the aggregate consumption PPP. The working paper version of this article reports tests of sensitivity to using the Deaton-Dupriez PPP (Ravallion, Chen, and Sangraula 2008). .1 Ravallion, Chen, and Sangraula 169 urban-rural differential in absolute poverty lines. However, there are other concerns that were not addressed. The weights attached to different commod­ ities in the conventional PPP rate are not appropriate for the poor (Ravallion, Datt, and van de Walle 1991), though it is not clear that using those weights entails a significant bias. 8 Yet another limitation is that the PPP is a national average; just as the cost of living tends to be lower in poorer countries, the PPP can be expected to be lower in poorer regions within a country, especially in rural areas. 9 For each country, the national poverty line was converted to 2005 inter­ national dollars using the individual consumption PPP from World Bank (2008). The 2005 PPP was not available for 11 of the 88 countries (mainly due to the poor ICP coverage in Latin America) and was deemed unreliable for one country (Zimbabwe).lo Allowing for missing PPPs and other data problems gave 75 linesY Appendix table A-I gives the precise poverty lines for each country; details on the sources are in the working paper version of this article (Ravallion, Chen, and Sangraula 2008). In no case do the sources overlap with Ravallion, Datt, and van de Walle (1991). The density function is given in figure 1. The poverty lines range from $19.05 to $275.71 a month, with a mean of $87.59 and median of $60.81 (figure 1). (The standard deviation is $66.22.) The mode is slightly under $50 a month. This article follows Ravallion, Datt, and van de Walle (1991) in using private consumption expenditure per capita from the national accounts as the measure of economic welfare (or, more precisely, household final consumption expenditure). The sample mean for private consumption expenditure is $209.40 a month ($6.89 a day) at 2005 PPP; 15 of the sample countries have consumption per capita of less than $60 per month, or about $2.00 a day. The poorest country by this measure is Malawi, at $1.03 a day. The mode of the national poverty lines is quite close to the mode of private consumption 8. Deaton and Dupriez (2009) estimated PPPs for the poor for a subset of countries with the required data. The results do not suggest that the implied reweighting has much impact on the consumption PPP. The working paper version discusses sensitivity of the international poverty line to the choice of PPPs (Ravallion, Chen, and Sangraula 2008). The Asian Development Bank (2008) has taken the further step of implementing special price surveys for Asian countries to collect prices on explicitly lower qualities of selected items than those identified in the standard rcp. Using lower quality goods essentially means lowering the poverty line. In terms of the impact on the poverty counts for Asia in 2005, the Asian Development Bank's method is equivalent to using a poverty line of about $1.20 a day by the methods described here; this calculation is based on a log-linear interpolation between the relevant poverty lines. 9. RavalIion, Chen, and Sangraula (2007) allow for urban-rural cost of living differences facing the poor and provide an urban-rural breakdown of the prior global poverty measures using the 1993 PPP. These estimates will be updated in the future work. 10. The 200S consumption PPP implies a poverty line of $6 a month, which is very hard to believe. 1 L One country, Madagascar, was dropped because of large inconsistencies in the data from various sources (national accounts aggregates reported by the World Bank and the International Monetary Fund). Using the World Bank's estimate of private consumption expenditure gives a poverty line almost three times mean consumption. 170 THE WORLD BANK ECONOMIC REVIEW FIGURE 1. Density Functions of Poverty Lines and Private Consumption per Capita at 2005 PPP 0.010 ,-~-~---~-------------, 0.008 0.006 ~ (/) c: c3 0.004 0.002 ........... _-..._-­ .................. -- 0.000 ·~-,--,--'i:__-.,.i~-..,..t---, ----------- i ,~ -100 0 100 200 300 400 500 600 700 800 r=- National poverty line ($ per person per month at 2005PPP) L:.-- Private consumpti0I1!~r:£Elr~~ ($ per ~onth at 2005 PPP) Source: Authors' analysis based on data in appendix table A-l. expenditure per capita, but otherwise the distributions are very different, with consumption showing a far greater spread (figure 1). The alternative to private consumption expenditure from the national accounts is mean household consumption or income from household surveys. However, in many cases the poverty line was calculated from such surveys, so any relationships between the national poverty lines and the survey means may well be spurious, being driven by common measurement errors. Consider, for example, the most popular method of setting a national poverty line, which values a predetermined food bundle and adds an allowance for nonfood spend­ ing based on the food Engel curve. Underestimation of nonfood spending in the survey will shift the Engel curve and automatically adjust the poverty line downward. The measurement error alone will generate a positive correlation between the poverty line and the survey mean. 12 (The overall direction of bias is ambiguous in theory, given that there will also be the usual attenuation bias when a regressor is measured with error.) Under the assumption that the measurement errors in the national accounts are largely independent of those in the surveys, private consumption expenditure is probably a better indicator. 12. The same would happen if the poverty line were derived by the alternative method of finding the total consumption expenditure level at which predetermined food-energy requirements are met on average. If nonfood spending is underestimated by the survey, the poverty line is automatically adjusted downward, reflecting the measurement error. A spurious correlation results. _ [&!IlL 23 IT 1M I . I; & $ • Ravallion, Chen, and Sangraula 171 FIGURE 2. National Poverty Lines and Log Private Consumption per Person for the Survey Year 300 0::­ 0.. 0.. • • LO o • ~ • • 10 200 .c i:: o • •• • • • • ~ § OJ • f 100 8. iii g ~ Z o 3 4 5 6 7 Log consumption per person at 2005 PPP Note: Fitted values use a lowess smoother with bandwidth = 0.8. Source: Authors' analysis based on data in appendix table A-I. Figure 2 plots the poverty lines against log consumption for the survey year. The least squares estimate of the elasticity of the poverty line to private con­ sumption expenditure is 0.655 (with a t-ratio of 13.68, based on a robust stan­ dard error).13 This elasticity estimate is significantly less than unity (t = 7.21), as used in relative poverty lines for many developed countries (see, for example, Eurostat 2005), although it is similar to some past estimates based on subjective poverty lines for developed countries. 14 However, figure 2 suggests that the economic gradient emerges strongly only once mean consumption is above a critical level. In a non parametric regression of the national poverty lines against log mean consumption,15 the elasticity of the poverty line to mean consumption rises from zero to around 0.7 at the highest level of mean con­ sumption (see figure 2).16 13. The estimate is quite robust to outliers; a median quantile regression gives 0.647 (t = 9.57). 14. Hagenaars and van Praag (1985) estimated an elasticity of 0.51 for eight European countries. Kilpatrick (1973) estimated an elasticity of about 0.6 for subjective poverty lines in the United States. 15. The nonparametric regression is Stata's locally weighted scatter plot smoothing method with the default bandwidth (0.8). Alternative bandwidths in the interval (0.2,0.9) were also tested. The mean of the predicted values in the poorest 15 countries ranged from $37.52 to $38.11 (although the regression line was clearly undersmoothed at bandwidths below about 0.5). 16. This elasticity was estimated by taking a simple moving average of the left- and right-side discrete differentials in logs at each data point along the nonparametric regression function in figure 2. 172 THE WORLD BANK ECONOMIC REVIEW The same pattern found by Ravallion, Datt, and van de Walle (1991) using the older compilations of national poverty lines is evident in figure 2, with the poverty line rising with mean consumption, but with a low initial elasticity. By interpretation, absolute poverty appears to be the dominant concern in poor countries, with relative poverty emerging at higher consumption levels. However, it is notable how high the overall elasticity is for developing economies. The economic gradient in the poverty lines comprises a component for food needs and one for nonfood needs, although this difference can be quantified only for a subset of the national poverty lines. For a subsample of 28 countries, complete data are also available for separating the food and nonfood com­ ponents of the national poverty lines. The mean food share at the poverty line is 0.564 (with a range of 0.260-0.794). The elasticity of the food component of the poverty line to mean consumption is 0.471 (t = 9.55), whereas the elas­ ticity of the nonfood component is almost twice as high, at 0.910 (t 8.97). (The overall elasticity is 0.679 (t = 11.02) for this subsample and 0.655 for the full sample.) So, the economic gradient in national poverty lines evident in figure 2 is driven more by the gradient in the nonfood component of the poverty lines (which accounts for about 60 percent of the overall elasticity), although an appreciable share is attributable to the economic gradient in food poverty lines. II. SETTING AN INTERNATIONAL POVERTY LINE BASED ON THE NATIONAL LINES Armed with the new compilation of national poverty lines, consider again the basic idea behind the $1 a day poverty line, which was chosen to be representa­ tive of poverty lines in poor countries. There are several ways of setting a new international poverty line consistent with this idea. The sample median poverty line is $60.81 a month, or almost exactly $2.00 a day; the sample mean is higher, at about $2.90 a day. However, the marked economic gradient shown in figure 2 implies that the mean or median will be well above the poverty lines found for the poorest countries. The poverty line for Malawi-with the lowest personal consumption expen­ diture per capita in the sample-is $26.11 a month. However, like all specific data points in a sample, this one is susceptible to measurement error, and the country-specific error term could be large. It is notable that even though the relationship in figure 2 is quite flat at low consumption, there is still a sizable variance. No doubt, idiosyncratic differences in the data and methods used in setting national poverty lines have a role; there are measurement errors and methodological differences between countries in how poverty lines are con­ structed, which can be interpreted as noise in the mapping from the underlying welfare space into the income space. Some averaging is clearly called for, as is normal in economic measurement. A better method is to use the expected t i Q lit '" Ravallion, Chen, and Sangraula 173 TABLE 1. Estimated Poverty Line for the Poorest Country for Various Parametric Models Predicted poverty line for the poorest country in 2005 PPP Specification dollars per month (C min = $31.34 for Malawi) -----~-~-~-- Zi = a-l- f3C i -I- 8; $31.04 (8.53) Zi"" aT f3tCi + f3 1 Cf + 8; $29.32 (6.59) Zj = aT f3 11n C j + f3t1n Cr + 8j $44.22 (6.89)" In Zj = a+ f3t1n C j + f:ltln C;- T fij $33.76; In Z = 3.52 (33.51) In Zi = a+ f31 Ci + f:lt C;- + 8j $32.63; In Z = 3.49 (47.16) Note: Numbers in parentheses are t-ratios based on robust standard errors. "The turning point In C = 4.04-above the lowest consumption. The predicted value of Zi at the turning point is $36.05 (t 13.61). Source: Authors' analysis based on sources described in Ravallion, Chen, and Sangraula (2008). value of the poverty line in the poorest country, based on how the poverty lines vary with mean consumption. Table 1 gives a number of parametric specifica­ tions (including those used by Ravallion, Datt, and van de Walle 1991; Ravallion 1994; Chen and Ravallion 2001) and the implied estimates of the poverty line for the poorest country. The estimates in table 1 raise three concerns. First, the results may be driven by the specific parametric form. Signs of this possibility include the much higher predicted national poverty line Z for the poorest country in the semi-log model-poverty line Zj regressed on a quadratic function of the log of personal consumption expenditure (in Cj ). But this is deceptive, since the turning point of the quadratic function is above the lowest consumption. This is clearly an artifact of the parametric form, since there is no sign in figure 2 of a negatively sloped segment at low private consumption expenditure per capita. If this spe­ cification is ignored, the results in table 1 suggest that a poverty line of around $1 a day at 2005 PPP is defensible if poverty in the world is measured by the standards of the poorest country in the world. Second, a parametric model need not estimate well at all levels of consump­ tion. For example, the linear regression of Zi on Ci has a very good overall fit, with a correlation of 0.995 with the fitted values in figure 2 and a correlation of 0.836 with the data. However, the linear projection based on this regression underpredicts the poverty lines for the poorest dozen or so countries. 17 The nonparametric regression in figure 2 provides a more flexible method of aver­ aging, given that the regression is ensured to have reasonably good fit over the full range of the data, including among the poorest countries. The predicted value of Malawi's private consumption expenditure per capita is $37.16 a month ($1.22 a day). 17. Based on the linear projection, the mean predicted Z for the poorest 15 countries (ranked by C) is $34.61. By contrast, the mean poverty line for the poorest 15 countries is $37.98, while the mean of the predicted values from the nonparametric regression is $37.89. 174 THE WORLD BANK ECONOMIC REVIEW The third concern is that focusing exclusively on the poorest single country in the sample could make the result vulnerable to measurement errors in con­ sumption. Arguably, it would be better to focus on a reference group of poor countries, with that reference group be defined as countries with personal con­ sumption expenditure per capita of less than some amount C*, say. The following empirical model of the national poverty lines in figure 2 takes these observations into account and allows for measurement errors and idiosyn­ cratic differences in the data and methods used in setting national poverty lines: (1) where Z* is the mean poverty line for the reference group (countries with Ct ::; C*), Ii takes the value one if i is a member of the reference group and zero other­ wise, f(C i ) == E[ZIC = Cil and E[ei I C = Cd = O. For continuity, Z* = f(C*). For internal consistency, the reference group must comprise countries for which Ci ::; C*. When this holds, the reference group can be said to be consistent. The reference group is the sampled countries with personal consumption expen­ diture per capita of less than $60 a month; in ascending order in terms of Ct , those countries are Malawi, Mali, Ethiopia, Sierra Leone, Niger, Uganda, Gambia, Rwanda, Guinea-Bissau, Tanzania, Tajikistan, Mozambique, Chad, Nepal and Ghana. Personal consumption expenditure for this group ranges from $31.34 to $56.90 a month, with a mean of $42.46 (or about $1.40 a day) and a median of $41.33. The mean poverty line is $37.98, or $1.25 a day (the median is $38.51). Under various parametric forms, the linear specification for f(Ct ) was as good as, or better than, others in terms of fit. is The estimated regression corre­ sponding to equation (1) is then (with t-ratios in parentheses based on robust standard errors): Zj = 37.983 Ii (12.55) + (19.388 (2.99) + 0.326 Ct )(l (11.15) - It) + 8i (2) R2 = 0.890, n = 74. The rising segment has a slope of about one-third. 19 The previously mentioned underprediction of the linear regression at low consumption is corrected for by using the $1.25 line as the lower bound. 18. The coefficient on a squared term in private consumption expenditure per capita was not significantly different from zero (t = 0.71). Regressing Z on a quadratic function of log consumption performed as well as the linear model in terms of RZ and gave a very similar estimate of Z*. The parsimonious linear model was therefore selected. 19. Because a common measurement error term appears in both variables, the use of the same PPP for converting both the poverty line and private consumption expenditure could create a spurious correlation. To check this, private consumption expenditure at 1993 PPP was used as the instrumental variable for private consumption expenditure at 2005 PPP (assuming the measurement errors are uncorrelated). This gave a slope of 0.347 (t = 8.42) with a slightly smaller sample (n = 70); the corresponding poverty line was $37.41 a month (t = 11.73). iJl'(B n old ,;; t. 81 JIi!iI t 31 Itt Ravallion, Chen, and Sangraula 175 To check whether the reference group is consistent, the estimated value of C* is calculated, such that Z* = ((C*), which gives C* = 59.50 (t = 3.26). So the choice of all countries with C; < $60 as the reference group is internally consistent with the estimate of equation (2). This estimation method is computationally convenient but has the econo­ metric drawback of treating the regressor Ii as data, which is incorrect since Ii is a function of C*, which depends on the parameters. A better way would be to use a suitably constrained version of Hansen's (2002) method for estimating a piecewise linear ("threshold") mode1. 20 This method gives Z* = 37.464 (t = 6.36) and a slope coefficient on Ci of 0.325 (t = 12.70) and C* = 59.31 (t = 1.82). These parameter estimates are very close to those in equation (2). The $1.25 line is also fairly robust to changes in the reference group. Taking the poorest 10 countries instead of the poorest 15 yields a mean poverty line of $37.27 a month ($1.22 a day) and taking the poorest 20 yields a mean poverty line of $38.33 ($1.26). However, these were not consistent reference groups, unlike that defined by the poorest 15 countries. While this article focuses on absolute poverty, the new data set on national poverty lines also points to a new schedule of relative poverty lines. With a little rounding off, Ravallion and Chen (2009) proposed a parsimonious sche­ dule of relative poverty lines based on the data in figure 2, with a lower bound of $1.25 a day but rising above a critical consumption level with a gradient of $1 in $3. More precisely, the Ravallion and Chen schedule of relative poverty lines (in dollars per day) is: (3) Zi R _ = max [ $1.25, $0.60 + 3 Ci] = $0.60 + max [C $0.65, 3 i ] . The lower bound of $1.25 is binding for the same 15 poorest countries used in setting the absolute line. The point at which the poverty line rises is at C = $1.95 per day. Ravallion and Chen (2009) discuss the theoretical rationale for relative poverty lines based on equation (3). This schedule of relative poverty lines has a high correlation with the fitted values in figure 2 (r 0.994) as well as with the data on national poverty lines (r = 0.836). Indeed, the precision in predicting the national poverty lines is slightly greater using equation (3) rather than the nonparametric regression in figure 2 (using the Stata program's default smoothing parameter).21 Furthermore, neither the fitted values from the 20. By this method, one essentially estimates equation (1) for each possible value of consumption in the data and picks the value that minimizes the residual sum of squares The variation on Hansen's model is that, in this case, the slope of the lower linear segment is constrained to be zero and there is no potential discontinuity at the threshold. We are grateful to Michael Lokshin for programming Hansen's method. 21. The standard deviation of the error is $36.13 for the relative poverty lines and $36.55 for the fitted values from figure 1. Note that a (sufficiently) less smoothed nonparametric regression would do berter than the piecewise linear model used here. 176 THE WORLD BANK ECONOMIC REVIEW nonparametric regression nor a cubic polynomial III C IS significant when added to a regression of Z on ZR.2 2 III. COMPARISONS WITH THE OLD "$1 A DAY" LINE The proposed new international poverty line has a lower value in the United States than the old line of Ravallion, Datt, and van de Walle (1991). The U.S. dollar value in 1993 of the new international poverty line of $1.25 a day is $0.92 a day-IS percent lower than the Chen and Ravallion (2001, 2004) poverty line of $1.08 a day at 1993 PPP. The $1.25 line in 2005 is equivalent to exactly $1.00 a day in the United States in 1996. Put another way, simply updating the old 1993 line for inflation in the United States would give a line of $1.45 a day in 2005,23 which is well above the poverty lines found in the poorest countries and significantly higher than the $1.25 line (t = 2.08; prob. = 4 percent). As the following discussion will make clear, these calculations are deceptive for two reasons. First, the underlying data on national poverty lines has improved, enabling use of a more representative sample of national lines than that used to set the $1.08 line at 1993 PPP. Second, the PPPs from different ICP rounds are not strictly comparable, and the new PPPs are likely to be a better guide to the cost of living in poor countries. As will be shown, these two effects work in opposite directions: the first raises the international poverty line whereas the second lowers it. The first difference between the proposed new international poverty line and the old one is in the underlying sample of national poverty lines. The effect of the new sample is to raise the international poverty line when assessed at a common set of PPPs. For the poorest 15 countries ranked by consumption per capita at 1993 PPP, the mean poverty line in the new sample is $44.19 ($1.45 a day). This compares to $33.51 ($1.10 a day), which is the mean for the eight countries in the old sample with consumption per capita below the upper bound of consumption for the poorest 15 countries in the new sample. This might be taken to suggest that there was an upward drift in the national poverty lines of poor countries over this period. This would seem implausible, however, as it appears to be quite rare for developing economies to increase the real value of their poverty lines over time. The more plausible explanation lies with the aforementioned differences between the old and new samples of national poverty lines. Making the sample more representative-with a much larger and more regionally balanced sample of developing economies and with both urban and rural lines for almost all countries-appears to have raised the 22. The joint F-test of the null hypothesis that the three parameters in the cubic function of Care all zero in the regression of Z on ZR gave F(3,69) = 0.14 (prob. = 0.93), while the t-test on the coefficient on the fitted values when added to the same regression was 0.44. 23. The ratio of the 2005 CPI for the United States to the 1993 CPI is 1.352. -'>- $ w b)4 $ J; Ravallion, Chen, and Sangraula 177 FIGURE 3. National Poverty Line in Local Currency Units Converted into Dollars for 1993 and 2005 for the 72 Countries with PPP Available for Both Years 360 320 :2 C 280 0 E ~ 240 a. a. a. 200 It') 0 0 N 160 iii Q) JE 120 » t: Q) > 0 80 a. 40 o 40 80 120 160 200 240 280 320 360 Poverty line at 1993 PPP ($/month) Source: Authors' analysis based on sources described in the text. poverty line. Conversely, one can conjecture that had the old sample had been more representative, the old international line would have been appreciably higher. The second difference between the new international poverty line and the old one is in the PPPs. There were substantial revisions to the PPPs in the 2005 ICP round relative to the 1993 round. Probably the most important difference for current purposes is that the 1993 ICP for developing economies used less rigorous standards for specifying the quality of goods and weaker supervision in poor countries, so that lower quality goods were priced than would have been found in the U.S. market. The following discussion focuses on this second difference using the new sample of national poverty lines. Some large changes in the PPPs are evident if the same national poverty line in local currency units is converted into dollars for both 1993 and 2005 and the results are then compared, as in figure 3 for the 72 countries in the data set with PPPs available for both years. It is notable that the 2005 ICP has tended to entail a downward revision in the dollar value of the lowest stratum of poverty lines. The implied revisions are substantial for poor countries. To see this, let pppr denote the true PPP exchange rate derived from the ICP round for date t. If the data were internally consistent, the PPP rate for a given country would 178 THE WORLD BANK ECONOMIC REVIEW change over time according to differences in the country's rate of inflation and that for the numeraire country, the United States, so that (4) where Dr is the true deflator for converting the country-specific poverty line to the PPP reference date, t. While equation (4) holds for the true values of all variables, the measurements are based instead on the observed values, PPP; and Df. To focus on the implications for the errors in the historical PPP data for developing economies, the 2005 PPP and the deflators are assumed to be accurate. The poverty lines are converted to a common currency using these observed data. Let Z;:; Zi D;/PPP; denote the calculated poverty line in PPP dollars in country i at date t where Zi is the poverty line in local currency for country i (at some country-specific date, which is implicit). Under these assumptions, the revision to the PPP for 1993 that is implied by the observed data can be readily derived as follows: pppi93 * (DOS /D93) ppp05 DOS /D93 us us i us us (5) J50~f/D93 ppp 93 = Z05/Z93 . ppp93 = I t t t t t The sample mean of this variable is 1.578 (with a standard error of 0.062; n = 72). Thus, a sizable underestimation of the 1993 PPP is implied by the new PPPs and inflation data. Furthermore, the extent of this underestimation tends to be greater for poorer countries. The implied values of PPPj3*IPPpj3 plotted against log consumption per capita at 2005 PPP show a marked negative gradi­ ent (figure 4). The correlation coefficient is -0.47, which is significant at the 1 percent level (t = -4.70). Among the poorest countries in personal consump­ tion expenditure, the data suggest that a marked upward revision is required to the 1993 PPPs. In other words, the 1993 ICP round underestimated the price level in these countries relative to that in the United States. This is consistent with the view that the 1993 ICP used a lower quality of goods in poor countries than would have been found in the U.S. market (say) because of looser standards of specifying the quality of goods and weaker supervision in poor countries, particularly the poorest countries. These observations are suggestive at best. The data problems are unlikely to be confined to the 1993 PPPs; errors are no doubt also present in the 2005 PPPs and the inflation rates. But these results are at least consistent with the interpretation that less rigorous specification and monitoring of quality stan­ dards in the 1993 ICP resulted in lower quality goods being priced in poor countries, leading to an underestimation of the PPP for many of the poorest countries or (equivalently) to underestimation of the true cost of living. Clearly, there are serious comparability problems across ICP rounds. Note, however, that the method of measuring global poverty used by Chen and lit jj p- II j Uljllliiill '" ""U I $ •. Uk 4 Ravallion, Chen, and Sangraula 179 FIGURE4. Implied Revisions to 1993 PPP Plotted against Log Private Consumption per Person at 2005 PPP 3.5 3.0 2.5 o 00 0 o 2.0 1.5 1.0 0.5 0.0 -t--'--T-'--~T-~I--~T--~I--~ 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Log consumption per person at 2005 PPP Note: Implied revisions to the 1993 PPP values (PPP93*) are calculated using the 2005 round and differential rates of inflation between 1993 and 2005; PPP93* is then normalized by the original estimate of the 1993 PPP rates (PPP93). Source: Authors' analysis based on sources and methods described in the text. Ravallion (2001, 2004, 2007) does not assume comparability of ICP rounds. The salient features of the method are that the international poverty line is con­ verted to local currency units in the ICP base year (using the same consumption PPP as was used for the national poverty lines) and is then converted to the prices prevailing in the relevant survey year using the best available CPI for that country. The PPP conversion is done only once, and all estimates are revised back in time. IV. CONCLUSIONS The original $1 a day poverty line aimed to assess poverty in the world as a whole by the standards of what poverty means in the world's poorest countries. This article has revisited this idea armed with a new set of national poverty lines for low- and middle-income countries, drawing on the World Bank's country-specific poverty assessments and the Poverty Reduction Strategy Papers prepared by the governments of the countries concerned. The new set of national poverty lines is both more up to date and more representative of developing economies, notably in Sub-Saharan Africa. These national poverty lines were converted to a common currency using the new set of household consumption PPP's estimated from the 2005 round of ICP price surveys. 180 THE WORLD BANK ECONOMIC REVIEW Because the 2005 ICP round implied substantial upward reVISIOns to the PPPs of the poorest countries, simply updating the old international poverty line for inflation in the United States gives a poverty line that is well above the lines found among the poorest countries at 2005 PPPs. Instead, a new international poverty line of $1.25 a day is proposed for 2005 (equivalent to $1.00 a day in 1996 U.S. prices), which is the mean of the lines in the poorest 15 countries in consumption per capita, based on the new compilation of national poverty lines. This new poverty line is fairly robust to different estimation methods. Using the new international poverty line proposed in this article, Chen and Ravallion (2008b) find that 1.4 billion people in 2005-25 percent of the population of the developing world-lived in poverty. That share was 52 percent 25 years earlier (in 1981) and 42 percent in 1990.24 However, Chen and Ravallion find that progress was highly uneven, both over time and across regions. If the trend is extrapolated forward, the developing world as a whole appears to be on track for attaining the first Millennium Development Goal. That is not the case, however, for developing economies excluding China. For those countries, the losses to the poor have roughly cancelled the gains, so that the number of people living below $1.25 a day stays at around 1.1-1.2 billion over 1981-2005. ACKNOWLEDGMENTS The authors have benefited from useful discussions, comments, and other help from Yonas Biru, Angus Deaton, Yuri Dikhanov, Olivier Dupriez, Francisco Ferreira, Alan Heston, Norman Loayza, Branko Milanovic, Halsey Rogers, Luis Serven, Changqing Sun, Eric Swanson, Fred Vogel, and seminar partici­ pants at the World Bank and conference participants at the World Institute for Development Economics Research, Helsinki, Finland. The authors also thank three anonymous referees for their helpful comments. 24. The set of countries is held constant over rime at the number of countries that have at least one household survey satisfying the quality conditions. The estimation method provides an estimate for each of these countries at each reference data point; for detail, see Chen and Ravallion (2008b). ~ (L 4Q g t U t ttA II; RavalJion, Chen, and Sangraula 181 ApPENDIX TABLE A-I. National Poverty Lines 2005 PPP dollars Consumption per capita per month for survey Poverty line per capita per Country Survey year year month Albania 2002 280.71 85.18 Argentina' 1999 641.90 183.07 Armenia 1998-99 174.84 73.36 Azerbaijan 2001 292.23 84.80 Bangladesh 2000 64.34 31.46 Belarus 2002 362.04 187.73 Benin 1999-2000 72.82 23.57 Bolivia a 2001 216.66 142.39 Bosnia and 2001 393.95 217.65 Herzegovina Brazil" 2002-03 465.45 180.14 Bulgaria 2001 445.70 100.77 Burkina Faso 2003 68.54 26.27 Cambodia' 2004 75.06 42.80 Cameroon 2001 112.96 69.62 Chad 1995-96 47.04 26.60 Chile' 2000 487.08 119.00 China 2002 120.78 25.89 Colombia" 1999 334.47 199.56 Congo Republic 2005 72.13 67.99 Cote d'Ivoire 1998 117.07 50.36 Djibouti 2002 111.70 95.61 Ecuador 2001 289.72 122.62 Egypt 1999-2000 225.68 53.43 Estonia 1995 431.16 102.78 Ethiopia 1999-2000 35.22 41.04 The Gambia 1998 40.88 44.92 Georgia 1997 182.79 111.24 Ghana 1998-99 56.90 55.65 Guinea Bissau 1991 45.12. 45.96 Hungary 1997 668.31 247.87 India 1999-2000 84.24 27.40 b Indonesia 1999 139.96 32.63 Jordan 2002-03 251.59 71.47 Kazakhstan 1996 213.41 95.32 Kenya 1997 112.80 84.71 Kyrgyz Republic 2003 109.85 60.81 Lao PDR 1997-98 32.10 Latvia 1995 370.11 137.91 Lesotho 1994-95 135.84 49.37 Macedonia FYR 1994 348.96 177.25 Malawi 2004-05 31.34 26.11 Mali 1988-89 31.96 41.89 -~~-~---~ (Continued) 182 THE WORLD BANK ECONOMIC REVIEW TABLE A-I. Continued 2005 PPP dollars Consumption per capita per month for survey Poverty line per capita per Country Survey year year month Mauritania 2000 99.63 68.16 Mauritius 1991-92 328.33 272.99 Mexico 2002 630.73 192.22 Moldova 2001 124.89 60.81 Mongolia 2002-03 80.55 57.88 Morocco 1998-99 167.73 55.33 Mozambique 2002-03 45.52 29.54 Nepal 2003-04 54.55 26.43 Niger 1993 39.34 33.35 Nigeria 1985 61.49 31.38 Pakistan" 1998-99 98.31 50.67 Paraguay 2002 222.27 192.14 Peru· 2000 326.61 76.10 Philippines 1988 134.17 46.02 Poland 1993 465.05 203.23 Romania 2001 397.77 125.57 Russian Federation 2002 455.72 132.67 Rwanda 1999-2001 41.33 30.17 Senegal 1991 78.92 19.05 Sierra Leone 2003-04 36.94 51.54 Sri Lanka 2002 233.05 45.38 Tajikistan 1999 45.49 58.83 Tanzania 2000-01 45.26 19.20 Thailanda 1992 243.52 57.58 Tunisia 1995 240.63 41.17 Turkey 2002 391.42 112.26 Uganda 1993-98 40.01 38.51 Ukraine 2002 254.62 109.43 Uruguay 1998 593.71 275.71 Venezuela RB 1989 492.30 224.73 Viemam 2002 81.18 32.52 Yemen 1998 76.37 65.37 Zambia 2002-03 60.40 39.69 - is not available. Note: For a summary of the methods used for each country and other details on the individual country estimates, see Ravallion Chen, and Sangraula (2008). The national poverty line is calcu­ lated as the weighted mean of the urban and rural poverty lines, using urban and rural real con­ sumption (or income) shares as the weights and the poverty lines as the deflators. aThe poverty line is an urban poverty line since the 2005 PPP is based on urban prices for that country. bThis rises to $31.25 using the adjustment for urban-rural cost of living differences in India used by Chen and Ravallion (2008b). Source: Authors' analysis based on sources described in Ravallion, Chen, and Sangraula (2008). '''MIL r. r e ! Ravallion, Chen, and Sangraula 183 REFERENCES Ackland, Robert, Steve Dowrick, and Benoit Freyens. 2006. "Measuring Global Poverty: Why PPP Methods Matter." Canberra: Australian National University. Ahluwalia, Montek. 1974. "Income Inequality: Some Dimensions of the Problem." In Hollis Chenery, Montek Ahluwalia, Clive Bell, John Duloy, and Richard Jolly, eds., Redistribution with Growth. New York: Oxford University Press. Asian Development Bank. 2008. Comparing Poverty across Countries: The Role of Purchasing Power Parities. Manila: Asian Development Bank. Chen, Shaohua, and Martin Ravallion. 2001. "How Did the World's Poor Fare in the 1990s?" Review of Income and Wealth 47(3):283-300. - - - . 2004. "How Have the World's Poorest Fared since the Early 1980s?" World Bank Research Observer 19(2):141-70. ---.2007. "Absolute Poverty Measures for the Developing World, 1981-2004." Proceedings of the National Academy of Sciences of the United States of America 104(43 ):16757 -62. - - - . 2008a. "China Is Poorer than We Thought, but No Less Successful in the Fight against Poverty." In Sudhir Anand, Paul Segal, and Joseph Stiglitz, eds., Debates on the Measurement of Poverty. Oxford, UK: Oxford University Press. - - - . 2008b. "The Developing World Is Poorer than We Thought, but No Less Successful in the Fight against Poverty." Policy Research Working Paper 4703. World Bank, Washington, DC. Deaton, Angus, and Alan Heston. 2008. Understanding PPPs and PPP-Based National Accounts. Princeton, New Jersey: Princeton University. Deaton, Angus, and Olivier Dupriez. 2008. "Poverty PPPs around the World: An Update and Progress Report." Development Data Group. World Bank, Washington, DC. Eurostat. 2005. "Income Poverty and Social Exclusion in the EU25." Statistics in Focus 0312005. Luxembourg: Office of Official Publications of the European Communities. Hagenaars, Aldi, and Bernard van Praag. 1985. "A Synthesis of Poverty Line Definitions." Review of Income and Wealth 31(2):139-54. Hansen, Bruce E. 2000. "Sample Splitting and Threshold Estimation." Econometrica 68(3):575-603. Kilpatrick, R. 1973. "The Income Elasticity of the Poverty Line." Review of Economics and Statistics 55(3):327-32. Pritchett, Lant. 2006. "Who is Not Poor? Dreaming of a World Truly Free of Poverty." World Bank Research Observer 21(1):1-23. Ravallion, Martin. 1994. Poverty Comparisons. Chur, Switzerland: Harwood Academic Publishers. - - - . 1998. Poverty Lines in Theory and Practice. Living Standards Measurement Study Paper 133. World Bank, Washington, DC. - - - . 2008a. "Poverty Lines." In Larry Blume, and Steven Durlauf, eds., The New Palgrave Dictionary of Economics. London: Palgrave Macmillan. - - - . 2008b. "On the Welfarist Rationale for Relative Poverty Lines." In Kaushik Basu, and Ravi Kanbur, eds., Social Welfare, Moral Philosophy and Development: Essays in Honour of Amartya Sen's Seventy Fifth Birthday. Oxford, UK: Oxford University Press. Ravallion, Martin, and Shaohua Chen. 2009. "Weakly Relative Poverty." Policy Research Working Paper 4844. World Bank, Washington, DC. http://go.worldbank.org/JER07Y9ABO (last accessed on June 17, 2009). Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2007. "New Evidence on the Urbanization of Global Poverty." Population and Development Review 33(4):667-702. - - . 2008. "Dollar a Day Revisited." Policy Research Working Paper 4620. World Bank, Washington, DC. http://go.worldbank.org/TOLT4IYYEO (last accessed on June 17, 2009). Ravallion, Martin, Gaurav Datt, and Dominique van de Walle. 1991. "Quantifying Absolute Poverty in the Developing World." Review of Income and Wealth 37(4):345-61. 184 THE WORLD BANK ECONOMIC REVIEW Summers, Robert, and Alan Heston. 1991. "The Penn World Table (Mark 5): An Extended Set of International Comparisons, 1950-1988." Quarterly Journal of Economics 106(2):327-68. World Bank. 1990. World Development Report: Poverty. New York: Oxford University Press. - - . 2008. 2005 International Comparison Program: Tables of Final Results. World Bank, Washington, DC. a Evidence on Changes in Aid Allocation Criteria Stijn Claessens, Danny Cassimon, and Bjorn Van Campenhout Have donors changed their aid-allocation criteria over the past three decades toward greater selectivity, a frequently stated goal of the international development community? Using data on how 22 donors allocated their bilateral aid among 147 countries over 1970-2004, the article finds that after the fall of the Berlin wall in 1989 and especially in the late 19905, bilateral aid responded more to poverty and the quality of the policy and institutional environment in the recipient countries. Furthermore, the sensitivity of aid allocation to the country's size and its debt burden has declined over time. These results are robust to different samples and model specifications, various econometric techniques, and alternative measures of institutional quality. While the specific factors causing these changes cannot be identified-these presumably include geopolitical and economic concerns and the many changes in the international aid architecture--donors still differ greatly in their selectivity. This suggests that further, multifaceted reforms are needed to ensure even greater selectivity of aid.JEL codes: 011, 016, 019 This article explores how country characteristics affect the way aid is provided by donor countries and how this has varied over time. Data on bilateral aid flows are relatively easily available for long periods of time for a large number of donors and recipient countries, allowing a combination of longitudinal and Stijn C1aessens (corresponding author), International Monetary Fund (USA) and University of Amsterdam; his email address is sclaessens@imf.org. Danny Cassimon, Institute of Development Policy and Management (lOB), University of Antwerp: his email addressisdanny.cassimon@ua.ac.be. Bjorn Van Campenhout, Institute of Development Policy and Management (lOB), University of Antwerp: his email addressisbjom.vancampenhout@ua.ac.be. The article was written in part while Claessens was at the World Bank. Earlier versions were presented at the GARNET conference, September 27-29, 2006, Amsterdam; two Economic and Social Research Council (ESRC) workshops (October 516, 2006 and May 2/3, 2007): the Oxford-Cornell conference New Directions in Development Assistance, Oxford, United Kingdom, June 11-12, 2007: and the ESRClWorld Economy and Finance Research Programme (WEF) conference Finance and Development, London, June 27-29, 2007. The authors thank conference participants, especially Badi Baltagi, and discussant Adeel Malik, as well as Ora Celasun, Rodney Ramcharan, and Antonio Spilimbergo for their comments. They also thank the journal editor and three anonymous referees for their useful suggestions and Ying Lin for help in preparing the data. The work was supported by the ESRClWEF program National and International Aspects of Financial Development (Award RES-156-2S-0009l. Additional results are available in the working paper version online at http://ssrn.coml abstract=997833 and at the journal website: http://wber.oxfordjournals.orgl. THE WORLD flANK ECONOMIC REVIEW, VOL. 23, No.2, pp. 185-208 doi:l0.l093/wber/lhp003 Advance Access Publication June 3, 2009 (g The Author 2009. 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@oxfordjournals.org 185 186 THE WORLD BANK ECONOMIC REVIEW cross-sectional approaches and making sound research easier. The general view is that aid is being allocated better recently, with greater emphasis on "deser­ ving" countries in "need," because of a combination of geopolitical and global economic trends, as well as international policy- and country-specific insti­ tutional and other changes that have brought greater transparency, better coordination, and greater alignment of policies and procedures. The article uses data on bilateral aid by 22 donors to 147 recipient countries over 1970­ 2004 to investigate how factors deemed to reflect need and merit affect aid allocation. 1 The results indicate that after the fall of the Berlin wall in 1989 and especially in the late 1990s, bilateral aid responded more to reCipient countries' economic needs and the quality of their policy and institutional environment and less to their size and external debt. These findings are important for several reasons. Aid flows are large, often more than 10 percent of a country's GDP and more than $100 per capita per year in some countries. Which countries received aid-poor or rich, deserving or less deserving-thus has important economic and social relevance. Studies on aid effectiveness such as the World Bank study Assessing Aid (1998) and the work of Burnside and Dollar (2000) showing that aid works better in good policy and institutional environments has led many policymakers to conclude that targeting aid to countries with more enabling environments maximizes overall aid effectiveness. Although the robustness of some of this research has been questioned,2 the findings have nevertheless reinforced the view that aid ought to be considered only for countries that are "deserving" and in "need." The consensus has been that much aid has not been allocated in this way, particularly in the recent past. With empirical studies as far back as the 1960s, research has shown that political and strategic interests rivaled concerns for growth, poverty reduction, and other economic objectives in aid allocation, at least until the early 1990s (Radelet 2006 and Easterly 2003 provide general litera­ ture reviews that also cover aid allocation). Notably, Alesina and Dollar (2000), confirmed by others, show that noneconomic factors, including geo­ political factors, greatly influence aid allocation, in addition to economic and development considerations. Since the mid-1990s, however, geopolitical and global economic changes and new research insights have altered the way official aid is provided-aid 1. Recipient countries' merit is proxied by their score on the World Bank's CPIA index. Three factors are used to reflect need for aid: poverty, proxied by CDP per capita; size, proxied by population; and inability to attract external financing relative to need, proxied by external debt burden. 2. Easterly, Levine, and Roodman (2004), using the same specifications, find that the Burnside and Dollar (2000) results do not stand up over a longer time period. Rajan and Subramanian (2008), correcting for the bias that aid typically goes to poorer countries or to countries after poor performance, find little robust evidence of a positive (or negative) relationship between aid inflows into a country and its economic growth. And Roodman (2007) highlights the general problems with econometric robustness in this area. $[ I _iI. j __ 114' _IILa ___ '! il4 j .II" 1_ • Claessens, Cassimon, Van Campenhout 187 scholars even speak of a paradigm shift. 3 Geopolitical changes such as the fall of the Berlin Wall in 1989 and the end of major Communist govern­ ments removed many of the geopolitical motivations for aid. Important economic changes include the end of central planning and the increase in private capital flows and globalization more generally, which have led to new development models and allowed for different forms of external financing. Partly in response to these forces, the forms and rules under which aid is being provided have changed at the multilateral and bilateral donor levels and at the individual recipient country level. Multilateral changes include a greater emphasis on coordination among donors and with recipient country priorities (the harmonization and alignment agenda put forward in the 2005 Paris Declaration), greater transparency, and the growing importance of alternative aid providers, such as private philanthropists engaged in health and environ­ mental issues. Individual donors have been changing their aid composition (the mix between project and program aid, for example), and many donors have been providing grants instead of loans. A greater openness in aid allocation is common, along with an aim for more selectivity and greater use of benchmarks and results-based allocations. These changes have been accompanied by changes in development approach, including stronger recipient country ownership of development programs (and not just by the government), greater use of Poverty Reduction Strategy Papers, the explicit incorporation of the Millennium Development Goals, and the enu­ meration of the objective of scaling up. Country-specific actions have included more debt relief, with almost all donors engaging in bilateral (official) debt reduction, in the latest round through the heavily indebted poor countries (HIPC) enhanced initiative. Additionally, the multilateral debt reduction initiat­ ive (MDRI) is under way. The goal of all these changes has been to increase the development efficiency and effectiveness of aid. The changes can also be expected to affect the amounts and distribution of bilateral aid flows. Several channels could be at work. Recipient countries that abide by the new paradigm should see them­ selves rewarded with more aid, and on more concessional terms. Institutional and policy changes should lead to fewer coordination problems among donors, resulting in better aid allocation. There should be less influence of historical, geopolitical, and other noneconomic or developmental factors in aid allo­ cations. Furthermore, official debt reduction may alter the effect of debt on aid allocation. Even with good polices in place, debt can deter aid flows. Overindebted but deserving countries may be less able to attract external 3. This development paradigm shift has been gradual, of course, and reality has often differed from the measures countries claimed to have taken or donors claimed to have supported (see Thomas and others 1991 for a review of the problems uncovered when structural adjustment loans were first evaluated in the late 19805). Nevertheless, some real changes did occur. 188 THE WORLD BANK ECONO'vlIC REVIEW financing and thus end up growing slower. And interactions between the quality of a country's policy and the composition of its debt burden can affect aid flows. Earlier research showed, for example, that donors continued to give new loans and grants to countries with poor policies and that were relatively more indebted to bilateral and multilateral financial institutions to prevent defaults on past loans and avoid having to admit to "mistakes" (Birdsall, Claessens, and Diwan 2003).4 These various effects might have changed over time, especially considering the large official debt reductions recently. Some recent studies on aid flows reveal (indirectly) that donors' selectivity toward country need and policies has improved over time (Berthelemy and Tichit 2004; Roodman 2005; Dollar and Levin 2006; Sundberg and Gelb 2006). Easterly (2007) expresses a contrarian view, finding no consistent evi­ dence of increased selectivity with respect to policies and only temporarily increased selectivity in the late 1990s with respect to corruption. The issue is thus unsettled, in part because few researchers have studied the effects of changes in aid architecture using disaggregated bilateral data. The main question this article addresses is whether changes have led over time to donors providing aid in a more rational manner. Specifically, it investi­ gates whether in recent periods donors have allocated aid with greater sensi­ tivity to recipient country income level and the quality of countries' policies and institutional environment. It examines the changes in sensitivities of aid allocation to country size and level of debt burden. The general finding is one of significant changes, with the characteristics that drive aid responding over time in "better" ways, especially to income level (poorer countries receive more aid) and country policies (better policies are rewarded with more aid). The small country bias seems to have declined, and debt burden seems to play a smaller role in determining aid flows. Although there is evidence of improve­ ments in selectivity for most donors, large differences among donors remain. This suggests a future research agenda on what drives some donors to reform their aid policies while others do not seem to be affected. The article is structured as follows. Section I describes the data and the methodology. Section II discusses the results and robustness checks. Section III considers some implications for further research. I. DATA A~D METHODOLOGY This section describes the data sources, variables, and methodology used in the study. 4. A study by Marchesi and Missale (2004), examining grants and net loans to a panel of 55 HIPCs and non-HIPCs during the 19805 and 19908, finds that total net transfers to HIPCs has been increasing with their debt level, as higher net loans from multilaterals and grants more than offset lower bilateral loans. Geginat and Kraay (2007) study whether IDA flows exhibited defensive lending (whether disbursements deviate from the CPIA-related formula for allocation, with higher allocation to countries with high IDA debt service). They conclude against defensive lending . • ru • Ii lJ!1 II. 4 Jf!"" . jl'.I$ Claessens, Cassimon, Van Campenhout 189 Data Sources and Variables Data on official development assistance (including debt reduction) for each reporting donor to each recipient country in a specific year come from the Organisation for Economic Co-operation and Development/Development Assistance Committee (OECDIDAC) Aid Statistics database (www.oecd.orgl daclstats). While the database does not include all bilateral donors (China, a recent donor, is not a reporting member, for example), it covers the bulk of international aid flows for 1970-2004. Recipient countries are restricted to developing countries (a few high-income countries also receive aid). The data are a three-dimensional panel of aid flows to 147 countries from 22 bilateral donors over the period. There are caveats, however. For example, classification of loans as official aid is based on a somewhat arbitrary cutoff as to their grant element (at least 25 percent), which is itself difficult to calculate. Also, despite adjustments, the quality of data on debt relief is poor. Table 1 provides more details on the variables used and their sources. The analysis uses actual disbursements (actual resources transferred) rather than commitments. The DAC statistics generally focus on the concept of net aid, which is total resources provided by donors as grants (including technical cooperation grants), loans, and debt relief, net of any loan principal repay­ ments. Unlike many earlier studies that use the net aid data directly, this study transforms the data into net aid transfers by also taking into account interest payments, thus deriving total net resource transfers. This concept of net trans­ fers, used in some other aid studies (Chang, Fernandez-Arias, and Serven 1999; Roodman 2005), is close to the economic concept of actual resources trans­ ferred, rather than being some accounting concept. It avoids treating interest payments differently from principal payments and receipts-important consid­ ering the many official debt restructurings that rescheduled interest payments and converted them into principal obligations. Thus, the total net aid transfer concept is defined as: Net aid transfer total (bilateral) official development assistance grants + total (bilateral) official development assistance loans extended to recipients official development assistance loan amor­ tization by recipients - interest paid by recipient Since the unit of interest for aid is the poor person, as in most studies, net aid transfer is scaled by the recipient population to get the annual bilateral net aid transfer per person (called "aid" for short). This dependent variable is then related to several independent variables. The main variable of interest, the need (or poverty) selectivity dimension of aid, is proxied by the recipient country's per capita income (in constant U.S. dollars) lagged one period (to limit the risk that aid flows are driving GDP per capita). Countries with poorer people are expected to receive more aid. The policy selectivity dimension of aid is investi­ gated using the World Bank Country Policy and Institutional Assessment (CPIA) score for the recipient country. This index, produced by World Bank ...... \0 o TABLE 1. Variables, Sources, and Descriptive Statistics (U.S. dollars, unless otherwise indicated) ..., Number of Standard ::I: m Variable Description Source observations Mean Median deviation Minimum Maximum ~ o ~ Dependent variable ,... Net aid transfer Net aid transfer per OECDIDAC Aid 95,921 2.36 0.008 42.5 -137.6 9,052 " '" >­ per capita capita­ Statistics database z observations used in x analysis '" (l o Nonzero observations OECDIDAC Aid 56,684 4.00 0.21 55.3 137.6 9,052 z Statistics database o ?:: All observations OECDIDAC Aid 105,512 2.15 0.001 40.5 -137.6 9,052 (l Statistics database ~ m Independent variables < Lagged GDP GDP per capita at World Bank, World 67,694 3,764 2,830 3,192 466 23,266 '" ~ per capita purchasing power Development parity rates in 2000 Indicators database prices, lagged 1 year Population Log (population) World Bank, World 105,512 2.8 million 5.0 million 11.7 million 19,700 1.3 billion Development Indicators data base CPIA Country Policy and World Bank, Country 66,154 3.46 3.57 0.88 0.72 6 Institutional Performance Rating Assessment score of International Development I Association-eligible countries Burnside-Dollar Policy Index created as World Bank, World 48,740 0.31 -0.18 1.27 5.47 3.08 I in Burnside and Development Dollar (2000) Indicators database 1 t Present value of The present value of World Bank, Global 75,768 182 J 03 328 0.0 6,510 external debt debt as a ratio to Development exports of goods Finance and services (percent) Net aid others Net aid transfer per OECDIDAC Aid 95,921 35.5 13.4 134 -129.2 9,567 capita provided by Statistics data base all other donors Donor sum of The sum of net aid OECDIDAC Aid 90,516 313 85.1 704 -17.0 11,189 nct aid transfers provided Statistics database transfers to all countries by the specific donor Lagged bilateral The sum of bilateral IMF Direction of 70,621 2.1 0.26 13.8 0.0 15.43 ~ ~ trade donor-recipient Trade database " II> country exports and §l ~ imports (percent), scaled by recipient country GOP, ~ lagged 1 year ~ ~ Note: Data for all variables are available for 1970-2004, except for the CPIA data, which are available only for 1977-2004 for most countries. The ;,j values are simple averages of individual bilateral average per capita flows. They differ from those in figure 1, which are weighted by population. ~ ~ ~ o ~ .... .... \0 192 THE WORLD BANK ECONOMIC REVIEW staff, is a composite rating of 16-20 aspects of countries' policies and insti­ tutions. 5 It is available for most countries in the sample and over a long period, from 1977 on. Another index of countries' policies (that of Burnside and Dollar 2000) is also used, for robustness. Studies have found that small countries get more aid per capita (for example, Alesina and Dollar 2000). This could happen for a variety of reasons. Small countries tend to be more open, and thus more vulnerable to external shocks, motivating more aid flows. Also, poor but large economies may have more opportunity to borrow in private capital markets, due to some economies of scale, making them less reliant on aid. Small countries may receive aid for pol­ itical economy reasons, say because they have disproportional representation in international organizations (for example, aid may be used to buy a favorable vote in the United Nations; see Kuziemko and Werker 2006). More generally, small countries are more easily swayed for a given amount of aid. A reduction in the sensitivity of aid to size thus suggests a move away from political economy reasons for aid flows. This small country effect is investigated using (the log of) recipient country population, as is generally done in the literature. To investigate whether debt burden affects new aid, countries' debt stocks relative to exports are included. As in Chauvin and Kraay (2005, 2007), the present value of debt stocks is used, instead of the nominal value, since nominal debt stocks can be misleading under the highly concessional interest rates of officialloans. 6 To check whether aid allocations have changed over time relative to need and policy selectivity measures, the sample is divided into three subsamples, 1970-89, 1990-98, and 1999-2004. The first period is similar to that exam­ ined in earlier studies and coincides with the period before the fall of the Berlin wall. The post-Berlin wall era is split into two periods to check whether relationships have changed more recently. The break point, 1998, coincides roughly with the start of the new literature on aid effectiveness and major changes in the international aid architecture (for example, the World Bank aid study of 1998 and the launch of the HIPC Debt Initiative and the Poverty Reduction Strategy Papers framework). Most of the other control variables are also commonly used in this literature: bilateral trade flows, to control for non-aid-related economic relations between countries; net aid transfers provided by all other donors (in the sample) to the same country, to control for aid coordination and possible complementarity or substitution among aid donors and flows; and total net aid transfers ptovided by a donor to all recipient countries, to control for the donor country's overall level of aid generosity. Some of these controls are extensive and create a bias toward finding no significant results for the main variables. For example, 5. More details are at: http/lsiteresources.worldbank.orglIDNResources/ CPIA2006Questionnaire.pdf. 6. For technical background, see Dikhanov (2004). 5,/ . 4i" iaU t I MIU,., Ciaessens, Cassimon, Van Campenhout 193 including the net aid provided by other donors to the same country may already capture policy selectivity if better countries receive more aid in general, not just from the donor examined. Methodology The panel data have three dimensions: donor, reCIpIent, and time. A fixed effects model is a natural candidate for an empirical model that tries to explain bilateral aid flows between donor i and recipient j at time t, using a matrix of explanatory variables, a fixed donor effect, a fixed recipient effect, and time dummy variables (Baltagi 2001). Including fixed effects for both donor and recipient accounts for any time-invariant historical, geographical, political, cul­ tural, or other influence that will lead to deviations from average aid flows. It thus takes into account that, say, Tanzania receives more aid than other similar countries or that Denmark gives more aid than similar donors. The year dummy variables are included to control for general changes over time unre­ lated to policy selectivity, need, size, and debt burden (for example, differences in global economic or financial conditions that increase or reduce the need for aid). Last, all regressions are estimated with standard errors adjusted for clus­ tering on the bilateral relationships. This three-way fixed effects model does not capture bilateral interactions, however. For example, if Denmark gives more aid to specific countries than other donors do or if some recipient countries receive less aid from specific donors, this would not be controlled for. For this reason pair-wise donor-reci­ pient dummy variables are also included? This also controls for former colo­ niallinkages, which is important since former colonial powers have been found to give more aid to their former colonies (Alesina and Dollar 2000). These dummy variables also control for the degree to which a recipient country can be considered geopolitically linked to a donor. Geopolitical links and other political motivations can drive aid flows, as when aid is given to induce favor­ able votes in the United Nations. As long as links are time-invariant, these dummy variables also cover specific strategic donor-recipient links, such as the United States and Egypt or Pakistan. Eliminating time-invariant effects in the fixed effects model is costly, however, because of the interest in how some marginal effects change over time. For example, do countries with certain characteristics-such as low levels of income per capita, which can be a slow moving, almost time-invariant fixed factor-receive more or less aid over time? To investigate such changes, the coefficients on the four variables of interest are allowed to change over time. Specifically, the four aid-determining variables-poverty (per capita income), 7. This leads to a model with fixed effects for donors, recipients, and donor-recipient pairs (and the time effects). Egger and Pfaffermayr (2003) show that this generalization of the three-way model is identical to a two-way model with only time and bilateral effects, which is what is estimated in this article. 194 THE WORLD BANK ECONOMIC REVIEW policy (ePIA), small country effect (population), and debt burden (present value of debt to exports)-are interacted with dummy variables for each of the three periods to capture structural breaks. This is done in one regression, thus keeping the coefficients for the other fixed effects (for donor, recipient, and bilateral) and for the other independent variables constant across the three periods. 8 This way, changes in each of the four relationships are analyzed con­ currently over time (there may have been changes in several dimensions over the same period) while keeping other factors constant. 9 A random effects model could be used instead of a fixed effects model. Rather than absorbing any time-invariant individual specific effects, the random effects model assumes that all explanatory variables are uncorrelated with the individual specific effects. This is unlikely to be the case in the current application, making the fixed effects methods preferred from an economic intuition perspective. However, Hausman tests were also conducted to help decide what model to use. The test chose the fixed effects over the random effects model. Nevertheless, the random effects panel regressions are also reported for robustness. One other issue facing all aid (as well as trade) studies is that for many donor-recipient country combinations aid flows are zero (in two-thirds of the current sample). This can introduce a selection bias,lO which can be accounted for by conducting a Tobit analysis or by first estimating a probit model to predict the chance of observing nonzero aid flows and then by including in a second regression the Heckman inverse Mill's ratio thus obtained. Or one can use only nonzero (or only positive) observations in a simple ordinary least squares (OLS) regression. Berthelemy and Tichit (2004) and Berthelemy (2006), using a three­ dimensional panel, show that for aid flows the differences are small between 8. An alternative would have been estimating the regressions separately for each period, but this would mean that the coefficients for the recipient country dummy variables and other fixed effects would be allowed to vary by period as well. This has large costs since the changes allowed in the fixed effects on a recipient, donor, and bilateral country basis are likely to capture some of the change behavior of interest with respect to income, policy, debt, and population. For example, the relationship between India and the United States might have changed over time for geopolitical reasons and for reasons of improved policies in India, but because the bilateral dummy variables could capture these changes, it would not be possible to differentiate between the two reasons. Also, because there are no specific predictions as to changes over time in the other control variables, it would be undesirable to allow them to vary over time. Keeping the other fixed effects and the other independent variable constant across periods makes it easy to test directly for the significance of the differences in period coefficients using a simple F-test. 9. Additionally, the four variables were interacted with year-by-year dummy variables to analyze the year-by-year evolution in sensitivities, which provided qualitatively similar results (see working paper version available online at http://ssrn.comlabstract=997833 and h!tJ:>:llwber.oxfordjournals.Ql1l{). 10. This can happen on the donor side if little is known about a recipient country and it therefore gets no aid and on the recipient side if there is no government interest in engaging with that parricular donor. In either case, no aid is being extended, but treating these observations as zero aid could bias the results. ., g, •• ! ,,9 ' Ai J" " 1; Claessens, Cassimon, Van Campenhout 195 fixed effects using nonzero observations only and Heckman, Heckman two­ stage using all observations, random effects, and OLS. Also, the trade literature has shown that zero flows do not have much impact on estimation results (Baldwin 1994; Frankel 1997). Testing explicitly for biases in trade flows using various techniques, Linders and de Groot (2006, abstract) conclude that "in the end, the results surprisingly suggest that the simplest solution, to omit zero flows from the sample, often leads to acceptable results, although the sample selection model is preferred theoretically and econometrically.» An intermediate approach is presented, however, that distinguishes a case where a donor never provided aid to a recipient from that where a donor pro­ vided aid, but not every year. Specifically, donor-recipient combinations with zero bilateral flows for the whole period are excluded, since it is more likely in these cases that a selection was made by the donor or recipient. Country pairs with zero observations that record nonzero aid flows at any time are retained, however, because for these pairs there is no (or less of a) selection issue. These donors might not have disbursed aid to the particular country every year due, perhaps, to the lumpiness of projects or the peculiarities of decision processes. This decision is based on the grant component of aid flows since the net debt components can have nonzero flows due to repayments, even when there is no active engagement by the donor in a specific year (debt repayment may con­ tinue long after a country graduates from aid dependency). This seems a more robust way of running the regressions. Nevertheless, all the regressions are run with all observations and with nonzero observations only, and the results are reported in the base regression in the robustness tests. II. EMPIRICAL RESULTS This section discusses the empirical results of the aid allocation analysis. It first provides some descriptive statistics and stylized facts and then presents a detailed discussion of the results for the main and robustness specifications. Descriptive Statistics Figure 1 shows the evolution of bilateral net aid transfers over time, measured in 2000 U.S. dollars, disaggregated by grant, loan, and debt relief components on a recipient country per capita basis.ll Net aid transfer increased in the 1980s, dropped in the mid-1990s, and recovered somewhat after about 1998, although total aid per capita in 2004 was still below the early-1990s peak 11. Grants are total bilateral grants, net of debt forgiveness grants. Loans equal net loan transfers (corrected for offsetting entries on debt relief), including interest payments, but net of interest payments forgiven. Debt relief sums debt forgiveness grants (net of offsetting entries on debt relief) and interest payments forgiven. Offsetting entries on debt relief are the amortization part of debt forgiveness and must be deducted to avoid double counting of amortization forgiveness in official development assistance (now and in future years). See IMF and World Bank (2007, box 4.1, p. 153) for further details on DAC debt relief accounting. 196 THE WORLD BANK ECONOMIC REVIEW Figure 1. Recipient Country per Capita Bilateral Net Official Development Assistance Transfers, 1970-2004 (2000 U.S. dollars) • Loans II Grants o Debt relief 1970 1973 1976 1979 1982 1985 1988 1991 1994 1997 2000 2003 Note: The values are weighted averages. They are the sum of all aid flows divided by the sum of all recipient countries' populations. Source: Authors' analysis based on Organisation for Economic Co-operation and Development (OECD) Development Assistance Committee Aid Statistics database (www.oecd.orgldadstats). in real terms. Overall, aid per capita remains within a fairly limited bound of $6-$8 per person for the whole period, with an outlier in the mid-1990s. The disaggregated aid data show that grants have replaced loans, with net loan transfers becoming negative in recent years. Debt relief largely accounts for the short-lived peak in aid volume in 1991 and for the most recent increase. Table 1 defines and provides some raw statistics for the variables. Data for all variables are available for 1970-2004, except for the CPIA data, which are available only from 1977 onward for most countries. Average net aid transfer was $2.40 per capita per year (in 2000 prices), but with large variations, from -$138 to $9,052 (aid to a small country that received a large amount of aid from a single donor in a single year). Excluding the cases with zero obser­ vations, average aid per capita per donor is $4. Of total net aid transfers, including the zero observations, grants per donor were the largest component, averaging $2.20 per capita per year, net loans per donor were $0.16 per capita .. t L Q 4- '-Ai t! J r Claessells, Cassimoll, Van Campellhollt 197 per year, and debt relief per donor was $0.04 per capita per year (not reported). For the explanatory variables, statistics are as expected and indicate large variations among countries. Recipients' GDP per capita (in 2000 prices) averages $3,800, varying from less than $500 to $23,266. The average population size is some 2.8 million, but the standard deviation is high, at 11.7 million. The smal­ lest country has 20,000 people and the largest (China) has 1.3 billion, and there are no countries in the segment between 300 million and 1 billion people. For this reason, population is used in log terms in the regressions. The average CPIA index is 3.46, but the index ranges from 0.72 to the CPIA maximum of 6. The debt burden, in present value terms, averages 182 percent of exports, but varies greatly as well. Total aid provided by other donors averages $35 per capita (of the specific recipient country) per year. Donors provide an average of $313 in net aid transfers per capita (of the specific recipient country) to all other countries in the same year. Bilateral donor-recipient country trade averages 2.1 percent of recipient country GDP, but again with large variation. Regression Results Table 2 presents the basic results. The sample includes 50,000 observations representing 2,384 specific donor-recipient combinations. Columns 1 and 2 present the fixed and random effects estimates for the whole period, keeping the coefficients for the four main variables constant. Columns 3 and 4 allow the coefficients for the four main variables to change for each of the three sub­ periods, again with fixed and random effects. 12 The discussion focuses on the fixed effects model, preferred by the Hausman (1978) specification test results. The random effects results are qualitatively similar for both the full sample and the three subperiods. The model with constant coefficients finds that the income level of the reci­ pient country matters (significant at the 1 percent level), with poorer countries receiving more aid. This suggests that donors do care about poverty. The size of the recipient country also matters, with larger countries receiving less aid per capita. On aggregate and over the whole period, donors are not taking into account the quality of the policy and institutional environment in the recipient country, as the CPIA is not significant. The total debt burden does not signifi­ cantly affect aid transfers, suggesting that neither concerns about debt over­ hang nor defensive lending drove aid flows over the whole period. Control variables show that the more aid a donor gives in general, the less it gives to any specific country, likely because the donor faces an overall budget constraint. And aid flows by one donor are positively affected by the aid of other donors (although the relationship is not statistically significant), hinting at complementarity among donors, possibly due to the signaling effects for the 12. The coefficients for the bilateral fixed effects and time dummy variables are not reported to save space. 198 THE WORLD BANK ECONOMIC REVIEW TABLE 2. Base Regression Results Base regression Period interactions (1) (2) (3) (4) Variable Fixed effects Random effects Fixed effects Random effects Lagged GDP per capita, 1970- -0.615'" -0.346'" 2004 (0.165) (0.0817) Lagged GDP per capita, 1970-89 -0.535*" -0.237'" (0.136) (0.0635) Lagged GDP per capita, 1990-98 -0.651'" -0.328'" (0.207) (0.0991) Lagged GDP per capita, 1999- -0.720'" -0.409'" 2004 (0.210) (0.107) Log (population), 1970-2004 -2.056 -0.815'" (1.345) (0.113) Log (population), 1970-89 -3.544** -1.194'" (1.506) (0.166) Log (population), 1990-98 -3.237** -0.893'" (1.518) (0.123) Log (population), 1999-2004 -3.020** -0.671' •• (1.535) (0.122) CPIA,1970-2004 0.0758 0.108 (0.0774) (0.0739) CPIA, 1970-89 -0.0344 0.0397 (0.105) (0.0927) CPIA, 1990-98 0.215 0.175 (0.135) (0.120) CPIA,1999-2004 0.919' , 0.772" (0.366) (0.314) Present value of debt, 1970-2004 -0.0141 0.0242 (0.132) (0.133) Present value of debt, 1970-89 -0.475' -0.436' (0.261) (0.259) Present value of debt, 1990-98 0.219 0.246 (0.154) (0.154) Present value of debt, 1999-2004 0.374 0.497 (0.376) (0.350) Net aid transfer other donors 5.197 7.830' -2.427 1.167 (4.001) (4.376) (3.559) (3.760) Donor sum of net aid transfers -0.294' -0.204 -0.290' -0.201 (0.164) (0.156) (0.163) (0.154) Lagged bilateral trade 12.07 20.57.... 11.79 20.31**' (8.881) (7.046) (8.S26) (6.979) Constant 34.70' 14.39*' • 49.27" 15.93'" I (20.S8) (1.924) (24.68) (2.101) Number of observations 49,S04 49,804 49,804 49,804 I Hausman test (stat and p-value) 0.000 98.97'" 0.000 131.14'" 'Significant at p < 0.1. I HSignificant at p < 0.05. I u'Significant at p < 0.01. Note: Numbers in parentheses are robust standard errors. Hausman specification test compares I the fixed effects and random effects. F-test results for differences in subperiod coefficients (fixed effects model, column 3): lagged GDP per capita (F-value: 1.S5; P = 0.16); log population (F-value: I 7.71; p = 0.005); CPIA (F-value: 3.13; p 0.04); and present value of debt (F-value: 2.30; p 0.09). Source: Authors' analysis based on data sources shown in table 1. I I & J. " n H t utili .! il. I! lUt t Claessens, Cassimon, Van Campenhout 199 quality of the recipient country policies or to better coordination. Donors give more aid to important trade partners (although the relationship is not statisti­ cally significant), perhaps because bilateral relationships are closer when trade is high or because donors tend to support (indirectly) their own exports to the recipient country.13 Looking at changes over time in the key relationships for the three subperiods shows an increase in the responsiveness of aid to recipient country income (in absolute terms) over the three subperiods, from -0.535 to -0.720 (all are highly significant). Although the F-test can reject only at the 16 percent level that these coefficients are different from each other, this is evidence that donors have become more focused on providing aid to the poorest countries rather than, say, to their political allies. The small-country bias has diminished over time, with the coefficient for population falling from - 3.544 to - 3.020. All these results are significant, and the F-test rejects (at the 0.05 percent level) that these coefficients are not different from each other. This decline in the small-country effect may reflect less interest by donors after the cold war to support small countries in, say, buying political favors such as votes in the United Nations. In general, it confirms the improvement in the quality of aid allocations. Aid becomes much more responsive to policy: the coefficient, negative and statistically insignificant in the first period, rises to 0.215 in the second period and to 0.919 and statistically significant in the most recent period. The F-test shows that the increase in sensitivity is statistically significant (at the 4 percent level). This confirms the growing sense that in recent years donors have deter­ mined their aid allocation much more on the basis of country policy and insti­ tutional environment. It also explains why the CPIA is not significant over the whole period, as aid becomes sensitive to policy and institutional environment only in the last period. The results also support the hypothesis that concerns among donors about countries' debt burdens have declined. Whereas in the early period, high debt deterred aid (the coefficient was negative and significant), in the two later periods aid was no longer negatively affected by recipient countries' debt burden, and the change in coefficients is statistically significant at the 10 percent level. This change is good news, revealing that debt burdens are no longer an obstacle to aid flows. It does not, however, say much about changes in defensive lending, which requires consideration of the composition of debt as well. 14 13. Bilateral trade can be scaled by donor GDP instead of recipient country GDP. A positive relation with aid could then be interpreted as evidence of strategic behavior and self-interest in aid allocation. Such relationships were commonly found for the 19705 and 19805. When the regression is rerun with trade scaled by donor GDP and with changes over time, a positive and statistically significant association is found in the first period as well. In the other periods, however, the coefficients are not significant and trend toward negative. 14. Regressions including debt composition variables shed more light on changes in defensive lending behavior by multilateral and bilateral creditors over time (see the online working paper version of this report). 200 THE WORLD BANK ECONOMIC REVIEW Robustness Tests Several robustness tests were conducted. First, the models are run using the policy index developed by Burnside and Dollar (2000) instead of the CPIA. The CPIA index is produced by World Bank staff and potentially suffers from endogeneity if staff adjust the CPIA to affect International Development Association (IDA) lending patterns-which are by design closely related to the CPIA scores-when there has been no real change in policies or institutional environment. This would lead to a false conclusion of increased selectivity. This bias would affect IDA flows most directly (not studied here), but not necessarily bilateral flows. The CPIA scores could also have been affected by the prospective lending behavior of other donors, with World Bank staff raising the CPIA scores for countries for which they expect more aid flows. While it is not clear whether these biases exist, and if they do, whether they have increased over time, they could nevertheless affect the regression results. To address this possibility, another policy index is used. The index developed by Burnside and Dollar (2000) and further described in Roodman (2007) uses three indicators of economic policy-the logarithm of 1 plus the inflation rate, budget balance as a percentage of GDP, and the Sachs-Warner (1995) trade openness variable (1, 0). The index is created using a linear combination of the three policy variables with weights of 6.85 for budget balance, -1.40 for inflation, and 2.16 for trade openness. The Burnside-Dollar index, while a more objective measure and less subject to biases, is not necessarily a better index than the CPIA. It does a poorer job of capturing the policy and institutional environment, since it is mostly outcome based. Thus inflation and the budget balance may change because of exogen­ ous shocks even when the policy and institutional environments do not. To overcome some of this variability, three-year averages are used for the three constituent indicators of the policy variable. Another disadvantage is that the data needed to create this index are not available for all countries, which reduces the sample for the regressions by about one-third, to some 33,000 observations and 1,644 specific donor-recipient combinations. When the regressions are run with the Burnside-Dollar indicator instead of the CPIA index and the coefficient is allowed to change by subperiod, the policy index is not statistically significant in the first period but becomes signifi­ cant in the second and third periods (table 3, column 1). This is similar to what happens using the CPIA index in the original regression (see table 2, column 3). The results for the other variables are different than for the CPIA, but this is due to missing observations in the Burnside-Dollar index, which make the samples different. When the regression is rerun with the CPIA index for only the subset of observations available for the Burnside-Dollar index (table 3, column 2), results are similar to those with the Burnside-Dollar index. Thus the differences between the results for the Burnside-Dollar ! t Wt. t I;;..,. TABLE 3. Robustness Tests (1) (2) (3) (4) (5) (6) Burnside-Dollar Excluding zero All Variable (2000) CPIA subset GMM Balanced observations observations Lagged GDP per capita, 0.331 *',* -0.357*"* -0.462*** -0.361 *** -0.849*** -0.480*** 1970-89 (0.0935) (0.101) (0.111) (0.105) (0.208) (0.123) Lagged GDP per capita, -0.306*** -0.326*** -0.660*** -0.354**" 1.010* ** -0.594*"" 1990-98 (0.0972) (0.110) (0.107) (0.116) (0.312) (0.191) Lagged GDP per capita, -0.270*** -0.331 *** -0.869*** -0.343*** 1.145**" -0.668*** 1999-2004 (0.0926) (0.108) (0.105) (0.113) (0.314) (0.196) Log (population), 1970-89 -0.581 -0.773 -9.767*** -1.799 -4.417* -3.355** (1.047) (1.107) (1.179) (1.392) (2.363) (1.370) Log (population), 1990-98 -0.356 -0.585 8.511 *** 1.721 3.869 -3.095** (1.059) (1.113) (1.176) (1.432) (2.408) (1.378) Log (population), 1999-2004 -0.213 -0.484 -7.860*** -1.576 -3.478 -2.919** (1.075) (1.127) (1.176) (1.421) (2.444) (1.392) (] Burnside- Dollar (2000), -0.0162 lr 1970-89 (0.0489) '" '" Burnside-Dollar (2000), 0.150* '" ;:!; ~ 1990-98 (0.0833) Q Burnside-Dollar (2000), 0.308' v. 1999-2004 (0.183) '" §" <:) CPIA, 1970-89 -0.0301 0.0689 -0.0395 -0.0950 -0.0220 .;:!; (O.0928) (0.0848) (O.0936) (0.155) (0.0970) ~ CPIA, 1990-98 0.0860 0.122 0.100 0.267 0.198 ;:!; (0.0818) (0.0999) (0.0741) (0.183) (0.127) Q 0.471 .. 0.527** 1.324**' ~ CPIA, 1999-2004 0.294 0.878** ~ (0.263) (0.208) (0.237) (0.498) (0.349) '" ;:: ~ Present value of debt, 1970-89 -0.569 -0.640 -0.0819 -0.910 -0.463 -0.395 <:) ~ (0.482) (0.519) (0.315) (0.628) (0.309) (0.245) tv (Continued) ...,. 0 N 0 N TABLE 3. Continued ...; ;: (1) (2) (3) (4) (5) (6) m Burnside-Dollar Excluding zero All >!! 0 Variable (2000) CPIA subset GMM Balanced observations observations " r Present value of debt, 1990­ 0.818** 0.598 0.0912 0.460 0.226 0.193 " '" ;.­ 1998 (0.409) (0.372) (0.140) (0.452) (0.209) (0.138) z ~ Present value of debt, 1999­ 1.052· 0.829* -0.241 0.642* 0.353 0.308 m () 2004 (0.595) (0.454) (0.515) (0.383) (0.464) (0.360) 0 Net aid transfer other donors 6.124** 6.690** 3.422** 10.25 ..... 2.535 2.030 z 0 (2.997) (2.891) (1.477) (2.956) (6.045) (3.190) ::: Donor sum of net aid transfer -0.262 -0.263 -0.137** -0.416** -0.308" -0.294· () (0.163) (0.163) (0.0549) (0.171) (0.176) (0.163) " m < Lagged bilateral trade -24.61** - 24.40"· 15.73*** 9.736 9.361 11.64 ;:;; (11.16) (11.12) (1.418) (7.553) (9.202) (8.745) >!! Lagged net aid transfer per 0.495"** capita (0.00492) Constant 5.447 8.913 -1.053*'" 31.09 57.86 47.19** (18.16) (19.01) (0.0824) (22.41) (39.84) (22.22) Number of observations 33,401 33,401 47,219 28,672 37,510 53,090 Number of donor-country 1,644 1,644 2,380 1,024 2,316 2,566 combinations ·Significant at p < 0.1. HSignificant at p < 0.05 . .... • Significant at p < 0.01. Note: Numbers in parentheses are robust standard errors. F-test results for differences in subperiod coefficients (Burnside-Dollar model, column 1): lagged GDP per capita (P-value: 1.08; p = 0.34); log pop (P-value: 3.27; p = 0.04); Burnside-Dollar (P-value: 1.84; p = 0.16); and present value of debt (P-value: 2.26; p = 0.10). Source: Authors' analysis based on data sources shown in table 1. Claessens, Cassimon, Van Campenhout 203 indicator (see table 3, column 1) and the CPIA index (see table 2, column 3) are likely attributable to the different samples. Because there can be important dynamics in aid determination, the next regressions include a lagged dependent variable and estimate the coefficients as a dynamic panel using first-differenced general method of moments (GMM) (Baltagi 2001). For example, because of stickiness in the adjustment of aid pol­ icies, aid flows in this period may relate to those in previous periods, even though country policies and other circumstances have changed. Also, aid pro­ jects may involve lumpy disbursements, leading to autocorrelation. The results are very similar to those in the basic regressions: the coefficient for the recipient country income level increases over the three periods, while those of population size decrease in both size and statistical importance (see table 3, column 3). However, the CPIA indicator is again statistically signifi­ cant only in the third period. The other control variables-aid of all the other donors in the sample, total aid from the same donor, and bilateral trade-also have the same sign and significance. The main difference is that the debt burden is now insignificant for all three periods. This robustness test again sup­ ports the conclusion that donors have become more selective. Another robustness test uses a balanced sample in the regressions, since the unbalanced sample used so far may have biases arising from its change in com­ position over time. For example, some poor transition economies in Eastern Europe entered the sample in the later periods; other countries "graduated" and received less aid as they became less poor. These factors may introduce some bias. Transition economies were relatively rich even while receiving aid, and graduating countries may have had better policies while receiving less aid over time for other reasons. Other countries could have data deficiencies, which could also bias the sample. Running balanced samples for the base regressions lowers the statistical sig­ nificance, in part because the sample size is much smaller (see table 3, column 4). However, most of the variables of interest (except sensitivity to income) have the same signs as in the base regression. Aid responsiveness to recipient GDP per capita is constant across periods, and while the coefficient on the policy variable becomes more positive over time, it is never significantly differ­ ent from zero. The same holds for size. Although the country size bias seems to diminish over the three periods, the individual coefficients do not differ signifi­ cantly from zero. Debt becomes statistically significant in the third period. The signs and statistical significance of the other control variables are similar to those in the base regression. To check the robustness of the treatment of observations with zero aid flows (dropping the cases in which a donor provided no aid to a recipient but keeping those in which a donor provided aid only in some years), all the regressions were also run with only the nonzero observations (37,510) and with all observations (53,090). The results are consistent with those of the base case (49,804 observations). The importance of GDP per capita in aid allocation 204 THE WORLD BANK ECONOMIC REVIEW rises in both regressions, while that of size declines (see table 3, columns 4 and 5). The CPIA index becomes statistically significant only in the third period, while the debt burden is not statistically significant in any period. Overall, and consistent with findings in the trade literature, these results make clear that the treatment of zero observations does not alter the main conclusions. The model was also estimated with a lagged dependent variable, but using a fixed effects model instead of GMM. Again, the results are qualitatively similar. The results also remain qualitatively the same in other robustness regressions-in terms of specific samples and using Hausman-Taylor regressions. (For more details on these robustness tests, see the appendix and the working paper version available online.) Finally, the panel regression results are dependent-in terms of their statisti­ cal advantages over other regression techniques-on a certain degree of data homogeneity. With much heterogeneity, the panel approach offers little gains and possibly some costs. Homogeneity can be considered in all three dimen­ sions: over time, across donors, and across reCIpIents. The base analysis investigated the time dimension, showing changes over time in how the key variables drive aid flows. It is easier to investigate donor homogeneity, since there are fewer donors than recipient countries. This was done by running the aid allocation regressions for groups of similar donors, such as the so-called like-minded donor group (including the Nordic countries, the Netherlands, and the United Kingdom) compared with the others. To test reci­ pient heterogeneity, aid allocation regressions were run for groups of similar recipients, by income level and region (Sub-Saharan Africa and other countries). Each time, two groups of countries were created and the coefficients were compared. Most results confirmed the base panel results, although generally with reduced statistical significance. The important exception is that the variables used to group countries are not as significant, which is to be expected. For example, when recipients are grouped by income level, income becomes less significant. (These results are available online in the appendix.) Changes over Time among Donors This study has documented a general improvement in aid allocation. Can it identify changes for individual donors that have contributed to this improve­ ment? Recent research (for example, BertheIemy 2006; Dollar and Levin 2006; Wood 2008) has highlighted differences among donors, with some donors behaving more altruistically and others focusing more on their geopolitical interests. The general impression is also that donors vary in how much they have improved the selectivity and quality of their aid. Whether these differences exist and whether they have changed over time-and if so, for which donors­ can also be analyzed within the study framework by estimating the elasticities of individual donors with respect to the four key selectivity measures. This is done within the panel approach, keeping all control variables the same for all " UP ilL_ .•IL _ iUI Ciaessens, Cassimon, Van Campenhout 205 TABLE 4. Donor-specific Sensitivities to Recipient Country Variables (average of three periods) Memo items Ratio of aid to Lagged Present GNI,1977­ GDP per Log value of 2004 (average Share of Donor country capita (population) CPIA debt percent) total aid Australia -0.38 -4.19 0.79 0.79 0.37 0.750.79 Austria -0.08 -1.51 0.22 0.24 0.21 0.590.24 Belgium -0.19 -1.52 0.29 -0.13 0.45 0.62-0.13 Canada -1.02 2.52 -0.32 -1.21 0.40 0.66-1.21 Denmark -0.20 -1.00 0.30 0.13 0.89 0.570.13 Finland -0.03 -0.89 0.18 0.18 0.38 0.580.18 France -2.01 12.53 1.57 -2.09 0.50 0.80-2.09 Germany -0.92 -6.74 0.71 -1.16 0.36 0.64-1.16 Greece 0.00 -1.20 0.14 0.16 0.17 0.390.16 Ireland -0.03 -1.05 0.13 0.46 0.25 0.480.46 Italy 0.39 1.91 -0.49 1.33 0.22 0.441.33 Japan 0.20 -1.22 1.22 2.03 0.27 0.722.03 Luxembourg -0.11 -2.82 -0.13 0.11 0.38 0.500.11 Netherlands -0.02 -1.68 0.01 -0.17 0.88 0.70-0.17 New Zealand -0.01 -0.78 0.18 0.12 0.27 0.780.12 Norway -0.53 -2.39 0.21 0.50 0.97 0.640.50 Portugal -0.05 0.65 0.42 -0.23 0.21 0.63-0.23 Spain -0.36 -0.51 0.70 0.82 0.19 0.650.82 Sweden -0.60 -4.88 0.10 0.61 0.86 0.690.61 Switzerland -0.13 -1.08 0.23 0.28 0.31 0.710.28 United Kingdom -2.24 16.42 0.97 1.60 0.33 0.58-1.60 United States -3.08 -5.21 -0.38 0.Q3 0.18 0.740.03 1970-89 -0.44 3.00 -0.03 0.44 1990-98 -0.53 -2.83 0.17 0.20 1999-2004 -0.59 -2.64 0.81 0.38 Average of -0.52 -2.82 0.32 0.05 individual donor coefficients for all periods Note: Results of regressions using the base regression model specification, but with donor specific, time period varying interactions (table 2, column 3). Source: Authors' analysis based on data sources shown in table 1. donors, but allowing the coefficients for each donor to differ and to vary over the three time periods. Large differences remain among donors (table 4). For GDP per capita, average sensitivity varies from - 3.08 for the United States to 0.39 for Italy, suggesting that aid from the United States is much more geared toward the poorest countries than is aid from Italy. For population, the sensitivity varies from -16.4 for the United Kingdom to 2.52 for Canada, suggesting that aid 206 THE WORLD BANK ECONOMIC REVIEW from the United Kingdom is more geared toward smaller countries than is aid from Canada. For the CPIA index, sensitivity varies between -0.49 for Italy and 1.57 for France, making France much more policy sensitive than Italy. Finally, for debt burden, average sensitivity varies from - 2.09 for France to 2.03 for Japan, suggesting that debt is more detrimental to aid flows for France than it is for Japan. While not all these coefficients are statistically significant and the results do not always correspond to general perceptions, the results do show large differ­ ences among donors. Table 4 also shows the relative importance of aid as a share of donor GNI and bilateral aid as a share of a specific donor's total aid, to determine whether donors who give more aid also tend to be more selective (which does not appear to be the case). The results also confirm the general improvement in selectivity, with average sensitivity for the 22 donors for the three periods showing an increase with respect to income and a sharp increase with respect to the CPIA index, less bias toward smaller countries, and less concern over debt burdens. (The magnitudes of the average of the individual donor coefficients are very similar to those in table 2, columns 3 and 4.15) III. CONCLUSIONS This study observed behavioral changes over time In actual aid flows toward what appear to be more optimal allocations across countries. Specifically, the roles of poverty and countries' policy and institutional environ­ ment increased while the effects of small size and the debt burden diminished. Most of these changes occurred in the 1990s and intensified in the more recent period. While these changes likely relate in part to reforms of the international aid architecture, it is unclear which institutional changes at the international or bilateral level have driven the changes in behavior. Long-standing multilateral financial institutions-such as the International Monetary Fund, World Bank, Paris Club, and consultative group meetings-have introduced many changes, which likely have affected the behavior of bilateral aid flows. More attention has also been paid to aid allocation beginning in the late 1990s, in part due to research begun in the mid-1990s. And changes such as the HIPC Debt Initiative and the Poverty Reduction Strategy Papers process diminished the influence of debt on donor flows and increased donor selectivity. While these and numerous other changes all likely influenced aid flows, studies, including this one, have not been able to document specific evidence of their impacts. 15. These regression results hold for most donors with respects to the CPIA index, with sensitivities higher in the late 19905 than before (results available in the working paper version). For the need dimension (GDP per capita), progress is less obvious as the increase in coefficients (in absolute value) is less consistent across donors, and significant differences remain. h I -t!I!!I:!I'lIt4 1\11". __ 4& ffil' n Claessens, Cassimon, Van Campenhout 207 Further precision in the institutional factors driving changes in behavior is important for understanding how to make the international aid system work better for developing countries, The constraint is the lack of good measures of changes in such factors as financial policies, transparency, and coordination at the donor country and international leveL Work on documenting institutional changes in a rigorous and quantitative way may help identify the most influen­ tial changes. However, this study observes-as other have-large remaining differences among donors in revealed selectivity that appear to be related to donors' institutional environments, This suggests that reforms will have to be multifaceted and include further changes to the political economy and account­ ability in donor countries as well. It would be desirable for future research to take into account the policy and institutional environment not only in recipient countries, but also in donor countries, and to consider how this affects selectivity. REFERENCES Alesina, A., and D. Dollar. 2000. "Who Gives Foreign Aid to Whom and Why?" Journal of Economic Growth 5(1):33-63. Baldwin, R.E. 1994. Towards an Integrated Europe. London: Centre for Economic Policy Research. Baltagi, B.H. 2001. Econometric Analysis of Panel Data. 2nd edition. Chichester, UK: John Wiley & Sons Ltd. Berthelemy, J.-C. 2006. "Bilateral Donors Interest vs. Recipient's Development Motives in Aid Allocation: Do All Donors Behave the Same?" Review of Development Economics 10(2):179-94. BertheJemy, l-e., and A. Tichit. 2004. "Bilateral Donors Aid Allocation Decisions: A Three-dimensional Panel Analysis." International Review of Economics and Finance 13(3):253-74. Birdsall, N., S. Claessens, and I. Diwan. 2003. "Policy Selectivity Forgone: Debt and Donor Behavior in Africa." World Bank Economic Review 17(3):409-35. Burnside, e., and D. Dollar. 2000. "Aid, Policies, and Growth." American Economic Review 90(4):847-68. Chang, e.G., E. Fernandez-Arias, and L Serven. 1999. "Measuring Aid Flows: A New Approach." Policy Research Working Paper 2050. World Bank, Washington, D.e. Chauvin, N., and A. Kraay. 2005. "What Has 100 Billion Dollars Worth of Debt Relief Done for Low-Income Countries?" UnpUblished Manuscript, World Bank (http://econ.worldbank.orglstaffl akraay). - - . 2007. "Who Gets Debt Relief?" JOl/rnal of the European Economic Association 5(2-3): 333-42. Dikhanov, Y. 2004. "Historical Present Value of Debt in Developing Economies: 1980-2002." World Bank, Development Data Group, Washington, D.e. Dollar, D., and V. Levin. 2006. "The Increasing Selectivity of Foreign Aid, 1984-2003." World Development 34(12):2034-46. Easterly, W. 2003. "Can Foreign Aid Buy Growth?" Journal of Economic Perspectives 17(3):23-48. -~~-. 2007. "Are Aid Agencies Improving?" Economic Policy: 22(52):633-78. Easterly, W., R. Levine, and D. Roodman. 2004. "Comment: Aid, Policies, and Growth." i\merican Economic Review 94(3):774-80. 208 THE WORLD BANK ECONOMIC REVIEW Egger, P., and M. Pfaffermayr. 2003. "The Proper Panel Econometric Specification of the Gravity Equation: A Three-way Model with Bilateral Interaction Effects.» Empirical Economics 2S(3): '571-S0. Frankel, lA. 1997. Regional Trading Blocs in the World Economic System. Washington, D.C.: Institute for International Economics. Geginat, c., and A. Kraay. 2007. "Does IDA Engage in Defensive Lending." Policy Research Working Paper 432S. World Bank, Washington, D.C. Hausman, lA. 1975. "Specification Tests in Econometrics." Econometrica 46(6):1251-71. IMF (International Monetary Fund). Various years. Direction of Trade database. Washington, D.C.: International Monetary Fund. www.imfstatistics.orgIDOT. IMF (International Monetary Fund) and World Bank. 2007. Global Monitoring Report 2007: Millennium Development Goals: Confronting the Challenges of Gender Equality and Fragile States. Washington, D.C.: IMF and World Bank. Kuziemko, I., and E. Werker. 2006. "How Much Is a Seat on the Security Council Worth? Foreign Aid and Bribery at the United Nations." Journal of Political Economy 14(6):905-30. Unders, Gert-Jan M., and Henri L.F. de Groot. 2006. "Estimation of the Gravity Equation in the Presence of Zero Flows." Tinbergen Institute Discussion Paper 2006-07213, Free University of Amsterdam, The Netherlands. Marchesi, S., and A. Missale. 2004. "What Does Motivate Lending and Aid to the HIPCs?" University of Milan, Centro Studi Luca d'Agliano Development Working Paper No. lS9. OECD (Organisation for Economic Co·operation and Deveiopment)IDAC (Development Assistance Community). Various years. Aid Statistics database. Paris: Organisation for Economic Co-operation and DeveiopmentfDeveiopment Assistance Community. www.oecd.orgldadstats. Rajan, R. G., and A. Subramanian. 200S. "Aid and Growth: What Does the Cross-Country Evidence Really Show?" Review of Economics and Statistics 90(4):643-65. Radelet, S. 2006. "A Primer on Foreign Aid." Working Paper 92. Center for Global Development, Washington, D.C. Roodrnan, D. 2005. "An Index of Donor Performance." Working Paper 67. Center for Global Development, Washington, D.C. - - - . 2007. "The Anarchy of Numbers: Aid, Development, and Cross-Country Empirics." The World Bank Economic Review 21(2):255-77. Sachs, Jeffrey D., and Andrew Warner. 1995. "Economic Reform and the Process of Global Integration." Brookings Papers on Economic Activity 1:1-11S. Sundberg, M., and A. Gelb. 2006. "Making Aid Work." Finance and Development 43(4):14-17. Thomas, V., A. Chhibber, M. Dailami, and J. de Melo. 1991. "Restructuring Economies in Distress: Policy Reform and the World Bank." New York: Oxford University Press for the World Bank. World Bank. 1995. "Assessing Aid: What Works, What Doesn't, and Why." New York: Oxford University Press. - - - . Various years. Global Development Finance. Washington, D.C.: World Bank. - - - . Various years. World Development Indicators database. Washington, D.C.: World Bank. - - - . Various years. "Country Performance Rating." Washington, D.C.: World Bank. http:// go. worldbank.orglALSSDP3T90. Wood, A. 200S. "Looking Ahead Optimally in Allocating Aid." World Development 36(7):1135-51. !!!lIt II i« It. W.lf$?U UI4 tL_' Does Education Affect HIV Status? Evidence from five African Countries Damien de Walque Data from the first five Demographic and Health Surveys to include HIV testing for a representative sample of the adult population are used to analyze the socioeconomic correlates of HIV infection and associated sexual behavior. Emerging from a wealth of country relevant results, some important findings can be generalized. First, succes­ sive marriages are a significant risk factor. Second, contrary to prima facie evidence, education is not positively associated with HIV status. However, schooling is one of the most consistent predictors of behavior and knowledge: education level predicts protective behaviors such as condom use, use of counseling and testing, discussion of AIDS between spouses, and knowledge about HIVIAIDS, but it also predicts a higher level of infidelity and a lower level of abstinence. JEL codes: 112, 012, 015 The HIVIAIDS epidemic is one the greatest challenges facing Africa. According to UNAIDS (2007), in 2007 20.9 million to 24.3 million people were infected with HIVIAIDS in Sub-Saharan Africa (about 67 percent of the world total), 1.5 million to 2.0 million died from the virus, and 1.4 million to 2.4 million became newly infected. The socioeconomic profile of the HIV/AIDS epidemic has been analyzed in the epidemiologic and the economics literature. Few of these studies used nation­ ally representative samples. Data sets that include the results of individual HIV tests are generally drawn from cohort studies limited to a specific area (Nunn and others 1994; de Walque 2003, 2007a; de Walque and others 2005), from surveillance data collected from pregnant women attending antenatal care clinics (Fylkesnes and others 1997; Kilian and others 1999), or from high-risk groups (Nagot and others 2002). Some of these data sets have only a limited number of sociodemographic variables, and most cannot claim to be representative. There are, however, a few studies that use more representative data sets. Fylkesnes and others (2001) compare results from surveillance data among pregnant women and from population-based surveys in Zambia and, Damien de Walque (corresponding author) is an economist in the Development Research Group at the World Bank; his email address is ddewalque@worldbank.ocg. THE WORLD BANK ECONOMIC REVIEW, VOL.23, No.2, pp. 209-233 doi:l0.l093/wberllhpOOS Advance Access Publication June 16,2009 © The Author 2009. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 209 210 THE WORLD BANK ECONOMIC REVIEW in both data sources, observe trends showing large declines in HIV prevalence associated with higher educational background and stable or rising prevalence associated with low education. This study goes further by using data from the first five Demographic and Health Surveys to include data on HIV testing for a nationally representative sample of the adult population. The data sets are from Burkina Faso (2003), Cameroon (2004), Ghana (2003), Kenya (2003), and Tanzania (2003-04), five African countries with HIVIAIDS epidemics of different proportions. HIV prevalence is substantially higher in Cameroon, Kenya, and Tanzania than in Burkina Faso and Ghana. The five data sets have very similar variables, allow­ ing easy comparisons across countries. l They also include a large set of socio­ demographic variables and numerous questions about sexual behavior and other practices and attitudes related to the AIDS epidemic? These five data sets are used to analyze the socioeconomic determinants of HIV infection in the general population, looking at the association between HIV status and urban status, marital status, education, wealth, and religion. Also analyzed are the associations between these factors and a large range of sexual behaviors and other practices and attitudes related to the HIV/AIDS epi­ demic, allowing for a better understanding of the channels through which socioeconomic variables can affect HIV infection. The analysis also focuses specifically on the relationship between education and sexual behaviors and attitudes associated with HIVIAIDS. The detailed description of the results provides substantial information rel­ evant at the country level and gives a comparative view of the HIVI AIDS epi­ demic. While some differences across countries are to be expected, the results across countries are surprisingly consistent. Having similar information about five African countries for the same period permits using both pooled and country-level regressions to assess which of the results can be generalized and are broadly relevant for policymakers engaged in the fight against the epidemic. Successive marriages is found to be a significant risk factor, as evidenced by the association with HIV prevalence and sexual behavior, suggesting that specific prevention efforts should be targeted to this group. Further, and contrary to prima facie evidence, education is not positively associated with HIV status. But schooling is one of the most consistent predic­ tors of behavior and knowledge. Education predicts protective behaviors such as condom use, use of counseling and testing, discussion between spouses, and knowledge about HIVIAIDS, but it also predicts a higher level of infidelity and a lower level of abstinence. It is possible that these associations, moving in 1. Buve, Camel, and Hayes (2001) describe an interesting multicenter study of risk factors for HIV infection in four cities in different African countries. 2. Gersovitz (2005) provides a useful discussion of the variables describing sexual behavior in Demographic and Health Surveys. " !(d . , I. lLUL 11 de Walque 211 opposite directions, cancel each other out and that, as a consequence, edu­ cation is not significantly associated with HIV status. However, while it is diffi­ cult to isolate a significant relationship between education and HIV in the overall population, a negative association can be found in urban settings, at least in regressions that pool the data from the five countries. Section I describes the data sets and the methodology. Section II covers the analysis of HIV status. Section III focuses on the association between education and a large range of sexual behaviors and other attitudes related ro the epidemic. Section IV presents implications for further research and for policymakers. I. DATA DESCRIPTION AND METHODOLOGY The five data sets used are very similar: four of them (Burkina Faso, 2003; Cameroon, 2004; Ghana, 2003; and Kenya, 2003) are standard Demographic and Health Surveys that also include data on HIV testing for a subsample of the population (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; and Kenya and ORC Macro 2004). The 2003-04 HIVIAIDS Indicator Survey for Tanzania is a lighter survey that focuses on HIV/AIDS, but for the purposes of this study, the rel­ evant variables are very similar (Tanzania and ORC Macro 2005). The independent variables used in the regressions are almost always the same: urban location, marital status, including polygamy and successive mar­ riages, education, proxies for wealth and poverty, and religion. 3 Summary stat­ istics are in table 1. Not shown in the tables but included in the regression are dummy variables for age, region, and ethnicity (except for Tanzania, for which the ethnicity variable is not available). The share of the urban population is much higher in Cameroon and Ghana than in the other three countries (table 1). Educational achievement, measured by the highest grade achieved, is generally higher for men than for women and is much lower in Burkina Faso than in the other countries. Several variables describe marital status. The omitted category comprises individuals who have never been married. Marriage is defined as being legally married or living with a partner with the intention of staying together and therefore covers both formal and informal marriage. The formerly married category includes widowed, divorced, and separated individuals. The proportion of widows and widowers is calculated as the share of all formerly married individuals and should be understood in the regressions as an interaction term with that vari­ able. Being in a polygamous union is also calculated as a share of all currently married individuals and is used as an interaction term in the analysis. But the 3. A previous version of this study also included male circumcision and female genital mutilation in the regressions (de Walque 2006). Controlling for these variables does not significantly modify the coefficients on the other variables, indicating that there was no omitted variable bias due to the noninciusion of the circumcision variables. IV ..... IV .., m ~ 0 ;<> ,... 1;;) '" ;.­ z ~ m () 0 z 0 TABLE 1. Summary Statistics for Independent Variables s:: () Burkina Faso Tanzania (2003­ ;<> (2003) Cameroon (2004) Ghana (2003) Kenya (2003) 04) '" <: m Men Women Men Women Men Women Men Women Men Women ~ , '! Variable (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Urban 0.240 0.216 0.573 0.547 0.448 0.484 0.253 0.250 0.302 0.308 t I (0.030) (0.027) (0.026) (0.027) (0.027) (0.027) (0.024) (0.023) (0.029) (0.029) i Currently married 0.559 0.773 0.507 0.672 0.532 0.623 0.508 0.600 0.531 0.635 i i (0.010) (0.011) (0.009) (0.009) (0.009) (0.010) (0.009) (0.007) (0.009) (0.009) Formerly married 0.018 0.038 0.091 0.087 0.060 0.092 0.041 0.101 0.054 0.118 l (0.003) (0.002) (0.004) (0.003) (0.004) (0.004) (0.004) (0.004) (0.003) (0.004) ~ Widoweda 0.172 0.559 n.a. 0.319 0.097 0.209 0.156 0.414 n.a. n.a. ! • (0.061) (0.036) (0.018) (0.019) (0.019) (0.031) (0.023) \ More than one marriage 0.221 0.098 0.252 0.173 0.253 0.190 0.130 0.051 0.178 0.145 (0.010) (0.004) (0.010) (0.006) (0.007) (0.007) (0.007) (0.003) (0.007) (0.006) Polygamous· 0.294 0.483 0.129 0.305 0.128 0.227 0.098 0.186 0.098 0.097 (0.014) (0.012) (0.008) (0.011) (0.008) (0.012) (0.008) (0.009) (0.007) (0.008) Education (years) 2.62 1.39 7.06 5.61 7.75 5.89 7.93 7.12 6.20 5.36 (0.218) (0.144) (0.160) (0.181) (0.178) (0.167) (0.139) (0.137) (0.108) (0.115) I Earth floor 0.493 0.573 0.313 0.454 0.091 0.113 0.448 0.595 0.644 0.673 (0.025) (0.025) (0.017) (0.021) (0.008) (0.010) (0.019) (0.020) (0.024) (0.024) Muslim 0.577 0,600 0.177 0.180 0.187 0.155 0.064 0.075 0.298 0.306 (0.022) (0.020) (0.016) (0.017) (0.017) (0.015) (0.008) (0.009) (0.020) (0.020) Catholic 0.249 0.231 0.396 0.376 0.155 0.144 0.266 0.252 0.326 0.310 (0.018) (0.014) (0.014) (0.014) (0.009) (0.009) (0.012) (0.012) (0.017) (0.016) Protestant 0.041 0.051 0.304 0.327 0.592 0.664 0.602 0.650 0.264 0.290 (0.004) (0.005) (0.012) (0.013) (0.017) (0.017) (0.014) (0.014) (0.015) (0.016) Other religion b 0.132 0.116 0.121 0.114 0.112 0.070 0.066 0.021 0.110 0.092 (0.013) (0.012) (0.007) (0.010) (0.008) (0.006) (0.007) (0.003) (0.016) (0.013) n.a. is not applicable because data were unavailable and so the variable was not included. Note: Numbers in parentheses are clustered standard errors. The residual categories are rural and never married. The data are weighted with the sample weights given by the data provider. 'Widowed is a mean taken on formerly married individuals and polygamous is a mean taken on currently married individuals. bInciudes animists and no religion in Burkina Faso; animists, no religion, and other religions (religions de !'eveil) in Cameroon; traditionalists and no religion in Ghana; and no religion in Kenya and Tanzania. In Ghana, other Christians are included under Protestants. Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 2005). I} ~ ~ " N ~ (;J 214 THE WORLD BANK ECONOMIC REVIEW mean for the variable of having been in successive marriages (not to be con­ fused with polygamy) is taken on the entire sample and can apply to both cur­ rently and formerly married people. More women than men are currently married, which can be explained either by polygamy (polygyny) or by the age differences between spouses, with men generally marrying later than women. The survey includes women of ages 15­ 49 in all surveys and men of ages 15-59 for Burkina Paso, Cameroon, and Ghana; ages 15-54 for Kenya; and ages 15-49 for Tanzania. 4 Widowhood, defined as having lost one spouse and not being remarried, is not recorded in the Tanzanian survey and in the Cameroon survey only a very limited number of men are widowers. Widows and widowers constitute a sub­ stantial portion of formerly married individuals, and there are usually more widows than widowers, either because women have a longer life expectancy or because they marry older men or because it is easier for men to remarry after the death of a wife. A large share of people have been in successive marriages, ranging from 5.1 percent of women and 13 percent of men in Kenya to 19 percent of women and 25 percent of men in Ghana. The cause of successive marriages can be either divorce or the death of the spouse. The data do not allow distinguishing between causes. Individuals whose spouse died are defined as widows or widowers if they did not remarry or as having been in successive marriages if they did remarry. The fact that more men than women have been in successive marriages suggests that it is easier for men to remarry. There are important variations in the share of married individuals in polyga­ mous union, ranging from 48 percent of women and 29 percent of men in Burkina Paso to 9.8 percent of men and 9.7 percent of women in Tanzania. The proportion of individuals living in a house with an earth floor is used as a measure of poverty (table 1). The wealth quintiles provided in the Demographic and Health Surveys can be misleading as a measure of wealth, particularly for African countries. There is generally too little variation in the lowest three (sometimes four) quintiles, and differences in living standards are difficult to distinguish at those quintiles. Instead of the wealth quintiles, the analysis uses two proxies for wealth and poverty: a set of indicators for durable goods as a proxy for wealth and the presence of an earth floor as a proxy for poverty, with durable goods as instruments, letting the data pick the weights. The share of the population living in a house with an earth floor is highest in Tanzania and lowest in Ghana. Men are generally less likely to live in a house with an earth floor, but that difference is significant only in Cameroon and Kenya. 4. Regressions were also run limiting the age range for men to 15-49 to maintain parallelism with women. The results, available on request, are very similar to those for the wider age ranges for men included in this article. .- L @j !lUll. 4i !f n .• ~ u_. U J 4 lib! I de Walque 215 Religious affiliations have been regrouped into four categories: Muslim (the omitted dummy variable in the regressions), Catholic, Protestant, and other religions. Other religions include animists and no religion in Burkina Faso; ani­ mists, no religion, and other religions (religions de l'eveil) in Cameroon; tradi­ tionalists and no religion in Ghana; and no religion in Kenya and Tanzania. In Ghana, other Christians are included under Protestants. Methodology and Potential Sources of Bias Though common in the epidemiologic literature, this study does not enter sexual behaviors and other variables as controls in the HIV infection regression or in regressions with other behaviors as the dependent variable. In a cross­ section analysis, the estimates derived from such regressions would suffer from reverse causality and endogeneity. For example, condom use could prevent HIV infection (expected negative association), but people who are HIV positive might be more likely to use condoms to protect their partners (reverse causality going from HIV to condom use and potentially driving a positive association). Moreover, condom use, sexual activities, and other AIDS-related practices are all choice variables that suffer from self-selection. One might, for example, expect individuals engaging in risky sexual behaviors to be more likely to use condoms because of their higher exposure. Instead, separate regressions are run, first with HIV status as the dependent variable (tables 2 and 3) and then with sexual behaviors and others attitudes and practices related to HIV/AIDS epidemic as the dependent variable (table 4, and tables S3-S5 in the online supplemental appendix). Similar independent variables are entered for each set of regressions: age, location (urban and regional dummy variables), marital status, wealth, education, religion, and eth­ nicity. Age is included since HIV infection has a distinct hump-shaped profile, increasing until the 30-40 age range and decreasing thereafter (figures 1 and 2). Location variables account for the fact that the risk of HIV infection depends on HIV prevalence in an individual's sexual network, which is location specific. At the same time, location might determine access to preven­ tion messages and methods such as condom use. Ethnicity and religion can also shape the sexual network and at the same time may influence the type of sexual practices considered desirable or acceptable by an individual. Marital status, including polygamy and successive marriages, is also an important determinant of the current and past sexual network. Finally, education and wealth are included as they can influence the shape of the sexual network, by conferring status and income. Education and wealth might also provide better access to prevention messages and methods and stronger incentives to avoid HIV infection. s The regressions with HIV as the dependent variable can be seen as the reduced form of a behavioral model in which age, location, marital 5. For an economic model linking education to incentives to avoid HIV infection, see de Walque (2007a). .... N 0\ .., ;:: '" li'! 0 ;e ""' 0 TAB LE 2. Determinants of Hrv Prevalence in Five Demographic and Health Surveys: Pooled Regressions '" )­ Z All with Tanzania All without Tanzania Age 15-29 with Tanzania Urban with Tanzania ~ Men Women Men Women Men Women Men Women '" ('J 0 z Variable (1) Linear regression model with earth floor, as a proxy for poverty, instrumented by other durable assets a Years of education -0.0003 (0.0006) (2) 0.0004 (0.0007) (3) -0.0002 (0.0006) 0 (4) (0.0008) (5) -0.0007 (0.0006) (6) -0.0008 (0.0009) (7) -0.0025" (0.0011) (8) -0.0026** 0 il:: ('J ;e m <: I t (0.0011) 1 Earth floor 0.0055 0.0081 0.0019 -0.0133 0.0056 -0.0276 0.0076 0.0436 '" li'! \ (0.0140) (0.0203) (0.0119) (0.0227) (0.0154) (0.0245) (0.0341) (0.0383) Urban Currently married Formerly married 0.0309**<- (0.0081) 0.0061 (0.0064) 0.0494*"" 0.0371"** (0.0098) -0.0021 (0.0073) 0.1152*** 0.0186" " (0.0075) 0.0062 (0.0067) 0.0155 0.0237"* (0.0097) 0.0059 (0.0089) 0.0800"-" 0.0231*'" (0.0078) 0.0095 (0.0074) 0.0296- 0.0182 (0.0118) 0.0097 (0.0077) 0.1076 ... • n.a. 0.0047 (0.0146) 0.0665*** n.a. -0.0027 (0.0120) 0.1250.... I (0.0130) (0.0136) (0.0114) (0.0195) (0.0175) (0.0192) (0.0230) (0.0221) Widowed n.a. n.a. 0.1983** 0.1103"" n.a. n.a. n.a. n.a. (0.0883) (0.0314) More than one marriage 0.0188** 0.0428"" 0.0112 0.0349· ... 0.0084 0.0572·" 0.0139 0.0516"· (0.0077) (0.0076) (0.0075) (0.0086) (0.0154) (0.0142) (0.0136) (0.0141) Polygamous -0.0095 0.0156'· -0.007 0.0167"· -0.0075 0.0287*" 0.0157 0.0214 (0.0106) (0.0065) (0.0111) (0.0071) (0.0276) (0.0117) (0.0288) (0.0163) ! t Protestant -0.0022 -0.0064 0.0033 -0.0033 0.0012 -0.0017 -0.01 -0.0212" (0.0051) (0.0058) (0.0052) (0.0065) (0.0059) (0.0073) (0.0094) (0.0117) i Muslim -0.0011 (0.0061) -0.0053 (0.0069) 0.0017 (0.0056) -0.0133"* (0.0065) 0.0026 (0.0069) -0.0023 (0.0079) -0.0067 (0.0121) -0.0234" (0.0139) I j i 1 ~' Other religion -0.0087 -0.0238** 0.0033 -0.0177* 0.0129* -0.0207 -0.0193 -0.0448** (0.0069) (0.0097) (0.0072) (0.0099) (0.0069) (0.0129) (0.0158) (0.0220) Observations 19,986 23,085 15,209 17,408 11,007 13,486 6,602 7,893 R-square 0.05 0.07 0.05 0.08 0.04 0.07 0.08 0.09 Marginal effects of probit estimations, with durable asset dummy variables controlling for wealth b (not shown) Years of education -0.0003 -0.0002 -0.0002 -0.0003 -0.0005 -0.0008" -0.0012** -0.0027*** (0.0004) (0.0004) (0.0004) (0.0005) (0.0003) (0.0005) (0.0006) (0.0008) Assets (2( 10)-Test 31.66 41.34 12.76 15.50 8.51 29.19 18.13 14.86 Prob > r 0.0005 0.0000 0.2377 0.1149 0.5791 0.0012 0.0528 0.1372 .. Significant at the 10 percent level. **Significant at the 5 percent level. ...... Significant at the 1 percent level. n.a. is not applicable because ethnicity and widowhood data were unavailable for Tanzania and urban location is not relevant for regressions 7 and 8 focusing on the urban sample. Regressions 1-4 were therefore run with and without Tanzania. Note: Numbers in parentheses are robust and clustered standard errors. HIV prevalence is the dependent variable. Data from the five surveys have been pooled. Dummy variables controlling for age, country and region, and ethnicity (regressions 3 and 4) are also included. The omitted dummy vari­ ables are rural, never married, and Catholic (see note in table 1). The data are weighted with the sample weights given by the data provider, multiplied by the country population divided by the sample size. 'To control for wealth/poverty, earth floor is included in the linear regression specifications and instrumented by the other durable assets: type of latrine, type of floor (earth floor or not), electricity, refrigerator, radio, television, bicycle, motorcycle, or car. bAli asset dummy variables are included (coefficients not shown, but the results of an F-test of joint significance are reported). Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 2005). ~ ~ .JS' ~ "" - N "-J N ~ 00 -l J: m ,;J 0 !;O t"' C) TABLE 3. Determinants of HIV Prevalence in Five Demographic and Health Surveys: Analysis by Country '" » z Burkina Faso (2003) Cameroon (2004) Ghana (2003) Kenya (2003) Tanzania 2004 :>; m n Men Women Men Women Men Women Men Women Men Women 0 Variable (1 ) (2) (3) (4) (5) (6) (7) (8) (9) (10) z 0 ::: Linear regression model with earth floor, as proxy for poverty, instrumented by the other durable assets' n !;O Years of 0.0006 -0.0023 -0.0004 -0.0009 0.0001 0.0004 -0.0002 -0.0021 -0.0006 0.0009 m <: education (0.0014) (0.0014) (0.0010) (0.0018) (0.0006) (0.0008) (0.0015) (0.0020) (0.0017) (0.0017) m Earth floor 0.0065 -0.0112 0.0186 - 0.009 0.0072 0.045 0.0223 0.0472 0.015 0.0192 ~ (0.0171) (0.0249) (0.0125) (0.0291) (0.0272) (0.0491) (0.0169) (0.0387) (0.0431) (0.0360) Urban Currently married 0.016 (0.0131) 0.0276*** -0.0154 -0.0076 (0.0096) 0.0156 0.0141 (0.0157) (0.0086) (0.0109) (0.0109) 0.0332** (0.0152) 0.0094 (0.0133) 0.0021 (0.0048) (0.0064) 0.0104 (0.0082) 0.002 -0.00.59 (0.0089) 0.0326* (0.0193) 0.0013 (0.01.50) 0.018 (0.0225) 0.0033 (0.0172) 0.0603** (0.0259) 0.0156 (0.0136) 0.0592*** (0.0228) -0.0082 (0.0129) I I Formerly married Widowed 0.02.5 (0.0229) 0.0867 -0.0009 0.0296 0.0109 (0.0279) (0.0140) 0.0 0.0919*** (0.0248) 0.0109 (0.0118) 0.1302**':' -0.0224 0.0310* (0.0159) 0.00.56 0.0067 (0.027.5) 0.3070'" 0.1028*** (0.0388) 0.1396** 0.0900*"· (0.029.5) n.a. 0.1228*** (0.0205) n.a. i .. (0.1024) (0.0402) (0.0000) (0.04.51) (0.0271) (0.0290) (0.1287) (0.0.556) t More than one marriage Polygamous 0.0088 (0.0138) 0.0214" 0.01.55 -0.0004 (0.0104) (0.0094) -0.0091 -0.0023 0.040.5*** (0.0111) 0.0045 0.0137 (0.008.5) -0.017.5 0.0288*** (0.0090) 0.0137 0.0122 (0.0216) 0.0127 0.0.563 (0.0346) 0.0364* 0.0346** (0.0174) -0.0033 0.0600*** (0.0168) 0.0216 I t I (0.0130) (0.00.59) (0.0129) (0.0109) (0.0136) (0.0101) (0.03052) (0.0195) (0.0254) (0.0162) t Protestant 0.0278 -0.0002 -0.0011 -0.0068 0.0075 0.0097 0.0028 -0.0009 -0.0146 -0.0144 i I (0.0235) (0.0111) (0.0074) (0.0096) (0.0055) (0.0079) (0.0097) (0.0115) (0.0108) (0.0120) Muslim 0.0119 0.0006 0.0114 -0.0005 0.0008 -0.0006 -0.0072 -0.0521 .. -0.0136 0.001 (0.0079) (0.0076) (0.0119) (0.0169) (0.0074) (0.0098) (0.0228) (0.0289) (0.0119) (0.0140) Other religion 0.005 -0.0102 0.0109 -0.0256*" 0.0102 0.0204 0.0204 0.0106 -0.0323"* -0.0369* (0.0106) (0.0069) (0.0084) (0.0116) (0.0134) (0.0124) (0.0175) (0.0514) (0.0149) (0.0196) Observations 3339 4164 4997 5085 3959 4919 2914 3240 4772 5665 R-square 0.06 0.04 0.06 0.1 0.04 0.D3 0.11 0.12 0.05 0.08 Marginal effects of probit estimations, with durable asset dummies controlling for wealth (not shown)b Years of 0 -0.0006* -0.0003 -0.0012 0 0.0002 0 -0.0008 -0.0009 0 education (0.0002) (0.0003) (0.0008) (0.0010) (0.OO02l (0.0005) (0.0009) (0.0013) (O.OOll) (O.OOll) Assets? 28.07 18.92 17.29 14.67 21.29 11.92 11.79 9.63 26.15 42.88 (10)-Test Prob >? 0.0018 0.0413 0.0681 0.1447 0.0191 0.2904 0.2255 0.4739 0.0035 0.0000 Mean HIV prevalence Mean HIV 0.0194 0.0182 0.0391 0.0662 0.0162 0.0270 0.0463 0.0868 0.0626 0.0769 (0.0031) (0.0027) (0.0030) (0.0043) (0.0022) (0.0024) (0.0051 ) (0.0064) (0.0047) (0.0052) Significant at the "10 percent level. ""'Significant at the 5 percent level. ** "Significant at the 1 percent level. n.a. is not applicable because data were unavailable and so the variable was not included. Note: Numbers in parentheses are robust and clustered standard errors. HIV prevalence is the dependent variable. Controls for age, region, and ethni­ city are also included; ethnicity and widowhood are not controlled for in Tanzania 2004 as the variable were not available. The omitted dummy variables are rural, never married, and Muslim (see note in table 1). The data are weighted with the sample weights given by the data provider. "To control for wealth/poverty, earth floor is included in the linear regression specifications and instrumented by the other durable assets: type of latrine, type of floor (earth floor or not), electricity, refrigerator, radio, television, bicycle, motorcycle, or car. bAll asset dummy variables are included (coefficients not shown, but the results of an F-test of joint significance are reported). Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 2005). ~ ~ j '" .... N \0 TAB L E 4. Condom Use and Extramarital Sex: Pooled Regressions IV IV o All with Tanzania All without Tanzania Age 15-29 with Tanzania .., Men Women Men Women Men Women ::c m Variable (1) (2) (3) (4) (5) (6) 11! o Determinants of using a condom at the last intercourse with spouse (married sample) ',... o" Years of education 0.0026·.... 0.0034..... 0.0027· .... 0.0039" u 0.0076 ...... 0.0041 ...... Linear (0.0009) (0.0006) (0.0009) (0.0007) (0.0020) (0.0010) '" ;,­ ~ l'<: Observations 10440 23883 7671 19895 2474 11524 m R-square 0.04 0.04 0.05 0.05 0.08 0.06 C"> o Years of education 0.0031··· 0.0026·... 0.0028· .... 0.0026 ...... 0.0071 ...... 0.0029 ...... ~ o Pro bit, marginal effects (0.0006) (0.0003) (0.0006) (0.0003) (0.0017) (0.0006) :::: Determinants of using a condom at the last intercourse if not with spouse (if nonmarital sex) C"> i" Years of education 0.0188·'" 0.0149· .. • 0.0170 .... • 0.0131 ...... 0.0176 ...... 0.0182""" m <: Linear (0.0029) (0.0029) (0.0033) (0.0030) (0.0037) (0.0036) m Observations 5530 5454 4177 4485 4303 4219 ~ R-square 0.14 0.13 0.14 0.1 0.14 0.12 Years of education 0.0215··.. 0.0224·... 0.0202 ... • 0.0212 ...... 0.0236 ...... 0.0278 ...... Pro bit, marginal. effects (0.0033) (0.0028) (0.0040) (0.0029) (0.0035) (0.0034) Determinants of having nonmarital sex in the last 12 months (currently married) Years of education 0.0008 0.0028·.... 0.0001 0.0033 ...... -0.0059 0.0052 ..... Linear (0.0013) (0.0005) (0.0013) (0.0005) (0.0037) (0.0009) Observations 11980 28736 8959 24518 3021 13839 R-square 0.12 0.11 0.15 0.13 0.13 0.15 Years of education 0.0036·... 0.0012·.... 0.0024"· O.OO13u" -0.0019 0.0023 ...... Pro bit, marginal effects (0.0012) Significant at the "10 percent level. (0.0002) (0.0012) (0.0002) (0.0035) (0.0004) I n Significant at the 5 percent level. ** • Significant at the 1 percent leveL Note: Numbers in parentheses are robust and clustered standard errors. Controls for age (dummy variables), urban location, marital status, religion, I I region, and ethnicity are also included (ethnicity and widowhood are not controlled for in Tanzania 2004 as the variable were not available). To control for wealth/poverty, earth floor (coefficient not shown) is included and instrumented by the other during assets: type of latrine, type of floor (earth floor or not), electricity, refrigerator, radio, television, bicycle, motorcycle, or car. In the probit specification, all asset dummy variables are included (coefficients not shown). The data are weighted with the sample weights given by the data provider, multiplied by the country population divided by the sample size. ! de Walque 221 FIGURE 1. Age Profile of HIV Prevalence in Men in Five African Countries I.~ I­ - .",,:' --------+---r'--.-.-.-~-..-.---~->.:_---.'"-----..-~..-.-.-.~~---__! . '"I -' I .," I • I I I k---­ /' -~~-;;;.~.~~~~-~~~.-~~~ -~-' 15-19 20-24 25-29 35-39 45-49 50-54 55-59 Age group Faso _____ Cameroon - ~ - Ghana - -.- - Kenya' .••• ­ Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 2005). status, religion, ethnicity, education, and wealth influence the composition of the sexual network, sexual practices, and prevention measures, which in turn can have an impact on the risk of HIV infection. Most of the individual characteristics used as regressors, with the exception of age and ethnic background, cannot be defined as completely exogenous vari­ ables. Location, marital status, education, wealth, and even religion are, at least to some extent, choice variables for individuals or their family. The data set does not offer sources of exogenous variations for those variables. Coefficients should therefore be interpreted with caution and as associations rather than causal effects. In particular, marital status and marital history are clearly endogenous variables, and their inclusion as regressors could be ques­ tioned. For example, the positive associations between HIV and widowhood and successive marriages are very likely to be influenced by reverse causality. As a robustness check, a specification is proposed that excludes marital status and marital history (table 52, model 2, in the supplemental appendix). The main result of the analysis-the absence of a gradient between education and HIV infection-does not change. 6 6. There is one negative coefficient, for Kenyan women, which is statistically significant at the 10 percent level. 222 THE WORLD BANK ECONOMIC REVIEW FIGURE 2. Age Profile of HIV Prevalence in Women in Five African Countries 14,-------------------------------------------------------, 12 t--.-----~ / , ' " , , , .. , .•......... ----,"':'/.L- ... ,~.~ .... __ ___ S;;;;_,,:-., 10 t - - - -..-.-.. ~.~---c;-'-/_/__;7"""'~-."'.;;:::--.-~.~.~---=::~'::~.~--~-~- - '! \' . , ~'" .~ 8 --.--.-T-tlt"'----,L--.-~------- -\- . - - ­ ! \ " i! " \ .... t - 6+--·------,hL-. .·--~-------·------~""""'::::::~\~. .--j I \ ~ 4 - / ~ -*----,---... ,-,< ----\­ I 0'" .........."""--_.... ..... .... 1 • .. . - --~--,---- ... 21"'~' .:-:-~-~/_. ~ . ~. .~------~----.-.---___l o I -- , -~ I 15-19 20-24 25-29 30-34 35-39 40-44 45-49 Age group Faso --a-- Cameroon - -.- - Ghana - .....- - Kenya Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 2005). Sexual behavior and other practices are all self-reported. This is an obvious but inescapable limitation. Diverging reports on self-reported behaviors between husbands and wives (for example more married men report using a condom in marriage or discussing AIDS with their spouse than do married women) lead to the suspicion that some behaviors are not truthfully reported. Gersovitz (2005) discusses the issue of self-reporting sexual behaviors in the Demographic and Health Surveys and shows several inconsistencies, in par­ ticular regarding virginity and the age at first sexual intercourse. Some discre­ pancies in reported sexual behavior between men and women in condom use and number of partners, for example, can potentially be explained by the fact that the extramarital partners of men with a high intensity of sexual activity, typically commercial sex workers, are not included or are underrepresented in the survey. De Walque (2007b), however, shows that the reports of men and women in couples are not always mutually consistent, even when, in theory, they should be, as in the case of condom use during the last sexual intercourse between the two interviewed partners. Polygamy, which is frequent in the studied countries, might also explain some of the reported discrepancies between married men and women. The self-reported feature of the sexual behavior variable would bias the analysis, especially if the reporting bias varies with education. For example if because of perceived social desirability more edu­ cated individuals are more likely to report that they are using condoms-even if . . _ $ 4 .. liJii '* * , ! de Walque 223 they are not-this would create a bias toward a positive association between condom use and education. One dependent variable that is not self-reported is HIV status, which is determined by an HIV test on a blood sample. Including HIV testing is one of the great advantages of the new Demographic and Health Surveys. However, some individuals who had been assigned to the sample for HIV testing refused to be tested or were absent. If the absence of a test is not random, this could be a source of bias. Table S1 deals with this issue and shows that acceptance of the test is somewhat less likely in urban areas. However, coverage of the HIV test is usually high (at 82-95 percent). Antiretroviral treatment is currently scaled-up in the five survey countries. Although the data do not allow analysis of that issue (though table S5a indicates that educated and, to a lesser extent, richer people are more likely to use volun­ tary counseling and testing services), access to treatment is expected to be easier in urban centers and for educated and richer people. If access to treatment keeps those people alive while its absence implies that poorer and less educated individuals in rural areas are more likely to die, this would bias upward the coefficients on education, wealth, and urban location in a regression with HIV status as the dependent variable. This should be kept in mind in the analysis, even if only a small share of HIV positive individuals is receiving treatment. Indeed, only a small percentage of the HIV positive population is medically eligible for treatment. There is a long interval between HIV infection­ seroconversion-and development of AIDS. For adults in Uganda, the median time from seroconversion to AIDS has been estimated at 9.4 years (Morgan and others 2002). Antiretroviral treatment is recommended only for individuals at the AIDS stage (generally, with fewer than 200 CD4 cells per cubic millimeter). In addition, access to treatment programs is recent and available to only a small percentage of medically eligible patients. Estimates of the proportion of people who are HIV positive and are receiving treatment varied from 0.66 percent in Tanzania to 5.35 percent in Burkina Faso.? The number of people on treatment was probably even smaller in 2003 and 2004, when the data for this study were collected. The regressions use the sample weights provided in the Demographic and Health Surveys, 8 and the standard errors are clustered at the enumeration-area level. Tables 3 and S1-S5 present results with country-specific regressions. However, since the Demographic and Health Surveys are defined similarly in each country, it is also interesting to pool the data from the five countries in a single regression, increasing the power of the analysis. The results of pooled 7. The estimates are 4 percent for Cameroon and Kenya and 1.3 percent for Ghana. These figures are calculated by the author and are based on data on treatment coverage from June 2005 (WHO and UNAID5 2005 and Burkina Faso Government 2006). 8. Table 52 presents as a robustness check of results from unweighted regressions and shows very similar results. 224 THE WORLD BANK ECONOMIC REVIEW regressions are included in table 2 (for HIV status) and table 4 (for sexual be­ havior). One condition for pooling the data across surveys is that the coeffi­ cients on the interaction between a country indicator and the variable of interest not be significantly different across countries. In the pooled regressions, the data are weighted by the sample weight provided in the survey, multiplied by the country population and divided by the sample size for each survey. II. HIV STATUS For both genders and in almost all countries, the age profile for HIV prevalence is hump shaped, first increasing with age and then decreasing (figures 1 and 2). The peak of HIV prevalence is at older ages in the countries with low overall HIV prevalence, such as Burkina Faso and Ghana. That peak is generally earlier for women than for men, with the exception of Burkina Faso. This is explained by women's tendency to initiate their sexual activity earlier than men do (table S4c, with an exception in Kenya) and by the higher estimated biologic probability of transmission from men to women than from women to men. The age profile seems to be more tilted toward older ages (to the right) in Tanzania and Ghana and for women in Burkina Faso while it is more tilted toward younger ages (to the left) in Cameroon and Kenya and for men in Burkina Faso. HIV prevalence is not a perfect measure of the current state of the epidemic since it is a stock variable, affected by past incidence and mortality rates. As noted, individuals who are HIV positive are asymptomatic for nine years on average before they get AIDS. Without treatment, individuals with AIDS die within one year on average. Therefore, lower HIV prevalence at older ages does not necessary mean that those birth cohorts were less likely to be infected but rather that a substantial portion of individuals in those birth cohorts who were HIV positive have already died. As a starting point for the analysis and for comparison with regression coef­ ficients in multivariate analyses, table 5 reports unadjusted means of HIV prevalence by education and by wealth and poverty levels from reports of the Demographic and Health Surveys (Burkina Faso Government and ORC Macro 2004; Cameroon Government and ORC Macro 2004; Ghana Government and ORC Macro 2004; Kenya Government and ORC Macro 2004; and Tanzania Government and ORC Macro 2005). These unadjusted means suggest that HIV infection generally increases with educational attainment, although the increase seems larger and more consistent from no education to some primary education than from some primary education to some secondary education or more. The unadjusted means also suggest that the risk of HIV infection increases with wealth. In every country, individuals living in a household with an earth floor (the proxy for poverty) are less likely to be HIV positive than those having another type of floor, who are considered less poor. That differ­ ence, however, is statistically significant at the 5 percent level only in Cameroon and Tanzania, for both men and women, and at the 10 percent meM ,f ~, M " J£.1 '"' TABLE 5. HIV Prevalence by Selected Characteristics: Unadjusted Means Burkina Faso (2003) Cameroon (2004) Ghana (2003) Kenya (2003) Tanzania (2004) Men Women Men Women Men Women Men Women Men Women Variable (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) No education 0.0173 0.0153 0.0213 0.0330 0.0126 0.0217 0.0225 0.0441 0.0416 0.0584 (0.0035) (0.0023) (0.0063) (0.0061) (0.0042) (0.0039) (0.0120) (0.0112) (0.0103) (0.0092) Primary education 0.0190 0.0305 0.0416 0.0713 0.0186 0.0335 0.0459 0.0986 0.0640 0.0809 (0.0071) (0.0109) (0.00S2) (0.0064) (0.0053) (0.0057) (0.0067) (0.0089) (0.0053) (0.0061) Secondary education or higher 0.0281 0.03 OS 0.0413 0.0799 0.0166 0.0274 0.0510 0.0817 0.0731 0.0925 (0.0103) (0.0109) (0.0040) (0.0064) (0.0029) (0.0033) (0.0076) (0.0098) (0.0163) (0.0160) Earth floor 0.0181 0.0130 0.0201 0.0458 0.0157 0.0216 0.0373 0.0721 0.0439 0.0551 (0.0042) (0.0026) (0.0037) (0.0059) (0.0061) (0.0058) (0.0069) (0.0069) (0.0042) (0.0053) No earth floor 0.0208 0.0252 0.0481 0.0833 0.0163 0.0277 0.0537 0.1061 0.0972 0.1161 (0.0043) (0.0051) (0.0040) (0.0054) (0.0023) (0.0026) (0.0068) (0.0103) (0.0095) (0.0093) Note: Numbers in parentheses are clustered standard errors. The data are weighted with the sample weights given by the data provider. Source: Author's analysis based on data from Demographic and Health Surveys (Burkina Faso and ORC Macro 2004; Cameroon and ORC Macro 2004; Ghana and ORC Macro 2004; Kenya and ORC Macro 2004; and Tanzania and ORC Macro 200S). i} ~ .s­ ~ '" N N V. 226 THE WORLD BANK ECONOMIC REVIEW confidence level for women in Kenya. The analysis In the remainder of this article goes beyond unadjusted means. The next analysis pools the observations for the five countries in regressions in which the dependent variable is HIV status (zero for HIV negative and one for HIV positive; table 2). The analysis controls for wealth using a group of assets as indicators of wealth rather than asset quintiles, which are not easily compared across countries. These assets are type of latrine, type of floor, avail­ ability of electricity, and ownership of a refrigerator, radio, television, bicycle, motorcycle, or car. For comparison with the coefficients from the linear model, results are also shown for a pro bit specification for cases where the dependent variable is binary (displaying the marginal effects of the probit coefficient). Two alternative methods are used to control for wealth or poverty. The linear models include the presence of an earth floor as a proxy for poverty and use the other durable goods as instruments, letting the data pick the weights. The pro bit speci­ fications use a set of indicators for durable goods as a proxy for wealth. Because the ethnicity9 and the widowhood variables were not available in the Tanzania survey, results are presented in table 2 with Tanzania and without (regressions 3 and 4). Table 2 also includes results for the entire age range (regressions 1-4, 7, and 8) and, because a previous study documented a negative relationship between education and HIV infection among women younger than 30 in rural Uganda (de Walque, 2007a),JO for the age group 15­ 29 (columns 5 and 6). Education does not seem to be significantly associated with HIV infectionY However, in regressions 7 and 8, which restrict the sample to urban areas, coefficients on years of education are negative and statistically significant. 12 This finding suggests that while it may be difficult to isolate a negative relation­ ship between education and HIV in the overall population, it can be measured in urban settings. It may be that the negative relationship between education 9. In the pooled regressions without Tanzania, the ethnicity indicators are interacted with the country indicators so that each country maintains its own set of ethnicity categories. 10. This was one of the first studies to show a negative gradient. Of 27 studies reviewed by Hargreaves and Glynn (2002), only one, on sugar estate workers in Ethiopia, reported a significantly negative association between HIV infection and education. Most of these studies, however, are in urban settings and based on data collected in the beginning of the 1990s, at an earlier stage of the epidemic. 11. An exception is among young women of ages 15-29, for whom more education is negatively associated with HIV infection (at the 10 percent significance level) under the probit specification, but not in the linear instrumental variables regression. The sample was also restricted to individuals under age 30 in the country-level regressions to determine whether the association between education and HIV status was different for younger individuals: only in the case of young women in Kenya is there a significant negative association between education and the risk of HIV infection (results available on request). 12. Regressions were also separated for the urban and rural samples in the country-level regressions, but no consistent or significant pattern was found (results available on request). This might due to the fact that the urban samples are rather small in each individual Demographic and Health Survey, increasing the interest in using pooled regressions. !J8t 4 @. t I . ~ de Walque 227 and HIV infection takes time to develop and is found earlier in cities, where information spreads faster and HIV prevalence is generally higher. Urban location tends to be positively associated with HIV infection, as does being formerly married. Widowed individuals are even more at risk (table 2, regressions 3 and 4). It is likely that marital disruption and widowhood are a consequence rather than a cause of HIV infection and that widowed individuals are more likely to be HIV positive because their partners died of AIDS. Especially among women, having been in successive marriages is also an impor­ tant risk factor. Since successive marriages can result from either the death of a spouse or divorce, identification is problematic for this association. The associ­ ation could be driven by reverse causality, as in the case of widowhood, or could be due to self-selection if individuals who are less likely to commit to one partner are also more likely to be infected by HIV. Nevertheless, this result, together with the finding that a substantial share of the population has been engaged in successive marriages, suggests that this group, especially women, could benefit from prevention programs targeted to their needs and vulnerabil­ ities. Being in a polygynous union appears to be a risk factor for women. Country-level regressions with HIV status as the dependent variable gener­ ally show results that are consistent with the pooled regression results (table 3). When a coefficient is significant in the pooled regression results, it generally has the same sign and significance in a substantial number of country-level regressions; for the remaining countries, it tends to be of the same sign but not significant, possibly because of sample size. For the variables for which the coef­ ficient was not significant in the pooled regressions, there are only a few excep­ tions in which the coefficient is significant in the country-level regressions. HIV prevalence is substantially higher in Cameroon (3.9 percent for men and 6.6 percent for women), Kenya (4.6 percent and 8.7 percent) and Tanzania (6.3 percent and 7.7 percent) than in Burkina Faso (1.9 percent and 1.8 percent) and Ghana (1.6 and 2.7 percent; bottom of table 3). Women are usually more likely to be HIV positive, except in Burkina Faso. Notice, however, that the age ranges in the survey are not the same for men and women except in Tanzania. Contrary to the unadjusted means (table 5), but consistent with the pooled regression results (table 2), there is no significant association between HIV infection and years of education. Analyzing the same countries and data sets, Fortson (2008) finds a positive association between education and HIV infec­ tion. However, she does not control jointly for education and wealth and there­ fore her coefficients on the education variables are likely to reflect the confounded influence of education and wealth. Fortson also argues that the relationship between HIV infection and education is not linear, and so she uses a quadratic function of education. To test the nonlinearity of the relationship, two specifications are included in which education is not constrained to be linear (table S2). Regression 4 includes education and education squared, as in Fortson (2008), and regression 5, which does not to assume a functional form, include indicators for the 228 THE WORLD BAt-;K ECOt-;OMIC REVIEW education categories (no education, at least some primary education, and at least some secondary education or above). Overall, while there is some evi­ dence of nonlinearity, it is difficult to conclude that there would be a strong association between education and HIV infection. HIV infection is positively associated with urban status for women in Cameroon, men in Kenya, and both men and women in Tanzania. This result is consistent with that for the pooled regression (table 2). There is a strong positive association between being formerly married and HIV infection for women in Cameroon, Ghana, Kenya, and Tanzania. This association is also found in the pooled regressions and could be due to reverse causality. Successive marriages seem to be an important risk factor. Consistent with the pooled regression results, having been in successive marriages is positively associated with HIV infection for women in Cameroon, Ghana, and Tanzania and for men in Tanzania. Polygyny does not seem to be associated with HIV infection, except in Burkina Faso, where the association is negative, and for women in Kenya, where the association is positive. The results for polygyny are different in the pooled regressions, which show a significant positive associ­ ation for women. It might be that the significance is higher in the pooled regressions because of the larger sample size. HIV test results are missing for some individuals who were randomly selected to be tested in the survey, because they refused to be tested or were absent or because of a technical problem. The proportion of people being tested (reported at the bottom of table Sl) is always above 82 percent, but is higher (above 92 percent) in Burkina Faso and Cameroon and for women in Ghana. Refusal to be tested is the main reason for the absence of a test. The absence of a test result, if not random, might cause a bias. Table Sl analyzes the determinants of the likeli­ hood to be tested if selected in the HIV sample of the surveys. Gersovitz (2007) also points out that some people, although randomly selected to be included in the sample, did not participate at all in the surveys (questionnaire and HIV test). Mishra and others (2008) conclude that the potential bias due to nonre­ sponse does not significantly affect the seroprevalence estimates obtained from nationally representative samples. However, for this analysis of the link between education and HIV infection, it is useful to verify whether the absence of an HIV test is correlated with education and other variables. Education does not seem to be associated with acceptance of the test (table Sl)Y People in urban areas are generally less likely to be tested. 14 13. In the marginal probit specification, there are two negative coefficients (men in Cameroon and in Kenya) that are significant at the 10 percent level. 14. Since HIV prevalence is generally higher in urban areas (tables 3 and 4), this could imply a slight downward bias in the estimates of overall HIV prevalence as well as a potential bias of the coefficients on urban location. This would require unobserved variables to be correlated, but not fully, with observed variables (HIV infection and urban location). However, overall, the facts that the coverage of the surveys is very good and that there is no association between education, the main variable of interest, and the probability of being tested in the survey, limits the scope for bias . u .. de Walque 229 III. SEXUAL BEHAVIORS AND ATTITUDES RELATED TO THE HIV/AIDS EPIDEMIC A range of sexual behaviors assumed to have an influence on the risk of HIV infection (condom use, extramarital sex, abstinence, virginity, and age at sexual initiation) are examined (tables 4 and S2-S5). Filmer (1998) and Blanc (2000) use earlier Demographic and Health Surveys to study the socioeconomic correlates of sexual behavior. Filmer generally finds a positive association between condom use outside marriage and education and urban status. Blanc finds that educated people are more likely to engage in nonregular sex but that they are also more likely to use condoms within those relationships. Two studies by de Walque (2003, 2007a) analyze sexual behaviors for a cluster of villages in rural Uganda. Those sexual behaviors are at the heart of most pre­ vention efforts, including the so-called "ABC" strategy (abstain, be faithful, or use a condom). The present study also analyzes the use of voluntary counseling and testing facilities, the probability that spouses will discuss AIDS, and the knowledge that an asymptomatic person can be HIV positive. Condom Use and Extramarital Sex The Demographic and Health Surveys ask respondents whether they used a condom during their last sexual intercourse and whether that intercourse occurred with a spouse or with another partner. Condom use is recommended in both cases, but not using a condom outside marriage is considered riskier. Because condom use at last intercourse differs widely according to whether the last intercourse was inside or outside marriage (compare the means in tables S3a and S3b), the cases are analyzed separately. Table 4 uses pooled regressions as in table 2 to look at the association between education and condom use during marital and nonmarital sex and extramarital sex. Education is always positively associated with condom use, both with the spouse and with an extramarital partner. IS But among women, education is also positively associated with extramarital sex. 16 Extramarital sex increases the risk of contracting HIV, but that risk can be mitigated with condom use. It might be because of these contradicting associations that edu­ cation is not significantly associated with HIV status. Examining the same three behaviors using country-level regressions finds that the positive association between education and condom use in marriage is robust and consistent with the pooled regression results (only for men in Burkina Faso and Kenya and for both genders in Tanzania is it not significant under the linear model; table S3). Education increases the likelihood of using 15. Extramarital sex includes all sexual relationships outside the union, regardless of whether an individual is married. 16. Under the pro bit specification, education is sometimes also positively associated with extramarital sex for men. 230 THE WORLD BANK ECONOMIC REVIEW a condom in extramarital relationships everywhere, except in Kenya (table S3b). Because condoms are an effective method of preventing HNIAIDS, these findings support increasing access to education as a path to addressing the HNIAIDS crisis in Africa. The country -level regressions also indicate that ed ucation is positively associated with extramarital sex in several countries, for both men and women (table S3c). This association is consistent with the pooled regression results in table 4. Other Sexual Behaviors and Attitudes Related to HIVIAIDS Abstinence is another strategy for avoiding AIDS. The analysis shows that in several countries more educated people are less likely to abstain from sex (table S4a). Education is negatively associated with abstinence for both men and women in Burkina Faso and Cameroon, for men in Ghana, and for women in Tanzania (linear model). The analysis of virginity is done for singles only since it is assumed that all ever-married individuals have had sexual activity (the data confirm this). Education is positively associated with virginity among women in Ghana, Kenya, and Tanzania (table S4b). Only among men in Cameroon is there a negative association between virginity and education. It is generally assumed that a later age at of sexual initiation is a way to prevent HIV/AIDS infection. 17 Education is always associated with a later sexual debut for women (table S4c). This relationship might also be affected by reverse causality since pregnancy frequently means that a girl has to drop out of school. Education is also positively associated with age of sexual debut for men in Tanzania, but educated men in Burkina Faso and Cameroon have earlier sexual experiences. Other attitudes and practices that are not sexual behaviors but are related to the HIVIAIDS epidemic were also examined. Education is always positively associated with obtaining information about one's HN status (table S5a), with an increased level of discussion about AIDS between spouses (table S5b), and with the knowledge that a healthy-looking person can be HIV posi­ tive (table SSc), used as an indicator of knowledge about the HIV/AIDS epidemic. From this analysis, education appears to be one of the most consistent predictors of behavior and knowledge. Education predicts protective behaviors such as condom use, use of counseling and testing, discussion of AIDS between spouses, and knowledge of HNIAIDS, but it also predicts a higher level of infidelity and a lower level of abstinence. It might be because of these contradicting associations that education is not significantly associated with HN status. 17. Gersovitz (2005) shows several inconsistencies in self-reported age at first sexual intercourse by comparing subsequent Demographic and Health Surveys in the same countries. The results of this analysis should therefore be treated with caution. 4 liM£!tdZl t 1 __ L 1.- L de Walque 231 IV. CONCLUSIONS The last wave of Demographic and Health Surveys in many countries, especially in Africa, includes HIV testing for a representative sample of the population. A very useful addition, this testing allows for a better assessment of the epidemic in each country. However, as antiretroviral treatment is scaled up in many countries, HIV prevalence will become an ambiguous indicator. If prevalence is increasing, it will it become difficult to ascertain whether that is due to a higher HIV incidence, and therefore a failure of prevention efforts, or to lower AIDS-related mortality, and therefore to the success of treatment pro­ grams. The development of nationally representative measures of HIV inci­ dence should therefore be encouraged. It would also be useful to include questions about antiretroviral treatment in the next wave of Demographic and Health Surveys. This article takes advantage of the HIV data in the Demographic and Health Surveys to study the socioeconomic determinants of HIV status and sexual behavior in five African countries and, because the variables are defined similarly, to use pooled data to draw some interesting generalizations. While having the results of HIV testing, an objective biomarker, is a benefit of the new wave of Demographic and Health Surveys, a limitation is that sexual beha­ viors are self-reported. Another shortcoming of the analysis is that each of the five data sets is a cross section and many of the variables are potentially endogenous, although the most obviously endogenous variables, such as sexual behaviors, were not used as regressors. Thus, the coefficients in this study should not be taken to imply causal relationships. But even if causal links cannot be established, some reported associations clearly show that some cat­ egories of the population are at greater risk and specific prevention interven­ tions should be directed to them. Several findings can be generalized and are important for policymakers engaged in controlling the HIVIAIDS epidemic. Successive marriages are a sig­ nificant risk factor, especially for women. Even if this result is due to self­ selection, it suggests a need for specific prevention efforts for that group. Contrary to the evidence derived from unadjusted means, education is not posi­ tively associated with HIV status. But schooling is one of the most consistent predictors of behavior and knowledge: educational achievement predicts pro­ tective behaviors such as condom use, HIV testing, discussion of AIDS between spouses, and knowledge about HIV/AIDS. However, it also predicts a higher level of extramarital sex and a lower level of abstinence. It is possible that these associations cancel each other out, thus explaining why education is not significantly associated with HIV status. However, while it is difficult to isolate a negative relationship between edu­ cation and HIV in the overall population, a negative association can be measured in urban settings (table 2, regressions 7 and 8). It could be that the negative gradient between HIV infection and education takes time to develop 232 THE WORLD BANK ECONOMIC REVIEW and is found earlier in cities, where information spreads faster and where HIV prevalence is generally higher. ACKNOWLEDGMENTS The author thanks Quy-Toan Do, Timothy Johnston, Ted Miguel, Mead Over, and seminar and conference participants at the World Bank, the University of California at Berkeley, and the International AIDS Economics Network Conference in Cuernavaca, Mexico, for useful discussions, and Rachel Kline for editorial assistance. SUPPLEMENTARY DATA A supplemental appendix to this article IS available at http:Uwber. oxfordjournals.orgl. REFERENCES Blanc, Ann K. 2000. The Relationship between Sexual Behavior and Level of Education in Developing Countries. Geneva, Switzerland: The Joint United Nations Programme on HIV/AIDS. Burkina Faso Government. 2006. "Comite Ministeriel de Lutte Contre Ie Sida." Ouagadougou. Unpublished report. Burkina Faso Government and ORC Macro. 2004. Enqu/ite Demographique et de Sante du Burkina Paso 2003. Ouagadougou: Institut National de la Statistique et de la Demographie. Buve, Ann, Michael CaraeJ, and Rea Hayes. 2001. "Multicentre Study on Factors Determining Differences in Rate of Spread of HIV in Sub-Saharan Africa: Methods and Prevalence of HIV infec­ tion." AIDS lS(suppl. 4):5S-514. Cameroon Government and ORC Macro. 2004. Enquete Demographique et de Sante du Cameroun 2004. Yaounde: Institut National de la Statistique. de Walque, Damien. 2003. "How Do Information and Education Affect Health Decisions? The Cases of HIVIAIDS and Smoking." PhD Dissertation, University of Chicago, Chicago, Illinois. ~~--. 2006. Who Gets AIDS and How? The Determinants of HIV Infection and Sexual Behaviors in Burkina Paso, Cameroon, Ghana, Kenya and Tanzania. Policy Research Working Paper 3844. World Bank, Washington, DC. ---.2007a. "How Does the Impact of an HIV/AIDS Information Campaign Vary with Educational Attainment? Evidence from Rural Uganda." Journal of Development Economics 84(2):686-714. - - - . 2007b. "Sem-Discordant Couples in Five African Countries: Implication for Prevention Strategies." Population and Development Review 33(3):501-23. de Walque, Damien, Jessica S. Nakiyingi-Miiro, June Busingye, and Jimmy A. Whitworth. 2005. "Changing Association between Schooling Levels and HIV -1 Infection over 11 Years in a Rural Population Cohort in South-West Uganda." Tropical Medicine and International Health 1O( 1O):993-100l. Filmer, Deon. 1998. The Socio-Economic Correlates of Sexual Behavior: A Summary of Results from an Analysis of DHS Data." In Martha Ainsworth, Lieve Fransen, and Mead Over, eds. Confronting AIDS: Evidence from the Developing World. Brussels: European Commission. Fortson, Jane G. 2008. "The Gradient in Sub-Saharan Africa: Socioeconomic Status and HIV/AlDS." Demography 45(2):303-22. .. $, i 11(11 Mt & t .J ; @ i. WAlk ; M1 I de Walque 233 Fylkesnes, Knut, Rosemary Mubanga Musonda, Kelvin Kasumba, Zacchaeus Ndhlovu, Fred Mluanda, Lovemore Kaetano, and Chiluba C. Chipaila. 1997. "The HIV Epidemic in Zambia: Socio-Demographic Prevalence Patterns and Indications of Trends among Childbearing Women." AIDS, 11(3):339-45. Fylkesnes, Knut, Rosemary M. Musonda, Moses Sichone, Zacchaeus Ndhlovu, Francis Tembo, and Mwaka Nonze. 2001. "Declining HIV Prevalence and Risk Behaviours in Zambia: Evidence from Surveillance and Population-Based Surveys." AIDS 15(7):907-16. Gersovitz, Mark. 2005. "The HIV Epidemic in Four African Countries Seen through the Demographic and Health Surveys." The Journal of African Economies 14(2):191-246. - - - . 2007. "HIV, ABC, and DHS: Age at First Sex in Uganda." Sexually Transmitted Infections 83(2):165-68. Ghana Government and ORC Macro. 2004. Ghana Demographic and Health Survey 2003. Accra: Ghana Statistical Service and Noguchi Memorial Institute for Medical Research. Hargreaves, James R., and Judith R. Glynn. 2002. "Educational Attainment and HIV-l Infection in Developing Countries: A Systematic Review." Tropical Medicine and International Health 7(6):489-98. Kenya Government and ORC Macro. 2004. Kenya Demographic and Health Survey 2003. Nairobi: Central Bureau of Statistics and Ministry of Health. Kilian, Albert H.D., Simon Gregson, Bannet Ndyanabangi, Kenneth Walusaga, Walter Kipp, Gudrun Sahlmuller, Geoffrey P. Garnett et al. 1999. ~Reductions in Risk Behaviour Provide the Most Consistent Explanation for Declining HIV-1 Prevalence in Uganda." AIDS 13(3):391-98. Mishra, Vinod, Bernard Barrere, R. Hong, and S. Kahn. 2008. "Evaluation of Bias in HIV Seroprevalence Estimates from National Household Surveys." Sexually Transmitted Infections 84(Sl):i63-i70. Morgan, Dilys, Cedric Mahe, Billy Mayanja, Martin J. Okongo, Rosemary Lubega, and James A.G. Whitworth. 2002. "HIV-1 Infection in Rural Africa: Is There a Difference in Median Time to AIDS and Survival Compared with That in Industrialized Countries?" AIDS 6(4):597-603. Nicolas, Nagot, Amadou Ouangre, Abdoulaye Ouedraogo, Michel Cartoux, Pierre Huygens, Marie Christine Defer, Tarnagda :lekiba, Nicolas Meda, and Philippe Van de Perre. 2002. "Spectrum of Commercial Sex Activity in Burkina Faso: Classification Model and Risk of Exposure to HIV." Journal of Acquired Immune Deficiency Syndromes 29(5):517-21. Nunn, Andrew J., Jane F. Kengeya-Kayondo, Sam S. Malabar, Janet A. Seeley, and Daan W. Mulder. 1994. "Risk Factors for HIV-1 Infection in Adults in a Rural Ugandan Community: A Population Study." AIDS 8(1):81-86. Tanzania Government and ORC Macro. 2005. Tanzania HIVIAIDS Indicator Survey 2003-04. Dar es Salaam: Tanzania Commission for AIDS and National Bureau of Statistics. UNAIDS (Joint United Nations Programme on HIVIAIDS). 2007. AIDS Epidemic Update: December 2007. Geneva: Joint United Nations Programme on HIV/AIDS. WHO (World Health Organization) and UNAIDS (Joint United Nations Programme on HIV/AIDS). 2005. Progress on Global Access to HIV Antiretroviral Therapy. An Update on "3byS": June 200S. Geneva: World Health Organization. www.who.intl3by5/fulIreportJune2005.pdf. I I I ' 1$ _, '" $, _as _ IN Ill;:· "".'eg • ,bJ A Cost- Benefit Analysis of Cholera Vaccination Programs in Beira, Mozambique Marc Jeuland, Marcelino Lucas, John Clemens, and Dale Whittington Economic and epidemiological data collected in Beira, Mozambique, are used to conduct this first social cost-benefit analysis for cholera vaccination in Sub-Saharan Africa. The analysis compares the net economic benefits of three immunization strat­ egies with and without user fees: school-based vaccination for school children only (age 5-14), school-based vaccination for all children (age 1-14), and a mass vacci­ nation campaign for all people older than one year. All options assume the use of a low-cost new-generation oral cholera vaccine. The analysis incorporates the latest knowledge of vaccine effectiveness, including new evidence on the positive externality associated with the resulting herd protection (both protection of unvaccinated individ­ uals and enhanced protection among vaccinated individuals arising from vaccination of a portion of the population). It also uses field data for incidence, benefits (private willingness to pay, public cost of illness), and costs (production, shipping, delivery, private travel costs). Taking herd protection into account has important economic implications. For a wide variety of parameters values, vaccination programs in Beira pass a cost-benefit test. Small school-based programs with and without user fees are very likely to provide net benefits. A mass vaccination campaign without user fees would result in the greatest reduction in the disease burden, but the social costs would likely outweigh the benefits, and such a program would require substantial public sector investment. As user fees increase, mass vaccination becomes much more attrac­ tive, and the reduction in disease burden remains above 70 percent at relatively low user fees. JEL codes: 11, H23, H4 --------- ------------------------- Cost-benefit analysis of vaccination programs is now rarely done anywhere in the world. Most health economists have given up on welfare-theoretic econ­ omic appraisals of health interventions such as vaccination programs, prefer­ ring to present cost-effectiveness calculations that use disability-adjusted life Marc Jeuland (corresponding author) is a doctoral candidate in the Department of Environmental Sciences and Engineering at the University of North Carolina at Chapel Hill; his email address is jeuland@email.unc.edu. Marcelino Lucas is director of the Ministry of Science and Technology of Mozambique; his email addressismarcelin@zebra.uem.mz. John Clemens is director of the International Vaccine Institute in Seoul; his email address is jclemens@ivi.int. Dale Whittington is a professor at the University of North Carolina at Chapel Hill and at the Manchester Business School, Manchester University, UK; his email address is dwhittin@email.unc.edu. THE WORLD BANK ECONOMIC REVlEW, VOL 23, 1'>0. 2, pp. 235 -267 doi:10.1093/wber/lhp006 Advance Access Publication July 22, 2009 © The Author 2009. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I TIlE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 235 236 THE WORLD BANK ECONOMIC REVIEW years (DALYs) or quality-adjusted life years (QALYs) as their measure of health outcomes. The conventional wisdom based on cost-effectiveness analysis is that cholera vaccination is not "cost effective" compared with other public health interven­ tions and that it is an unlikely candidate to receive scarce public funds, except ,I I during cholera outbreaks or in refugee camps (Murray, McFarland, and Waldman 1998; Naficy and others 1998; Jamison and others 2006). The limit­ I ations of cost-effectiveness analysis are well known, however. Such analysis cannot provide finance ministries the economic information needed for cross­ sectoral comparisons with other public investments, such as education or I I infrastructure. The cut-off levels used to determine what is a "cost-effective" intervention have no theoretical justification. Nonhealth benefits of policy I I interventions (such as time savings) are typically ignored, as are private costs (such as queuing and travel costs to obtain health-related goods). Finally, cost­ I effectiveness analysis does not address the rationale for government involve­ ment. Simply because an intervention is judged to be cost-effective does not mean that there is a good reason why government should be involved in its provision; private markets provide many goods and services that would pass a cost-effectiveness test. This article uses cost- benefit analysis to show that there is a strong econ­ omic case for cholera vaccination in Beira, Mozambique. It also shows that it is both possible and desirable to use standard economic methods to appraise complex health interventions in developing countries and that health econom­ ists do not have to rely on cost-effectiveness analysis and its ad hoc health outcome measures of DALYs and QALYs. Section I presents background information on cholera, the new-generation cholera vaccines, and the latest epidemiological evidence for the cholera vacci­ nation externality. Section II describes the analytical framework, including the vaccination program options to be examined and the basis for the herd protec­ tion model. Section III develops the cost- benefit model and explains the under­ lying assumptions. Key results of the model are summarized in section IV. Section V discusses the value of these findings to policymakers. I. CHOLERA AND NEW-GENERATION CHOLERA VACCINES Cholera is a serious, persistent disease across large portions of Mozambique. The World Health Organization reported that Mozambique had the world's highest number of cholera cases (20,080) in 2004, and Mozambique consist­ ently appears among the worst hit countries for other years (WHO 2005). During the most recent outbreak, in 2006, there were at least 5,800 cases nationwide, affecting 22 districts and 4 provinces in the center of the country. The largest city suffering from endemic cholera and high incidence rates is Beira (population of 550,000), where a series of epidemics broke out in 2001­ 03, with more than 3,800 cases for three years in a row. Most cases occur Jeuland, Lucas, Clemens, and Whittington 237 shortly after seasonal flooding between the end of the rainy season in January and February and the advent of the cool, dry season in May and June (Lucas and others 2005). Cholera is most commonly transmitted through consumption of contami­ nated water or through food that has come into contact with such water. Large infectious doses are usually required to cause severe cholera. Newly infected people pass through an incubating stage averaging 3-4 days, followed by an infectious state averaging about 10 days (Barua and Greenough 1992; Sack and others 2004). Especially in endemic areas, many infected people are asympto­ matic, due to natural immunity acquired during previous infection or ingestion of smaller doses of infectious organisms. In symptomatic individuals, cholera causes acute dehydration, which can lead to death in the absence of prompt treatment with intravenous rehydration therapy. With proper treatment, case fatality rates from cholera infection are low (typically less than 1 percent; Ryan 2000). Case fatality rates may reach 20-50 percent in refugee camps, after natural disasters, or in other situations where health systems are unable to deliver treatment promptly and effectively (Naidoo and Patric 2002). Beira, however, has a dedicated, single-purpose Cholera Treatment Center with diagnostic as well as treatment capabilities, all delivered free of charge to the sick. Due to mass publicity campaigns, residents know and use the center, and the risk of death from cholera is low. Nevertheless, policymakers and resi­ dents remain very concerned about cholera infection. People want to avoid the pain and suffering associated with the disease, as well as the (small) risk of death and the indirect costs associated with lost time and labor. New-generation vaccines offer a way of controlling cholera in areas like Beira, where modern water and sanitation infrastructure remains largely unaf­ fordable for government and the urban poor. In late 2003, the International Vaccine Institute and Medecins sans Frontieres conducted a pilot community­ based cholera vaccination trial in one neighborhood of Beira, using health centers and schools as vaccination outposts. The purpose of the trial was to test the effectiveness of a two-dose regimen of the Dukoral™ oral, killed, whole-cell vaccine (rBS-WC), given two weeks apart (Lucas and others 2005).1 This work followed an epidemiological study of cholera incidence among different age groups. The research program in Beira also included economic studies measuring the public and private costs of illness associated with cholera and the private demand for vaccines and private costs incurred in acquiring them (such as travel costs and time spent traveling to the vaccination clinic or waiting to be seen). The results of those studies are reported elsewhere (Lucas and others 2007; Jeuland and others 2008). The findings from this suite of 1. These research activities were conducted through the International Vaccine IllStitute's Diseases of the Most Impoverished program, which works to accelerate the development and introduction of new-generation vaccines against cholera, typhoid fever, and shigellosis in developing countries. The program involves a number of parallel epidemiological and social science studies, as well as vaccine technology transfer activities. 238 THE WORLD BANK ECONOMIC REVIEW epidemiological and economic research activities, combined with recently pub­ lished findings on the herd protection effects (when vaccination of a portion of the population confers protection on unvaccinated individuals and enhanced protection of vaccinated individuals) of cholera vaccination from Bangladesh (Ali and others 2005; Longini and others 2007), provide a rich new evidence base that is used in this welfare-theoretic economic appraisal (cost-benefit analysis) of larger-scale vaccination program options in Beira. In the only widely cited cost-benefit analysis of cholera vaccination, Cookson and others (1997) find that cholera vaccination passes a cost-benefit test in the base case for vaccination of a confined population in an outbreak­ prone area of northern Argentina. However, the study did not measure demand for the vaccine and instead used avoided cost of illness as a measure of benefits. The main reason that cholera vaccination was found to be an attrac­ tive economic proposition was the unusually high avoided cost of illness ($622 per case2 ) because of high hospitalization rates, helicopter transport of infected individuals, and extensive medical therapy. Only in very special instances will the use of an avoided cost of illness measure of benefits result in positive net benefits, however, because cholera is relatively easy and inexpensive to treat. The total cost of illness of a cholera episode is generally much lower than the high values reported by Cookson and others (see also Poulos and others 2008). The cost-benefit analysis by Cookson and others (1997) was also conducted before field evidence was available on the herd protection from cholera vacci­ nation. Ali and others (2005) recently reanalyzed incidence data for vaccinated and unvaccinated people from baris (neighborhoods) in Matlab, Bangladesh, that had varying levels of vaccination coverage. They showed that cholera inci­ dence in the first year following the vaccination program was highly dependent on coverage rates in the baris for members of control groups receiving a placebo rather than the vaccine. Because the cholera vaccine offers less than 100 percent protection, herd protection was also found among vaccine recipi­ ents, but the effect was weaker than among nonvaccinated individuals. This article shows that these findings potentially have important implications for cost benefit analysis of vaccination programs. II. ANALYTICAL FRAMEWORK FOR THE COST-BENEFIT ANALYSIS OF CHOLERA VACCINATION PROGRAMS This section describes the baseline conditions, the three program options, and the approach to modeling herd protection in this paper. The Status Quo, Baseline Conditions Cost-benefit analyses calculate the costs and benefits of a policy intervention as a change from status quo or baseline conditions. The baseline condition in 2. All monetary amounts reported in this article are presented in 2005 US dollars. i 0 J lit K% - . Jeuland, Lucas, Clemens, and Whittington 239 Beira is that cholera vaccines are not sold on demand-there is no free, com­ petitive market for cholera vaccines. This situation is the norm in developing countries because governments often prohibit importing vaccines for sale by private healthcare providers, a policy endorsed by the international public health community. There are three main reasons why governments in developing countries prevent free market sales of many vaccines. One reason is that governments want to be monopso­ nists, exerting market power to negotiate low, bulk purchase prices from inter­ national vaccine suppliers. In many cases, governments also implement their strategy by securing donor support for vaccine purchases. Another reason is that governments want to ensure that vaccines are safe and effective, and regu­ latory hurdles to vaccine licensing are often high. A third reason is that govern­ ments often have an explicit policy that vaccines must be provided free of charge; charging for vaccines is often illegal. 3 Thus, this analysis compares various government or donor-sponsored vacci­ nation programs with administered prices (including the provision of free vaccines) rather than with prices in a market with private providers. Government-sponsored vaccination programs with administered prices are the most realistic options in Mozambique and most other developing countries. Neither governments nor donors are contemplating a free market for vaccine provision in Mozambique or elsewhere. In addition, information on the demand curve for cholera vaccines in a market with many private providers, needed to assess costs and benefits of the market sale of vaccines, does not exist. This article relies on private vaccine demand curves estimated with data from stated preference (hypothetical market) surveys that describe an institutional setting in which the government administers a vaccination campaign and charges specific user fees for the vac­ cines. Demand might be quite different in a private market with multiple provi­ ders. Also, vaccine demand will be heavily influenced by social marketing and health promotion campaigns associated with government-sponsored programs, as has been the case in Beira in the past. Private vaccine providers might well launch such campaigns to increase awareness of the benefits of new-generation cholera vaccines, but in the current situation in Beira and elsewhere, people expect to receive this health information from public sector health professionals. 3. It would be a relatively straightforward application of cost-benefit techniques to evaluate the costs and benefits of changing government policy to permit market sales of cholera vaccines. Assessing the costs and benefits of such vaccine "deregulation" requires having a private demand curve for cholera vaccines and evidence on the magnitude of the positive externality associated with herd protection. Then the area under the demand curve would serve as a measure of the direct benefits to individuals vaccinated. Accounting for the positive externality from herd protection would require determining what proportion of people among unvaccinated groups would receive the indirect herd protection and what their willingness to pay would be for this ex ante risk reduction. 240 THE WORLD BANK ECONOMIC REVIEW Program Options This article therefore estimates the costs and benefits of three targeted one-time, government-administered vaccination programs in Beira, compared with the baseline condition in which cholera vaccines are not available to anyone: • Option 1: school-based vaccination of school children (ages 5-14); • Option 2: school-based vaccination of all eligible children (ages 1-14); and • Option 3: community-based vaccination of all eligible people (ages 1 through adult; the vaccine cannot be given to infants younger than age 1). Costs and benefits of each option are calculated over a range of prospective government-administered user fees. The motivation for targeting different age cohorts is policymakers' interest in designing cholera vaccination programs particularly for children, who are most vulnerable to dying of cholera and have the highest disease incidence. Age-group specifications are based on practical considerations. School-age chil­ dren are typically easy to reach through targeted school-based immunization (option 1). Option 2 might be more difficult to implement because younger children would have to be brought to the school vaccination site by parents or guardians. Option 3 would require the most extensive planning and awareness­ raising efforts in order to encourage adults to bring themselves and children to community vaccination sites. The analysis assumes that all the vaccination program options would use a whole-cell-type oral cholera vaccine similar to the one now produced in Vietnam that can be purchased for about $0.50 a dose (The more expensive Dukoral™ vaccine manufactured by Crucell in the Netherlands sells for at least $3-$4 a dose). A single two-dose vaccine regimen protects against cholera for about three years. The exact level of this effectiveness depends on the proportion of the total population that is vaccinated, because of the herd protection effect. Modeling Herd Protection To incorporate the herd protection externality into the cost-benefit calcu­ lations, mathematical relationships for both indirect effectiveness (l1u, among the unvaccinated) and total effectiveness (l1v, among the vaccinated) are speci­ fied as functions of population coverage using the data in Ali and others (2005). Specifically, two exponential curves are fitted to the data, imposing the constraint that protection among the unvaccinated must be less than or equal to protection among the vaccinated (see appendix A for parameter assumptions and appendix B for a summary of the data and additional details on this math­ ematical model). When combined into an overall protection curve representing :tI it S, t d.. " " . " b . n I~f, MUflU., _ It, WI _, Ji • feuland, Lucas, Clemens, and Whittington 241 FIGURE 1. Model of Overall Cholera Vaccine Effectiveness as a Function of Vaccination Coverage Rate in Year 1 in Herd Protection Scenario, and Comparison with Results from Epidemiological Studies ..­ 100 -­ tfl. 5 c 75 ~ ~ 41 41 50 c 'u u ~ 25 ... 1ii ~ 0 o o 20 40 60 80 100 Population coverage (%) - Total effectiveness (vaccinated) - - - - - - - Indirect effectiveness (unvaccinated) - - - Overall effectiveness o MJdel from Longini and others (2007): overall effectiveness x Data from Ali and others (2005): overall effectiveness Note: These estimates for effectiveness were adjusted downward by 17 percent in year 3 to represent the vaccine's waning protection over time. No adjustment was made to the assumed effectiveness in year 2. Source: Authors' analysis based on data from a cholera vaccination trials and epidemiological and economic research in Beira, Mozambique; Ali and others (200S); and Longini and others (2007); see text for details. average effectiveness in the general population, these mathematical functions yield effectiveness levels for the first year following vaccination that are very similar to results from epidemiological model simulations conducted by Longini and others (2007) (figure 1). Two limitations of this use of the herd protection data from Matlab, Bangladesh, should be highlighted. First, data are for the year following vacci­ nation only, whereas cholera immunizations provide about three years of pro­ tection (Thiem and others 2006; Trach and others 1997). In the cost-benefit calculations, the protective effects for the third year are thus adjusted down­ ward by 17 percent to simulate the typical reduction in field effectiveness of the vaccine over time, estimated in vaccine trials to be from 60 percent in years 1 and 2 to 50 percent in year 3. The second difficulty is the specificity of the herd protection effect to the Matlab setting. Conditions such as population density, in- and out-migration, geography, and natural immunity were somewhat different from those in Beira. 242 THE WORLD BA:-.!K ECO:-.!OMIC REVIEW Detailed information on how such factors might influence herd protection was not available, and the effects observed in Matlab were used without modifi­ cation. As Longini and others (2007) argue, in populations with lower natural immunity than that of Matlab, which experiences a low level of cholera infec­ tion year-round, with periodic larger epidemic outbreaks, higher coverage rates would be needed to achieve the observed levels of overall vaccine effectiveness. Nonetheless, it seems reasonable to expect that natural immunity would also be high in Beira, where cholera is also endemic and incidence rates appear to be even higher than in Matlab (Deen and others 2008). III. THE COST - BENEFIT MODEL For the cost-benefit calculations, Beira's population was subdivided into six groups: vaccinated and unvaccinated individuals in three age groups (young children ages 1-4 years, school children ages 5-14, and adults ages 15 and older). The costs and benefits of cholera vaccination for each of these six groups were calculated for each program option. The calculations included infants younger than age 1 under the assumption that infants received the benefits of herd protection from the vaccinated population even though they were too young to receive the vaccine themselves. The following equation was used to calculate the net benefits of a specific vaccination program option to the total population of Beira: Net benefits = [Total private benefits + Total public benefits]- [Total costs] = [WTP(vaccinated) + WTP(unvaccinated) + Public COl] - [Total costs] 3 = LN;. {qi(Pv) . WTP;,v(Pv, 1]v) + [1 - qi(Pv)]' WTPj,u(Pv, 1]u)} i=1 3 3 + LNi · a;[qi(pv)]' COli - LN;. qi(Pv) . (cv + cp,i), ;=1 i=1 (1) where N j is the number of individuals in the population in each of the three age groups i; qi( Pv) is the fraction of individuals in each age group i that would choose to be vaccinated with two doses at user fee (price) Pv for the two-dose regime; WTPi ,v(Pv,1]v) is the average willingness to pay for the vaccine protec­ tion for each individual in age group i as a function of Pv and total effective­ ness of the vaccine 1]v among the subset of households that choose to have household members in this age group vaccinated; WTPi,u( Pv, 1]u) is the average willingness to pay for the two vaccine doses for each individual in age group i as a function of Pv and indirect vaccine effectiveness 1]u among the subset of ¥d #Jill J;\II 4 x, __ )I. If. II . $ Utt 1 ~_ lif J I II!!. • Jeuland, Lucas, Clemens, and Whittington 243 TABLE 1. Typology of Costs and Benefits Considered in the Beira Cost­ Benefit Analysis Model Benefit Cost Private: Benefits to vaccinated persons Cost of vaccine production Externality 1: Cost of illness savings to public Cost of vaccine delivery (shipping, cold chain, health system from reduced cholera cases wastage, outpost rental, awareness-raising, (not due to herd protection) vaccine administration) Externality 2: Additional protection of Private cost of vaccine acquisition (time spent vaccinated persons due to herd effects traveling and queuing) Externality 3: Protection of unvaccinated persons due to herd effects Externality 4: Cost of illness savings to public health system from reduced cholera cases (due to herd protection) Source: Authors' analysis. households that do not choose to have household members in this age group vaccinated; ai is the number of cases of illness avoided per person in age group i due to the vaccination program option; COli is the discounted cost of illness incurred by the government or public health care system per case of illness for a person in age group i; Cv is the total monetary cost of vaccination per person for production, shipping, and delivery of two doses; and Cp,i is the additional private cost of vaccination per person in age group i due to traveling to out­ posts and waiting in line to receive the vaccine. The first three terms of equation (1) describe the benefits of the vaccination program option to the population, while the final term corresponds to the costs of implementing the program option. On the benefit side, the first term includes the direct private and indirect herd protection benefits to vaccinated people. The second and third terms describe two other types of benefits: indirect protection of nonvaccinated people and reduced costs to taxpayers through avoided public health system costs (table 1). If there were no herd protection effect, total benefits would be limited to the private benefits to vaccinated individuals plus the benefits to taxpayers from reduced public treatment costs among those protected by vaccination. Under the assump­ tion that individuals are unaware of the herd protection effect, the private demand for the vaccine remains the same with or without herd protection. WTP;,v(Pv,1]v) is greater than Pv, and WTPi ,u(Pv,1]u) is less than PV' As Pv increases, fewer people are willing and able to purchase vaccines; q; thus decreases. At the same time, the private valuation WTPi,v (per vaccinated indi­ vidual) increases in Pv because only people with WTPi,v greater than Pv buy the vaccine. WTPi,u also increases in Pv, albeit at a slower rate, because raising prices excludes people with increasing private willingness to pay from partici­ pation in the vaccination program. For example, charging a price of $1 for the vaccines excludes some people with nonzero willingness to pay who would not 244 THE WORLD BANK ECONOMIC REVIEW be excluded by a program without user fees. The result is that average willing­ ness to pay for vaccine protection among the unvaccinated, excluded individ­ uals is higher in the program charging user fees than in the one without them. WfP;,v and WTP;,u encompass all the private benefits from vaccination, includ­ ing private avoided cost of illness and the value of reduced mortality risks, pain and suffering, and lost productivity due to illness. For public goods and goods not traded in markets (in this case, cholera vac­ cines), stated preference methods are often used to obtain empirical estimates of average willingness to pay and coverage rates at different prices (Hanemann 1994; Carson, 2000; Whittington and others 2002, 2009; Lucas and others 2007). Ideally, such measures would also capture the effect of different vaccine effectiveness levels 7J on demand. Varying levels of effectiveness would alter the prevalence of the disease, and the health economics literature argues that disease prevalence can have an important impact on demand for preventive goods (Ahituv, Hotz, and Philipson 1996; Philipson 1996; Gersovitz 2000; Gersovitz and Hammer 2003). In other words, the estimates of willingness to pay for a preventive treatment should be prevalence elastic. However, prior work on household demand for cholera vaccines in Beira does not enable esti­ mation of this prevalence elasticity (Lucas and others 2007). Coverage levels were thus assumed to depend only on price: (2) where qi,O is the fraction of people who would acquire the two-dose vaccine regime in age group i at a price of zero if everyone was informed of the cam­ paign (obtained using an econometric model predicting demand in age group i at a price of zero; see Lucas and others 2007); and /3p,; is the regression coeffi­ cient on price obtained from the models for age group i. In practice, equation (2) is written for the; households in a stated preference sample and includes a set of explanatory variables X other than price that also influence demand for vaccines. For simplicity, the expression shown in equation (2) is used in the cost-benefit calculations, assuming households to be "average" with respect to these other characteristics. Average willingness to pay per two-dose vaccine regime-among the vaccinated-is then expressed as the consumer surplus obtained from buying Ni x qi two-dose vaccinations at price Pv divided by the number of vaccinations Njxq; plus the actual expendi­ ture on each vaccination Pv (or, alternatively, as the average area under the demand curve for the subset of people with willingness to pay of Pv or higher): S;:' N; . q;,o . exp( -/3p,; . Pv)dpv (3) = .. '--+Pv Ni . qi,O . exp( -/3p.i . Pv) = (l//3p,J + Pv' ,4 )I k a Xi x: • Jeuland. Lucas, Clemens, and Whittington 245 Estimates of willingness to pay calculated using equation (3) thus assume that willingness to pay is independent of vaccination coverage. This willingness to pay measure derived using equation (3) can only be used to calculate WTPj,v(Pv,7]v) in equation (1) because it describes the average willingness to pay among vaccinated people in age group i, not the willingness to pay for pro­ tection among the unvaccinated WTPj,u( Pv,l'Ju)' If all residents of Beira are assumed to have nonnegative willingness to pay for cholera vaccines, the average willingness to pay for all people in Beira-not just the vaccinated subsample-should provide insights about WTPi,u(Pv,l'Ju)' This average willingness-to-pay measure is simply the area under the private demand curve for the population4 : (4) - - - J"" qi,O . exp(-{3p,i . Pv)dpv = WTPi(Pv) = (qi'O). -(3 . ° p,l WTPi,v(Pv,l'Ju) was then calculated as follows: V7Ju < 0.55 otherwise Demand among the unvaccinated is assumed to he directly proportional to effectiveness up to the "normal" level of protection cited in the literature on cholera vaccines (roughly 55 percent over three years). For example, a person willing to pay $1 for the vaccine but who would remain unvaccinated at Pv = $1.50 would also he willing to pay $0.50 for a vaccine that offered 27.5 4. It is possible to use respondents' average willingness to pay for cholera vaccination to estimate the value of a statistical life. This derived estimate of the value of a statistical life can then be compared with estimates in the literature using other methods to check the plausibility of the willingness-to-pay estimates. In interpreting the following value of a statistical life calculation and comparison, it should be noted that research on the value of a statistical life in developing countries is limited, and estimates may not be directly comparable. Assume that all private benefits are from reduced private cost of illness and averted mortality. The cost of illness study in Beira found a private cost of illness of $21.60 for children and $17.10 for adults (Poulos and others 2008). To calculate the present value of avoided private cost of illness from cholera vaccination, it is assumed that these benefits accrue over three years, that the vaccine is roughly 55 percent effective over three years, and that the real discount rate is 3 percent. The ex ante avoided private cost of illness in each age group is then simply its three-year cost of illness multiplied by the observed annual incidence rate (table 2). The ex ante avoided private costs of illness are thus $0.29 for young children ages 1-4 (range of $0.15-$0.58); $0.09 for school children ages 5-14 ($0.05-$0.19); and $0.10 for adults ages 15 and older ($0.05-$0.20). These amounts are subtracted from the average willingness to pay for vaccine protection to estimate the ex ante value of the mortality risk reduction from cholera vaccination: $1.60 for children ages 1-4 ($1.30-$1.80); $1.20 for children ages 5-14 ($1.10-$1.30); and $1.10 for adults ages 15 and older ($1.00-$1.20). Finally, to convert these ex ante values of mortality risk reduction to values of a statistical life, the cholera case fatality rate is assumed to be 1 percent. The resulting values of a statistical life are $11,000 for children ages 1-4 ($4,300-$25,000), $ 25,000 for school children ages 5-14 ($11,000-$54,000), and $18,000 for adults ages 15 and older ($8,000-$38,000). In reality, the cholera case fatality rate is to 246 THE WORLD BANK ECONOMIC REVIEW percent protection through indirect effects. Beyond 55 percent effectiveness, it is assumed that demand will not change with increasing indirect vaccine effec­ tiveness, similar to the behavior of demand among the vaccinated. Equations (3) and (5), for willingness to pay among vaccinated and unvacci­ nated individuals in age group i, rely on several restrictive assumptions. First, they assume that respondents had a realistic understanding of vaccine effective­ ness based on the language in the contingent valuation survey scenario. To avoid having to explain the concept of vaccine effectiveness in percentage terms to very poor, often illiterate, respondents in the contingent valuation survey in Beira, respondents were told that the hypothetical vaccine would provide "strong protection" in the first year, with declining effectiveness in years two and three. Second, because the contingent valuation survey did not allow measurement of the elasticity of demand with respect to effectiveness, demand among the vaccinated was assumed to be unresponsive to rising total vaccine effectiveness, T'/v' To the extent that demand would increase in response to the increased benefits from herd protection, WTP i,v( Pv, T'/v) may have been underestimated, especially at high coverage (corresponding to a zero or very low price). For the same reasons, WTP i ,lI(Pv,T'/lI) may also have been underestimated for high cov­ erage rates. The cost- benefit calculations presented here are thus probably con­ servative, in that they may underestimate private benefits at high coverage levels, when herd protection is highest. Third, it was assumed that respondents were unaware of herd protection effects when answering the contingent valuation survey questions. s There was no mention of a possible herd protection effect in the contingent valuation scenario, and it seems unlikely that respondents considered this a possibility. In equation (1), the third term characterizes the cost savings to the public health system resulting from implementation of the vaccination program option. Because treatment costs in publicly funded health care institutions are not paid by private individuals, the costs would not be captured in the willingnesHo-pay measures of benefits. The number of cases of illness avoided per person in age group i due to a particular vaccination program option, ai, depends on the fraction of people covered, qi, and therefore also on Pv' For a given vaccination program, as Pv increases and qi decreases, aj also decreases. Thus higher user charges for vaccines translate into a reduction in public cost of illness savings from vaccination. The number of avoided cases per person in highly uncertain and may be considerably lower than 1 percent, which would mean that these derived values of a statistical life estimates are too low. These values for households in Beira are similar to those found in stated preference studies among low-income households in Bangladesh and India (Maskery and others 2008; Bhattarcharya, Alberini, and Cropper 2007). 5. Otherwise, they might have been responding strategically to the contingent valuation questions, anticipating that they could receive indirect benefits from vaccination without having to pay anything. tili ; t! (L[IU" If t j 1.14 " • leuland, Lucas, Clemens, and Whittington 247 each age group ai can then be expressed as: where 7Jv(') and 7Ju(') are again the total and indirect protection levels over a vaccine's duration (in years), d, among vaccinated and unvaccinated individ­ uals, as functions of the overall coverage level in all three age groups, q(pv), as shown in figure 1, and Ii is the cholera incidence rate in age group i (expressed in cases per person per year). The final term of equation (1) describes the costs of vaccination. In some situations vaccination costs may exhibit economies of scale due to high fixed set-up costs. It is also possible that in some settings vaccination costs could exhibit diseconomies of scale as it becomes more and more difficult to vacci­ nate hard-to-reach members in the pool of unvaccinated individuals. The detailed information needed to estimate the costs of vaccination programs in Beira as a function of coverage was not available, so for simplicity, variable vaccination costs were assumed to be independent of coverage level, and no fixed cost for campaign set-up was included. Finally, other research has shown that the additional private cost of vacci­ nation per person arising from traveling to health outposts and waiting in line to receive immunizations is an important determinant of vaccine demand (Jeuland and others 2008). The private components of vaccine cost (cp,;) were also assumed to be the same for all three vaccination program options. Parameter Values and Sensitivity Analysis The values of the parameters in the cost-benefit model (equation (1)) are sum­ marized in table 2. More discussion of these parameter values and sensitivity ranges, as well as their sources, is provided in appendix A. To examine the effect of uncertainty in these parameter values on the cost­ benefit results, three types of sensitivity analysis were undertaken. The first explored the influence of individual model parameters on total net benefits. The upper and lower bounds for the parameters were chosen to reflect reason­ able values based on evidence in the published literature. When available, 95 percent confidence intervals were used; in other cases, judgment was required in specifying these plausible parameter ranges (see appendix A). The second sensitivity analysis studied the effect of simultaneous changes in several par­ ameters. To investigate such changes, "best" and "worst" case scenarios were specified, corresponding to combinations of the extreme parameter values used in the one-way sensitivity analysis, and the net benefits for these extreme scen­ arios were determined. The joint probability of all of the parameters in the benefit-cost calculations taking the best (or worst) case value from the perspec­ tive of maximizing (or minimizing) the net benefits of a vaccination program option, however, is extremely low. To better understand the probability of different economic outcomes, Monte Carlo simulations were conducted. In this TABLE 2. Description of Model Parameters, with Ranges of Uncertainty ~ oc Sensitivity analysis -l Parameter Base case Worst case Best case Notes/source :r: m ~ Population P 550,000 550,000 550,000 na o Incidence Ii (cases per 1,000 Deen and others (200S) " t" o per year) Ages <5 8.S 4.4 17.6 " )­ z Ages 5-14 2.9 1.4 5.7 '" m 1.9 7.7 C'\ Ages 15+ 3.8 o Public cost of illness COl Poulos and others (200S) z o ($ per case) ::: Ages 1-4 13.0 0 26.0 C'\ Ages 5-14 13.0 0 26.0 " m <: Ages 15+ 14.5 0 29.9 Delivery cost ($ per dose) 0.5 2.0 0.3 Lauria and Stewart (200S) '" ~ Production cost 0.6 0.8 0.5 Thiem and others (2006) ($ per dose) Private cost of vaccination cp 0.2 0.4 0.1 Jeuland and others (200S) ($ per dose)" Vaccine effectiveness: no herd 55 40 70 Thiem and others (2006), protection (%) Trach and others (1997) Extent of herd protection effect Reduced by 17% Reduced by 33 % No reduction Adjusted from Ali and in year 3 in all years over time others (2005); see Fig. 1 Coverage if vaccine is free Lucas and others (2007): q;,o(%) b Predicted intercept of Ages 1-4 53 40 86 demand at price 0 Ages 5-14 59 44 65 i Ages 15+ 61 46 67 1 1 i Slope of demand curve f37 Lucas and others (2007): ! Ages 1-4 0.2S 0.40 0.24 Predicted slope of demand Ages 5-14 Ages 15+ 0.47 0.52 0.57 0.62 0.28 0.36 with respect to price I . I f . Per capita willingness ro pay Average per capita willingness WfP; to pay from exponential Ages 1-4 1.9 1.0 3.6 demand model (equation (3)) Ages S-14 1.3 0.8 2.3 Ages lS+ 1.2 0.7 1.9 Per vaccine willingness to pay Average willingness to pay per WTPi •v vaccine, from Ages 1-4 3.6 2.S 4.2 Ages .5 -14 2.2 1.8 3.6 Ages 1S+ 1.9 1.6 2.8 ';;;­ :;: Discount rate 8 (%) 3 10 0 na E"'" na, not applicable. fi­ t-< aEstimated from data collected after the mass vaccination campaign in Beira in 2003 and 2004 with no user fees; children are assumed to have half the :;: private costs of adults. va bBase case coverage level is estimated based on results obtained in surveys with "time-to-think"; lower bound is 2.5 percent lower; and upper bound is the no "time-to-think" estimate. ~ CRange obtained from econometric demand models with 9.5 percent confidence interval on the price coefficient. '" ~ Source: As shown in table. ~ .... ~ j: o ;:t tv ~ \0 250 THE WORLD BANK ECONOMIC REVIEW third type of sensitivity analysis, the probability distribution for each parameter in the model was assumed to be triangular. 6 The bounds for the parameters were set to the low and high values shown in the best and worst cases of Table 2; the base case values shown in Table 2 were assumed to be the most probable outcome of each parameter. IV. RESULTS Cost- benefit results are presented for program options 1-3, with and without the herd protection externality, for three levels of user fees ($0, $1, and $2.20). The economic results without user fees are discussed first, followed by the out­ comes for programs with user fees. Programs with No User Fees The greatest reduction in the burden of disease occurs when no user fees are charged, because both vaccination coverage and the number of people pro­ tected are highest. For all three program options, herd protection greatly increases the reduction in disease burden, but the economic impact of indirect protection varies across program options (table 3). The economic impact of indirect protection is especially large for the two school-based program options. If herd effects are ignored, the net benefits of both school-based pro­ grams are approximately zero (-$16,000 for option 1 and $22,000 for option 2); they rise to $454,000 and $489,000 when herd protection is incorporated in the analysis. The effect of herd protection on the community-based option 3 program is much more modest and does not result in positive net benefits (-$101,000 without herd protection and -$28,000 with it). With herd protection, the benefit-cost ratios for the three options without user fees are 3.3 for option 1, 2.8 for option 2, and 1.0 for option 3, signifying that the benefits per dollar spent are higher for option 1 than for options 2 and 3, even though the total number of vaccinations and the reduction in disease burden are smallest with option 1. To examine the sensitivity of these results for programs without user fees to changes in parameter values, the net benefits of the program options were cal­ culated with and without herd protection using one-way sensitivity analysis, the high and low ranges specified in table 2, and the Monte Carlo simulations. Several observations emerge from this sensitivity analysis. First, the worst-case outcomes result in negative net benefits for all programs, irrespective of whether herd protection is included. For example, including herd protection, the net benefits in the worst case are -$113,000 for option 1, -$183,000 for 6. Triangular distributions are characterized by linearly increasing probability between the lower bound and most probable values, and linearly decreasing probability between the most probable and the upper bound values. d .• 1M t; w w ZA , M. J; I , t.1 • TABLE 3. Outcomes for Vaccination Programs with and without User Fees, in the Base Case and Sensitivity Scenarios (units in thousands of 2005 US$ unless otherwise indicated) Option 1 School children Option 2 All children (Ages 5-14) (Ages 1-14 yrs) Option 3 All ages (Ages 1 +) No herd Herd No herd Herd No herd Herd Programs protection protection protection protection protection protection No user fees Number of vaccinations (thousands) 82 82 112 112 322 322 Reduction in disease burden (%) 6 45 13 57 31 87 Public cost of illness savings 5 42 10 53 30 82 Costs a 181 181 246 246 708 708 Net public costs b 176 139 235 192 679 626 Net benefits -16 454 22 489 -101 28 Benefit-cost ratio" 0.9 3.3 1.1 2.8 0.9 1.0 Worst case net benefits (benefit-cost ratioe ) -262 (0.3) -113 (0.7) -339 (0.3) -183 (0.6) -1,094 (0.3) -1,084 (0.3) ~ ;: Best case net benefits (benefit-cost ratioe ) 29 (1.3) 959 (10.1) 75 (1.5) 1,189 (9.4) 147 (1.3) 1,218 (3.9) ii;'" ;:: Probability that net benefits> Od (%) 15 99 + 29 99 + 6 15 ."'­ Median of net benefits -67 350 -58 370 -330 260 t'-' ;: User fee = $1.00 " Number of vaccinations (thousands) 52 52 74 74 199 199 J; (J Relative reduction in disease burden 63 70 69 76 63 87 c;;­ (% of program with no fees) ~ ~ Public cost of illness savings 3 29 7 40 19 71 §! Costs a 114 114 162 162 437 437 i:l ;:: Net public costs b 59 33 81 48 220 167 "'­ Net benefits 42 375 92 459 144 273 ~ Benefit-cost ratio" 1.3 4.0 1.5 3.6 1.3 1.5 a' ~. Worst case net benefits (benefit-cost ratio") -113 (0.5) -18 (0.9) -150 (0.5) -41 (0.9) -472 (0.4) -424 (0.5) ..... Best case net benefits (benefit-cost ratio") 51 (1.9) 632 (11.6) 97 (2.1) 847 (11.0) 220 (1.9) 1,306 (6.5) 0 ;:: Probability that net benefits> Od (%) 55 100 68 100 50 70 IV v. (Continued) .... TABLE 3. Continued N '"" N Option 1 School children Option 2 All children (Ages 5-14) (Ages 1-14 yrs) Option 3 All ages (Ages 1 +) -I :r t"M No herd Herd No herd Herd No herd Herd Od (%) 87 100 94 100 86 100 Median of net benefits 38 240 79 330 140 360 aNot including private costs of vaccination. bCa!culated as the total cost-revenues from vaccine sales-public cost of illness avoided. Some of these net public costs are negative; for example, the programs are cost saving, because revenues and avoided public cost of illness outweigh the costs of vaccination. "The benefit-cost ratio is dimensionless and defined as the ratio of benefits (willingness to pay for protection avoided public cost of illness) to the cost of acquiring vaccines in the program (whether private or through public subsidy). dBased on the outcomes of 10,000 Monte Carlo trials. Source: Authors' analysis based on data described in the text. • leuland, Lucas, Clemens, and Whittington 253 option 2, and -$1.08 million for option 3). However, the Monte Carlo simu­ lations suggest that poor economic outcomes are likely only for the community-based program (85 percent of simulations have negative net benefits), whereas the school-based program options have positive net benefits in 99 percent of the simulations. The worst-case outcomes depend on the con­ vergence of an unlikely set of factors: low herd protection, low demand for vaccines, low cholera incidence and avoided cost of illness savings, high vacci­ nation cost, and high private cost for obtaining vaccines. Second, the parameter uncertainty contributing most to the range of econ­ omic outcomes is that associated with the cost of vaccination, for all three pro­ grams. Figure 2 presents the results of the one-way sensitivity analyses for these options, accounting for herd protection (the equivalent diagrams ignoring herd protection are available in the online supplement at http://wber.oxfordjournal­ s.org/). Vaccination cost is particularly important for the community-based program option because it delivers the most vaccines. Third, as shown in figure 2, the influence of the extent of herd protection decreases as program size increases. For option 1, this parameter is the third most important in influencing the range of final outcomes; for option 2, it is fifth; and for option 3 it is not among the top eight parameters. This results from the interaction between cholera vaccine herd protection and the demand for protection from the disease in the population affected by each program option. In option 3 every individual with nonzero demand for cholera vaccines is assumed to be vaccinated. The remaining unprotected individuals have zero willingness to pay for protection, so that the economic benefits of indirect pro­ tection are very low (they include only public cost of illness savings and protec­ tion to infants, who cannot be vaccinated). In option 2 all children for whom willingness to pay for protection is nonzero are vaccinated, but adults and infants excluded by the program design are indirectly protected at a level of more than 40 percent, and the willingness to pay for protection among these excluded individuals is nonzero. The per person impact of this externality increases further in program option 1, as another program-excluded group of unvaccinated people (young children) with nonzero average willingness to pay for protection are also protected at no cost (the overall indirect effect is now about 25 percent). The importance of this indirect effect is further shown by the fact that the parameters of the demand function for adults are second in importance in all three programs, even though they are vaccinated only in option 3. Programs with User Fees The cost-benefit results for cholera vaccination program options are also calculated at two different user fee levels: a subsidized price of $1 for the two-dose regime and $2.20, which corresponds to the full cost of production plus delivery in the base case (see table 3). Based on the demand curve from Lucas and others (2007), slightly fewer than 40 percent of people in Beira are 254 THE WORLD BANK ECONOMIC REVIEW FIGURE 2. Tornado Diagrams Showing the Eight Parameters Most Important in Determining Variation in Net Benefit Outcomes for Programs without User Fees, Including Herd Protection for Program Options 1-3 Option 1 Total net benefits (2005 US$) 200,000 400,000 600,000 800,000 Vaccine cost; production + delivery ($) Demand slope (adults ages 15+) Extent of herd protection (fraction of observed effect) Demand slope (school children ages 5-14) Demand intercept (school children ages 5-14) Demand intercept (adults ages 15+) Demand intercept (young children ages 1-4) Demand slope (young children ages 1·4) Option 2 Total net benefits (2005 US$) o 200,000 400,000 600,000 800,000 Vaccine cost: production + delivery ($) Demand slope (adults ages 15+) ~, i I J Il'~$ ,.: Demand slope (school children ages 5-14) -{).57 '-{).28 Demand intercept (adults ages 15+) Extent of herd protection (fraction of observed effect) 0.7 Demand intercept (young children ages 1-4) Public cost of illness (adults ages 15+) Demand slope (young children ages 1-4) Option 3 Total net benefits (2005 US$) -1 ,200.000 -800,000 -400,000 0 400,000 Vaccine cost: production + delivery ($) 5.6 la l 6 . Demand slope (adults ages 15+) I, -0.62"'.-{).36 Time costs ($) ! 0.8.0.2 Demand slope (school children ages 5-14) .,: -{).57., -{).28 Public cost of illness (adults ages 15+) I 0"30 Cholera incidence (adults ages 15+) ; 1.9j 7.7 Vaccine duration (years) 2 ~ 4 Demand slope (young children ages 1-4) --:():~!-:()~~~ Source: Authors' analysis based on data from a cholera vaccination trials and epidemiological and economic research in Beira, Mozambique; see text for details. i1$ t .41 ;lEN 5 4 feuland, Lucas, Clemens, and Whittington 255 estimated to be willing to pay a price of $1 for the two-dose cholera vaccine regime in a community-based program; that share drops to about 20 percent at the $2.20 price. In the base case, net benefits for all three program options are positive for both these user fee levels, even if herd protection is ignored. Overall disease reduction levels drop significantly for the smaller school-based programs. There is a crucial difference between the economics of option 3 and the two school-based programs when herd protection is included. Whereas net benefits of option 3 increase as the user fee increases, net benefits decrease for the two school-based programs. Figure 3 shows in more detail how social costs, benefits, and net benefits change over a range of user fees from zero to $5, assuming base case par­ ameter values, for all three program options, with and without herd protec­ tion. When herd protection is not included, net benefits are maximized very close to the full economic cost of vaccination, for all three programs ($2.60, including private costs), because public cost of illness savings are relatively small. Also, for option 3, net benefits peak very close to this full economic cost of vaccination when herd protection is included (the maximum is at roughly $2.45). This is because benefits to the unvaccinated continue to increase as the user fee increases well past the full cost of vaccination, such that total benefits stay relatively high even as cost decreases. In contrast, total benefits tend to fall faster than total costs for the school-based options (note, however, that net benefits do increase at first for option 2 up to a user fee of about $0.15), because herd protection in the untargeted popu­ lation is lost as coverage declines. In option 1, for example, only 3 percent of the population are vaccinated when the user fees are $1 and 2 percent when they are $2.20, so the herd protection is low. Because of the very different sizes of the program options, it is easier to compare them by examining the benefit-cost ratios, which give an indication of the magnitude of benefits of vaccination relative to its costs. These calcu­ lations show that the smallest program achieves the highest benefit-cost ratios. This stems from the fact that the small programs both target people for whom household demand for cholera vaccines is high and take advantage of the fact that marginal herd protection benefits increase most quickly at low coverage levels. All school-based programs have a benefit-cost ratio of about 3 or higher in the base case, indicating that these programs are economically very attractive. The sensitivity analysis of the programs with user fees shows that the impor­ tance of the cost parameter decreases as the user fee goes up (figure 4). Targeting individuals with higher demand for the vaccine begins to offset the downside risk associated with cost uncertainty. School-based programs are likely to pass a cost-benefit test in Beira even for very pessimistic cost assump­ tions. High vaccination costs would have to be combined with unfavorable values of other parameters for school-based programs to fail. For the community-based programs, however, the risk is quite high, although it 256 THE WORLD BANK ECONOMIC REVIEW FIGURE 3. Total Costs, Benefits, and Net Benefits As a Function of User Fees for All Three Program Options, with and without Herd Protection No herd protection, program option 1 "Average" herd protection. program option 1 0.0 1.0 2.0 3.0 4.0 5.0 0.0 10 2.0 3.0 4.0 S.O User fee ($) User fee ($) No herd protection, program option 2 "Average'" herd protection, program option 2 800 --------~--~----- . ~.,. - ~"- O+-----T------~----r_-- 0.0 10 2.0 3.0 4.0 5.0 User fee ($) User fee ($) No herd protection, program option 3 "Average" herd protection, program option 3 BOO _ BOO 51 56 1.28 1.48 Source: Ali and others (2005). 264 THE WORLD BANK ECONOMIC REVIEW epidemiological models. Their conclusions are based on a reanalysis of first­ year incidence data among vaccinated and unvaccinated people from baris with different levels of coverage in Matlab, Bangladesh. For simplicity and use in the cost-effectiveness models, their data were interpreted and applied (table B-1, columns 1, 3, and 4), as explained below. The average coverage rates shown were calculated based on their data (column 2). The data however, apply only to first-year herd protection effects. Because oral cholera vaccine efficacy declines over time, use of the data for longer-term cost-effectiveness and policy analysis requires the adjustments described in the text of this article so as not to overstate the health benefits of vaccination. Herd Protection as an Exponential Function of Coverage The incidences for vaccinated and unvaccinated populations are modeled with a set of two differential equations. The first predicts incidence among the vacci­ nated (V) as a function of coverage (x) only. The second, for the incidence among the unvaccinated (U), is similar but specifies that incidence can never be higher among the vaccinated subgroup than among the unvaccinated: (B-1) Ix,v Ioov . exp( -kvx), (B-2) Ix,u = Io,u . exp( -kvx) + (Io,u - Io,v) . exp( -kux) where kv and ku are rate constants. The parameters for equation (B-1) can be estimated with a simple ordinary least squares regression of the coverage data in table B-1 on the log of incidence rates among the vaccinated. The intercept (Io,vl is 4.5 cases per 1,000 (p 0.02) and the rate constant (k y ) is -0.02 (p = 0.09). Io,v would be the incidence for a hypothetical vaccinated individual if no vaccines were given to the population-that is, if there were no herd protection at all. As shown, incidence Ix,y declines at a decreasing rate as coverage rates increase. The R2 for the regression is 0.66, which is similar to the R2 for a linear model (0.69). Given the estimated parameters for Io,v, the parameters of equation (6) for avoided cases are then estimated with a simple nonlinear least squares model. The R 2 for this model is 0.97 (higher than the 0.86 obtained for a linear model), and the intercept (lo,u) is significant at the 10 percent level (p 0.07), though the rate constant (ku) has a p-value of 0.14. Io,u is the baseline inci­ dence among the unvaccinated-it represents incidence among unvaccinated individuals given a coverage level of zero. Without loss of generality, equation (B-2) can be rewritten as: (B - 3) Ix.u Io,u . exp( -kux). Figure B-1 plots the observed MarIab data against the exponential fits. As cov­ erage increases, the incidence for unvaccinated individuals approaches that for the vaccinated subgroup. M I X . II!! Q4I'1 It j leuland, Lucas, Clemens, and Whittington 265 FIGURE B-1. Observed Incidences in Matlab, Bangladesh, and Predictions from Exponential Fit 16 14 ~ .9:! 12 _ _ Observed incidence among vaccinated people Q. 0 ~ 8. 10 _ _ Observed incidence among unvaccinated people 0 0 C!. ..... ... 8 , . - Predicted incidence among vaccinated people - - - - Predicted incidence among unvaccinated people I 8. Q) t.I 6 c: Q) '0 '13 4 .E 2 -. ""lio .;. -.. 0 0 20 40 60 80 100 Coverage (%) Source: Authors' analysis based on data from Ali and others (2005). The herd protection among the vaccinated (7Jx,v) and the unvaccinated (7Jx,u) can be expressed as the difference between the baseline incidence (lo,u) and the incidence in the group in question, as shown in equations (B-4) and (B-5); IO'll lo,v ' exp( -kvx) Io,v (B-4) 7Jx,v = ----~-I------ = 1 - I ' exp( -kvx ), O,ll _ O,ll lo,u -lo.ll ,exp( -ku x ) 1 (k-) (B 5) 7Jx,u I·· = - exp - uX , O.u ACKNOWLEDGMENTS This research program works to accelerate the development and introduction of new-generation vaccines against cholera, typhoid fever, and shigellosis, The results of the program support public decision-making on immunization pro­ grams for cholera, typhoid fever, and shigellosis. The research in Beira, Mozambique, could not have been completed without the work of the many individuals associated with the data collection effort in Mozambique, including Juvenaldo Amos, Alberto Baptista, Avertino Barreto, Fauzia Ismael, Arminda Macuamule, and Francisco Songane; members of the Medecins Sans Frontieres and World Health Organization teams Sonia Ampuero, Philippe Cavailler, Claire-Lise Chaignat, Philippe Guerin, Pierre Kahozi-Sangwa, Bruno Lab, Dominique Legros, Claude Mahoudeau, Margaret McChesney, Thomas 266 THE WORLD BANK ECONOMIC REVIEW Nierle, and Valerie Perroud; and IVI team members Mohammed Ali, Jacqueline Deen, Andrew Nyamete, Mahesh Puri, Lorenz von Seidlein, and Xuan-Yi Wang. The authors also thank Joseph Cook, Donald Lauria, and Brian Maskery for their contributions to the model developed to assess the economics of vaccination programs and the modeling of the herd protection effect, and the journal editor and three anonymous referees for helpful comments. FUNDING This work was supported by the Diseases of the Most Impoverished program, administered by the International Vaccine Institute with support from the Bill & Melinda Gates Foundation. REFERENCES Ahituv, A., V. Hotz, and T. Philipson. 1996. "The Responsiveness of the Demand for Condoms to the Local Prevalence of AIDS." The Journal of Human Resources 31(4):869-97. Ali, M., M. Emch, L. von Seidlein, M. Yunus, D.A. Sack, M. Rao, J. Holmgren, and J. Clemens. 2005. "Herd Immunity Conferred by Killed Oml Cholem Vaccines in Bangladesh: A Reanalysis." The Lancet 366(9479):44-9. Barua, D., and W.B. Greenough. 1992. Cholera. New York: Plenum. Bhattacharya, Soma, Anna Alberini, and Maureen L. Cropper. 2007. "The Value of Mortality Risk Reductions in Delhi, India." Journal of Risk and Uncertainty 34(1):21-47. Carson, R.T. 2000. "Contingent Valuation: A User's Guide." Environmental Science and Technology 34(8):1413-8. Cookson, S.T., D. Stambouli an, J. Demonte, L. Quero, C.M. DeArquiza, A. Aleman, A. Lepetic, and M.M. Levine. 1997. "A Cost-Benefit Analysis of Programmatic Use of CVD 103-HgR Live Oral Cholera Vaccine in a High-Risk Population." International Journal of Epidemiology 26(1):212-9. Deen, J.L., L. von Seidlein, D. Sur, M. Agtini, M.E.S. Lucas, A.L. Lopez, D.R. Kim, M. Ali, and J.D. Clemens. 2008. "The High Burden of Cholera in Children: Comparison of Incidence from Endemic Areas in Asia and Africa." PLoS Neglected Tropical Diseases 2(2):e173. doi:10.13711 journal. pntd.0000173. Gersovitz, M. 2000. "A Preface to the Economic Analysis of Disease Transmission." Australian Economic Papers 39(1):68-83. Gersovitz, M., and J. Hammer. 2003. "Infectious Diseases, Public Policy, and the Marriage of Economics and Epidemiology." The World Bank Research Observer 18(2):129-57. Hanemann, W.M. 1994. "Valuing the Environment through Contingent Valuation." Journal of Economic Perspectives 8(4):19-43. Jamison, D., J. Breman, A. Measham, G. Alleyne, M. Claeson, D.B. Evans, P. Jha, A. Mills, and P. Musgrove eds. 2006. Disease Control Priorities in Developing Countries. New York: Oxford University Press. Jeuland, M., M. Lucas, J. Deen, N. Lazaro, and D. Whittington. 2008. "Estimating the Private Benefits of Vaccination against Cholera Using Travel Costs." Working Paper. Department of Environmental Sciences and Engineering, University of Notth Carolina, Chapel Hill. Lauria, D., and J. Stewart. 2008. "The Costs of Vaccination Programs.» Working Paper. Department of Environmental Sciences and Engineering, University of North Carolina, Chapel Hill. _4 ill 4 ? Jeuland, Lucas, Clemens, and Whittington 267 Lauria, D.T., B. Maskery, C. Poulos, and D. Whittington. 2009. "An Optimization Model for Reducing Typhoid Cases in Developing Countries without Increasing Public Spending.» Vaccine 27(10):1609-2l. Longini, l.M., A. Nizam, M. Ali, M. Yunus, N. Shenvi, and J.D. Clemens. 2007. "Controlling Endemic Cholera with Oral Vaccines." PLoS Medicine 4(11):1776-83. Lucas, M., M. Jeuland, M. MacMahon, D. Whittington, and A. Nyamete. 2007. "Private Demand for Cholera Vaccines in Beira, Mozambique." Vaccine 25(14):2599-609. l.ucas, M.E.S., ].L. Deen, L. von Seidlein, X.-Y. Wang, J. Ampuero, M. Puri, and M. Ali and others. 2005. "Effectiveness of Mass Oral Cholera Vaccination in Beira, Mozambique." New England Journal of Medicine 352(8):757-67. Maskery, B., Z. Islam, J. Deen, and D. Whittington. 2008. "An Estimate of the Economic Value That Parents in Rural Bangladesh Place on Ex-ante Mortality Risk Reductions for Their Children." Working Paper. Department of Environmental Sciences and Engineering, University of North Carolina, Chapel Hill. Murray, J., D.A. McFarland, and R.J. Waldman. 1998. "Cost-Effectiveness of Oral Cholera Vaccine in a Stable Refugee Population at Risk for Epidemic Cholera and in a Population with Endemic Cholera." Bulletin of the World Health Organization 76(4):343-52. Naficy, A., M.R. Rao, C. Paquet, D. Antona, A. Sorkin, and J.D. Clemens. 1998. "Treatment and Vaccination Strategies to Control Cholera in Sub-Saharan Refugee Settings-A Cost-Effectiveness Analysis." Jama-Journal of the American Medical Association 279(7):521-5. Naidoo, A., and K. Patric. 2002. "Cholera: A Continuous Epidemic in Africa." The Journal of the Royal Society for the Promotion of Health 122(2):89. Philipson, T. 1996. "Private Vaccination and Public Health: An Empirical Examination for U.S. Measles." The Journal of Human Resources 31(3):611-30. Poulos, c., A. Riewpaiboon, J.F. Stewart, A. Nyamete, S. Guh, J. Clemens, and S. Chatterjee and others.2008. "Costs of Illness Due to Endemic Cholera." Working Paper. Research Triangle Institute, Research Triangle Park, NC. Ryan, E.T. 2000. "Mortality, Morbidity, and Microbiology of Endemic Cholera among Hospitalized Patients in Dhaka, Bangladesh." The American Journal of Tropical Medicine and Hygiene 63( 1): 12-20. Sack, D.A., R.B. Sack, G.B. Nair, and A.K. Siddique. 2004. "Cholera." The Lancet 363(9404):223-33. Sinha, A., O. Levine, M.D. Knoll, F. Muhib, and T.A. Lieu. 2007. "Cost-Effectiveness of Pneumococcal Conjugate Vaccination in the Prevention of Child Mortality: An International Economic Analysis." The Lancet 369(9559):389-95. Thiem, V.D., J.L. Deen, L. von Seidlein, D.G. Canh, D.D. Anh, and J.K. Park. 2006. "Long-Term Effectiveness against Cholera of Oral Killed Whole-Cell Vaccine Produced in Vietnam." Vaccine 24(20):4297-303. Trach, D., J. Clemens, N. Ke, H. Thuy, N. Son, D.G. Canh, P.Y. Hang, and M.R. Rao. 1997. "Field Trial of a Locally Produced, Killed, Oral Cholera Vaccine in Vietnam.» The Lancet 349(9047):231-5. UNICEF (United Nations Children's Fund). 2003. Reference Vaccine Price List. New York (www.unicef.orglsupply/index_immunization.html). Whittington, D., O. Matsui-Santana, ].J. Freiberger, G. Van Houtven, and S. Pattanayak. 2002. "Private Demand for a HIV/AIDS Vaccine: Evidence from Guadalaiara, Mexico." Vaccine 20(19/20):2585 -91. Whittington, D., D. Sur, J. Cook, S. Chatterjee, B. Maskery, M. Lahiri, C. Poulos, and others. 2009. Rethinking cholera and typhoid vaccination policies for the poor: Private demand in Kolkata, India. World Development 37(2):399-409. WHO (World Health Organization). 2005. "Cholera, 2004." Weekly Epidemiological Record 80(31):261-8. I I I I I I I I I _, 4 ) if! $l $ ,tb .*' Do Exporters Pay Higher Wages? Plant-level Evidence from an Export Refund Policy in Chile Ivan T. Kandilov The impact of increased export activity on plant wages is estimated in a developing country context. To avoid potential endogenous selection problems, the empirical analysis benefits from exogenous variation in exports induced by a policy experiment­ an export subsidy system implemented in Chile in 1986. Analyses using data from a panel survey of Chilean manufacturing establishments show that while the export subsidy had only a modest positive impact on the industrywide relative high-skilled wage, it significantly increased the plant-level relative high-skilled wage in medium-size establishments, which arc most likely to take advantage of the subsidy and enter the export market. JEL codes: 015, F16, J30, F14 According to the factor proportions theory of international trade (Heckscher-Ohlin), the relative high-skilled wage is expected to fall after trade liberalization in a developing country that is abundant in low-skilled labor. However, most empirical studies focusing on developing countries offer the opposite conclusion-wage inequality (between high- and low-skilled workers) actually rises with entry into the world market (Beyer, Rojas, and Vergara 1999; Hanson and Harrison 1999a, b; Gindling and Robbins 2001; Goldberg and Pavcnik 2007).1 One potential cause for the gtowth in inequality following trade liberalization is that the increase in export activity may lead to a rise in exporters' relative high-skilled wage. The superior performance of exporters over firms producing only for the domestic market is a well-established empirical regularity in international trade (Bernard and Jensen 1997, 1999). Firms selling abroad tend to be larger and Ivan T. Kandilov is an assistant professor of agricultural and resource economics at North Carolina State University; his email address is ivan_kandilov@ncsu.edu. The author thanks Charlie Brown, Jim Levinsohn, Daniel Hamermesh, John DiNardo, Joshua Linn, Peter Debaere, and Amy Kandilov for all their support, comments, and helpful conversations. The author also thanks the journal editor, three anonymous referees, and seminar participants at the University of Michigan, Indiana University, Wellesley College, George Washington Universiry, Miami University of Ohio, and North Carolina State University for useful comments. He is also grateful to Jim Levinsohn for sharing the data. 1. Robertson (2004) is an exception-he finds that the relative skilled wage in Mexico fell after the country joined the North American Free Trade Agreement in 1994. THE WORLD BANK ECONOMIC REVIEW, VOL. 23, No.2, pp. 269-294 doi:l0.1093/wberllhp004 Advance Access Publication June 6, 2009 The Author 2009. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development I THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 269 270 THE WORLD BANK ECONOMIC REVIEW more productive and to pay higher wages, especially to high-skilled workers. It is also documented that exporters are "better" than nonexporters even before entering foreign markets-Bernard and Jensen (1999, p. 3) find that" ... good firms do become exporters. , .. " Hence, any attempt to identify the effect of exports on wages is hindered by endogenous selection. The trouble from an econometric perspective is that the larger, more productive, and higher-wage plants are the ones choosing to enter the export market. This selection process is described theoretically in Melitz (2003). A cross-sectional study relating plants' export behavior to wages may therefore arrive at the wrong conclusions about the effects of exports on wages. This article shows empirically that increased export activity in a developing country can lead to a higher relative high-skilled wage. 2 The impact of increased foreign sales on wages is estimated by taking advantage of exogenous variation in exports induced by an export subsidy that the government of Chile introduced in 1986. The Chilean policy allowed producers in industries export­ ing less than a government-specified threshold level to receive a subsidy of 10 percent of the value of their exports. This interesting policy experiment helps the research design avoid potential endogenous selection issues that could bias the estimate of the effect of exports on wages. To investigate the impact of the subsidy on the relative white-collar wage, a Chilean plant-level manufacturing panel dataset covering 1979-96 is used. Levinsohn (1999), Pavcnik (2002, 2003), and Lopez and Namini (2006) have used shorter versions of the same panel. No one, however, has yet exploited eli­ gibility for the export subsidy as a source of exogenous variation in export activity. Levinsohn (1999) and Pavcnik (2002) used the data for 1979-86 to investigate plant-level employment and productivity changes, respectively, in response to tariff reductions in the early 1980s. Pavcnik (2003) employed the 1979-86 panel to explore skill upgrading in Chilean plants. Lopez and Namini (2006) used the 1990-99 panel to provide evidence on self-selection into exporting markets. To circumvent the endogenous selection problem, previous research has used exchange rate shocks as a source of exogenous variation in exports. Verhoogen (2008) takes advantage of the Mexican peso devaluation of 1994 to test his theory of trade, wage inequality, and product quality upgrading. As there is no cross-sectional variation in the export incentive, he evaluates the differential impact of the Mexican peso devaluation on the wages of larger and more pro­ ductive plants. He finds evidence that following the crisis of 1994, white- and blue-collar wages as well as the skill premium increased more in more pro­ ductive plants than in less efficient ones. 2. The terms high-skilled and white-collar labor, as well as low-skilled and blue-collar labor, are used interchangeably, The data section contains an outline of the broad subcategories included in the white- and blue-collar labor aggregates used in the Chilean data. t, _ l 4. Kandilov 271 This article uses the cross-sectional nature of the Chilean export subsidy to evaluate its impact by comparing wages in two groups of industries before and after one group was granted access to the subsidy. The main empirical findings suggest that industry exports increased strongly in response to the subsidy. While the average plant-level relative white-collar wage rose only slightly fol­ lowing adoption of the export subsidy, the evidence suggests that the impact of the subsidy was quite uneven across plants. The relative white-collar wage rose significantly more in medium-size establishments-those most likely to enter the export market in response to the subsidy-than in large plants, which are more likely to increase their already existing export sales, or in small plants, which are least likely to take advantage of the export incentive. More broadly, this article contributes to the empirical literature on inter­ national trade and plant-level performance. Most of the studies in this literature find little change in plant behavior in response to increased international trade. Levinsohn (1999) finds little impact of trade liberalization on labor markets; Clerides, Lach, and Tybout (1998) estimate no change in firms' cost structure due to increased export activity; Bernard and Jensen (1999) document no differ­ ences in productivity and wage growth for plants when they start exporting. Using an employer-employee matched data set from Morocco, Fafchamps (forthcoming) also fails to find evidence that exporters pay higher relative skilled wage, conditional on plant size. This article, along with Verhoogen (2008), is an exception to many previous investigations, perhaps because the empirical analysis here benefits from exogenous variation in the export incen­ tive. It finds evidence that as exports increase, plant-level relative high-skilled wages rise for establishments that are most likely to enter the export market. The article is organized as follows. Section I highlights some theoretical reasons why firms might increase (relative) wages as they start serving foreign markets. Section II describes the data and presents summary statistics. Section III outlines the Chilean export subsidy and its use in identifying the impact of increased export activity on wages. Results and discussion follow in section IV. Section V offers some concluding remarks. I. WHY MIGHT EXPORTERS PAY HIGHER WAGES? Theoretically, three mechanisms could potentially explain why rising export activity would lead to higher wages: quasi-rents due to industry- and exporting-specific skills, labor quality upgrading, and efficiency wages. Any of these channels could contribute to higher wages in the export-subsidized manu­ facturing industries in Chile. If the export subsidy increases exports without reducing production for dom­ estic consumption, demand for workers in industries granted the subsidy will rise. Because of industry-specific skills and training, workers in the export­ subsidized industries would extract quasi-rents as demand for their services increases but their supply is relatively fixed, at least in the short run. Even with 272 THE WORLD BANK ECONOMIC REVIEW substitution away from domestic market production toward exports, the presence of both industry- and export-specific skills-such as knowledge of a foreign language or ability to operate sophisticated, new, or foreign machinery-would raise demand for workers with such skills, allowing them to extract quasi-rents. Another reason for higher wages in the subsidized industries could be labor force upgrading as a result of increased demand for the exporting-specific skills. With non-negligible separation costs, upgrading of the quality of labor hired is likely to take place in the medium to long run and to result in higher wages for workers in industries granted the export subsidy. Yet another reason for higher wages in exporting establishments might be that exports require greater worker effort. Exporters would pay higher wages to reduce shirking and to increase care in production. This argument, in the spirit of Shapiro and Stiglitz (1984), involves imperfect monitoring and is developed theoretically in Verhoogen (2008), whose general framework is con­ sistent with all three mechanisms discussed here. The model assumes differ­ ences in preferences for product quality between developed and developing countries, with consumers in developed countries valuing quality more than consumers in developing countries (see also Murphy and Shleifer 1997). Production of quality in the model is sensitive to workers' effort, especially to high-skilled workers' effort. Therefore, to start exporting, firms in developing countries like Chile need to upgrade output quality and hence to pay higher (efficiency) wages, especially to high-skilled workers, to elicit greater effort. In this environment, an export subsidy raises exports and product quality and leads to higher wages and a higher skill premium. Finally, there are reasons to expect that the subsidy could provide a greater incentive to enter foreign markets for large and medium-size plants than for small establishments. Empirical research (Roberts and Tybout 1997) has shown that nonexporters incur substantial sunk costs to enter the export market; these costs are incorporated theoretically in Melitz (2003) and in Yeaple (2005). Because larger and more productive establishments are more profitable in the export market, they are more willing to pay the sunk costs to enter foreign markets in response to an incentive such as the Chilean export subsidy. Further, recent theoretical advances (Verhoogen 2008; Bustos 2007) show that there is a greater effect on wages for plants that enter the export market (extensive margin) than for plants that increase the volume of already positive exports (intensive margin) in response to an export incentive. Because large establishments are more likely to have been exporters before the subsidy is introduced than are small and medium-size plants, wages in large establish­ ments are less likely to be affected than wages in medium-size establishments, which are most likely to enter the export market following the subsidy. Overall, the introduction of an export subsidy would be expected to boost wages and the relative high-skilled wage as a result of increased export activity. However, the effect on wages may be uneven across plants, with the greatest impact in establishments that enter the export market. us .-,* t t ... g - j tilU Kandilov 273 II. DATA The empirical analysis is based primarily on the Chilean Annual National Industrial Survey (ENIA), a panel survey carried out by the National Statistical Institute of Chile, for the 18 years from 1979 to 1996. The survey covers the universe of all Chilean manufacturing establishments with at least 10 employ­ ees. Surveying about 4,800 plants a year, it includes information on approxi­ mately 11,000 plants, for a total of more than 80,000 plant-year observations over the time span of the panel. The National Statistical Institute updates the survey annually by incorporating establishments founded during the year and excluding plants that stopped operating for any reason. 3 An establishment is not necessarily a single-plant firm. However, it is reason­ able to assume that most plants in the survey are single-plant firms. The data show that only about 5 percent of establishments purchase materials from other establishments within the firm, and Pavcnik (2002) suggests that about 90 percent of the Chilean manufacturing firms have only one plant. For each establishment, the survey collects data on production, value added, sales, employment and wages, exports, investment, energy consumption, balance sheets, and other plant characteristics. The white-collar labor aggre­ gate, unlike its wage counterpart, is divided into three subcategories: white­ collar executives, white-collar administrative workers, and white-collar pro­ duction workers. The data on blue-collar workers, again unlike its wage counterpart, is subdivided into blue-collar production workers and blue-collar nonproduction workers.4 Data on plant-level exports were collected only after 1990, and balance sheet information, which includes the value of capital stock, is available only for 1980 (or 1981).5 A plant's capital stock for years after 1980 is calculated by the perpetual inventory method based on initial book value in 1980 (or 1981) and only for plants present in the survey in 1980 (or 1981).6 Plant location was collected only for the first three years of the survey, but more than half of manufacturing output in Chile is produced in the Santiago metropolitan region. All plants are classified into 29 three-digit inter­ national standard industrial classification (lSIC, revision 2) industries and then further subdivided into 89 four-digit ISIC industries. As expected, there are big differences between exporters and nonexporters (table 1). Consistent with previous research (Aw and Batra 1999), Chilean 3. If an establishment contracts in size to fewer than 10 employees, it is no longer included in the Chilean survey and would be considered a «false" exit. Fajnzylber and Maloney (2005) report that for a similar dataset from Colombia, which does not censor for plants with fewer than 10 employees, about 36 percent of plants that fall below the la-employee cut-off continue to operate and would be considered false exits. 4. Nearly 90 percent of blue-collar workers are production workers, and more than 75 percent of white-collar workers are executives or administrative workers (table 1). S. Establishment age is not recorded. For all plants present in the first year of the survey (1979), age is imputed by assigning an age of one in 1979. Age is therefore measured with error. 6. See Liu (1993) for detailed description of the construction of the capital measure in this dataset. N " .j>. TABLE 1. Summary Statistics on Plant Characteristics ..; :I; Plant characteristic All plants (1979-96) Nonexporters (1990-96) Exporters (1990-96) No subsidy (1985) Subsidy (1985) '" t:! ~ Log white-collar wage Log blue-collar wage 6.20 5.51 6.23 5.62 6.92 5.97 6.14 5.29 5.90 5.21 0 )d ,... " I I Log relative white-collar wage 0.69 0.61 0.95 0.85 0.69 '" ;.- j z i Log total employment 3.67 3.56 4.75 3.67 3.54 ;-: White-collar labor (fraction of total) Log total sales Age (years) 0.21 9.95 7.19 0.21 9.81 10.24 0.26 11.74 10.72 0.20 9.80 5.98 0.19 9.66 6.23 '" (") 0 z 0 il: I White-collar employees White-collar executives White-collar administrative White-collar production workers White-collar female workers 18.82 0.18 0.58 0.24 0.32 13.72 0.21 0.58 0.21 0.38 53.67 0.16 0.55 0.28 0.28 17.15 0.19 0.55 0.26 0.23 14.22 0.17 0.61 0.22 0.34 (") )d '" <:: '" t:! I 1 # Blue-collar employees 53.44 40.27 139.11 55.00 43.07 Blue-collar production workers 0.88 0.89 0.88 0.90 0.86 Blue-collar nonproduction workers 0.12 0.11 0.12 0.10 0.14 Blue-collar female workers 0.19 0.19 0.21 0.10 0.22 Fraction exporters' (1990-96) 0.22 Exports (fraction of sales) (1990-96) 0.28 FrExpSub jt 0.96 (0.10) 0.87 (0.15) 1.00 (0.00) FrExpSubW;t 0.83 (0.31) 0.43 (0.33) 1.00 (0.00) Number of observations 78,321 25,218 28,382 1,139 2,860 t Note: Numbers in parentheses are standard deviations. All numbers are averages. Values for subcategories are fractions of the main category. aExporters are defined as plants with positive sales abroad. I Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey (ENIA); see text for details. iI i i -~ Kandilov 275 manufacturing exporters are larger than nonexporters, as measured by sales or total employment, and they pay higher wages, especially to white-collar workers? Although comparisons of average wages between exporters and non­ exporters are suggestive of the estimate of the effect of exports on wages, such comparisons may be contaminated by endogenous selection if initially higher-wage plants choose to enter the export market. The next section pre­ sents the econometric strategy that deals with such selection issues. III. ECONOMETRIC STRATEGY To identify the effect of increased export activity on wages, this article takes advantage of the exogenous variation in exports induced by the Chilean export subsidy implemented in 1986 and still in effect. The industry- and plant-level econometric models are based on the comparison of wages between two groups of plants before and after one group is granted access to an export subsidy. In essence, this is a difference-in-differences framework, with the group of plants given access to the export subsidy in 1986 and thereafter serving as the treatment group and the others as the control group. Data Preparation On December 19, 1985, the Chilean government established a system of refunds for exporters of nontraditional exports. 8 This policy was adopted partly in response to the severe recession of 1982-83 in Chile-the government sought to diversify exports and stimulate growth through export orientation. Intended to refund duties paid by exporters on imported materials, the program instead simply refunds a percentage of the value of exports-10 percent at the time the system was established. Only products whose export value averaged less than US$2.S million for the two years before 1985 qualified for the rebate at its inception in 1986. 9 The Ministry of Economic Affairs, Development, and Reconstruction was responsible for establishing an annual list of products excluded from the benefit for the following year-referred to here as the "exclusion list." This list remained unchanged for the first three years and changed little thereafter. In 1989, the government expanded eligi­ bility, offering a 5 percent rebate to a small number of previously excluded pro­ ducts. In 1992, a 3 percent export refund was implemented for a few product codes previously excluded from the 10 percent and 5 percent rebates. Because these subsidies were much smaller and applied to only a few product codes, they are ignored. If the export subsidy had an effect on wages, this exclusion 7. Plant-level average wage is a measure of daily wages in constant 1979 Chilean pesos per worker, as information is available only on days of operation, but not on hours. 8. Congress of Chile (Congreso Nacional de Chile) Law 18480 (www.congreso.c1). 9. This threshold (later applied for years after 1986) increased progressively from US$2.5 million to $5 million to $7.5 million toward the end of the sample period. 276 THE WORLD BANK ECONOMIC REVIEW will bias the estimates slightly toward zero, since a few plants that received access to a small export subsidy are classified as ineligible. A concordance was employed to translate products on the exclusion list, which identifies products by the six-digit harmonized system (HS), into indus­ tries at the four-digit ISIC classification used in the Chilean Annual National Industrial Survey (Hoekman, Mattoo, and English 2002). Twenty-one four­ digit ISIC industries contain at least one product that is ineligible for the export subsidy, and the average for this set of 21 industries is seven excluded products. All products in the remaining 68 industries are allowed the 10 percent refund for their exports. This information is used to construct an industry measure of export subsidy eligibility, FrExpSub, which is equal to the fraction of products in each indus­ try eligible for the subsidy. For the 68 industries that produce only eligible pro­ ducts, FrExpSub is equal to one. For the 21 industries that produce at least one ineligible product, FrExpSub is some fraction between zero and one, with an average of 0.87 and a standard deviation of 0.15. Although only a small fraction (1 - 0.87 = 0.13) of product codes is ineligi­ ble for the export subsidy on average, a much larger fraction of output is likely excluded because not all manufacturing products in the six-digit HS classifi­ cation are produced. Also, the eligibility cut-off for the export subsidy is based on the volume of exports, not on the share of output exported. This means that the subsidy rule excludes the product lines with some of the largest volumes. Thus, if one could calculate the fraction of output ineligible for the export subsidy in the 21 industries with at least one product line excluded from the export rebate, it would be larger than 0.13 (the average fraction of products excluded from the subsidy). Unfortunately, no output data are avail­ able at the six-digit HS level; otherwise, calculating the fraction of output in a four-digit ISIC industry truly excluded from the export subsidy would be trivial. However, based on the dollar amount limit in the subsidy rule and the fact that exports constitute about 0.28 of total sales for the average exporting establishment (table 1), an approximate statistic of 0.50 can be computed for the average output excluded from the export subsidy in the set of 21 industries with at least one ineligible product line. 10 A benefit of performing the analysis at the more aggregated four-digit ISIC level is that plants may change their (six-digit HS) product selection to benefit from the export subsidy, but in the data, plants (almost) never change their four-digit ISIC 10. The method used to calculate the fraction of output in industries classified as ineligible that comes from product lines truly excluded from the export rebate can be illustrated with the following example. If the three product lines a, b, and c all map into industry A, classified as ineligible for the export subsidy, and both a and b, but not c, are excluded from the refund, there must be at least U5$2.5 million worth of exports of each product a and b. 5ince establishments servicing foreign markets export about 0.28 of their output (see table 1), total production of a and b must be at least U5$9 million each, and therefore there must be at least U5$18 million worth of output in industry A that is truly excluded from the export subsidy. 4$ 4 AiA xu fU; ,U,jI Kandi/au 277 industry of operation. Thus, even if a plant in an industry classified as ineligible (FrExpSub E [0,1)) changes product selection to potentially benefit from the export subsidy, the estimated impact would understate the true effect, since a posi­ tive effect of increased export activity on (relative white-collar) wages is expected. While there are no output data at the six-digit HS product level, export data for eligible and ineligible products can proxy for their output. Thus, for each four-digit ISle industry, the fraction of exports of eligible products in total exports can be computed, FrExpSubW, and used as a proxy for the fraction of output in each industry eligible for the export subsidy. This measure can poten­ tially provide more precise industry eligibility information than does the measure based on the fraction of eligible products, FrExpSub. A complication is that export data at the six-digit HS product level are avail­ able starting only in 1990, four years after the subsidy was implementedY Because exports respond to the incentive offered by the subsidy, using exports from 1990 as weights to construct the subsidy eligibility from 1986 onwards is problematic. 12 For this reason, the industry subsidy eligibility variable using the 1990 product-level exports as weights is constructed, but it is used as a robustness check and not in the baseline specification. The summary statistics for the new subsidy eligibility variable, FrExpSubW, show that about 43 percent of exports (output) are eligible for the subsidy in the set of 21 indus­ tries with at least one ineligible product line (see table 1). This fraction (0.43) is very similar to the 0.50 computed using FrExpSub, as outlined above. Pre-existing differences (as of 1985, the year before the subsidy was implemented) between establishments in the industries eligible for the subsidy and those not eligible are presented in table 1. Establishments in the subsidized industries tend to start out with slightly lower white-collar, blue-collar, and relative white-collar wages. They are also smaller, as measured by total employment or sales. The pre- and post-subsidy trends in exports and total sales for both eligible and ineligible industries suggest that the growth rates of exports and total sales were similar before 1986 when the export subsidy was implemented (see table 2). The growth rate of total sales for the eligible indus­ tries was somewhat lower than that of ineligible industries, but the difference is quite small (-0.02 and 0.01). After 1986, however, exports grew much faster for the eligible industries than for the industries with excluded products. But total sales grew at very similar rates, suggesting substitution of sales away from domestic into foreign markets for the subsidy-eligible industries. 11. The six-digit HS product-level export data are from the United Nations Comtrade database (http://comtrade.un.org). 12. This procedure is problematic only for the 21 industries that contain at least one ineligible product line. For the remaining 68 industries, the treatment measure would be equal to one with or without weighting by the export fraction of eligible products. 278 THE WORLD BANK ECONOMIC REVIEW TAB L E 2: Average Annual Growth Rates of Industry Exports and Total Sales before and after the Export Subsidy Pre-subsidy (1980-85) Post-subsidy (1986-93) --.~.-.--- Subsidy status Exports Total sales Exports/sales Exports Total sales Exports/sales No subsidy 0.10 0.01 0.10 0.17 0.16 0.02 Subsidy 0.10 -0.02 0.13 0.43 0.15 0.28 Note: All numbers are averages. Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey (ENIA) and export data from the World Bank and United Nations; see text for details. Model and Analysis Formally, the identification strategy at the industry level can be written as: (1) where log Wit is the natural logarithm of the white-, blue-, or the relative white­ collar average wage in industry j (four-digit ISIC) in year t, t = 1979, 1980, ... , 1996, and 11986 is a dummy variable equal to one for 1986 and thereafter and equal to zero otherwise. Industry, J..Lj, and year, 'Tr, fixed effects, are included to absorb interindustry wage differentials and aggregate shocks. As already defined, FrExpSubj is the fraction of products in each industry eligible for the subsidy. The coefficient of interest is a1, which represents the impact of increasing the share of subsidy-eligible products in a given industry on the average industrywide wage. Further, to take advantage of the detailed plant-level panel data, the follow­ ing econometric model is specified: (2) where log Wilt is the natural logarithm of the white-, blue-, or the relative white­ collar average daily wage for plant i, in industry j (four-digit ISIC) in year t, t 1979, 1980, ... , 1996. Relevant plant-level controls, such as plant age and the fraction of female employees, are included in the vector ~jt. Plant-specific time-invariant heterogeneity is captured in the plant fixed effect, Ai' The coefficient of interest here is /31, which represents the impact of increasing the share of subsidy-eligible products in a given industry on the average plant-level wage. With an export incentive such as the subsidy, the least productive (smallest) plants, which had likely served only the domestic market, would not enter the export market because it would not be profitable enough to do so (Melitz 2003). Hence, wages in small establishments would not necessarily be affected by the export policy. At the other end of the spectrum, the most productive j -. Q. $ ¥JI!JJ Kg ill.! ,d r u d_ II@ iii a Kandilov 279 (largest) plants would find exporting much more rewarding. Many of them were probably already serving international markets before the subsidy was introduced. The new export incentive would likely induce the largest establish­ ments to increase their already positive exports (intensive margin). Recent theoretical developments (Verhoogen 2008; Bustos 2007), however, suggest that wages (and productivity) would be affected mainly in establishments that enter the export market (extensive margin) as a result of the export incentive. In that case, the large plants that had previously served foreign markets might not experience large wage effects following the subsidy. There is an additional reason why wages in very large establishments might not respond to the incentive. Because the eligibility cut-off is in volume of exports at the product level, a plant would not be eligible for the subsidy if it exports more than the cut-off volume. This is most likely to happen with large establishments, which tend to be the large exporters. If small, medium size, and large plants manufacture the same products, all plants would then be ineli­ gible. If, however, small and medium-size establishments manufacture a differ­ ent set of products than do large plants within the same four-digit ISle industry, only the large plants would be ineligible for the export rebate and their wages would not change following the subsidy. To assess the within-industry heterogeneity that might arise in response to the export subsidy, industry-level eligibility is interacted with plant size, proxied by the logarithm of the plant's initial employment. 13 In particular, the following specification is estimated: logwijt ='YIFrExpSubiIt986 + 'Y2FrExpSubi 10g(Employment)ii (3) + 'Y3FrExpSubiIi986 10g(Employment)ij + Ai + 'Tt + (ljt· As discussed, the impact of the export subsidy might not be linear in plant size. In particular, the export subsidy might not affect small and large plants as much as medium-size plants. To this end, two additional specifications are esti­ mated. One is similar to equation (3) but includes a quadratic term in plant size. Another explicitly splits the sample into three groups of plants based on initial employment-small (10-16 employees), medium (17-32 employees), and large (more than 32 employees): log Wjjt = ih SmaU;FrExpSubjit986 + 52Medium;FrExpSub; it986 (4) + 53Large;FrExpSubjI1986 +Xijt 54 + ,uJ*Small; + ,ur*Mediumi + ,u!'*Largej + ~*Smalli + ~*Mediumi + f*Largej + Ai + 'Tt + I/Jijt, 13. Initial rather than contemporaneous employment is used because it is exogenous with respect to the export subsidy. The two measures are highly positively correlated, with a correlation coefficient of about 0.80. The results are similar if initial sales are used instead as a proxy for plant size. 280 THE WORLD BAl'iK ECONOMIC REVIEW where Smail, Medium, and Large are the three plant size group indicators. 14 To check how robust the results are to the size category cut-offs used, another version of specification 4 that uses the (within-industry) deciles of the size dis­ tribution as cut-offs is also estimated. Problems with heteroskedasticity and serial correlation arise in a plant-level panel data setup such as this one, but the consequences from potential serial correlation can be even more severe if the identification involves difference-in-differences with multiple time periods, where the main variable of interest varies by industry rather than by plant. To solve this problem, this study follows Bertrand, Duflo, and Mullainathan (2004), who recommend cal­ culating robust standard errors clustered by industry, not by industry-year cell or by plantY For the industry-level regressions specified in equation (1), stan­ dard errors are also clustered by industry. IV. RESULTS This section examines the effect of the export subsidy on exports and wages, Effects of the Subsidy on Exports The survey did not record plant-level exports before 1990, which makes it impossible to estimate the impact of the subsidy on plant-level exports since the subsidy was introduced in 1986. However, industry-level (four-digit ISIC) export data for the period before and after the introduction of the subsidy are available from the World Bank (Nicita and Olarreaga 2001).16 The following equation is estimated for industry-level exports: (5) log Exports jt tP1FrExpSubih986 + ILj + 'Tt + ILjtimet + Yijt' which includes an industry-specific time trend, IL/timet> to further capture any differences in export trends across industries. As expected, when industries gain access to the export subsidy, their exports rise (table 3, regression 3.1). To control for industry-specific business cycle fluctuations in exports, a full set of interactions between the economywide GDP growth rate and the industry 14. The plant size cut-offs divide the initial size distribution of all plants into three equal parts. The average establishment-level exports (in 1990) for large plants are more than US$7 million-higher than the product line cut-off for the export subsidy throughout the sample period. It is therefore unlikely that large plants that export predominantly in one product category would qualify. The average establishment-level exports for medium-size and small plants are about 10 times smaller than those of large plants, well below the cut-off. If medium-size and small plants manufacture products that are different from those produced and exported by large establishments, the introduction of the export subsidy will strengthen the export incentive for them. 15. This estimator of the variance-covariance matrix is consistent in the presence of any correlation pattern within industries over time. 16. The export data from the World Bank cover 1980-96 (Nicita and Olarreaga 2001). Mirrored exports (Chilean exports as recorded by export partners) are used because they tend to be more accurate. U4 lit 2U $1 ", ru ." TABLE 3. Impact of the Export Subsidy on Industry-level Exports and Total Sales Log exports Log sales Log exports/sales Variable 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 FrExpSub;t * 11986 1.79** (0.91) 1.63* (0.90) -0.10 (0.21) 0.13 (0.22) 1.89* * (0.95) 1.75** (0.90) FrExpSubWi / [1986 0.55** (0.25) 0.01 (0.13) 0.56* (0.32) Industry-specific Yes Yes Yes Yes Yes Yes GDP growth-rate interaction R2 0.91 0.93 0.93 0.97 0.97 0.97 0.86 0.88 0.88 Number of 1,080 1,080 1,080 1,080 1,080 1,080 1,080 1,080 1,080 observations *Significant at the 10 percent level; **significanr at the 5 percent level; and ***significanr at the 1 percent level. Note: Numbers in parentheses are robust standard errors clustered by industry. All regressions contain a full set of year and (four-digit ISIC) industry dummy variables, as well as (four-digit ISIC) industry-specific time trends. Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey and export data from the World Bank and United Nations; see text for details. ~ ;:: E:-: §"' ty ... 00 282 THE WORLD BANK ECONOMIC REVIEW dummies is included (regression 3.2). The estimates from regressions 3.1 and 3.2 are quite similar. The estimated coefficient of 1.79 (regression 3.1) is positive and statistically significant, implying that if an industry's fraction of eligible pro­ ducts rose 0.10, or 10 percentage points (the sample standard deviation of FrExpSub;t; see table 1), exports in that industry would rise 17.9 percent. 17 While this estimate is statistically significantly different from zero, it is not sig­ nificantly different from 10 percent, implying that the hypothesis that the elas­ ticity of exports with respect to FrExpSubjt is unity cannot be rejected. The results with the alternative subsidy eligibility measure, FrExpSubW;t, (regression 3.3) imply that if an industry's fraction of eligible output rose 0.31, or 31 per­ centage points (the sample standard deviation of FrExpSubWjt ; see table 1), exports in that industry would rise 17.1 percent. This estimate is quite similar to that obtained using FrExpSub;t. It is informative to see how the expansion of exports affected overall indus­ try sales. As the export subsidy was introduced, the producer price for exports increased relative to the producer price in the domestic market. This could, for example, induce some substitution away from production for the domestic market and toward production for the international market. Another possi­ bility is that there would be no substitution but only an increase in exports. This can happen if the production decisions for the domestic market and for exports are decoupled. 18 The estimates of the effect of the export subsidy on total industry sales (see table 3, regressions 3.4-3.6) are close to zero, indicating that the export subsidy had little impact on industry output. Hence, domestic industry sales must have declined almost as much as industry exports rose, leading to a nearly one-for-one production substitution. Supporting this conclusion are the last three specifications in table 3 (regressions 3.7-3.9), which report the esti­ mated impact of the export subsidy on the ratio of industry exports to sales. The coefficients are very similar to the effect of the subsidy on exports for regressions 3.1-3.3. Effects of the Subsidy on Wages While imprecisely estimated, the effect of the export subsidy on the industry white-collar wage is close to zero (table 4). The estimated coefficients are -0.02,0.15, and 0.10 (regressions 4.1-4.3). There is a small negative effect of the export subsidy on the industry blue-collar wage (regressions 4.4-4.6). Consequently, the results indicate a moderate positive effect on the relative white-collar wage (regressions 4.7-4.9). The estimates imply that if the frac­ tion of industry products eligible for the export subsidy increases 0.10 17. As pointed out earlier, excluding 13 percent of products in an industry from the export subsidy (FrExpSubjt = 0.13) is equivalent to excluding about 50 percent of industry output. Therefore, the impact of increasing the industry fraction of products eligible for the subsidy by 0.10 is roughly equivalent to increasing the industry output eligible for the subsidy by 0.50, or 50 percentage points. 18. That is the theoretical set-up in Verhoogen (2008). IUt4 v u _,. It-.it TABLE 4: Impact of the Export Subsidy on Industry-level Wages Log white-collar wage Log blue-collar wage Log relative white-collar wage Variable 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 FrExpS ubit *11986 -0.02 (0.54) 0.15 (0.45) -0.20 (0.60) -0.49 (0.30) 0.18 (0.16) 0.63** (0.30) FrExpSubWit" I 1986 0.10 (0.17) 0.00 (0.14) 0.10 (0.12) Industry-specific GDP Yes Yes Yes Yes Yes Yes growth-rate interaction R2 0.95 0.97 0.97 0.96 0.98 0.98 0.52 0.65 0.65 Number of observations 1,210 1,210 1,210 1,210 1,210 1,210 1,210 1,210 1,210 'Significant at the 10 percent level; **significant at the 5 percent level; and ***significant at the 1 percent level. Note: Numbers in parentheses are robust standard errors clustered by industry. All regressions contain a full set of year and (four-digit ISIC) industry dummy variables, as well as (four-digit ISIC) industry-specific time trends. Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey and export data from the World Bank and United Nations; see text for details. ~ :t ~ 0­ <:: N OQ VJ 284 THE WORLD BANK ECONOMIC REVIEW (FrExpSub jt = 0.10), the relative industry white-collar wage would rise 1.8-6.3 percent. Alternatively, if the fraction of industry output eligible for the subsidy increases 0.31 (FrExpSubW;t = 0.31), the relative white-collar wage would rise 3.1 percent. The next set of results, from equation (2), estimate the effect of the export subsidy on the average plant-level wage using the detailed plant-level data from the Chilean Annual National Industrial Survey. Similar to the estimates with aggregate industry data, the results indicate that the export subsidy has almost no impact on the average plant-level white-collar wage (table 5, regression 5.1). The next specification (regression 5.2) includes a number of covariates­ plant age and age squared (see Brown and Medoff 2003) and the fraction of female white-collar employees. The estimated nonlinear effect of plant age shows that wages increase with age but at a decreasing rate. Additionally, having a higher fraction of female employees is associated with a lower plant-level wage. The coefficient on the export subsidy, however, does not change much from that in regression 5.1. To assess the impact of labor composition on the average plant-level white­ collar wage, the next specification explicitly controls for the fraction of white­ collar administrative workers and executives (regression 5.3).19 Plant-level white-collar labor composition may change in response to the export subsidy. For example, a firm may choose to substitute one type of white-collar labor for another, substituting a relatively less expensive type for one that has become relatively more expensive as a result of the subsidy. As expected, plants with a larger fraction of white-collar executives also have higher white-collar wages. The estimate of the effect of the export subsidy, however, while slightly higher, is very similar to the previous estimates (regressions 5.1 and 5.2). Using the alternative measure of subsidy eligibility, FrExpSubWjt, produces similar results (regression 5.4). Equation (2) is next re-estimated using the blue-collar plant-level wage as a dependent variable. The results reveal that the export subsidy had a small nega­ tive impact on blue-collar wages (see table 5, regressions 5.5-5.8). The effect is similar to that obtained with industry-level data. The results further indicate that the export subsidy had a small positive impact on the relative white-collar wage (see table 5, regressions 5.9-5.11). The estimates imply that if the frac­ tion of industry products eligible for the export subsidy rose 0.10 (FrExpSub jt = 0.10), the average plant-level relative white-collar wage would rise 1.51-1.57 percent. Alternatively, the results imply that if the fraction of industry output eligible for the subsidy increases 0.31 (FrExpSubWjt = 0.31), the relative white-coUar wage would rise 0.90 percent. Overall, the industry-level and the average plant-level results show that adoption of the export subsidy was followed by a mild increase in the relative white-collar wage, driven mostly by the small decline in the blue-collar wage. 19. The omitted category is white-collar production workers. Nt. • •• "J AII$4 ) ( i U _$I . , II _ti It _ k TABLES. Impact of the Exports Subsidy on Plant-level Wages Log white-collar wage Log blue-collar wage Log relative white-collar wage Variable 5.1 5.2 5.3 5.4 5.5 5.6 5,7 5.8 5.9 5.10 5.11 FrExpSubjt"I 1986 -0.008 0.006 0,011 -0.159 -0.146 -0.147 0.151 0,157 (0.073) (0.072) (0.071) (0.106) (0.103) (0.103) (0.099) (0.094) FrExpSubWj/11986 -0.013 -0,041 0.029 (0.039) (0.040) (0.057) Plant age 0.080'" 0.082**' 0.048**' 0.080*** 0.080'" 0.044'" 0.001 0.003 (0.003) (0.003) (0.004) (0.003) (0.003) (0.002) (0.003) (0.004) Plant age squared -0.001*"* -0.001'" -0.001**' -0.001'" -0.001'" -0.001"" -0.001' -0.001' (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) White/blue-collar -0.024 -0.020 -0.023 -0.199'" -0.191 u* -0.196'" 0.022 0,027 female workers (0.037) (0.034) (0.034) (0.045) (0.037) (0.037) (0.038) (0.038) White-collar 0.145**' 0.146'" executives (0,015) (0.016) White-collar 0.Q15 0.012 administrative (0.012) (0.013) workers Blue-collar -0.094'" ·-0.095'" non-production (0.031) (0.032) workers R2 0.98 0.99 0.99 0.99 0.98 0.99 0.99 0.99 0.67 0.68 0.68 Number of 78,321 78,321 78,321 78,321 78,321 78,321 78,321 78,321 78,321 78,321 78,321 observations 'Significant at the 10 percent level; **significant at the 5 percent level; and ***significant at the 1 percent leveL Note: Numbers in parentheses are robust standard errors clustered by industry. All specifications contain a full set of year dummy variables and plant ~ ;:,: fixed effects. Tbe omitted category is white-collar production workers. ~ 0­ Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey and export data from the ~ World Bank and United Nations; see text for details. N oc 0.324 (0.194) for OI, and -0.097 (0.081) for 03. The estimates for the relative white-collar wage are quite similar. These results indicate that medium­ I size establishments experienced an increase in the white-collar wage as a result I I I I I • " _Jt n . , Kandilat} 291 of the export subsidy. The magnitudes imply that if the average industry's frac­ tion of products eligible for the subsidy rose 0.10 (the sample standard devi­ ation), high-skilled wages in medium-size plants would increase 0.032, or 3.2 percent. White-collar wages are estimated to decrease 0.97 percent in the large plants and by 2.10 percent in the small plants. To check how robust the results are to the three size category cut-offs, another version of specification 4 is estimated that includes 10 size groups using the (within-industry) deciles of the size distribution as cut-offs. 21 The results support the findings for the three size categories (table 7). The positive impact of the export subsidy on both white-collar and relative white-collar wages is particularly strong for the fourth, fifth, and sixth deciles, which include the medium-size establishments. While data limitations make it impossible to test whether exports of plants in those three deciles rose in response to the subsidy, the behavior of total sales can be examined. Consistent with the previous findings (table 6), regression 7.4 in table 7 documents that total sales for plants in the fourth, fifth, and sixth deciles rose on average more than sales in the bottom three or the top four deciles in response to the export subsidy. While the industrywide estimates implied that the increase in industry exports was not accompanied by a rise in industry sales (table 3), the plant-level results suggest that establishments most likely to enter foreign markets and increase exports in response to the subsidy did experience an increase in total sales (table 7). V. CONCLUSION Using the exogenous variation in export activity brought about by an interest­ ing policy experiment-an export subsidy-this study identifies the effect of increased exports on plant-level wages in the Chilean manufacturing industry. Avoiding potential endogenous selection, the research design uses plant-level panel data to show that while there was only a mild increase in the relative white-collar wage for the average plant as a result of the subsidy, the relative white-collar wage in smaller (medium-size) establishments-those most likely to take advantage of the subsidy and enter foreign markets-rose more sub­ stantially following implementation of the export subsidy. This evidence high­ lights the importance of skill in the export process, and it conforms to previous work documenting a rise in the relative high-skilled wage following trade liber­ alization in developing countries. While the relative white-collar wage rose in response to the export subsidy in smaller (medium-size) plants, the relative employment of white-collar workers in these establishments did not change much. In the short run, in part due to large separation costs, incumbent white-collar workers may have gained 21. Size groups 1-10 are also based on initial employment level. Size group 1 contains the smallest establishments, and size group 10 the largest. tv \0 tv -I :t m >1i o TABLE 7. Impact of the Exports Subsidy on Plant-level Wages by Plant-size Group ;1i Size group 8*FrExpSubj ,*11986 0.400 (0.189) -0.467** (0.193) 0.066 (0.217) 0.933 (0.606) Size group 9*FrExpSubj t" 11986 -0.194 (0.222) -0.572**· (0.183) 0.378 (0.282) - 0.932 * (0.494) Size group 10*FrExpSub;,* 11986 -0.092 (0.195) -0.313**" (0.102) 0.220 (0.237) 0.582 (0.476) R2 0.99 0.99 0.68 0.96 Number of observations 78,321 78,321 78,321 78,321 I *Significant at the 10 percent level; **significant at the 5 percent level; and h *significant at the 1 percent level. Note: Numbers in parentheses are robust standard errors clustered by industry. All specifications contain a full set of year dummy variables and plant fixed effects, as well as full sets of year and industry dummy variables interacted with the plant size categories. Size groups 1-10 are based on initial employment level, with size group 1 containing the smallest establishments, and size groupJ 0 the largest. Source: Authors' analysis based on data from the National Statistical Institute of Chile's Annual National Industrial Survey and export data from the World Bank and United Nations; see text for details. I Kandilov 293 higher wages because of either industry- and exporting-specific skills they possess or an effort-elicitation mechanism put in place to ensure output quality suitable for foreign markets. In the long run, plant-level relative white-collar wages in smaller (medium-size) establishments may have increased because of labor quality upgrading as well. FUNDING Financial support from the Rackham fellowship at the University of Michigan is gratefully acknowledged. REFERENCES Aw, Bee Yan, and Geeta Batra. 1999. "Wages, Firm Size, and Wage Inequality: How Much Do Exports Matter?" In D. Audretsch, and R. Thurik, eds., Innovation, Industry Evolution, and Employment. Cambridge: Cambridge University Press. Bernard, Andrew, and J. Bradford Jensen. 1997. "Exporters, Skill-upgrading, and the Wage Gap." Journal of International Economics 42(1-2):3-31. - - . 1999. "Exceptional Exporter Performance: Cause, Effect, or Both?" Journal of International Economics 47(1):1-25. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. "How Much Should We Trust Difference-in-Differences Estimates?" Quarterly Journal of Economics 119( 1):249-75. Beyer, Harald, Patricio Rojas, and Rodrigo Vergara. 1999. "Trade Liberalization and Wage Inequality." Journal of Development Economics 59(1):103-23. Brown, Charles, and James Medoff. 2003. "Firm Age and Wages." Journal of Labor Economics 21(3):677-98. Bustos, Paula. 2007. "Multilateral Trade Liberalization, Exports, and Technology Upgrading: Evidence on the Impact of MERCOSUR on Argentinean Firms." Center for Research on International Economics (CREI), Department of Economics and Business, Universitat Pompeu Fabra, Barcelona. Cierides, Sofronis, Saul Lach, and James Tybout. 1998. "Is Learning by Exporting Important? Micro-dynamic Evidence from Colombia, Mexico, and Morocco." Quarterly Journal of Economics 113(3):903-47. Cortazar, Rene. 1997. "Chile: The Evolution and Reform of the Labor Market." In S. Edwards, and N.C. Lustig, cds., Labor Markets in Latin America: Combining Social Protection with Market Flexibility. Washington, D.C.: Brookings Institution Press. Fafchamps, Marcel. (forthcoming). "Human Capital, Exports, and Earnings." Economic Development and Cultural Change. Fajnzylber, Pablo, and William Maloney. 2005. "Labor Demand and Trade Reform in Latin America." Journal of International Economics 66(2):423-46. Gindling, T., and Donald Robbins. 2001. "Patterns and Sources of Changing Wage Inequality in Chile and Costa Rica during Structural Adjustment." World Development 29(4):725-45. Goldberg, Pine\opi, and Nina Pavcnik. 2007. "Distributional Effects of Globalization in Developing Countries." Journal of Economic Literature 45(1):39-82. Hanson, Gordon, and Ann Harrison. 1999a. "Trade Liberalization and Wage Inequality in Mexico." Industrial and Labor Relations Review 52(2):271-88. - - . 1999b. "Who Gains from Trade Reform? Some Remaining Puzzles." Journal of Development Economics 59(1): 125-54. 294 THE WORLD BANK ECONOMIC REVIEW Hockman, Bernard, Aaditya Mattoo, and Philip English. 2002. Development, Trade, and the WTO: A Handbook. Washington, D.C.: World Bank. Levinsohn, James. 1999. "Employment Responses to International Liberalization in Chile." Journal of International Economics 47(2):321-44. Liu, Lili. 1993. "Entry-exit, Learning, and Productivity Change: Evidence from Chile." Journal of Development Economics 42(2):217-42. Lopez, Ricardo, and Julian Namini. 2006. "Random versus Conscious Selection into Export Markets­ Theory and Empirical Evidence." Indiana University, Department of Economics, Bloomington. Melitz, Marc. 2003. "The Impact of Trade on Intra-industry Reallocations and Aggregate Industry Productivity." Econometrica 71(6):1695-725. Murphy, Kevin, and Andrei Shleifer. 1997. "Quality and trade." Journal of Development Economics 53(1):1-15. Nicita, Alessandro, and Marcelo Olarreaga. 2001. "Trade and Production, 1976-1999." Policy Research Working Paper 2701. World Bank, Washington, D.C. Pavcnik, Nina. 2002. "Trade Liberalization, Exit and Productivity Improvements: Evidence from Chilean Plants." Review of Economic Studies 69(1 ):245-76. - - . 2003. "What Explains Skill Upgrading in Less Developed Countries?" Journal of Development Economics 71(2}:311-28. Roberts, Mark, and James Tybout. 1997. "The Decision to Export in Colombia: An Empirical Model of Entry with Sunk Costs." American Economic Review 87(4):545-64. Robertson, Raymond. 2004. "Relative Prices and Wage Inequality: Evidence from Mexico." Journal of International Economics 64(2):387-409. Schank, Thorsten, Claus Schnabel, and Joachim Wagner. 2007. "Do Exporters Really Pay Higher Wages? First Evidence from German Linked Employer-Employee Data." Journal of International Economics 72(1):52-74. Shapiro, Carl, and Joseph Stiglitz. 1984. "Equilibrium Unemployment as a Worker Discipline Device." American Economic Review 74(3):433-44. Verhoogen, Eric. 2008. "Trade, Quality Upgrading and Wage Inequality in the Mexican Manufacturing Sector." Quarterly Journal of Economics 123(2):489-530. Yeaple, Stephen. 2005. "A Simple Model of Firm Heterogeneity, International Trade, and Wages." Journal of International Economics 65(1):1-20 . . . ..._ _ _.."' ~ , ......._ _ _ _..."',_ _ t .. ......, " , ............."''''.......... ,~ '''''_,... ....''''_'[_ '''' ,., .................t..._ _ _ _ . ..... ,_,'''......,. ._ _ _ _ _ ,~ f).,J!lJ _ : ':I~~4'~~,'.- The Determinants of Funding to Ugandan Nongovernmental Organizations Marcel Fafchamps and Trudy Owens Original Ugandan data collected by the authors are used to examine the determinants of funding to local nongovernmental organizations (NGOs). Success in attracting grants from international donors depends mostly on network effects. NGOs that raise in-kind resources locally tend to be young and managed by someone who is simul­ taneously employed elsewhere. There is some evidence of crowding out: NGOs that receive grant funding are less likely to obtain resources locally, whether in cash or in kind. But this seems to be primarily the result of selection. Once NGO-fixed effects are controlled for, there is no evidence that NGOs receive less revenue from fees and donation after obtaining a grant. These results suggest that donors regard Ugandan NGOs as subcontractors of their development efforts, not as charitable organizations in their own right. JEL codes: 019, 012 Nongovernmental organization (NGO) involvement in development has been increasing in recent years (Edwards and Hulme 1995; Hulme and Edwards 1997)-partly because of dissatisfaction with government delivery of public services. International NGOs as well as bilateral and multilateral donors have sought to channel more development funding through local NGOs, causing the sector to grow rapidly in developing economies. But it is unclear whether donors, through their funding, encourage a local charitable sector or local NGOs are simply subcontractors for international development agencies. NGOs in poor countries are presumed to be charitable organizations, that is, operating with an altruistic or a philanthropic purpose shared by their members and promoters. Much of donors' dissatisfaction with governments' public service delivery originates in concerns over corruption. Civil servants running government schools and health centers are assumed to be motivated by self­ interest, explaining why they divert resources from the public (Reinikka and Svensson 2003; Lindelow, Reinikka, and Scensson 2003). But NGOs are Marcel Fafchamps (corresponding author) is a professor in the Department of Economics at Oxford University; his email addressismarcel.fafchamps@economics.ox.ac.uk. Trudy Owens is a lecturer in the School of Economics and the Centre for Research in Economic Development and International Trade at the University of Nottingham; her email addressistrudy.owens@nottingham.ac.uk. THE WORLD BANK EOONOMIC REVIEW, VOL 23, No.2, pp. 295-321 doi:10. 1093/wber/lhp001 Advance Access Publication March 11, 2009 © The Author 2009. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / TIlE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 295 296 THE WORLD BAKK ECOKOMIC REVIEW considered less selfish and thus less likely to divert funds-a belief that underlies the switch in donor funding. Several researchers have doubted that NGO motives in poor countries are first and foremost charitable (Edwards and Hulme 1995; Platteau and Gaspart 2003). But these doubts are generally based on a few case studies. No one has investigated these issues using a large representative sample of NGOs. Given the growing importance of local NGOs and their potential for delivering ser­ vices, this gap needs to be filled. This article examines the factors that influence local NGOs' capacity to attract external resources using a nationally represen­ tative survey of 300 NGOs in Uganda. NGOs obtain resources in several ways. Some resources are raised in cash, for example, financial grants, and membership fees; others are raised in kind, for example, volunteer work and complimentary use of equipment and facili­ ties. In Uganda, international grants are by far the major source of funding for domestic NGOs (Barr, Fafchamps, and Owens 2003). For small NGOs, mem­ bership fees and donations are important, suggesting that local NGOs that receive donor funding may be different in some fundamental sense from NGOs that attract voluntary contributions from nationals. To investigate this, the factors that influence local NGOs' capacity to successfully obtain grant funding are examined and contrasted with the determinants of voluntary con­ tributions in cash and in kind. This issue is first approached from a reduced form perspective to examine whether the ex ante characteristics of NGOs receiving grant funding are the same as those that do not. NGOs receiving external funding differ markedly from those that do not: they are much more likely to be part of an inter­ national network and to be managed by an educated, well-connected leader. Grant recipients on average raise fewer resources domestically. The investigation then turns to whether donor funding displaces voluntary contributions from nationals, that is, whether international funding is a comp­ lement or substitute for local charity. Local NGOs may be genuinely altruistic organizations whose effectiveness is enhanced by external funding. Externally funded NGOs would be expected to expand and attract more local resources. Local NGOs may act as subcontractors for international donors, in which case local funding does not matter. Local NGOs may also be altruistic, but external funding may crowd out their willingness to give. The issue is difficult to investigate, especially given the challenge of collect­ ing data on NGOs. An instrumental variable approach yields evidence that grant recipients raise fewer resources locally, notably from member fees and contributions. This suggests crowding out. But in a similar analysis using NGO-fixed effects, the evidence of crowding out disappears. This suggests that grant recipients are NGOs that are, on average, less likely to receive local contributions. Taken together, the evidence suggests that grants from external donors do not encourage a local charitable sector. Many local NGOs seem to be created simply Fafchamps and Owens 297 to obtain grant funding. 1 This interpretation is reinforced by the numerous Ugandan NGOs that have a shadowy existence when they do not receive an exter­ nal grant. For instance, of the roughly 1,700 NGOs registered in Kampala at the time of the survey, only a quarter could be located. Grants do not appear to go to NGOs that would raise funds on their own; instead, they go to a few well-educated, well-connected organizations and individuals skilled at writing grant applications. Observing that grant recipients do not raise local resources does not imply that they deliver services poorly. But it calls into question the assumption that underlies the switch away from government services: if local NGOs are not driven by an altruistic motive, why should they be expected to behave in a less opportunistic manner than civil servants? There may be other reasons why donors prefer private service delivery, such as better control, faster response to emergencies, or the promotion of a specific message or agenda. But based on the evidence presented here, it would be foolish to rely on altruism to econom­ ize on monitoring. Donors seem to understand this welL Survey results indicate that NGOs are subject to extensive donor monitoring. Given the growing number of NGOs in Africa, increased funding from donors, and NGO potential for delivering services to the poor, a thorough analysis of the sector is overdue. Evidence suggests that there has been strong growth of NGOs in the region (Anheier and Salamon 2006; Wallace, Bornstein, and Chapman 2007). Presumably partly in response to this growth, several African countries­ Kenya, Nigeria, Uganda, and Zimbabwe-have recently implemented new NGO monitoring or regulation frameworks. This article is thus pertinent to all African countries seeing NGO growth. A major reason for the limited work to date is a lack of representative data on NGO structures, finances, and activities. Access to such information is extremely difficult to obtain due to government sensitivities. The Ugandan government, however, has been willing to support NGO surveys. Because Uganda is a good example of a growing and dynamic African NGO sector, its insight can help guide policy across the continent. The article is organized as follows. Section I presents the conceptual frame­ work that underlies the empirical analysis. A simple model is constructed in which a local NGO receives external funding from an altruistic donor. Section II presents the data, which come from a survey of Ugandan NGOs. Section III discusses the empirical analysis. And section IV summarizes the findings and presents some areas for further research. I. CONCEPTUAL FRAMEWORK Most NGO funding in Uganda comes from international donors (Barr, Fafchamps, and Owens 2005), so understanding fundraising by local NGOs 1. At the time of the survey, Uganda had only 400 registered (for-profit) firms but 3,500 registered NGOs. 298 THE WORLD BANK ECONOMIC REVIEW requires understanding what motivates international donors to channel devel­ opment assistance through local NGOs. Altruism, Crowding Out, and Efficiency Donors may wish to avoid channeling all their assistance through government agencies for several reasons, including corruption, instability, and ideological and political differences. To bypass the government, donors can use private (for-profit) subcontractors or local not-for-profit NGOs. In Uganda, the number of local NGOs has skyrocketed to 20 times as many registered NGOs as registered firms. Many registered NGOs exist only on paper, but this never­ theless suggests that donors prefer to channel funds through local NGOs than through private firms. 2 The question is why. One possible reason is local NGOs have more expertise in delivering the ser­ vices that interest donors. Although this is important in some cases and needs to be controlled for, according to Barr, Fafchamps, and Owens (2005), local NGOs in Uganda are often young, and most adopt a holistic approach, without any strong specialization by activity or region. Another possible reason is that channeling funds through not-for-profit organizations prevents misappropriation of development funding. Barr, Fafchamps, and Owens (2005) cast serious doubt on this premise as well. NGOs in Uganda do not file a tax return and are subject to little or no govern­ ment scrutiny regarding profit distribution. Donors monitor grant recipients, but they could just as easily monitor for-profit subcontractors. It is thus unclear whether channeling funds through NGOs provides any advantage in this respect. This leaves one important possibility: local NGOs are altruistically motiv­ ated and thus less subject to moral hazard. An NGO that cares about the welfare of the beneficiaries of development assistance is less likely to divert funds. Furthermore, local NGOs may provide a cheaper service because they can access a workforce, equipment, and buildings at less than market price. Barr, Fafchamps, and Owens (2005), for instance, have shown that many local NGOs in Uganda employ volunteers and use buildings and equipment on a complimentary basis. They also raise local funding in the form of membership fees and local donations. Channeling development assistance through charitable organizations is subject to another incentive problem, dubbed "crowding out" in the literature. Crowding out occurs when outside funding reduces local charitable contri­ butions. For instance, suppose that each $ of outside funding reduces local con­ tributions by 8$, with 8 < 1. This means that each $ of outside funding 2. Many registered NGOs have no actual existence and were created to attract donor funding that never materialized. But if donors had sought instead to channel funds through registered for-profit firms, the same behavioral process of wishful creation would instead have increased the number of registered firms. This did not happen, suggesting that donors have targeted NGOs-or at least are consistently perceived to have. Fafchamps and Owens 299 generates (1 - B)$ of additional spending on beneficiaries. Although the mech­ anism is different, crowding out is similar to fund diversion because it is an implicit tax on development assistance. The two sources of funds are easily embedded within the same stylized model. Consider an altruistic organization, hereafter called the NGO. 3 This organization is made up of members and promoters who serve a target beneficiary. Beneficiary welfare is denoted by V(t, z), where t is the cost to the NGO of the service provided to beneficiaries and z is an exogenous NGO characteristic that denotes how competent the NGO is in serving the beneficiary group. It is assumed that fJVlfJt> 0, fJVf{Jz > 0, and fJV 21fJtfJz > O. The first two assumptions mean that the welfare gain to beneficiaries increases with the size of the transfer and with NGO competence. The third assumption means that more competent NGOs are more productive, that is, an incremental transfer t generates a higher increase in beneficiary welfare when NGO competence z is higher. The NGO starts with a stock of resources T, which for now is taken as given. This stock includes the financial resources of members and promoters as well as the value of their time. The NGO must decide how much of T to allo­ cate to the beneficiary target group. The remaining is consumed by the organiz­ ation (that is, by members and promoters). The decision problem facing the NGO can be written: 4 (1) max V(t, z) t + wU(T - t) subject to t ::; T, 3. In this model, altruism and joy of giving are basically equivalent, so the distinction between them is not emphasized here. For a discussion, see Ribar and Wilhelm (2002) and the references cited therein. 4. NGOs may also raise local private funds in addition to donor grants. The literature on charitable contributions has typically couched the discussion of crowding out in terms of public versus private outside funds (Ribar and Wilhelm 2002; Andreoni and Payne 2003). This is due largely to the literature's focus on developed countries, where charitable contributions from the general public are common. A distinction has been drawn between altruism-that is, concern for the utility of the beneficiary population-and joy-of-giving-which does not depend on beneficiary welfare. Free riding among altruistic benefactors reduces voluntary contributions as the number of benefactors increases. Ribar and Wilhelm (2003) show that when altruism is the sole reason for giving, for many functional forms and parameter values public funds crowd out private contributions one for one--that is, one additional dollar of public money reduces private contributions by one dollar. In this model free riding does not arise because, by construction, there is a single contributor. With multiple private contributors, free riding is another source of crowding out, in which case the distinction between altruism and joy-of­ giving becomes relevant. Here (JJU(T-t) can be regarded as a reduced form summarizing the equilibrium of the private contribution game. The model can also allow for active fundraising by the NGO. Modeling this in detail would take too much space, so only a few essential observations are included here. Imagine that the NGO has a (probabilistic) production function for obtaining grants and private funds. Fundraising takes time and effort from NGO promoters, thereby subtracting from t. When the NGO has no grant, the opportunity cost of promoter time is low, and the NGO devotes more effort to raising private funds. When the 300 THE WORLD BANK ECONOMIC REVIEW where w is a welfare weight measuring how much the NGO cares about the welfare of its promoters. s Let t(T, z, w) denote the NGO decision on the amount of transfer it makes to target beneficiaries. It is easy to show (see Proposition 1 in Appendix) that organizations with more resources (higher T) or more altruism (lower w) give more, that is, have a higher t, while more com­ petent organizations (higher z) give less. Whether there is crowding out depends on the sign of 8V2 /{)t 2 • If 8V2 /8t 2 < 0, the amount given t increases less than proportionally with NGO resources T(see Proposition 2 in Appendix). There is crowding out: adding external funding G to NGO resources Ttranslates into additional transfers of less than ° G to beneficiaries. It is natural to assume {)V2 /{)t 2 < whenever the marginal welfare gain falls with t, perhaps because of satiation or because of increasing marginal costs in the production of services. In contrast, if 8V2 /8t 2 > 0, external funding has a multiplier effect, that is, dtldT> 1. This will be the case if there are threshold effects in consumption­ for instance, because the utility of beneficiaries rises faster than cost over a certain range-or if there are increasing returns in service delivery-for instance, because setup costs are fixed. When this happens, NGO members and promoters respond to external funding by volunteering more of their own resources, because they are more productive in achieving their altruistic goal. To summarize, whether transfers to beneficiaries increase more or less than proportionally with external resources depends on the sign of {)2V/{)t 2 and, hence, on whether marginal delivery costs are increasing or decreasing. The model is suf­ ficiently general to encompass situations in which crowding out is so large that part of the external funding G is appropriated by the NGO, that is, when t < G. Since dtldw < 0, diversion of funds is more likely if the NGO is less altruistic-as would be the case, for instance, if the NGO is actually a for-profit entity. 6 The difficulty is identifying NGOs that are competent-so that they can provide the service in a cost-effective manner-but also altruistic enough not to divert external funds for personal consumption. So far it has been assumed that the donor observes the characteristics T, z, and wand effort t of the NGO. In practice, donors are not fully informed about the type and effort of grant applicants. Donors may seek to observe effort t through monitoring. As Barr, Fafchamps, and Owens (2005) documented for Uganda, this can be accom­ plished in a variety of ways, such as reporting requirements, field visits, surveys of benefactors, and audits, that are all costly. Monitoring diverts resources that NGO receives a grant, the opportunity cost of the promoter's time rises, thereby reducing private fund raising effort. This is another source of crowding out. Again, wU(T t) incorporates this effect. S. It is assumed that U(·) is increasing and concave-that is, that the marginal utility of consumption falls with consumption or U" < O. 6. Therefore, w can alternatively be seen as measuring how little guilt or shame NGO promoters would feel from diverting outside funds. A dishonest promoter would not mind setting t < G, thereby diverting outside funds toward personal consumption. Altruism and dishonesty are thus two sides of the same coin. Fafchamps and Owens 301 could otherwise be devoted to beneficiaries. 7 It is thus in donors' interest to economize on monitoring. This can be accomplished by selecting more altruistic grant recipients. But how to do so is unclear because NGOs may portray themselves as more altruis­ tic than they actually are. Donors are thus expected to be conservative in choosing grant recipients and to display a strong preference for NGOs with which they have worked in the past or for individuals with whom they have previously dealt in other NGOs. Local NGOs may also raise funds from donations or user fees. The incentive issues surrounding these funds are similar to those affecting donor grants. The main difference is that local funders may be better able to observe NGO com­ petence (z) and altruism (w). Contributions from NGO members are ambigu­ ous because they can serve as a payment for service or as a user fee. Without going into the details how user fees are set, revenue from user fees is an increasing function of NGO output t (an NGO that produces nothing receives no user fees). To the extent that a grant enables an NGO to produce more, it also boosts revenue from user fees. In these data, it is difficult to distinguish between user fees and charitable con­ tributions because user fees are often recorded as membership fees and NGO members are typically beneficiaries (Barr, Fafchamps, and Owens 2005). Without detailed information on the conditionality of membership fees, it is impossible to separate the fee-for-service element from charitable giving. Still, income from membership fees is likely to increase with grant income, thereby generating a multiplier effect that goes in the direction opposite to crowding out. Testing Strategy The empirical objective here is to identify the factors that affect NGO capacity to raise internal and external funds and resources. Let internal resources, in cash and in kind, be denoted by C i and external grants be denoted by G;. There are two steps in the process. First, reduced forms are estimated and C; and G; are regressed on various NGO characteristics Qi that proxy for their competence (z), wealth (T), and level of altruism (w): (2) (3) If, as they often claim, donors rely on NGO altruism to minimize incentive problems, the same variables would be significant in regressions (2) and (3), that is, factors that make it more likely that an NGO raises internal funds should also explain success in raising external funds. 7. This is true whether the monitoring cost is borne by the donor (field visit) or by the grant recipient (reporting). Cost minimization should allocate monitoring tasks between donor and recipient. 302 THE WORLD BANK ECONOMIC REVIEW Inference from comparing regressions (2) and (3) relies on the absence of omitted variable bias. It is conceivable, for instance, that NGOs specializing in different activities may be forced to seek different sources of funding. If local donors are unwilling to fund certain activities, NGOs may have to turn to external donors. If factors affecting the choice of activity are correlated with characteristics Q, this may confound inference. Potentially, a serious concern in other settings, it is unlikely to be a serious source of bias for Uganda. The overwhelming majority of surveyed NGOs remain unspecia­ lized, adopting a holistic approach to development (Barr, Fafchamps, and Owens 2005). So if local funds for certain types of activities were limited, most surveyed NGOs could find an activity that fits local interests. In fact, this is precisely what most do with respect to international donors: the choice of activity is thus best conceived as driven by the availability of funds, not driving it. Even if NGOs are not altruistic, it may still make economic sense for donors to prefer them over for-profit subcontractors. This point was initially made by Hausman (1980), who argues that in markets where the quantity or quality of service cannot be verified, organizations that cannot distribute profits provide a more trustworthy alternative. If external donors regard NGOs as subcontrac­ tors and do not expect NGO promoters to contribute or to raise private funds locally, only their competence matters; their wealth and altruism are irrelevant. In this case, variables measuring wealth and altruism would not be significant in the external resource regression (3), though they might be significant in the internal resource regression (2). Second, the extent of crowding out is tested. To this effect, t(T + G, Z, w) ­ G for grant recipients is compared with t(T, z, w) for nonrecipients. Following Ribar and Wilhelm (2002) and Andreoni and Payne (2003), voluntary contri­ butions C; to NGO i by members and promoters are regressed on whether the NGO is a grant recipient G; and on a set of control variables Qi: (4) Finding 1'1 < 0 is prima fade evidence of crowding out. One difficulty with this approach is the possibility of endogeneity bias; NGOs that were unsuccessful in raising grant funding may put more effort into generating local and internal resources to keep the organization going. To correct this possibility, G; is instrumented using variables that affect grant allo­ cation but not crowding out, such as the factors that affect the probability of receiving a grant independent of beneficiary considerations. One such factor is how connected the NGO is. Because of asymmetric information, NGOs may be more likely to receive grants from donors who are closer to them socially or contractually. Variables proxying for this are used to instrument access to grants. Fafchamps and Owens 303 Another possible source of bias in regression (4) is unobserved heterogeneity. To see how this can affect inference crowding out, suppose that donors are attracted to NGOs that are less involved in raising internal or local funds. This could be because such NGOs devote more attention to courting international donors and are more receptive to their needs. In this case, there would be a negative relationship between Ci and C i in regression (4), even after instru­ menting. But this relationship would be due to reverse selection by donors. To investigate this possibility, an NGO fixed effect version of regression (4) is esti­ mated: (5) taking advantage of the fact that each NGO was asked to provide income state­ ments for two consecutive years. This is equivalent to testing whether increased grant income Cit from 1 year to the next is associated with reduced internal funds Cit. Control variables Qi drop out of the regression because they are time invariant; their effect is captured by the fixed effect Uj. Regressions (2-5) are complementary. Suppose that regressions (2) and (3) show that altruism affects local fundraising, but not success in grant appli­ cation. And suppose that 1'1 < 0 but 61= O. This implies that there is no crowding out at the level of the individual NGO; receiving a grant does not reduce local contributions. But since 1'1 < 0, it also implies that donors allocate grants to NGOs that average fewer local contributions and are not particularly altruistic. Over time, this can have dramatic implications for the structure of the NCO sector because donor behavior affects NGO entry. If having a chari­ table purpose and collecting local contributions are not a prerequisite for obtaining a grant, new NGOs will not be particularly altruistic and will not seek local funds. The NGO sector will be reduced to a mere extension of devel­ opment assistance. II. THE DATA In 2002, Barr, Fafchamps, and Owens undertook the first nationally represen­ tative survey of NGOs in Uganda. The survey, initially proposed by a group of Ugandan NGOs, was organized by the World Bank in collaboration with the Office of the Prime Minister of Uganda, with funding provided by the Japanese government and the World Bank. The survey was conducted by the Center for the Study of African Economies of Oxford University in collaboration with International Development Consultants in Kampala. The survey collected information on each NGO's activities, its sources of funding, and its personnel, including characteristics of its leader. A two-step sample selection process was used. The first step identified districts for data col­ lection. The capital Kampala was included because of its importance as a base for many NGOs. In addition, 14 of the country's 56 remaining districts were 304 THE WORLD BANK ECONOMIC REVIEW randomly selected. 8 A random sample of NGOs was then selected-100 from Kampala and 200 from the 14 other districts. 9 For sampling purposes, an NGO belonged to the district in which its headquarters were located. To draw a random sample of NGOs, a list of all active NGOs in the selected districts was constructed using the records of the NGO Registration Board in the Ministry of Internal Affairs. 10 As of December 2000, approximately 3,500 NGOs were registered, though not all were operational, so the registers for the selected districts were updated and verified before sampling. l l A sample of 100 NGOs was then drawn randomly from the 451 Kampala-based NGOs that could be verified. For the other districts, a self-weighting sample of 200 NGOs was randomly selected from the verified list. The combined stratified sample (Kampala plus districts) roughly represents the national situation. Further details on the sampling procedure can be found in Barr, Fafchamps, and Owens (2003).12 The authors then cleaned the data. Given the heterogeneity of the dataset, two outliers were identified: one is an NGO much older than others and another is a large international NGO with much more abundant resources. Excluding these outliers does not noticably alter the results. III. EMPIRICAL ANALYSIS The empirical analysis consists of three parts: univariate analysis, reduced form regressions, and testing for crowding out. Univariate Analysis Based on the data, C, is constructed as a measure of financial contributions to the NGO from members through fees and donations. This information is avail­ able only for a subsample of the dataset (199 respondents) that agreed to provide financial accounts. But data on the number of full-time paid and 8. The 14 selected districts were Arua, Busia, Iganga, Jinja, Kabale, Kassese, Kibaale, Lira, Luwero, Mbale, Mbarara, Mukono, Rakai, and Wakiso. One district (Gulu) that was initially included in the list was subsequently replaced because of the lack of security in the region. 9. The overall sampling proportion required to yield a sample of 200 was calculated by dividing the proposed sample size by the number of NGOs in all the districts during the exercise. This sampling proportion was then multiplied hy the number of NGOs in each district to yield a self-weighting sample. 10. The registry does not include the Catholic Church, the Church of Uganda (Anglican), and the Uganda Muslim Supreme Council, three organizations that have been operating in the country for many years. These organizations are thus omitted from the survey in spite of their large size. 11. See Barr, Fafchamps, and Owens (2005) for a detailed discussion of the results of this verification exercise. 12. A detailed questionnaire was designed and pretested in Uganda by the authors. The survey was conducted through face-to-face interviews between enumerators and an NGO representative-usually the leader of the NGO. The enumerators and their supervisors received a week of training on the questionnaire and on interviewing techniques before the survey began. A copy of the questionnaire can be found in Barr, Fafchamps, and Owens (2003). Fafchamps and Owens 305 voluntary staff during last 12 months and whether the NGO has complimen­ tary use of equipment or vehicles are available for the full sample. Barr, Fafchamps, and Owens (2005) have shown that these are important resources, especially for nongrant recipients. For G j , two measures are used: the value of grant funding received in the last fiscal year (in Ugandan shillings) and a dummy variable that takes the value 1 if the NGO received a grant in the 12 months before the survey. Information on funding is available only for the respondents that provided financial data. The qualifications and experience of the NGO leader are used as measures of NGO competence z. Qualification variables include age, edu­ cation, and work experience. Because the NGO leader is nearly always its pro­ moter, the wealth and parental background of the NGO leader and whether the leader has a relative living abroad are used as control variables for wealth T. The wealth of the NGO leader cannot be used directly because it may be subject to reverse causation due to crowding out or fund diversion. Altruism w is proxied by a dummy variable that takes value 1 if the NGO has a religious affiliation. Many international donors, being secular organiz­ ations, are reluctant to facilitate religious proselytizing by funding churches' social activities. It is, however, reasonable to expect religious organizations to be more altruistic, at least toward their followers. This is indeed what the evi­ dence suggests. At the time of the Uganda NGO survey, focus group interviews conducted in the communities that NGOs serve showed that when the NGO leader has a religious title, the NGO is more likely to be perceived as altruistic (Barr and Fafchamps 2006). When donors care a lot about being perceived as altruistic, they may overcome their secular leanings and choose to operate through religious organizations. This is the approach taken by Reinikka and Svensson (2003), who use religion as a proxy for altruism in their examination of a micro-level dataset on primary health care facilities in Uganda. They find that "working for God" matters: workers and leaders of religious not-for-profit health care facilities have intrinsic motivations to serve poor people. This seems to be the case in Uganda's NGO sector as well. Religious NGOs in Uganda are expected to be more altruistic, and religious NGOs are thus expected to be more successful at raising charitable funds locally. A dummy variable for having a female NGO leader is included to capture various confounding effects associated with gender, including the possibility that female leaders are more altruistic. To proxy for favoritism, dummy vari­ ables are used to indicate whether the local NGO is a subsidiary of the donor and whether the NGO is a member of a Ugandan NGO network. Finally, several variables on previous and current work experience are included as additional measures of competence and to indicate how connected the NGO leader is. Presumably, NGOs that are better connected have a better chance of securing grant funding. Table 1 describes the regressors for the whole sample as well as a break­ down between grant recipients and nonrecipients. A simple t -test of the Vol 0 0'\ TABLE 1. Descriptive Statistics ..., Nonrecipient Recipient Total l: "' ~ Number of Number of Number of 0 ~ Statistic Mean observations Mean observations Mean observations t-test p> It I "" Cl NCO leader competence '" ;> z Age 41.41 76 41.31 201 41.34 277 0.084 0.933 ;0<: Education 14.93 75 15.85 207 15.61 282 -2.318 0.021 '" (1 0 Length of time with NCO 4.69 78 6.89 212 6.30 290 -3.400 0.001 z Previously worked for NCO 0.49 78 0.44 211 0.45 289 0.774 0.440 0 ~ Previously worked for government 0.37 78 0.51 209 0.47 287 -2.052 0.041 (1 Current employment with an NCO 0.24 79 0.41 208 0.36 287 -2.670 0.008 ~ Current other employment 0.79 79 0.53 212 0.60 291 3.994 0.000 "' <: NCO leader wealth '" ~ Wealthy family 1.79 79 1.79 200 1.79 279 -0.003 0.998 Relative lives abroad 0.36 80 0.44 204 0.42 284 -1.133 0.258 Altruism Religious affiliation 0.35 77 0.28 207 0.30 284 1.151 0.251 Female 0.20 80 0.26 215 0.24 295 -0.995 0.320 Favoritism Subsidiary of foreign NGO 0.05 78 0.17 215 0.14 293 -2.580 0.010 Network 0.51 78 0.79 213 0.72 291 -4.887 0.000 Other Age ofNGO 6.41 80 11.14 215 9.86 295 -3.044 0.003 Number of staff 86.56 80 98.66 215 95.38 295 -0.138 0.890 NGO wealth 18,960 80 14,561 215 15,754 295 0.427 0.670 Proportion that raise voluntary 0.98 80 0.91 215 0.93 295 1.887 0.060 contributions Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). Fa{champs and Owens 307 difference between grant recipients and nonrecipients is also reported: NGO leaders are more likely to have a substantially higher level of education, to have more work experience, to have previously worked for the government, and to have other current employment with an NGO. They are less likely to have any other current employment. Grant recipients are also older, more likely to be a subsidiary of a foreign NGO, and more likely to belong to a Ugandan NGO network. These findings suggest that personal contacts matter. NGO leaders that receive grants tend to be better connected. The experience and qualifications of the NGO and its leader also matter, suggesting that grant funding goes to more competent NGOs. In contrast, the wealth and parental background of the NGO leader do not show a systematic relationship with grant recipient status. This is the first evidence that donors regard local NGOs more as subcontrac­ tors than as altruistic partners. Grant recipients are less likely to raise voluntary contributions from members and local private donors (the difference is not large but is statistically significant). This is because most Ugandan NGOs raise some contributions from members. In aggregate, grants account for about 80 percent of total NGO funding in Uganda, whereas internal and local funding from private con­ tributors accounts for less than 3 percent. 13 But NGOs differ widely in the pro­ portion of their funding that comes from local private hands because most grant funding goes to only a few NGOs; the majority of Ugandan NGOs receive small grants or none at all. Section I hinted that if donors rely on NGO altruism, they should monitor them less. It follows that donors should use evidence of altruism, such as voluntary contributions by members and promoters, to decide how closely to monitor grant recipients. To investigate this idea, the analysis turns to whether donors are more likely to monitor NGOs for which voluntary contributions Ci are zero. To this effect, NGOs required to supply monthly and half-yearly financial accounts are exam­ ined. Two groups of grant recipients are compared: those that receive only a grant and no voluntary contributions and those that receive both. Recipients that receive both are less likely to have to report financial accounts. The differ­ ence is significant at the 1 percent level for monthly reports and at the 10 percent level for half-yearly reports. This suggests that, consistent with model predictions, NGOs that depend on grant funding have more stringent monitor­ ing requirements. Reduced Form Regressions Inference based on univariate comparisons can be misleading because explana­ tory variables often interact with each other. This section turns to multivariate analysis, estimating reduced forms of regressions (2) and (3). First, the 13. The remaining 17 percent is from business income. 308 THE WORLD BANK ECONOMIC REVIEW determinants of success in obtaining a grant with only characteristics related to the NGO, excluding the characteristics of the NGO leader, are considered. The first column of table 2 shows estimates from a probit and confirm several of the univariate findings: the likelihood of receiving a grant increases with the age of the NGO, when the NGO is a subsidiary of a foreign NGO and when the NGO belongs to a Ugandan network of NGOs. While the first finding may be indicative of NGO experience, the other two probably capture the role of personal contacts in accessing grant funding. These findings suggest that donors have difficulties identifying NGOs they can trust and thus rely on networks to screen grant recipients. As section I points out, this would result in repeated interaction to economize on screening and monitoring. This is indeed what the data suggest: of 161 surveyed NGOs that reported having ever received a grant, only 9 had never received one before. The NGO age effect is nonlinear, peaking at around 3 years of experience and falling thereafter. A significantly positive age coefficient emerges when the squared age term is dropped. A religious affiliation has a negative but not significant effect. Other variables, such as whether the NGO targets the poor or is based in the capital city Kampala, have no significant effect. Results are robust to the exclusion of outliers. 14 Next, NGO leader characteristics are included. The results in the second column of table 2 suggest that grant attribution is driven mostly by acquain­ tance, with no evidence that competence matters. The age and education of the leader are not significant and experience (proxied by length of tenure in the surveyed NGO and by previous experience in another NGO) has a negative influence on the likelihood of obtaining a grant. NGOs whose leader works at another NGO have a higher likelihood of obtaining a grant, a finding consist­ ent with the idea that contacts playa role in obtaining grants. As predicted by the model, wealth indicators have a negative effect: NGOs whose leaders had wealthy parents and who have a regular job elsewhere are less likely to have obtained a grant. The fourth column of table 2 shows similar results using grant revenue as the dependent variable, though grant revenue is available for only two-thirds of the respondents. The estimation uses the log(grant revenue + 1) to avoid losing zero observations and a tobit estimator is used to account for censoring. Results are by and large similar to those of column 2. The main difference is that being based in Kampala raises grant income, suggesting that NGOs based in the capital tend to receive larger grants. The only other variable that remains significant is current employment with another NGO. NGOs whose leader is employed by another NGO also seem to receive more grant funding. 14. One NGO in the sample has been in existence for over SO years. When this observation is dropped, the age squared term is no longer significant in the grant regression but remains significant in other regressions. One NGO is an outlier with respect to grant income; excluding it from the sample has little difference. % • M 11, k4 <- S@ TABLE 2. Determinants of Success in Obtaining a Grant 1 if received grant, 0 otherwise (probit) Log of grant revenue (tobit) Variable Coefficient t-value Coefficient t-value Coefficient t-value Coefficient t-value NCO characteristics Log NGO age 1.282 4.81 1.902 4.49 8.973 4.94 8.516 3.73 Log NGO age squared -0.224 -2.95 -0.293 -3.12 1.651 -3.34 -1.527 -2.68 Religious affiliation -0.325 -1.46 -0.411 1.57 -0.391 -0.30 0.685 0.47 Subsidiary of foreign NGO 0.614 1.73 0.137 0.34 3.593 2.24 4.168 2.22 Belongs to a network 0.828 3.61 0.910 3.65 2.689 2.01 2.549 1.74 Headquarters in Kampala 0.333 1.43 0.198 0.68 2.580 2.00 2.230 1.52 Targets the poor 0.071 0.38 0.096 0.41 -0.853 0.79 -0.705 -0.58 NCO leader characteristics Female -0.242 -0.92 1.102 0.75 Log age of leader -0.397 0.62 1.183 0.34 Log education of leader 0.539 1.20 2.329 0.82 Log length of time with NGO -0.505 -2.19 -1.018 -0.99 Previously worked for government 0.041 0.17 1.878 1.35 Previously worked for another NGO -0.542 -2.26 -1.727 -1.43 Currently works for another NGO 0.819 2.75 2.510 1.92 ;:p Currently has other employment 0.538 -2.12 -1.341 -1.09 ~ From a wealthy family -0.417 -2.01 -0.224 -0.20 . ;:s­ ~ Relative lives abroad 0.411 1.60 0.924 0.67 Constant R2 -1.285 -4.05 0.182 0.06 -7.227 -3.31 8.222 -0.56 ..,.., "::l' '" ;:t 0.276 0.387 0.069 0.079 Number of observations 278 229 190 164 0 ~ ;:t Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). '" V> 0 \0 310 THE WORLD IIA!-JK ECONOMIC REVIEW These results are then compared with those NGOs raising internal and local resources. Three indicators of local and internal funding are considered: revenue from fees and donations, proportion of full-time workers who are vol­ unteers, and whether the NGO receives complimentary use of equipment or vehicles from other sources. The first captures the main sources of internal and local finance, which is quite small in terms of aggregate funding. The other two capture in-kind resources. Volunteers account for 54 percent of full-time workers and 71 percent of part-time workers in the sector as a whole, so the contribution is non-negligible. A quarter of all NGOs use vehicles belonging to others and a quarter have complimentary use of equipment (such as compu­ ters) that does not belong to them. The same reduced form regressions are estimated for all three indicators, with and without NGO leader characteristics. Table 3 shows that the factors influencing internal and local resources are quite different from those influen­ cing grant funding. Being a subsidiary of a foreign NGO has a negative effect on local funding and volunteers, contrary to grant funding, which has a posi­ tive effect. NGO age has a large negative effect on volunteers and complimen­ tary use of equipment, suggesting that these are temporary palliatives used by young NGOs, not permanent ways of funding operations. Religious NGOs and NGOs that target the poor use fewer volunteers, a finding that is hard to recon­ cile with the idea of an altruistic motive for volunteering but that is consistent with volunteering being a way of jump-starting an NGO before it receives a grant. NGO leader characteristics also have a very different effect on local resources. Longer tenure at the current NGO is associated with more revenue from fees and donations, suggesting that experience is important in raising funds locally. Having an outside job has a positive effect on volunteering and complimentary use of equipment, two findings that are again consistent with efforts to jump-start an NGO with limited resources. This reduced form analysis suggests that the factors associated with success in attracting grant funding are quite different from those associated with raising resources internally or locally. Grant funding seems to be influenced largely by network effects-being a subsidiary of an international NGO, belonging to an NGO network, or having a leader who works for another NGO. Volunteers and complimentary equipment, in contrast, seem to be resources that young NGOs mobilize to jump-start operations, perhaps in the hope of obtaining grant funding later. Only fees and donations from local private sources depend on NGO leader experience. Testing for Crowding Out Next the analysis estimates regression (4) to determine whether NGOs that receive grants generate fewer voluntary donations of time and money. As in table 3, the dependent variables are revenues from fees and donations, pro­ portion of full-time workers who are volunteers, and a dummy variable for TABLE 3. Determinants of Success in Attracting Local Funding and Resources Proportion of volunteers in workforce 1 if use of equipment or vehicle Log of fees and donations (tobit) (ordinary least squares) (probit) Variable Coefficient t-value Coefficient t-value Coefficient t-value Coefficient t-value Coefficient t-value Coefficient t-value NGO characteristics Log NCO age -2.911 -2.60 -3.521 2.58 -0.163 -3.34 -0.128 1.97 -0.355 -1.70 0.637 -2.17 Log NCO age squared 0.600 1.85 0.652 1.88 0.022 1.63 0.010 0.68 0.048 0.81 0.112 1.52 Religious affiliation 0.436 0.48 0.606 0.63 -0.167 -3.69 -0.142 -2.86 0.110 0.60 0.017 0.08 Subsidiary of foreign NCO -4.961 -3.72 5.712 -3.87 -0.193 3.00 -0.186 -2.77 0.059 0.25 0.029 -0.11 Belongs to a network -0.133 -0.15 -0.669 -0.71 -0.038 -0.81 -0.007 -0.14 -0.025 -0.14 -0.138 -0.67 Headquarters in Kampala -3.777 -4.08 3.050 -3.15 -0.077 1.55 -0.015 -0.26 -0.294 1.54 -0.371 -0.16 Targets the poor -1.455 1.92 -1.381 -1.75 -0.076 -1.88 -0.089 -1.98 -0.173 -1.11 -0.146 -0.80 NGO leader characteristics Female -0.666 -0.68 0.000 0.00 -0.361 -1.68 Log age of manager 0.754 0.36 -0.086 -0.77 -0.436 0.89 Log education of manager -2.076 -0.55 -0.074 -0.55 0.757 1.57 Log length of time with NCO 1.754 0.51 0.018 0.51 0.144 0.92 Previously worked for -0.811 0.56 0.026 0.56 -0.213 1.08 government Previously worked for another -0.775 0.68 0.031 0.68 -0.207 -1.15 NCO Currently works for another -0.291 0.14 0.007 0.14 -0.046 -0.23 ;,r ~ NCO ::!"­ ;;, Currently has other 1.204 3.06 0.148 3.06 0.414 2.22 ;i employment """ to ;;, From a wealthy family 0.386 -1.46 -0.062 1.46 -0.080 -0.46 ;:t ..... Relative lives abroad -1.447 -0.05 -0.002 -0.05 -0.098 -0.49 0 Constant 9.433 7.34 10.960 2.40 0.851 16.35 1.292 2.40 0.349 1.43 0.357 0.16 ~ R2 0.069 0.100 0.245 0.305 0.033 0.019 ;:t Number of observations 190 164 274 225 278 229 '" Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). .... w ..... 312 THE WORLD BANK ECONOMIC REVIEW complimentary use of equipment and vehicles. The grant variable is a dummy variable that takes value 1 if the NGO has ever received a grant. The dependent variables are first regressed on a grant-funding dummy vari­ able and a series of control variables. These control variables include the same NGO characteristics as in table 3 as well as a series of NGO leader character­ istics. Table 4 shows a negative conditional correlation between grants and all three categories of voluntary contributions. The grant variable is significant in the fees and donations regression and nearly significant at the 10 percent level in the volunteers regression. To address the endogeneity of the grant variable, variables are needed that predict grant funding but are conditionally uncorrelated with receiving local donations and resources. Without a controlled or quasi-experiment there are no truly exogenous instruments, so suitable instruments must be found in the data. The previous section showed that variables that proxy for how socially con­ nected the NGO leader is may serve as instruments since they predict receiving grants but not raising local contributions. The following variables are thus used as instruments: the length of time the leader has been with the NGO, whether the leader previously worked for the government, whether the leader has other employment, and whether the leader has a relative living abroad. These variables may help the NGO obtain the necessary contacts with inter­ national donors, but once NGO characteristics are controlled for, they do not appear to help the NGO raise local private funds. But the possibility that they do cannot be rejected a priori. Results should thus be interpreted with a grain of salt, given the dearth of evidence. The instrumenting regression is shown in table 5.15 The instruments are jointly significant, but the F-statistic is less than 10, which indicates a weak instrument problem. Table 6 shows instrumented regression results for voluntary contributions in cash and in kind. An instrumental variable tobit is estimated for fees and donations, an instrumental variable linear regression for share of volunteers, and an instrumental variable probit for complimentary use of equipment and vehicles. The growing literature on coping with weak instruments recommends using corrected confidence intervals. Since the instruments in the regressions are weak, the estimated p-value from a corrected likelihood ratio test proposed by Moreira (2001) is also reported. 16 Several specification tests are also con­ ducted and the exogeneity of the grant variable is rejected in all three regressions. 17 Overidentification restrictions are tested for the second 15. Because the number of observations varies across estimations, there are three instrumenting regressions. To save space, only the one for fees and donations is shown. The other two are qualitatively similar. 16. The Moreira p-value correction exists only for linear models. The values reported in table 5 are thus based on linear implementations of the regressions. 17. A Wald test is used for the first (tobit) and third (probit) regressions, and a Hausman test for the second (linear) regression. «idl ft •• , .t w J,,4Mt* Ii! A II! 4 ," tl tU TABLE 4. Determinants of Success in Raising Voluntary Contributions Proportion of volunteers in Log of fees and donations workforce (ordinary least 1 if use of equipment or (wbit) squares) vehicle (probit) Variable Coefficient t-value Coefficient t-value Coefficient t-value NCO characteristics Received a grant -4.021 -3.84 -0.101 1.61 0.272 -1.16 NCO characteristics LogNGO age -0.133 -0.11 0.10 -1.69 -0.40 -1.57 Log NCO age squared 0.050 0.15 om 0.49 0.07 1.09 Religious affiliation 0.155 0.17 -0.18 -3.74 0.Q3 0.15 Subsidiary of foreign NCO 5.293 -3.71 -0.18 -2.78 0.03 0.12 Belongs to a network 0.421 0.44 -0.01 -0.17 -0.03 -0.15 Headquarters in Kampala -3.021 3.25 --0.Q3 -0.44 -0.40 -1.97 Targets the poor 1.798 -2.32 -0.Q7 1.67 -0.10 -0.57 NCO leader characteristics Female -1.355 1.44 0.02 -0.49 -0.39 -1.85 Log age of leader 0.416 -0.24 -0.03 -0.28 -0.29 -0.70 Log education of leader -2.372 -1.56 -0.08 -0.59 0.55 1.57 Previously worked for another NCO -1.451 1.87 0.00 -0.06 -0.14 -0.78 « -;::-. Currently works for another NCO 0.486 0.58 0.05 1.01 0.10 0.52 if From a wealthy family 0.067 -0.1 -0.05 1.15 0.05 -0.34 .g Constant 18.581 2.43 1.25 2.47 0.37 0.21 '" I::. Number of observations 164 229 229 ~ Note: Due to space constraints, only the first instrumenting regression is included. The results for the other two are nearly identical, differing only o because of sample size and estimation method. ~ ;;; Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). w .... w 314 THE WORLD BANK ECONOMIC REVIEW TABLE 5. Instrumenting Regression on Receiving a Grant Variable Coefficient t-value NCO characteristics LogNGO age 4.950 4.98 Log NGO age squared -0.094 -3.71 Religious affiliation -0.071 1.04 Subsidiary of foreign NGO 0.119 1.30 Belongs to a network 0.206 3.07 Headquarters in Kampala 0.094 1.34 Targets the poor -0.054 -0.93 NCO leader characteristics Female -0.210 -0.30 Log age of leader -0.176 1.12 Log education of leader 0.006 0.05 Previously worked for another NGO -0.128 -2.25 Currently works for another NGO 0.145 2.35 From a wealthy family -0.105 1.97 Instruments Log length of time with NGO -0.082 -1.65 Previously worked for government 0.096 1.46 Currently has other employment -0.057 -0.97 Relative lives abroad 0.131 1.98 Constant 1.025 1.59 Centered R2 0.404 F-statistic p-value Joint F-test of instruments F(17,146) 5.81 (0.000) Number of observations 164 Note: Due to space constraints, the instrumenting regression is shown for fees and donations only. The results for the other two instrumenting regressions are nearly identical, differing only slightly because of sample size and estimation method. Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). regression, which is linear, and they are not rejected: the Sargan X2 statistic has a value of 3.92 and a p-value of 0.271. Admittedly, in the absence of truly exogenous instruments these test results offer only partial reassurance. Instrumented regression results confirm that grants are negatively correlated with higher local resources: the instrumented grant variable has a negative sign in all three regressions. The effect is large for fees and donations and significant at the 5 percent level for all three dependent variables. Similar results are obtained if NGO leader characteristics are omitted. ls Despite the difficulty of obtaining the information required for such a test and the absence of 18. Because revenue from fees and donations is reported only for a subset of respondents, reporting bias is addressed using a Heckman selection model. Results are not reported here to save space. The key finding is that when selection is controled for, the grant variable remains negative and significant. The selection equation also suggests that selectivity bias is not an issue. ill P" ,! 2._ !!R'llJl!WJlU*- ,t! i ukt « TABLE 6. Determinants of Success in Raising Voluntary Contributions, instrument Results Proportion of volunteers in Log of fees and donations workforce (ordinary least 1 if use of equipment or (tobit) squares) vehicle (probit) Variable Coefficient t-valuc Coefficient t-value Coefficient t-value Received a grant p-value corrected for weak instruments a 13.943 -2.76 (0.003) -0.662 1.95 (0.024) -2.789 - 2.06 (0.029) NCO characteristics Log NGO age 4.074 1.59 0.137 0.91 0.559 0.91 Log NCO age squared -0.782 -1.33 -0.037 -1.20 -0.096 -0.76 Religious affiliation -0,480 -0.39 -0.217 3.30 0.214 -0.76 Subsidiary of foreign NCO -4.048 -2.22 -0.129 -1.49 0.260 0.69 Belongs to a network 2.182 1,45 0.125 1.26 0.440 1.08 Headquarters in Kampala -1.564 -1.19 0.004 0.06 -0.285 -1.07 Targets the poor -2.174 -2.14 -0.065 1.24 -0.079 -0.35 NCO leader characteristics Female 1.023 -0.86 -0.022 -0.36 -0.537 -2.02 Log age of leader -0.693 -0.30 0.093 0.73 -0.781 1.40 Log education of leader -2.010 1.03 0.031 0.25 1.101 2.17 Previously worked for another NGO -2.814 -2.40 0.046 -0.76 0.511 1.94 ~ Currently works for another NGO 1.858 1.44 0.139 1.81 0.440 1.40 ~ ;::­ From a wealthy family -1.027 1.06 0.097 1.80 -0.232 1.05 Constant Centered R2 22.439 0.282 2.19 1.331 0.019 2,47 1.636 0.70 ~ ~ Number of observations 164 229 229 \:)., o aBased on the conditional likelihood ratio test proposed by Moreira (2001) and implemented in Stata using condivreg. ~ Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). ~ - (.;.> v. 316 THE WORLD BANK ECONOMIC REVIEW TABLE 7. Fixed Effect Estimation Revenue from fees Revenue from and donations Revenue from fees donations Variable Coefficient t-value Coefficient t-value Coefficient t-value Log grant revenue 0.066 0.98 0.135 2.76 -0.054 -0.89 Year dummy (2000 1) -0.383 -1.94 1.070 0.75 -0.522 2.93 Overall R2 0.12 0.08 0.09 F test that fixed effects = 0, p-value 0.000 0.000 0.000 Number of observations 352 352 352 Source: Authors' calculations based on survey data from Barr, Fafchamps, and Owens (2003). rigorous evidence on this issue in developing economies, these results provide valuable-if impressionistic-information. The findings reported in tables 4 and 6 are consistent with crowding out: NGOs that obtain grant funding appear to raise fewer resources locally. But the findings may be misleading because they do not control for unobserved het­ erogeneity. To investigate this possibility, NGO fixed effects are introduced. Respondents were asked to provide retrospective income data for 2000 and 2001. Regression (4) can be estimated, and the hypothesis that NGOs receive fewer fees and donations from private sources after receiving a grant to be tested. Table 7 shows the results from an NGO fixed effect regression of revenue from fees and donations on grant revenue. Increased grant revenue is associated with increased income from fees but not with increased donations. The total net effect on contributions from private sources is not significant. It therefore appears that once unobserved heterogeneity among NGOs is accounted for, the evidence of crowding out disappears. The contradiction between the two sets of results suggests that the evidence of crowding out in tables 4 and 6 is due in fact to a selection effect: NGOs that are on average more successful at obtaining grants from international donors are significantly less likely to raise local resources. But once an NGO receives a grant, there is no evidence that it reduces internal funding. If anything, the income it generates from membership fees increases. This is probably because grant revenue enables the NGO to offer more services to members, and in exchange it receives more user fees. This interpretation is reinforced by the observation that most NGOs offer services to their members (Barr, Fafchamps, and Owens 2005). At the same time, income from donations-probably a concept closer to altruistic contributions-does not fall with grant income, suggesting that crowding out is not present at the individual NGO leveL Alternative explanations for a negative correlation between grants and local funding are possible without necessarily blaming NGOs' lack of altruism. One is that some NGOs work on a task or issue for which the possibility of raising local funds is limited, for example, certain types of advocacy work (for the '" Fafchamps and Owens 317 environment, women, and the like). To pursue a specific agenda, these NGOs must rely on grant funding. Although this phenomenon may arise in some countries, it is unlikely to account for the pattern observed in Uganda. The overwhelming majority of NGOs surveyed remain unspecialized, adopting a holistic approach to development (Barr, Fafchamps, and Owens 2005). So if local funds for certain types of activities were limited, the overwhelming majority of NGOs surveyed could find in their large portfolio of self-professed interests an activity that suits local benefactors. In fact, this is precisely what most NGO do with respect to international donors. Another possibility is that NGOs reduce local fundraising from user fees to increase beneficiary demand for their services. This could generate a negative relationship between grant funding and income from user fees. As in table 6, the opposite occurs: grant recipients collect more user fees, presumably because grant income enables them to undertake activities for which user fees can be collected. To summarize, the analysis here suggests that the NGOs that seek grants and are good at obtaining them differ from those that are less successful at securing grants. It seems as though international donors do not seek out the most altruistic and charity-minded NGOs when allocating grants. Combined with the earlier results that proxies for altruism are not correlated with secur­ ing grants, donors may indeed see local NGOs more as subcontractors than as local charitable organizations to be encouraged by outside assistance. IV. CONCLUSION This article examines the determinants of internal and external funding for NGOs in Uganda. Statistically, the results are not very strong: the sample size is small, and the data show a lot of measurement error, making inference diffi­ cult. A controlled or quasi-experiment cannot be used to address causal infer­ ence issues in a completely convincing way. Despite these shortcomings, the results provide valuable, even if tentative, evidence on a topic characterized by an abundance of unsubstantiated claims and a dearth of hard evidence. Success in securing grant funding depends primarily on networking, for example, whether the NGO is member of an NGO network or umbrella organ­ ization, whether it is a subsidiary of a foreign NGO, and whether the NGO leader works at another NGO. This may be because donors find it difficult to screen local NGOs and so tend to rely on networks to access relevant infor­ mation. NGO experience matters, but peaks after only 3 years. Variables proxying for NGO leader competence are not significant, and leader experience and wealth reduce the likelihood of obtaining a grant. Donors more closely monitor NGOs that raise no local resources and tend to provide grants repeat­ edly to the same NGOs. Different factors are associated with raising local resources, either through vol­ unteers, member fees and donations, or complimentary use of vehicles and equip­ ment. Very young organizations, often led by someone with regular employment 318 THE WORLD BANK ECONOMIC REVIEW elsewhere, seem to resort to volunteers and complimentary equipment. Leader experience appears to matter only in raising funds from fees and local donations. A cross-section analysis yields evidence of crowding out: Ugandan NGOs that receive grants raise fewer resources locally. But the same analysis with NGO fixed effects causes the evidence of crowding out to disappear. Instead, income from member fees increases when an NGO receives more grant funding. Donations from members, in contrast, remain unchanged. This suggests that grant recipients do not reduce local funding after receiving a grant. The evidence of crowding out from cross-section regressions is probably due to a selection effect: donors select NGOs that are on average less involved in raising local resources, which would happen if donors regard NGOs as (for­ profit) subcontractors of their development efforts. These findings contradict the reason often given to justify channeling development funds through NGOs, namely that they are more altruistic than government agencies and thus are less likely to divert development funds for personal gain. There may be reasons besides altruism for channeling development assistance through NGOs rather than government agencies. For instance, NGOs may have a lower cost of service delivery, donors may have a better control over spending and activities, or donors may seek to further a philosophical or ideological objective that they could not pursue through secular government agencies. In Uganda, most NGOs are extremely small and un specialized (Barr, Fafchamps, and Owens 2005). It is thus unlikely that they offer a lower cost of delivery since they cannot capture returns to scale and to specialization. 19 But because they are more flexible and can be activated more quickly than govern­ ment services, NGOs may be well suited for relief operations and for small, localized, or unconventional interventions. This is consistent with Barr, Fafchamps, and Owens (2003), who report that Ugandan NGOs focus on light interventions rather than on the long-term delivery of curative health and fu11­ time education. Tighter financial control over development assistance may also explain why donors prefer NGOs. Barr, Fafchamps, and Owens (2003) have shown that Ugandan NGOs are subject to numerous forms of monitoring by grant agencies. These issues deserve further investigation. ACKNOWLEDGMENTS The authors thank Professor Sam Tulya-Muhika, Kintu Nyago, and their team of enumerators for assisting in the data collection. They are also grateful to Peter Ssentongo from Uganda's Office of the Prime Minister and Mary Bitekerezo of the World Bank's Uganda Country Office in Kampala for their 19. This does not apply to the Catholic Church, the Church of Uganda, and the Uganda Muslim Supreme Council, which are all very active in delivering social services but are not registered as NGOs under Ugandan law. t 1 4 jJ , 2 t J, I 1(" , dttu t-.l", $ " Uk n 1 Fafchamps and Owens 319 assistance and comments. They are also very grateful to Abigail Barr for her very valuable comments and assistance in data collection, cleaning, and prep­ aration. The support of the U.K. Economic and Social Research Council is gratefully acknowledged; this work is part of the program of the Council's Global Poverty Research Group. FUNDING Funding this research was given by the World Bank and the Japanese government. ApPENDIX Proposition 1. Organizations with more resources (higher 1') or more altruism (lower w) give more-have a higher t-while more competent organizations (higher z) give less. Proof. The 11rst- and second-order conditions (SOC) for an interior optimum are of the form: 8V~!2~ _ wU'(T t) = 0, 8t (A -1) Ef2v w U" < O. 8t2-+ Using simple comparative statics, dtldT, dtldz, and dtldw can be written as: &V UII) d t - w Uf! d T = 0, ( 8i2 + w (A - 2) dt wU" dT= SOC> O. 2 8 V U")d t ( --+w 8t2 U' dw= 0, (A - 3) dt U' d~= SOC < O. Ef2 V ") d Ef2v + 8t8z dz = ( 8t 2 + wU t 0, (A - 4) dt = 82~y/8t8z < 0 dz SOC . 320 THE WORLD BASK ECONOMIC REVIEW As anticipated, organizations with more resources (higher T) or more altruism (lower (0) give more, while more competent organizations (higher z) give less. Proposition 2. If fi2V/at 2 < (> )0, the amount given t increases less (more) than proportionally with NGO resources T. Proof. Totally differentiating the first-order condition yields: dt wU" dT = a2 v71.itT+wUII (A - 5) 1 I I It follows that I I (A - 6) ~={<1 if (a V jat2 ) < 0 2 2 I dT > 1 if ({j2 V j at ) > 0 I Since U" < O. I I I , I REFERENCES I Andreoni, James, and Abigail Payne. 2003. "Do Governemnt Grants to Private Charities Crowd Out I Giving or Fund-raising?" American Economic Review 93(3);792-812. Anheier, Helmut K., and Lester M Salamon. 2006. "The Nonprofit Sector in Comparative Perspective." I In Walter W Powell, and Richard Steinberg, eds. The Nonprofit Sector: A Research Handbook. New I Haven, CT: Yale University Press. Barr, Abigail, and Marcel Fafchamps. 2006. "A Client-Community Assessment of the NGO Sector in Uganda." Journal of Development Studies 42(4);611-39. Barr, Abigail, Marcel Fafchamps, and Trudy Owens. 2003. "Non-Governmental Organizations in Uganda." Report to the Government of Uganda. Oxford University, Centre for the Study of African Economies, Oxford, UK. --.2005. "The Governance of Non-Governmental Organizations in Uganda." World Development 33(4):657-79. Edwards, M, and David Hulme. 1995. Non-Governmental Organizations: Performance and Accountability: Beyond the Magic Bullet. London: Earthscan. Hausmann, H. 1980. "The Role of Nonprofit Enterprise." Yale Law Journal 89:835-901. Hulme, 0, and M. Edwards. 1997. NGOs, States and Donors: Too Close for Comfort? Basingstoke, UK: Macmillan. Lindelow, Magnus, Ritva Reinikka, and Jakob Svensson. 2003. "Health Care on the Frontlines: Survey Evidence on Public and Private Providers in Uganda." Policy Research Working Paper. World Bank, Washington, DC. Moreira, Marcelo. 2001. "Tests with Correct Size when Instruments Can Be Arbitrarily Weak." Working Paper 37. Berkeley: Center for Labor Economics, University of California. Platteau, Jean-Philippe, and F. Gaspart. 2003. ~The Risk of Resource Misappropriation in Community-Driven Development." World Development 31 (10); 1687 -703. .. 4. ttt P UM; '. CF/K + CF/K" SOE 0 1.00 1.37 ;:e p-value "'­ 0.32 0.24 Note: All specifications include firm and year fixed effects. Numbers in parentheses are standard errors. "Significant at the 10 percent level. *" Significant at the 5 percent level. *' S ~ ;:e l.'! ~'**Significant at the 1 percent level. w Source: Authors' analysis based on data from two World Bank surveys of Czech firms in 2003 and 2004 and Bureau van Dijk Electronic Publishing w w (2005); see text for details. 334 THE WORLD BANK ECONOMIC REVIEW Regression 3 examines whether the link between cash flow and investment differs between multinational suppliers and other firms. The model includes a dummy variable that takes the value of 1 in each year in which the firm supplies an MNC operating in the Czech Republic and 0 otherwise. The dummy variable is also interacted with cash flow. If firms with linkages to mul­ tinationals find it easier to obtain credit, the sum of the coefficients on cash flow and the interaction term should not be statistically significant. While cash flow continues to bear a positive and statistically significant coefficient, the interaction term is negative and statistically significant at the 1 percent level. The F-test indicates that the hypothesis that the sum of the two coefficients is equal to 0 cannot be rejected, suggesting that, unlike nonsuppliers, multina­ tional suppliers do not face liquidity constraints. Neither labor productivity nor its interaction with cash flow reaches conventional significance levels. The supplier dummy variable is not statistically significant, suggesting that multina­ tional suppliers do not differ in their investment behavior from other firms. 8 Next, the analysis tests for whether the finding that multinational suppliers are less credit constrained is due to firms being exporters rather than to their being multinational suppliers. Exporting firms may be less credit constrained because of a steady stream of income from more creditworthy foreign custo­ mers, and their experience dealing with foreign buyers may better position them to become multinational suppliers. Potential firm-level determinants of investment behavior (size, age, and debt level) are also controlled for. The findings are robust to these additional controls. The coefficient on the interaction between the multinational supplier dummy variable and cash flow remains negative and statistically significant at the 1 percent level. As before, the F-test suggests that multinational suppliers do not face liquidity constraints. In contrast, exporters appear to be as liquidity constrained as other Czech firms. The interaction term is not statistically significant, and the F-test rejects the absence of a link between investment and cash flow. 9 The likely expla­ nation is that many Czech firms that continued to sell to their Slovak customers after Czechoslovakia split in 1993 are considered to be exporters, yet their Slovak buyers are unlikely to be more creditworthy than Czech buyers. This also explains why such a high percentage of observations in the sample pertain to exporters. 10 The additional controls for size, age, and debt level do not appear to be statistically significant. Lizal and Svejnar (2002) find that state enterprises in the Czech Republic were facing soft budget constraints in the 1990s. As there are only 19 state enterprises in the sample, many of which were privatized during the period considered, there is little concern that their presence affects the main findings. Nevertheless, 8. Some differences may be captured by firm fixed effects included in the model. 9. Excluding the supplier dummy variable and its interaction with cash flow from the model would not change this conclusion. 10. The Slovak Republic is the second largest export market for Czech firms. I,«M bi :. _ *1 j_ 1m javorcik and Spatareanu 335 regression 5 adds a state enterprise dummy variable and its interaction with cash flow. Neither variable appears to be statistically significant, but as expected the F-test cannot reject the hypothesis that state enterprises are not credit con­ strained. The finding on multinational suppliers remains unchanged. The last column in table 2 includes all the controls listed in equation (1) and confirms the earlier conclusions. The cash flow variable has a positive and stat­ istically significant coefficient, and its interaction with the multinational sup­ plier dummy variable is negative and significant at the 1 percent level. Based on these coefficients and the F-test, Czech firms in general appear to be liquid­ ity constrained, but multinational suppliers do not. As before, the results suggest that state enterprises may be subject to soft budget constraints. Are Future Multinational Suppliers Less Credit Constrained? As mentioned, it is possible that less liquidity constrained firms self-select as suppliers to MNCs. Because multinational customers tend to have higher requirements for quality, technological sophistication, and on-time delivery than domestic buyers in developing and transition economies, becoming a mul­ tinational supplier is likely to be associated with some fixed cost for local firms. Thus, it may well be the case that only firms not facing liquidity con­ straints are able to become multinational suppliers. This possibility is examined by checking whether multinational suppliers appear to be less liquidity con­ strained than other firms before they start their contracts with multinationals, as estimated by the following model: lit! Kit-l = f30 + f31 11Sit!Sit-l + f32 CFit! Kit- 1 + f33CFitjKit-1 * Supplierit + f34Supplierit + f3 S CFit/Kt-l * 1 yr beforeit + f361 yr beforeit (2) + f37CFit!Kt-l * 2 yrs beforeit + f382 yrs beforeit + f39CFit!Kjt-l * In(VAjL) + f3lOln(VAjL) + f311In(Sizeit) + f312 ln (Agejt) + f313Debt ratioit + Vi + Vt + Uit where 1 yr beforeit equals 1 at time t if firm i will become a multinational sup­ plier at t + 1, and 0 otherwise, and 2 yrs before,t equals 1 at time t if firm i will become a multinational supplier at t + 2, and 0 otherwise. A sum of f32 and f37 equal to 0 would indicate that multinational suppliers were not credit constrained two years before starting their relationship with an MNC. A sum of f32 and f3s equal to 0 would suggest that multinational suppliers were not facing credit constraints one year before starting their relationship with an MNC. Either or both findings would suggest self-selection of unconstrained firms into becoming multinational suppliers. The estimation results of equation (2) are presented in table 3. Regression 1 looks at whether multinational suppliers were liquidity constrained one year TABLE 3. Current Suppliers, Future Suppliers and Nonsuppliers w w 0'; Variable (1) (2) (3) (4) (5) .., L\.Sales 0.085* ** (0.032) 0.084'" (0.032) 0.078" (0.033) 0.078** (0.033) 0.063* (0.034) ;r m CF/K 0.482·" (0.048) 0.482 **. (0.048) 0.432*** (0.051) 0.432"* (0.051) -0.326 (0.636) >l1 CF/K • 2 yrs before --0.26 (0506) -0.234 (0.506) - 0.232 (0.507) - 0.258 (0.498) 0 ;:0 CF/K * 1 yr before -0.510'" (0.133) -0.515'" (0.133) -0.465'" (0.134) -0.465'" (0.134) -0.622*** (0.164) r-< t:I CF/K * Supplier --0.439*** (0.067) -0.440'" (0.067) -0.399'" (0.068) -0.400'" (0.069) -0.395 u > (0.111) 2 yes before -0.048 (0.102) -0.065 (0.102) -0.065 (0.102) -0.061 (0.100) "' ;,. z 1 yr before 0.02 (0.079) -0.014 (0.087) -0.037 (0.087) -0.037 (0.087) -0.023 (0.086) :>: Supplier 0.D3 (0.073) -0.01 (0.084) -0.032 (0.084) -0.032 (0.088) -0.034 (0.083) CF/K * In(VA/L) 0.008 (0.007) 0.008 (0.007) 0.011 (0.008) 0.011 (0.008) -0.015* (0.009) '" n 0 In(VA/L) -0.009* (0.005) -0.009 (0.005) -0.013** (0.007) -0.013" (0.007) -0.002 (0.007) z 0 Debt/K 0.011 (0.007) 0.011 (0.007) 0.015'> (0.007) !:: In(Employment) -0.069 (0.059) -0.069 (0.059) -0.028 (0.058) n -0.064 (0.069) ;:0 In(Age) -0.072 (0.070) -0.073 (0.070) m Supplier' Year 1999 -0.009 (0.066) < m Supplier' Year 2000 0.010 (0.072) >l1 Includes interactions of CF/K with two-digit industry fixed effects Number of observations 1382 1382 1359 1359 1359 Number of firms 314 314 307 307 307 R2 0.19 0.19 0.16 0.16 0.22 F-test CF/K + CF/K > Supplier 0 0.46 0.44 0.26 0.25 1.25 p-value 0.50 0.51 0.61 0.62 0.26 CF/K + CF/K'l yr before 0 0.04 0.06 0.06 0.06 2.09 p-value 0.83 0.80 0.80 0.80 0.15 CF/K + CF/K ' 2 yrs before= 0 0.19 0.15 0.16 0.53 p-value 0.66 0.70 0.69 0.47 Note: All specifications include firm and year fixed effects and a constant. Numbers in parentheses arc standard errors. "Significant at the 10 percent level. * *Significant at the 5 percent level. ** *Significant at the 1 percent level. Source: Authors' analysis based on data from two World Bank surveys of Czech firms in 2003 and 2004 and Bureau van Dijk Electronic Publishing (2005); see text for details. • Javorcik and Spatareanu 337 before they started their relationship with an MNC. As before, the coefficient on cash flow is positive, though slightly larger, and statistically significant at the 1 percent level. The interaction terms between the multinational supplier dummy variable and cash flow and between future supplier and cash flow are both negative and statistically significant at the 1 percent level. F-tests suggest that, unlike Czech firms in general, neither current nor future multinational suppliers face liquidity constraints. Regression 2 considers the two-year period before starting a relationship with an MNC. The interactions of cash flow with 1 yr before and supplier remain negative and statistically significant. The coefficient on the interaction with 2 yrs before is negative, though not statistically significant. F-tests cannot reject the hypothesis that multinational suppliers are not liquidity constrained and that this lack of constraints is already present in the two-year period before becoming a supplier. Regression 3 shows that the findings are robust to controlling for firm size, age, and debt level. In sum, the findings are suggestive of unconstrained firms self-selecting into becoming multinational suppliers. To take into account a currency crunch that took place in the Czech Republic in 1999-2000 following a banking crisis (see Pruteanu 2004), an interaction of the supplier dummy variable with a dummy variable for year 1999 (and 2000) is added to the specification. Doing so will shed light on whether multinational suppliers were affected differently by the credit crunch: multinationals with their global distribution networks are less affected by changes in the Czech market and thus less likely to adjust their relationships with their suppliers. As evident from regression 4, however, there is no indi­ cation of any different investment behavior among multinational suppliers than among other firms during the credit crunch period. Neither interaction term is statistically significant. Other conclusions remain unchanged. To account for the possibility that firms in growing sectors might be more likely to be both multinational suppliers and not liquidity constrained, inter­ actions between dummy variables for two-digit NACE codes (18 in total) and the cash flow variable are added. Only two of these interaction terms are stat­ istically significant (furniture; computer, electronic, and optical products). The results confirm the previous findings that suppliers to MNCs are not liquidity constrained and that the effect is already present two years before signing a contract with an MNC. This specification also finds a significant positive coef­ ficient on the debt variable and a significant negative coefficient on the inter­ action between cash flow and labor productivity. One may wonder about the results of F-tests based on the interaction of cash flow and labor productivity as well as the interaction of cash flow and current (or future) supplying status. F-tests taking into account the average labor productivity among current (or future, as appropriate) suppliers support the earlier conclusions: both current and future multinational suppliers do not appear to be credit constrained. 338 THE WORLD BANK ECONOMIC REVIEW Finally, additional robustness checks (not reported to save space) show that the conclusions are not affected by dropping observations with negative values for cash flow or by including industry-year fixed effects. Another way to shed light on the link between credit constraints and multinational supplying status is to estimate a probit model that aims to explain the supplying status with the lagged liquidity ratio, gross profit (logged), and debt (normalized by capital). Supplierjt is the dependent variable. Liquidity ratio is defined as the difference between current assets and current liabilities divided by total assets. This specification also controls for firm size (number of employees), age, and labor productivity (all in logs) as well as three-digit industry and year fixed effects. The results show a positive and statistically significant link between lagged liquidity ratio, lagged gross profit, and the probability of being a multinational supplier (table A-I). Coefficients on debt, employment, and labor productivity are not statistically significant. As this last finding is somewhat puzzling, firm performance was also measured using total factor productivity estimated by the sector-specific production function (ordinary least squares or the Olley­ Pakes 1996 method). Once liquidity ratio, gross profit, and debt are controlled for, firm productivity is not a statistically significant predictor of supplying status. Finally, the data also indicate that younger firms are more likely to supply MNCs. l l In sum, the findings suggest that firms not facing liquidity constraints self-select into becoming multinational suppliers. This is consistent with the observation that to obtain contracts from MNCs firms need to meet the strin­ gent req uirements of multinational customers and that only firms with access to financing may be able to do so. The survey data are in line with these con­ clusions. Most suppliers make improvements within the 12-month period before signing a contract with an MNC. The most frequent changes include improvements to product quality, staff training, and productivity enhance­ ments. Many of these changes are probably made to obtain ISO certifications. More than 40 percent of suppliers reported being required by prospective mul­ tinational customers to obtain ISO certification. As the certification process is quite costly, usually involving the services of a specialized consulting firm, it would not be surprising if only firms that were not liquidity constrained were able to complete it. Robustness Checks To eliminate the possibility that the findings could be driven by MNCs extending credit to future suppliers, the 15 Czech firms that reported receiving some financial help from their multinational customers were removed from the 11. In a probit model predicting the decision of Czech firms to become multinational suppliers rather than the decision to supply MNCs in a given year, liquidity ratio and firm size were the main predictors of the decision to become a multinational supplier. # J ) ,Jt. t )@ if, J .. & & . Javorcik and Spatareanu 339 sample. The results confirm the earlier pattern. Multinational suppliers were not liquidity constrained two years before supplying an MNC, and they remained unconstrained while supplying the multinational (table 4, regressions 1 and 2). To examine whether the findings are due to the possibility that future multi­ national suppliers have a lower credit risk because of a contract with an MNC, Czech suppliers that reported that having a relationship with an MNC helped them obtain financing are dropped from the sample. Eliminating these 24 firms does not affect the results (regressions 3 and 4). The finding that multinational suppliers are less credit constrained is thus confirmed, and the evidence suggests that less constrained firms self-select into becoming multinational suppliers. Instrumental Variable Approach With the evidence suggesting self-selection by less credit constrained firms into supply relationships with MNCs and the possibility that some explanatory variables are endogenous, the final step is to apply an instrumental variable approach. The analysis uses the generalized method of moments (GMM) system estimation (proposed by Blundell and Bond 1998) and instruments for sales growth, labor productivity, supplier status, cash flow, and cash flow inter­ actions with supplier status and with labor productivity. The GMM estimator combines a differenced and a level equation. Lagged levels of endogenous vari­ ables are used as instruments for contemporary differences, and lagged differ­ ences are used as instruments for the level equation. Several additional instruments are also used. Firms whose managers speak a foreign language or who have worked for foreign companies before are likely to be better positioned to obtain contracts from multinationals. Thus, dummy variables reflecting these two characteristics are used as instruments for supply status. Level of language proficiency was determined by whether the manager can conduct business negotiations in a foreign language or can understand a business agreement in a foreign language, as reported in surveys. As exporters may find it easier to become multinational suppliers because of their experience of dealing with foreign customers, the second lag of exporting status is also used as an instrument. As it is also likely that proximity to MNCs facilitates business relationships, the instrument set includes proxies for the presence of multinationals in the same industry and in downstream industries. The share of sector output pro­ duced by foreign firms is the proxy for the presence of MNCs in the same sector. It is calculated by weighting the output of each firm f in sector i (Yft ) by the share of the firm fs equity owned by foreigners (Foreign shareft) and dividing it by the total output of sector j: E f for allfEj Foreign shareft * Yft (3) MNCs in the same sector it E f for allf~;Y;----- TABL E 4. Excluding Suppliers Benefiting from Multinational Assistance w ~ Excluding firms reporting easier access to credit o Excluding firms receiving financial assistance from because of their relationship with multinational multinational corporations corporations ...; Variable :::: en (1) (2) (3) (4) »1 o 6.Sales 0.076** (O.O,B) 0.068** (0.035) 0.104"*" (0.034) 0.098*** (0.035) ,... " CF/K 0.419*** (0.049) 0.353*** (0.052) 0.492*" (0.049) 0.442'-" (0,OS2) o CF /K " 2 yrs before - 0.23 (0.505) -0.193 (0.505) -0.45 (0.631) -0.46 (0.631) '" ;;. Z CF/K ' 1 yr before --0.471*'" (0.134) - 0.411 *** (0.134) -0.530"** (0.135) -0.480"· (0.136) :>< CF/K " Supplier -0.410*** (0.071) -0.358**' (0.072) -0.454* ** (0.069) -0.416*" (0.070) 2 yrs before - 0.091 (0.106) -0.11 (0.106) -0.002 (0.121) -0.018 (0,121) '" n o 1 yr before - 0.059 (0.094) -0.083 (0.094) 0.028 (0.099) 0.002 (0.099) z o Supplier -0.05 (0.088) -0.074 (0.088) 0.038 (0.097) 0.012 (0.098) z:: CF/K *In(VA/L) 0.01 (0.007) 0.014' (0,007) 0.007 (0.008) 0.01 (0.008) n In(VA/L) Debt/K 0.008 (0.005) 0.012* (0.006) 0.007 (0,007) -0.008 (0.005) -0.015*" (0.007) 0.012 (0.007) " m < In(Employment) -0.065 (0.058) -0.110' (0.062) '" :E! In(Age) - 0,056 (0.071) -0.088 (0.075) Intercept 0.103 (0.092) 0.511 (0.327) 0.061 (0.099) 0.742** (0.346) Number of observations 1311 1288 1267 1244 N umber of firms 299 292 290 283 R2 0.16 0.13 0.21 0.17 F-test CF/K + CF/K * Supplier = 0 0.02 (J.Ot 0.32 0.15 p-value 0.90 0.94 0.57 0.70 CF/K + CF/K * 1 yr before '= 0 0.15 0.19 0.08 0.08 p-value 0.70 0.66 0.78 0.78 CF/K CF/K * 2 yrs before = 0 0.14 0.10 0.00 0.00 p-value 0.71 0.75 0.95 0.98 Note: All specifications include firm and year fixed effects. Numbers in parentheses are standard errors. "Significant at the 10 percent level. "*Significant at the 5 percent level. ,,* * Significant at the 1 percent level. Source: Authors' analysis based on data from two World Bank surveys of Czech firms in 2003 and 2004 and Bureau van Dijk Electronic Publishing (2005); see text for details. .. Javorcik and Spatareanu 341 The proxy for the presence of multinationals in downstream sectors (sectors supplied by firm i operating in sector j) is defined following Javorcik (2004) as: . ~ LfforallfEk Foreign shareft *Yft Potential MNC customersit = ~ a ik * -~---~-- ...--- . kifkh LfforallfEk Yft (4) The proportion of sector j's output supplied to a downstream sector k based on the 1999 input-output matrix of the Czech Republic (ajk) is used to weight multinational presence in each downstream sector k. As the formula indicates, inputs supplied within the sector are not included. Thus, the greater the foreign presence in sectors supplied by industry j and the larger the share of output supplied to industries with a multinational presence, the higher is the value of the variable. 12 The calculations are based on all firms included in the AMADEUS database, not just the firms in the sample. Cash flow interactions with the instruments mentioned above are used to instrument for the inter­ action of cash flow with the multinational supplier dummy variable. Table 5 lists the instruments included in a given specification. The number of observations in GMM regressions is smaller than in the previous specifications. Because the model is expressed in first differences an additional year of data is lost. Further years of data are lost because the instru­ ments are based on second and further lags. While the results should be treated with caution because of the small number of observations, they are nevertheless informative. The Hansen test for overiden­ tification restrictions shows that the null hypothesis cannot be rejected at con­ ventional significance levels (see table 5). The Arellano-Bond test shows that the null hypothesis of no second-order serial correlation also cannot be rejected. These specification tests suggest that the regressions yield consistent estimates. The GMM results suggest that supplier status has no significant impact on a firm's liquidity constraints, once self-selection is taken into account. The inter­ action term between cash flow and supplier status is not statistically significant in any of the regressions (or in many other regressions estimated but not reported here to save space). In all specifications, the F-test rejects the absence of a relationship between cash flow and investment for multinational suppliers. As expected, the cash flow variable remains statistically significant in all regressions, suggesting that domestic firms are liquidity constrained. In summary, the evi­ dence suggests that suppliers differ from nonsuppliers in liquidity constraints, but the effect appears to be due to self-selection rather than to a relationship with an MNC leading to an easing of the supplier's financial constraints. 12. To illustrate the meaning of the variable, suppose that the sugar industry sells half of its output to jam producers and half to chocolate producers. If no multinationals are producing jam but half of all chocolate production comes from foreign affiliates, Potential MNC customersjt will he calculated as ! follows: * 0 ·d * :): !. w TABLE 5. Generalized Method of Moments Regressions "'" tv Variable (1) (2) (3) (4) (5) ..., :t IjK lagged 0.119" (0.051) 0.119" (0.051) 0.111" (0.050) 0.121*' (0.051) 0.124" (0.052) ASales 0.021 (0.054) 0.017 (0.055) 0.010 (0.055) 0.015 (0.053) 0.017 (0.054) '" l!i CFjK 0.322*" (0.067) 0.323'" (0.066) 0.336'" (0.068) 0.324"" (0.066) 0.323'" (0.066) 0 :­ :z: In(VAIL) -0.015'" (0.005) -O.DlS'" (0.005) -0.016'" (0.005) -0.015' ,. (0.005) - 0.D15·'· (0.005) ~ Debt/K 0.013 (0.012) 0.013 (0.012) 0.012 (0.012) 0.013 (0.012) 0.012 (0.011) '" Cl 0 In(Employment) -0.016 (0.021) -0.018 (0.022) -0.019 (0.020) -0.012 (0.021) -0.D1 (0.021) In(Age) -0.002 (0.002) -0.002 (0.002) -0.002 (0.002) -0.002 (0.002) - 0.002 (0.002) :z: 0 ::: , , Intercept 0.238 (0.166) 0.246 (0.166) 0.241 (0.153) 0.201 (0.149) 0.195 (0.145) Cl Number of observations 728 728 728 728 728 ~ Number of firms Additional instrumental 243 CF jK,_ 2' Manager's 243 CF/K, _2' Manager's 243 CF jK,_ 2"Manager's 243 CF/K'-2 "Potential MNC 243 CF/K,_ 2' Manager's '" < ;;; t 1 j variables foreign language foreign language foreign language customers, .2 foreign experience l!i CF/K,__ 2'Potential MNC CFjK,_z"Potential MNC Potential MNC Potential MNC CF /K, _2'Potential MNC j customers,_2 customers,_ 2 customers,_l customers,_ 2 customers,_ 2 ~. Potential MNC MNCs in the same CF/K,-2'MNCs in the Potential MNC * Potential MNC customers,_2 customers, __ 1 sector,_2 same sector'_l customers,_ 2 I CF/K,_2'MNCs in the same sector ,- 2 Exporter'_2 MNCs in the same sector,_2 CF/K,2'MNCs in the same sector,. 2 I MNCs in the same Expotter'_2 MNCs in the same J sector,_2 sectort_2 Exporter t -2 F-test CF/K + CFIK " Supplier = 0 17.58 17.75 17.64 18.3 18.27 p-valuc 0.00 0.00 0.00 0.00 0.00 AR(l) test p-value 0.03 0.03 0.Q3 0.03 0.03 AR(2) test p-value 0.94 0.95 0.98 0.93 0.93 Hansen test p-value 0.91 0.91 0.94 0.92 0.92 Note: Numbers in parentheses are standard errors. *Significant at the 10 percent level. ,. *Significant at the 5 percent level. * **Significant at the 1 percent level. Source: Authors' analysis based on data from two World Bank surveys of Czech firms in 2003 and 2004 and Bureau van Oijk Electronic Publishing (2005); see text for details. " • Javorcik and Spatareanu 343 IV. POLICY IMPLICATIONS Many countries around the world strive to attract FDI, believing that foreign investors not only bring capital but also serve as a channel of knowledge trans­ fer across international borders. Policymakers, expecting some of this knowl­ edge to result in externalities that benefit domestic producers, are willing to offer often generous incentive packages to foreign investors. For instance, 59 of 108 countries surveyed by the World Bank reported offering some type of incentives for FDI in 2004 (Harding and Javorcik 2007). A recent survey of the empirical literature on spillovers from FDI concludes that such spillovers are most likely between MNCs and their local suppliers (Gorg and Greenaway 2004). Thus, understanding what factors allow local firms to become suppliers to MNCs could have strong implications for under­ standing knowledge spillovers and public policy choices. Two main findings emerge from the study. First, in contrast to Czech firms in general, which face financial constraints, multinational suppliers do not appear to be liquidity constrained. Second, the data suggest that the lack of liquidity constraints is present before firms enter into a supplier relationship with MNCs, which is consistent with unconstrained firms self-selecting into supplying multinationals. Caution is required, however, in interpreting these findings. While the findings are robust to a number of controls that may be driving both access to credit and the ability of firms to supply multinationals, the possibility remains that the list of controls is incomplete. Further, even though the results suggest that well­ functioning credit markets are important in facilitating business relationships between local firms and MNCs, they do not suggest that a well-developed finan­ cial market is a sufficient condition for such relationships. Many other factors, such as a certain level of sophistication of the local manufacturing sector, a match between the skill endowment of the host economy and the sourcing needs of MNCs, and a good business environment, may be needed in order for these relationships to materialize. Thus, the findings could plausibly be generalized to other upper middle-income economies, but probably not to low-income econonues. ACKNOWLEDGMENTS The authors thank Thorsten Beck, Steve Fazzari, Jose Luis Groizard, Leonardo Iacovone, Yue Li, Inessa Love, Jan Svejnar, three anonymous reviewers, partici­ pants in the workshops Regional and Micro-level Effects of Globalization in Tiibingen, FDI and the Consequences in Ghent, Eastern Economic Association Annual Meetings in New York City, Midwest Conference on Economic Theory and International Trade in Columbus, OH, and the L1COS seminar at Catholic University Leuven for helpful comments and suggestions. • 344 THE WORLD BANK ECONOMIC REVIEW FUNDING The authors are grateful to the World Bank's Research Support Budget for financial assistance for the project "Vertical Relationships between Multinationals and Local Firms in the Czech Republic." ApPENDIX A TABLE A-I. Pro bit Model Predicting a Firm's Supplying Status Variable Liquidity ratio 0.743*** (0.191) 0.830*** (0.223) 0.823*** (0.248) 0.809* ** (0.253) lagged In(Gross profit) 0.079* * (0.037) 0.083* (0.049) 0.151" * (0.061) 0.129** (0.057) lagged Debt/K lagged 0.131 (0.236) 0.074 (0.248) 0.446 (0.317) 0.428 (0.325) In(Employment) 0.095* (0.050) 0.075 (0.069) 0.094 (0.066) 0.106 (0.067) lagged In(Age) lagged 0.054 (0.086) -0.D35 (0.090) -0.470*** (0.132) -0.480"** (0.135) In(VA/L) 0.014 (0.019) lagged In(Total factor 0.105 (0.320) productivity) lagged In(Total factor 0.27 (0.195) productivity Olley-Pakes) lagged Intercept -0.561 (0.604) 1.444*" (0.677) -1.223 (0.823) 1.983**" (0.725) Number of 1350 1051 949 887 observations Note: All specifications include industry and year fixed effects. Numbers in parentheses are robust standard errors. "Significant at the 10 percent leveL "·Significant at the 5 percent leveL 'f"" Significant at the 1 percent level. Source: Authors' analysis based on data from two World Bank surveys of Czech .firms in 2003 and 2004 and Bureau van Dijk Electronic Publishing (2005); see text for details. REFERENCES Alfaro, Laura, Areendam Chanda, Sebnem Kalemli-Ozcan, and Selin Sayek. 2004. "FDI and Economic Growth: The Role of Local Financial Markets." Journal of International Economics 65(2):89-112. - - . 2006. HoUi Does Foreign Direct Investment Promote Economic GroUlth? NBER Working Paper 12522. Cambridge, MA: National Bureau of Economic Research. Beck, Thorsten, Asli Demirgurc-Kunt, and Ross Levine. 1999. "A New Database on Financial Development and Structure." World Bank Policy Research Working Paper 2146. World Bank, Washington, D.C. Blalock, Garrick, and Paul J. Gertler. 2008. "Welfare Gains from Foreign Direct Investment through Technology Transfer to Local Suppliers." Journal of International Economics 74(2):402-21. ¢$ 4 iJ t J u, d. J • Javorcik and Spatareanu 345 Blundell, Richard, and Steve Bond. 1998. "Initial Conditions and Moment Restrictions in Dynamic Panel Data Models." Journal of Econometrics 87(1):115-43. Bureau van Dijk Electronic Publishing. 2005. AMADEUS: A Pan-European Database of Comparable Financial Information for 9 Million Public and Private Companies. Brussels. Chaney, Thomas. 2005. Liquidity Constrained Exporters. Working Paper, Department of Economics, University of Chicago. CzechInvest Factsheet No.3, January 2002. Prague: CzechInvest. Fazzari, Steven M., R. Glenn Hubbard, and Bruce C. Petersen. 1988. "Financing Constraints and Corporate Investment." Brookings Papers on Economic Activity 1:141-95. - - - . 2000. "Financing Constraints and Corporate Investment: Response to Kaplan and Zingales." Quarterly Journal of Economics 115(2):695-705. Gelos, R. Gaston, and Alejandro Werner. 2002. "Financial Liberalization, Credit Constraints, and Collateral: Investment in the Mexican Manufacturing Sector." Journal of Development Economics 67(1):1-27. Gorg, Holger, and David Greenaway. 2004. "Much Ado about Nothing? Do Domestic Firms Really Benefit from Foreign Direct Investment?" World Bank Research Observer 19(2):171-97. Harding, Torfinn, and Beata S. Javorcik. 2007. "Developing Economies and International Investors: Do Investment Promotion Agencies Bring Them Together?" Policy Research Working Paper 4339. World Bank, Washington, D.C. Harrison, Ann, and Margaret McMillan. 2003. "Does Direct Foreign Investment Affect Domestic Firms' Credit Constraints?" Journal of International Economics 61(1}:73-100. Harrison, Ann E., Inessa Love, and Margaret S. McMillan. 2004. "Global Capital Flows and Financing Constraints." Journal of Development Economics 75(1):269-301. Hoshi, Takeo, Anil Kashyap, and David Scharfstein. 1991. "Corporate Structure, Liquidity, and Investment: Evidence from Japanese Industrial Groups." Quarterly Journal of Economics 106(4):33-60. Javorcik, Beata S. 2004. "Does Foreign Direct Investment Increase the Productivity of Domestic Firms? In Search of Spillovers through Backward Linkages." American Economic Review 94(3): 605-27. Javorcik, Beata S., and Mariana Spatareanu. 2005. "Disentangling FDI Spillover Effects: What Do Firm Perceptions Tell Us?" In Theodore Moran, Edward Graham, and Magnus Blomstrom, eds., Does Foreign Direct Investment Promote Development? Washington, D.C.: Institute for International Economics. - - . 2008. "To Share or Not To Share: Does Local Participation Matter for Spillovers from FDI?" Journal of Development Economics 85(1-2):194-217. Kaplan, Steven, and Luigi Zingales. 1997. "Do Investment-Cash Flow Sensitivities Provide Useful Measures of Financing Constraints?" Quarterly Journal of Economics 112(1):169-215. ---.2000. "Investment-Cash Flow Sensitivities Are Not Valid Measures of Financing Constraints." Quarterly Journal of Economics 115 (2): 707 -12. Konings, Josef, Marian Rizov, and Hylke Vandenbussche. 2003. "Investment and Credit Constraints in Transition Economies: Micro Evidence from Poland, the Czech Republic, Bulgaria and Romania." Economic Letters 78(2):253-58. Lizal, Lubomir, and Jan Svejnar. 2002. "Investment, Credit Rationing, and the Soft Budget Constraint: Evidence from Czech Panel Data." Review of Economics and Statistics 84(2):353-70. Manova, Kalina. 2006. Credit Constraints in Trade: Financial Development and Export Composition. Cambridge, MA: Harvard University, Department of Economics. Modigliani, Franco, and Merton Miller. 1958. "The Cost of Capital, Corporation Finance, and the Theory of Investment." Americml EcOtlOmic Review 48(3):261-97. Moran, Theodore. 2001. Parental Supervision: The New Paradigm for Foreign Direct Investment and Development. Washington, D.C.: Institute for International Economics. • 346 THE WORLD BANK ECONOMIC REVIEW Moran, Theodore, Edward Graham, and Magnus Blomstrom. 2005. Does Foreign Direct Investment Promote Development? Washington, D.C.: Institute for International Economics. Olley, Steven G., and Ariel Pakes. 1996. "The Dynamics of Productivity in the Telecommunications Equipment Industry." Econometrica 64(6):1263-97. Pruteanu, Anca. 2004. "Was There Evidence of Credit Rationing in the Czech Republic?" Eastern European Economics 42(5):58-72. Rajan, Raghuram, and Luigi Zingales. 1998. "Financial Development and Gf()'wth." American Economic Review 88(3):559-86. t . 4 " · , , ' " ' "- ___ ::>~- "->.,::::-, ,>,:>'/:t:,':'--::;"<':_< .' Tb~~d'Va#tagelot~tl~~rs: '" .. S7~"Y9~r p~l'~cfej~~Ci~~~~.f~s~r~~~\ler~fo~.. Vwt OXf~tdJoufn~~~WebSteai~~jouI'nalS.~rg; . llekctthis joutnal~thelisttorea~ its hQ~age, aJld. dic),to1l44~ce A,cc~. JlQo~kthepageandrli~arly chetkfor the latest accepted papers; ~~lt#Oro.jOpr~~~9~; ... "v ,,_,,", __'v. .<-'.," ~_ ,,' __ ,_",' . __ ,_.,',;"),','-- ___ .: ,_, J '';:, ~,<:~~/' OXFORD JOURNALS • @t • .JiiI Q K !!II ¢ Forthcomingpapers in THE WORLD BANK ECONOMIC REVIEW • Macroeconomic Stability and the Distribution of Growth Rates Vatcharin Sirimaneetham andJonathan R. W Temple • Political Accountability and Regulatory Performance in Infrastructure Industries: An Empirical Analysis Farid Gasmi, Paul Noumba, and Laura Recuero Virto ISSN-13: 978-0-19-957452-11 I Aysegul Akin-Karasapan 08192 WASHINGTON DC 9 78019 574520