80519 Volume 26 • Number 1 • 2012 ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE) THE WORLD BANK ECONOMIC REVIEW Volume 26 • 2012 • Number 1 THE WORLD BANK ECONOMIC REVIEW Empirical Evidence on Satisfaction with Privatization in Latin America Céline Bonnet, Pierre Dubois, David Martimort, and Stéphane Straub Skills, Exports, and the Wages of Seven Million Latin American Workers Irene Brambilla, Rafael Dix-Carneiro, Daniel Lederman, and Guido Porto How to Deal with Covert Child Labor and Give Children an Effective Education, in a Poor Developing Country Alessandro Cigno Resource Windfalls and Emerging Market Sovereign Bond Spreads: The Role of Political Institutions Rabah Arezki and Markus Brückner When Should We Worry about Inflation? Raphael Espinoza, Hyginus Leon, and Ananthakrishnan Prasad Is Economic Volatility Detrimental to Global Sustainability? Yongfu Huang Pages 1–163 The Discriminatory Nature of Specific Tariffs Sohini Chowdhury www.wber.oxfordjournals.org 2 THE WORLD BANK ECONOMIC REVIEW editors Alain de Janvry and Elisabeth Sadoulet, University of California at Berkeley assistant to the editor Marja Kuiper editorial board Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on August 19, 2013 Harold H. Alderman, World Bank (retired) Caroline Freund, World Bank Chong-En Bai, Tsinghua University, China Paul Glewwe, University of Minnesota, Pranab K. Bardhan, University of California, USA Berkeley Philip E. Keefer, World Bank Thorsten Beck, Tilburg University, Justin Yifu Lin, World Bank Netherlands Norman V. Loayza, World Bank Johannes van Biesebroeck, K.U. Leuven, William F. Maloney, World Bank Belgium David J. McKenzie, World Bank Maureen Cropper, University of Maryland, Jaime de Melo, University of Geneva USA Ugo Panizza, UNCTAD Asli Demirgüç-Kunt, World Bank Nina Pavcnik, Dartmouth College, USA Jean-Jacques Dethier, World Bank Vijayendra Rao, World Bank Quy-Toan Do, World Bank Martin Ravallion, World Bank Frédéric Docquier, Catholic University of Jaime Saavedra-Chanduvi, World Bank Louvain, Belgium Claudia Paz Sepúlveda, World Bank Eliana La Ferrara, Università Bocconi, Italy Jonathan Temple, University of Bristol, UK Francisco H. G. Ferreira, World Bank Dominique Van De Walle, World Bank Augustin Kwasi Fosu, United Nations Christopher M. Woodruff, University of University, WIDER, Finland California, San Diego The World Bank Economic Review is a professional journal used for the dissemination of research in development economics broadly relevant to the development profession and to the World Bank in pursuing its development mandate. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantita- tive development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. Consistency with World Bank policy plays no role in the selection of articles. The Review is managed by one or two independent editors selected for their academic excellence in the field of development economics and policy. The editors are assisted by an editorial board composed in equal parts of scholars internal and external to the World Bank. World Bank staff and outside researchers are equally invited to submit their research papers to the Review. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 26 † 2012 † Number 1 Empirical Evidence on Satisfaction with Privatization in Latin America 1 Ce´ line Bonnet, Pierre Dubois, David Martimort, and Ste´ phane Straub Skills, Exports, and the Wages of Seven Million Latin American Workers 34 Irene Brambilla, Rafael Dix-Carneiro, Daniel Lederman, and Guido Porto How to Deal with Covert Child Labor and Give Children an Effective Education, in a Poor Developing Country 61 Alessandro Cigno Resource Windfalls and Emerging Market Sovereign Bond Spreads: The Role of Political Institutions 78 Rabah Arezki and Markus Bru ¨ ckner When Should We Worry about Inflation? 100 Raphael Espinoza, Hyginus Leon, and Ananthakrishnan Prasad Is Economic Volatility Detrimental to Global Sustainability? 128 Yongfu Huang The Discriminatory Nature of Speci�c Tariffs 147 Sohini Chowdhury SUBSCRIPTIONS:A subscription to The World Bank Economic Review (ISSN 0258-6770) comprises 3 issues. Prices include postage; for subscribers outside the Americas, issues are sent air freight. Annual Subscription Rate (Volume 26, 3 Issues, 2012): Institutions—Print edition and site-wide online access: £176/ $264/E264, Print edition only: £162/$242/E242, Site-wide online access only: £147/$221/E221; Corporate—Print edition and site-wide online access: £264/$395/E395, Print edition only: £242/$362/E362, Site-wide online access only: £219/$330/E330; Personal—Print edition and individual online access: £43/$64/E64. US$ rate applies to US & Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary subscriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as �rm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con- sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. The current year and two previous years’ issues are available from Oxford University Press. Previous BACK ISSUES: volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oup.com. Tel: þ44 (0)1865 353907. Fax: þ 44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oup.com. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, Tokyo, 4-5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: custserv.jp@oup.com. Tel: þ 81 3 5444 5858. Fax: þ 81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in February, June, and October, by Oxford University Press for the International Bank for Reconstruction and Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals postage paid at Cary, NC and at additional mailing of�ces. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to working with the global community to bring the highest quality research to the widest possible audience. Oxford Journals will protect the environment by implementing environmentally friendly policies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on environmental and ethical policies. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales, ADVERTISING: Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ 44 (0)1865 354767; Fax: þ 44(0)1865 353774; E-mail: jnlsadvertising@oup.com. DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the respective authors and contributors and not of the International Bank for Reconstruction and Development/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. The World Bank Economic Review is printed on acid-free paper that meets the minimum require- PAPER USED: ments of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT # 2012 The International Bank for Reconstruction and Development/THE WORLD BANK All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA. Empirical Evidence on Satisfaction with Privatization in Latin America ´ line Bonnet*, Pierre Dubois†, David Martimort‡, and Ce ´ phane Straub§ Ste Since the 1980s, privatization of formerly state-owned �rms has been extensively implemented by governments across Latin America. Despite the fact that most evalu- ations of the process fail to �nd signi�cant adverse ef�ciency and welfare effects, there has been a strong surge in public discontent with such policy in the region. This paper performs a systematic empirical analysis of the determinants of such dissatisfaction, using survey data from Latinobarometro covering 17 countries over the period 1998- 2008, complemented by country level data on macroeconomic, political, and insti- tutional aspects as well as data on privatization. Dissatisfaction appears to respond to absolute and relative welfare effects, and we �nd a robust U-shaped effect in term of education and income levels, with individuals in the middle of such distributions being more critical with the outcome of privatizations. JEL Classi�cation: L33, D83 Since the 1980s, privatization of formerly state-owned �rms has been exten- sively implemented by governments across Latin America, with infrastructure sectors (water, transport, energy and telecommunications) generating most of the proceeds (See Bortolotti and Siniscalco, 2004, Chapter 2). As a matter of fact, the World Bank’s Private Participation in Infrastructure database shows that for the period 1990 to 2004, Latin America and the Caribbean has been the leading region in the world in terms of number of projects (1062 of a total * Celine Bonnet, Toulouse School of Economics (GREMAQ, INRA), celine.bonnet@tse-fr.eu. † Pierre Dubois, Toulouse School of Economics (GREMAQ, INRA, IDEI), pierre.dubois@tse-fr.eu. ‡ David Martimort, Paris School of Economics, martimor@parisschoolofeconomics.eu. § ´ phane Straub (corresponding author), Toulouse School of Economics (ARQADE and IDEI), Ste Tel: (33) 561128529, stephane.straub@univ-tlse1.fr. We are grateful to Paulina Beato for having initiated this research and the Inter-American Development Bank for its �nancial support. We thank Jim de Melo and three anonymous referees, Tim Besley, Thierry Magnac, Stephane Saussier, Federico Weinschelbaum, Anne Yvrande and seminar participants at Toulouse, Paris 1, San Andres-Buenos Aires and EBRD, London for useful comments. The usual disclaimer applies. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 1 – 33 doi:10.1093/wber/lhr037 Advance Access Publication July 14, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 1 2 THE WORLD BANK ECONOMIC REVIEW of 2976) and investment �gures (US$ 392 bn. of a total of US$ 871 bn.).1 At the country level, �ve Latin American countries feature in the top ten ranking by number of projects (Brazil, Argentina, Mexico, Chile and Colombia), and three in terms of aggregate investment (Brazil, Argentina, Mexico). Within the region itself, there are also signi�cant variations across countries, from the very active ones like Bolivia, Peru, Brazil, Argentina and El Salvador in which accu- mulated proceeds as of 1999 ranged between 8 and 20% of GDP, to laggards like Uruguay, Paraguay, Costa Rica and Ecuador, in which virtually no privati- zation took place (Lora and Panizza, 2002). Given the scale of the privatization wave, evaluating this process has become an important challenge both for practitioners and scholars. A number of researchers have undertaken this dif�cult task focusing on several important aspects of the process, including its macroeconomic impact, �rm- and sector- level ef�ciency, employment, speci�c social outcomes like health, income distri- bution, poverty and welfare. To date, most studies found neutral to positive effects, with the possible exception of speci�c cases of price increase and layoffs in privatized �rms.2 However, since the beginning of the 2000s, opinion surveys from Latin America have revealed a profound and growing dissatisfac- tion with privatization, a situation that has already created a backlash against this policy, including popular protests, riots and governments in some countries making or being elected on pledges for a return to state-provided public services.3 Understanding this contrast between the generally positive economic evaluations and the striking evolution of negative public opinions on the priva- tization process therefore constitutes quite of a challenge both for policy- makers and researchers. The objective of this paper is to perform a systematic empirical analysis of the causes of public discontent with privatization in Latin America, using over 130,000 survey data observations from Latinobarometro covering 17 countries over the period 1998-2008, complemented by country level data on macroeco- nomic, political, and institutional aspects as well as data on the extent of priva- tization. The joint use of those two sources of data is necessary because the causes of public discontent are expected to be linked both to individual aspects (income, asset holdings, employment and social status, education, beliefs) and to “environmental� factors (size of the privatized sectors, economic cycle, insti- tutional quality). We unveil the mechanisms behind the observed determinants found to be signi�cant. More precisely, we ask to which extent the growing dissatisfaction simply results from a standard assessment of the effect of priva- tization on a combination of individual and group level welfare. 1. Note, however, that these investment �gures must be taken with some caution, as they represent commitments rather than actual spending. 2. We briefly review this literature in Section 2 below. See Martimort and Straub (2009) for a more detailed discussion and references. 3. The cases of Argentina and Bolivia are in order. Bonnet, Dubois, Martimort and Straub 3 We provide evidence on how the expressed level of dissatisfaction differs by level of income, education and along other socioeconomic divides, and argue that it partly reflects relative income considerations.4 In a nutshell, we �nd that dissatisfaction can be explained by a mix of individual characteristics that point to categories of individuals who have suffered or bene�ted less than others from privatizations. In particular, the effect of education, socioeconomic variables and assets variables signal a rather robust U-shaped effect in term of education and income levels, with individuals in the middle of such distri- butions being more critical with the outcome of privatizations. While the nature of our data does not allow us to systematically distinguish pure welfare effects from relative income concerns, which could for instance be linked to an unequal distribution of gains across social classes, we indicate in the discussion of the results why and when it is likely that both aspects are at play.5 A similar mix of absolute and relative income effects helps understand the outcome in terms of employment status, with public sector employees, unem- ployed and home workers categories corresponding to lower satisfaction levels, and private sector employees and, to some extent, students to higher approval rates. Indeed, a combination of direct welfare losses for some categories ( public sector employees, unemployed) and informational effects (to the extent for example that privatization signals a shift toward more competitive job market practices) for others, seems relevant. Finally, beliefs may also matter. Indeed, the respondents’ assessments of pri- vatization is strongly correlated with their views on the economic situation, their political preferences and the level of trust in society. Individuals forming more pessimistic evaluations of the economic situation are also less satis�ed with privatizations, but so are those placing themselves more to the left of the political spectrum and having more pronounced preferences for democracy and a lower level of trust in others. We document these patterns but stop short of performing a full-fledged econometric analysis of these issues due to data limitations. The paper is organized as follows. Section 2 offers an overview of existing evaluations of the privatization process in Latin America and then reviews con- tributions that have addressed, directly or indirectly, the issue of dissatisfaction with privatizations. Section 3 presents the data we are using. Section 4 intro- duces the basic econometric models, including estimations on individual data, 4. Individual answers may rely on anything from purely sel�sh considerations to completely altruistic ones, as well as capture, among others, fairness concerns, concerns for one’s (or one’s group) relative position in society, and experienced vs. revealed utility that is sensitive to the timing of economic effects (e.g., see Senik (2004), Ravaillon and Lokshin (2001) and Kahneman and Thaler (1991)). Note that the use of subjective survey data raises issues relative to the interpretation of individual answers (Clark et al. (2005), Ravaillon and Lokshin (2000) and Bertrand and Mullainathan (2001)). These issues are discussed in the body of the paper. 5. See Hopkins and Kornienko (2004) for a theoretical approach to this issue, and Senik (2004) for empirical evidence using Russian data. 4 THE WORLD BANK ECONOMIC REVIEW aggregate data and pseudo panel �xed effects and presents the results. Section 5 summarizes the main lessons from the analysis and concludes. I. LITERATURE REVIEW As mentioned above, the �rst wave of evaluations of the privatization process reached mostly positive conclusions regarding the impact on output and ef�- ciency, while stressing that negative redistributive effects, when presents, were limited. First, a large majority of the studies focusing on ef�ciency found improvements in �nancial and operating performance. Chong and Lope ´ z-de-Silanes (2004) document improvements in pro�tability (net income to sales, operating income to sales), in operating ef�ciency (cost per unit, sales to assets and sales to employee ratio), and in output in 7 countries (Argentina, Bolivia, Brazil, Chile, Colombia, Mexico, Peru). They also show that there was signi�cant labor retrenchment in privatized �rms in most countries, with the median country experiencing a 24% cutback and values up to 50 and 60% in the case of Peru and Mexico respectively. This evolution has in general been accompanied by increased investments and extensions in service coverage. Andres et al. (2007) review the impact of privatization for 181 �rms in 3 sectors (telecommunications, electricity distribution, water and sewerage) across 15 countries in Latin America, controlling for pre-privatization and transition-period trends. Again, the conclusion is that there were improvements in operating performance and productivity, mostly because of the important reduction in the workforce, a tendency to price increases, improvements in quality indicators such as distributional losses in water and electricity and per- centage of incomplete calls in telecoms, but no signi�cant impacts on output and coverage.6 From a macroeconomic point of view, the combination of incoming sales proceeds and the reduction in transfers to public �rms generally strengthened the �scal position of governments (Davis, Ossowski, Richardson and Barnett, 2000). Some countries, like Argentina and Mexico, used part of the privatiza- tion proceeds to withdraw a portion of domestic and foreign debt to cover pension costs (Kikeri and Nellis, 2002). Finally, tax proceeds also increased, as newly privatized �rms generated substantial pro�ts (Estache, Guasch and Trujillo, 2003). In terms of the social and welfare impact of privatization, Galiani, Gertler and Schargrodsky (2003) is one of the �rst paper providing convincing evi- dence of a positive impact of privatization on social indicators, in this case health. It shows that infant mortality from waterborne diseases declined by 8% 6. Other contributions include La Porta and Lo ´ pez-de-Silanes (1999), Megginson and Netter (2001), and Kikeri and Nellis (2002). This last paper reports that signi�cant labor reductions are mainly observed in the sub-group of non-competitive �rms, while competitive sectors have in general managed to maintain aggregate employment to its previous level. Bonnet, Dubois, Martimort and Straub 5 in Argentinean localities, which privatized water services and by as much as 26% in the poorest areas. McKenzie and Mookherjee (2003) is a systematic effort to evaluate the effect of privatization programs on social indicators like employment, income distribution, poverty and some aggregate welfare measure, by expenditure per capita deciles and by country/sector. It summarizes the results from 4 countries (Argentina, Bolivia, Mexico and Nicaragua) for which detailed studies have been carried out. Overall, the paper shows positive evolutions in all these dimensions and little evidence of negative effects on poverty or income distri- bution, thanks in particular to the extension in service coverage that in some cases compensated price increases. The exceptions are employment, with cut- backs within the privatized industries ranging from 30 to 75%. This is not sur- prising, as overstaf�ng is a well-known feature of many Latin American public �rms, but comes with the quali�cation that the layoffs were small in relation to the total labor force (7 to 9% in Nicaragua, 2% in Argentina, 1% in Mexico, 0.13% in Bolivia). As for prices, there is mixed evidence with prices going down in about half of the cases. Given these stylized facts, a number of explanations have already been put forward in the policy literature to understand the current dissatisfaction trend with privatizations in Latin America. We may classify these contributions into four categories: † Welfare considerations: A �rst line of research has tried to assess the impact of privatizations on prices, quality of the services and employ- ment, on the welfare of different groups and therefore on the evolution of satisfaction. Given the conclusions summarized above showing that such effects were mostly positive, this raises the question of whether some negative effects of privatization were not picked by these studies. In this respect, some partial evidence can be found regarding deteriorat- ing quality, or improvements in quality which did not compensate for price increases, in particular when price cap regulation was used (McKenzie and Mookherjee, 2003, Estache, Guasch and Trujillo, 2003, Nellis, Menezes and Lucas, 2004). Job losses and the deteriorating quality of working conditions (longer hours worked, lower job security and social bene�ts) are other important channels through which real welfare losses might have materialized for some subsets of the popu- lation (McKenzie and Mookherjee, 2003, Lo ´ pez-Calva and Rosello´ n, 2002). † Macroeconomic landscape: The impact of the business cycle, including the possible disruptive effect of large macroeconomics shocks, devalua- tions, etc., has sometimes been deemed responsible for the waning support for pro-market reforms (e.g. by Lora and Panizza, 2002, and Panizza and Yan ˜ ez, 2005). However, the direct impact of privatizations on the business cycle remains unclear, with opinions ranging from those attributing the rise in economic instability to privatizations, to more 6 THE WORLD BANK ECONOMIC REVIEW positive ones considering that they contributed to limit the effect of external shocks. It is therefore dif�cult to assess whether the correlation between the fall in economic activity and dissatisfaction with privatiza- tions is due to a direct negative welfare impact, to a gap between actual and expected performances, or to a change in beliefs somehow linked to the evolution of the overall economic situation. Answering this question would require making assumptions on the macroeconomic effects of pri- vatizations and on the structure of errors in individual judgments that are bound to be speculative. † Political economy: Another trend of the literature has tried to link the negative appraisal of privatization to distributional concerns. The basic idea is that, although privatization might come with ef�ciency gains, the effects of projects renegotiations and cancellation (Guasch, Laffont and Straub, 2008), corruption, and the lack of transparency of the process introduce distributional concerns among groups. Martimort and Straub (2009) offer a theory of how the degree of corruption that prevails in a society responds to changes in the ownership structure of public service providers. Privatization, even though it fosters investments in infrastruc- ture, might also open the door to more corruption. The public dissatis- faction towards privatization is then crucially affected by changes in the degree and pattern of corruption, as the public perception and awareness are modi�ed when corruption changes in nature. Indeed, corrupt activi- ties mainly consist of siphoning public budgets under public ownership whereas they amount to raising regulated prices under private ownership. Martimort and Straub (2009)’s model thus helps understand the fact that popular dissatisfaction with the process is especially high among middle class consumers, who bear the bulk of the cost generated by corrupt deals after privatizations, and therefore perceive themselves as the big losers in the allocation of ef�ciency gains. This interpretation is in line with Shirley (2005) and Nellis and Shirley (2005), who argue that citizens’ discontent in many cases stems from the perception of an unfair distribution of gains, biased in favor of elites. In an empirical paper using only three waves of the Latinobarometro surveys, Checchi et al. (2009) �nd that disagreement with privatization is most likely when the respondent is poor, privatization was massive and quick, involved a high proportion of public services as water and electri- city, and in countries where there is high inequality of income. A robust non-linear relationship between socioeconomic status and dissatisfaction with privatization suggests that middle-to-low income households, with a median level of nine years of education, perceive to have suffered from privatization. This result is again broadly consistent with recent empiri- cal research in Latin America that points to distributional concerns in the implementation of privatization policy because the consequences of the corresponding changes in tariffs were not adequately addressed by policy-makers and regulators. Bonnet, Dubois, Martimort and Straub 7 † Beliefs: However, it must be noticed that the �ndings in these papers may also be consistent with an alternative story in which an increase in the perception of corruption, or more generally of some unfair distri- bution of the gains from privatizations, may undermine trust in market reforms and induce a shift in beliefs, as conjectured for example by Di Tella and MacCulloch (2009), who argue that observing corruption causes people to become more left-wing. Indeed, recent contributions have highlighted the crucial role of beliefs in the expression of opinions on policy or social issues, both at the theoretical and at the empirical level.7 Using an interesting natural experiment in Argentina, Di Tella, Galiani and Schargrodsky (2007) show for instance that a simple change in land tenure status can induce important changes in individual pro- market beliefs, even in the absence of any signi�cant welfare change. It has been an open question to determine whether the rise in discontent with privatization in Latin America was due to a more general shift in beliefs against free-market policies, to some type of “reform fatigue� that would alter the support for what is perceived to be a liberal policy agenda, or to the fact that results from this policy did not match the expectations of certain categories of agents.8 In the case of privatization, Earle and al. (2003) show that the privati- zation policy design in the Czech Republic in the 1990s was instrumental in shifting popular beliefs in the usefulness of this policy. Indeed, indi- vidual directly involved in the process through both restitution and voucher programs were much more supportive and displayed higher faith in market reforms. Di Tella et al. (2008) show that beliefs about the bene�ts of water privatization in Argentina in the 1990s were strongly affected by negative government propaganda, but that this effect was conditional on not having gained access to water in the process. With respect to this general literature, our paper makes several original con- tributions. First of all, the analysis builds on a much larger data set spanning opinions in the region from 1998 to 2008. This allows us to obtain more robust results, in particular using a pseudo panel technique to address the issue of anchoring effects that are known to weaken results drawn from subjective surveys. Second, we relate the empirical results to a number of new underlying theor- etical explanations. For example, as in Checchi et al. (2009), we �nd a U-shaped relationship between education or income and the level of dissatisfac- tion. We argue that they reflect both the perception of direct welfare effects but also important relative income considerations among different social groups as modeled in Martimort and Straub (2009). We also enrich the description of the role of the socioeconomic status, by suggesting that privatization has dis- tinct signaling effects depending on individuals’ employment status, with private employees being especially sensitive to the pro-competitive message 7. Piketty (1995), Di Tella and McCulloch (2005), Benabou and Tirole (2006). 8. Panizza and Yan ˜ ez (2005), Lora and Olivera (2005). 8 THE WORLD BANK ECONOMIC REVIEW privatization conveys. Similarly, while we do �nd some effect of the business cycle as in Panizza and Yan ˜ ez (2005), we also show that dissatisfaction with privatizations cannot simply be explained by a reaction to deteriorating econ- omic conditions. Indeed, we �nd very different results when looking at opinions about the country economic situation for example. I I . D ATA Latinobarometro provides a series of yearly household surveys since 1995. Each year, a representative panel of individuals is asked a list of questions. Individuals are not re-interviewed every year and the data are more like a rotating representative panel. The data used here cover the period 1998 to 2008, except 1999 when the survey was not carried out, with coverage rising to 17 Latin American countries after 1996. For each country, there are approximately between 600 and more than 1000 respondents. This means a total of up to 130,000 observations, across 17 countries and 10 years. The survey includes one question about the level of satisfaction with services that have been privatized. It was asked each year (with some variations) between 1998 and 2008, but does not differentiate by sectors. There is a ques- tion differentiating by sectors, but it was only asked in 1995 and 1998. We use the only question that has suf�cient intertemporal coverage (1998 to 2008, except 2004). It asks respondents to indicate whether they strongly agree / agree / disagree / strongly disagree with the statement that privatizations have been bene�cial to the country. Additionally, the survey contains a full set of individual characteristics: demographics, assets, access to public services. It also contains answers to a host of subjective questions capturing individual opinions on several aspects like democracy, institutions, laws, politics, citizen participation, public policies, poverty, other socioeconomic subjects, international relations and general values. However, because there have been frequent changes in the layout of the survey, many of these questions are not available across a suf�cient number of time periods and cannot be exploited empirically. The Latinobarometro data from successive years were stacked together and then merged with country level data for the period 1998 to 2008 from a variety of sources. This includes data from the World Bank PPI database on the amount of privatization proceeds by country and sectors, aggregate governance Indicators from the Political Risk Service’s International Country Risk Guide, democracy and autocracy indicators from the Polity 4 database, and generic country level data from the World Bank World Development Indicators. Details about the sources and descriptive statistics are in the Appendix. Bonnet, Dubois, Martimort and Straub 9 III. ECONOMETRIC MODELS AND EM P I R I CA L R ES ULT S Figure 1 represents the evolution of the percentage of respondents in each country that (strongly) agree with the fact that privatizations have been ben- e�cial to the country. The graph con�rms the sharp decrease of the average satisfaction with priva- tization from 1998 to 2005 with a peak of dissatisfaction around the years 2002-2003. Some countries, such as Bolivia, Ecuador, El Salvador, Mexico, Paraguay and Venezuela, exhibit a U-shaped pattern, with satisfaction increas- ing again in 2007-2008. On the other hand, during the 1998-2008 period, our privatization proceeds variable shows that the proceeds decreased gradually from 1998 to 2003, before peaking up again after that, although not reaching the level of the 1990s. The main drivers of this U-shaped evolution were Brazil, Columbia, Mexico among others. On the other hand, Argentina, Guatemala and Peru show a peak at the end of the 1990s but did not register new privatizations subsequently, while the reverse pattern holds for Ecuador, which registered important proceeds in 2008. Interestingly, this global evolution therefore exhi- bits a negative correlation with that of satisfaction. In what follows, we use different methods to test the determinants of satis- faction or dissatisfaction with privatization, given the household survey data F I G U R E 1. The Evolution of Satisfaction 10 THE WORLD BANK ECONOMIC REVIEW available and the aggregate country-year information, starting with simple indi- vidual data. Methodology and Results Using the Individual Data Denoting by yict the opinion about privatization of individual i in country c and year t, and Xict the vector of his characteristics, gct a country-year �xed effect representing the �xed component across individuals that affects the opinion about privatization in country c and year t (as for example the average influence of a media campaign), and e ict an unobserved individual deviation of individual opinion on privatization, we assume that the individual opinion is determined by the following equation: yict ¼ X0ict b þ gct þ 1ict : ð1Þ Without loss of generality, we can also assume that yct is determined by observed country-year characteristics Sct and unobserved ones hct such that gct ¼ S0ct d þ hct : Then we can also re-write yict ¼ X0ict b þ S0ct d þ hct þ 1ict : ð2Þ The individual survey opinion about privatization allows us to estimate the model at the individual level and thus identify the parameters b and gct from the �rst speci�cation or b and d from the second one after assuming that Eðhct jXict ; Sct Þ ¼ 0, in addition to the �rst necessary assumption Eð1ict jXict ; Sct Þ ¼ 0. If this assumption cannot be made, then one can estimate the �rst speci�cation and then, after estimating the country-year �xed effects gct, regress these effects on characteristics Sct of the country and period with only Eðhct jSct Þ ¼ 0. In what follows, we present the results from both approaches. Table 1 presents the estimation of model (2) on individual data. The depen- dent variable is equal to 1 if the individual either agrees or strongly agrees with the fact that privatizations have been bene�cial to the country and 0 if he/she either disagrees or strongly disagrees. Assuming that the error term is normally distributed, one can estimate such discrete choice model by maximum likeli- hood using the usual probit model. The list of individual characteristics Xict, includes demographics (sex, age, marital status, education and occupation), wealth characteristics captured by asset ownership (TV, fridge, computer, washing-machine, car, secondary house, tenancy status), and access to basic services (drinking water, hot water, sewerage).9 The country level 9. Telephone access could also be added to the list, but this variable is not available for 2005. Estimations not shown here show that it is not signi�cant when included. T A B L E 1 . Probit estimations with individual data Probit (1) (2) (3) (4) (5) Demographics Sex 2 0.0330*** 2 0.0351*** 2 0.0349*** 2 0.0351*** 2 0.0226* (0.00928) (0.00933) (0.00951) (0.00954) (0.0121) Age 2 0.00255*** 2 0.00280*** 2 0.00272*** 2 0.00271*** 2 0.00196*** (0.000408) (0.000396) (0.000430) (0.000431) (0.000524) Couple 2 0.0241** 2 0.0307*** 2 0.0195* 2 0.0196* 2 0.0207 (0.0100) (0.0102) (0.0106) (0.0106) (0.0132) Education respondent 2 0.0672*** 2 0.0794*** 2 0.0701*** 2 0.0693*** 2 0.0812*** (0.0203) (0.0148) (0.0199) (0.0200) (0.0230) Education respondent (sq) 0.00889*** 0.00950*** 0.00912*** 0.00899*** 0.0102*** (0.00222) (0.00172) (0.00220) (0.00220) (0.00254) Employment status Public sect. employee 2 0.0783*** 2 0.0667*** 2 0.0825*** 2 0.0830*** 2 0.116*** (0.0160) (0.0157) (0.0168) (0.0168) (0.0194) Private sect. employee 0.0215 0.0209 0.0165 0.0162 0.0166 (0.0138) (0.0132) (0.0138) (0.0138) (0.0159) Unemployed 2 0.0714*** 2 0.0421** 2 0.0563*** 2 0.0550*** 2 0.0672*** (0.0179) (0.0169) (0.0184) (0.0180) (0.0228) Retired 0.0374 0.0455** 0.0414* 0.0413* 0.00135 (0.0238) (0.0229) (0.0246) (0.0246) (0.0286) At home 0.0191 0.0177 0.0181 0.0181 0.00900 (0.0141) (0.0136) (0.0145) (0.0145) (0.0188) Student 2 0.0286 2 0.0447** 2 0.0290 2 0.0281 2 0.0724*** (0.0187) (0.0181) (0.0194) (0.0194) (0.0224) Asset ownership Tv 2 0.0235 2 0.0310 2 0.0147 2 0.0156 2 0.0144 (0.0233) (0.0195) (0.0231) (0.0230) (0.0318) Fridge 0.00826 0.00606 0.00374 0.00212 0.00414 Bonnet, Dubois, Martimort and Straub (0.0184) (0.0174) (0.0205) (0.0204) (0.0271) (Continued ) 11 TABLE 1. Continued 12 Probit (1) (2) (3) (4) (5) Computer 0.0435** 0.0590*** 0.0409** 0.0416** 0.0399** (0.0175) (0.0148) (0.0170) (0.0170) (0.0187) Wash 0.0577*** 0.0576*** 0.0547*** 0.0539*** 0.0640*** (0.0168) (0.0145) (0.0160) (0.0160) (0.0181) Car 0.0801*** 0.0715*** 0.0857*** 0.0855*** 0.0765*** (0.0110) (0.0101) (0.0110) (0.0111) (0.0126) Secondary house 0.0550*** 0.0578*** 0.0546*** 0.0549*** 0.0586*** (0.0137) (0.0130) (0.0137) (0.0137) (0.0163) Home owner 0.00948 0.0158 0.00962 0.00952 2 0.00462 (0.0122) (0.0115) (0.0120) (0.0121) (0.0138) Access to services Drinking water 2 0.0687*** 2 0.0465** 2 0.0610*** 2 0.0602*** 2 0.0777*** (0.0217) (0.0192) (0.0223) (0.0224) (0.0266) THE WORLD BANK ECONOMIC REVIEW Hot water 0.0628*** 0.0527*** 0.0545*** 0.0559*** 0.0604*** (0.0202) (0.0147) (0.0184) (0.0182) (0.0195) sewerage system 2 0.0282* 2 0.0419*** 2 0.0342** 2 0.0347** 2 0.00761 (0.0162) (0.0158) (0.0170) (0.0170) (0.0190) Country level var. GNI per capita 0.00645 0.00521 2 0.00637 (0.0245) (0.0245) (0.0299) GDP growth 2 1 0.0170** 0.0169** 0.0153** (0.00675) (0.00684) (0.00667) Privat. proceeds (106) 1.394e þ 06 2 298,878 1.466e þ 06 (5.014e þ 06) (6.543e þ 06) (5.674e þ 06) Corruption 0.0604 0.0614 0.0574 (0.0419) (0.0424) (0.0437) Bureaucratic quality 0.123 (0.246) Democracy index 2 0.00293*** (0.00101) Opinions variables Better situation 2 0.0908*** (0.0122) Future situation 2 0.0870*** (0.0118) Left/right 0.0244*** (0.00388) Law con�dence 2 0.144*** (0.00869) Trust 0.115*** (0.0165) Democracy preference 2 0.0257** (0.0128) Country �xed effects Yes No Yes Yes Yes Country 2 Year �xed effects No Yes No No No Observations 130,914 130,914 122,134 122,134 73,754 Robust standard errors in parentheses (clustered at the country level). Coef�cients signi�cant at 10%: *; 5%: **; 1%: ***. Variables coding. Dependent variable: 1 if individual either agrees or strongly agrees with fact that privatizations have been bene�cial to the country, 0 if he/she either disagrees or strongly disagrees. Demographics: sex (0 ¼ man, 1 ¼ women); age (years); couple (0 ¼ living in couple, 1 ¼ single); education of respondent (1 ¼ illiterate; 2 ¼ basic incomplete; 3 ¼ basic complete; 4 ¼ secondary, medium, technical incomplete; 5 ¼ Secondary, medium, technical complete; 6 ¼ superior incomplete; 7 ¼ superior complete). Employment status (reference category is independent workers): public sector employee/ private sector employee/unemployed/retired/at home/student (1 ¼ yes, 0 ¼ no). Asset ownership: tv/fridge/computer/wash machine/car/secondary house/ home owner (1 ¼ yes, 0 ¼ no). Access to services: drinking water/hot water/sewerage system (1 ¼ yes, 0 ¼ no) Country level variables: GNI per capita (in US$); GDP growth 2 1 (lagged growth in %); privatization proceeds (accumulated proceeds as a % of GDP); corruption (PRS ICRG index, ranges from 0 (not corrupt) to 6 (highly corrupt)); Inflation 2 1 (lagged inflation in %); bureaucratic quality (PRS ICRG index, ranges from 0 (low quality) to 4 (high quality)); democracy index (0 2 10 scale, from less to more democratic); Unemployment 2 1 (lagged unemployment in %). Opinions variables: better situ- ation (1 ¼ better, 2 ¼ equal, 3 ¼ worse); future situation (1 ¼ better, 2 ¼ equal, 3 ¼ worse); left/right (ranges from 0 (extreme left) to 10 (extreme right); law con�dence (1 ¼ very high, 2 ¼ high, 3 ¼ low, 4 ¼ very low); trust (1 ¼ yes, 0 ¼ no); democracy preference ( 2 1 ¼ prefers authoritarian regime, 0 ¼ indifferent, 1 ¼ prefers democracy). Bonnet, Dubois, Martimort and Straub 13 14 THE WORLD BANK ECONOMIC REVIEW characteristics Sct are related to the macroeconomic environment10 (income per capita, lagged GDP growth), governance (corruption, quality of the bureauc- racy), the political environment (a democracy index) and the level of privatiza- tion proceeds. Finally, we also introduce individual opinion variables on several aspects, including how people place themselves on a left-right political spectrum, trust in the law, in other members of society, and assessments of the present and future economic situation, both at the personal and collective levels. Standard errors are clustered at the country-year level. To summarize, Table 1 shows that women are less satis�ed by privatizations as well as older people, people living in couple, public sector employees, unem- ployed and students (although this last variable is not systematically signi�- cant). Moreover, there is a U-shaped relationship between the degree of satisfaction and the level of education, meaning that the less satis�ed with pri- vatizations are those with medium education.11 Actually, the effect of the edu- cation level and its square imply that it is decreasing up to the education level 3 to 3.5, which is just below the average of the distribution in the sample and corresponds to complete basic education ( primary school) or slightly above.12 Table 1 also shows that being richer, in the sense of holding certain assets (computer, washing-machine, car, secondary house), corresponds to a higher level of satisfaction with privatizations. These categories make up 17, 48, 30, and 12% of the sample respectively, and can be interpreted as representative of the top end in terms of income. A similar result holds for people having access to hot water (43% of the sample). On the other hand, individuals who report not having access to drinking water and sewerage systems appear to be more satis�ed on average than the rest of the population. This is a relatively small subset (10 and 25%), likely to capture the bottom of the income distribution, i.e., individuals who might have gained, or expect to gain access to public ser- vices through privatizations. When opinion variables are introduced in column 4, they also appear to be correlated with satisfaction about privatizations. For example, the more they are to the left in terms of political preferences, and the less they trust other people in society, the less individuals are satis�ed with privatizations, while a higher level of trust in the judicial system corresponds to higher satisfaction. Moreover, the more people perceive that the situation of the country has dete- riorated, and the more pessimistic they are about the future of the country, the 10. Lagged values are relevant since the surveys are typically carried out around the middle of the year. 11. A similar pattern emerges when using the socioeconomic level of the respondent, ranging from 1 to 5, as evaluated by the person carrying out the survey. This indicates that the less satis�ed with privatizations are the “middle class� people, with a reversal point around 3.5 (the average of this level in the sample is 2.8 and the median is 3). We do not include this variable systematically in our estimations, however, because it is missing for 2002. 12. Note that this is a fairly low level by international standard. Bonnet, Dubois, Martimort and Straub 15 less satis�ed they are as well. Note however that, although most of these results make intuitive sense, the inclusion of such opinion variables on the right-hand side of the estimations is the source of speci�c econometric problems that make the interpretation of the results dif�cult. We discuss this issue in Section 3 below. Finally, one can look at the effect of country level variables on satisfaction. First of all, the level of income per capita is consistently positive, although not signi�cant. Looking at the effect of the economic cycle, higher growth in the year before the interview has a signi�cant and positive effect on satisfaction with privatizations, with each additional point implying around 1.6% higher satisfaction. The effect of the amount of accumulated proceeds from privatiza- tions is positive but statistically insigni�cant, so if anything it seems to be the case that individuals like privatization more in years in which their countries have increased their proceeds from privatizing more than the average across countries. Concerning corruption and the quality of the bureaucracy, the results show that the more corruption there is in the country and the lower the quality of the bureaucracy, the lower is the overall satisfaction with privatizations.13 However, both variables fail to be statistically signi�cant. The index of democ- racy, on the other hand, is negative and signi�cant, suggesting that dissatisfac- tion is stronger in more democratic environments. Estimations in column 2 include country-year �xed effects, and are strictly similar to those in column 1. Moreover, when year �xed effects are introduced as well, results not shown here to save space show that both the results on indi- vidual variables and on country-level aspects remain unchanged. Note that the (country-speci�c) time trend shows that satisfaction decreased signi�cantly over the period.14 One important concern might be that answers to the privatization question in fact capture some general discontent with economic policies or the state of the economy for example. To discard this possibility, we run similar esti- mations with alternative answers to opinion questions as the dependent vari- able. Results available from the authors show that other opinion variables do not exhibit the same correlations than the level of satisfaction with privatiza- tion. For example, using the opinion about the country economic situation as 13. Note that for institutional variables a higher score corresponds to less corruption and to a better bureaucracy. 14. Assuming again normally distributed error terms, one can estimate model (2) using an ordered probit estimation when the dependent variable is the ordered response about whether privatizations have been bene�cial, from disagree strongly ( ¼ 1) to agree strongly ( ¼ 4). The results, available from the authors, con�rm most of the �ndings of the probit model. It is interesting to get the same results with either one model or the other because it shows the robustness of the �ndings. In principle, the ordered probit model is more ef�cient than the probit model, which uses less information about the respondents opinion on privatization, but the probit model is also more robust to misclassi�cation of respondents between agree and very agree or disagree and very disagree. 16 THE WORLD BANK ECONOMIC REVIEW the dependent variable in (2), we do not �nd the same effects of asset owner- ship or education variables. Another important aspect is the possibility that opinions on privatization may be correlated with the government position on that policy, and that changes of political ideology of the government may have an impact on these opinions. For example, Di Tella et al. (2008) have shown that some group of the population may alter their judgement on privatization as a result of govern- ment propaganda. Country �xed effects would capture such effects as long as no fundamental ideological change occurs during our period of study. On the other hand, a political change, for example from a pro-privatization govern- ment to one that is openly opposed to such a policy, may induce a surge in anti-privatization opinions, and this in turn may be stronger for some speci�c groups. Particularly relevant here is the surge of leftist governments across Latin America in the last 10 to 15 years. This movement, which is often con- sidered to have started with Chave ´ z accession to power in Venezuela at the beginning of 1999, is considered to include the elections of Evo Morales in Bolivia, Rafael Correa in Ecuador, and Daniel Ortega in Nicaragua, all in 2006, and to a lesser extent of Luiz Ina ´ cio Lula da Silva, in Brazil in 2002, Nestor Kirchner in Argentina in 2003, Tabare ´ Va´ zquez in Uruguay in 2005, and Fernando Lugo in Paraguay in 2008. For the period 1998-2008 included in our data, the governments of Chave ´z in Venezuela in 1999 and Kirchner in Argentina in 2003 can be considered to have embodied signi�cant changes in the of�cial discourse towards privatiza- tion policy, and in both cases these elections were the result of a serious politi- cal and economic crisis in which the anti-liberal sentiments were already very high, leading us back to 1998 or before in the case of Venezuela, and around 2000–2001 for Argentina. Additionally, in Bolivia, Ecuador and Nicaragua, Figure 1 shows that contrary to what the ideology story seems to suggest, satis- faction actually increased after 2006. This leaves us with relatively few signi�- cant within-country change in our sample. In any case, the use of country �xed-effects is capturing any effect on opinions that would stem from the (unchanging) political orientation of the government, while country-year �xed effects take care of within country changes. Table 2 presents some further speci�cations using individual data. First of all, in column 1, we interact employment categories with the education level. Most of the effects of employment categories now appear to concentrate on higher education categories. In column 2, we interact the main individual variables of interest with a dummy that discriminates between richer countries (Argentina, Brazil, Costa Rica, Chile, Mexico, Panama, Uruguay, Venezuela) and poorer ones (Bolivia, Colombia, Ecuador, El Salvador, Guatemala, Honduras, Nicaragua, Paraguay, Peru). Results on the interacted variables are represented side by side. In terms of education, the U-shaped effect is preserved in both sub- samples, although its statistical signi�cance is stronger for poorer countries. Bonnet, Dubois, Martimort and Straub 17 T A B L E 2 . Probit estimations with individual data (6) (7) Interaction with country wealth Probit Poor Rich Demographics Sex 2 0.0325*** 2 0.0346*** (0.00932) (0.00920) Age 2 0.00258*** 2 0.00275*** (0.000410) (0.000415) Couple 2 0.0229** 2 0.0219** (0.0101) (0.0102) Education respondent 2 0.0633*** 2 0.0845*** 2 0.0466** (0.0207) (0.0277) (0.0237) Education respondent (sq) 0.00903*** 0.0114*** 0.00594** (0.00224) (0.00306) (0.00268) Employment status Public sect. employee 0.0598 2 0.0296*** (0.0454) (0.00948) Private sect. employee 2 0.0730** 0.0212*** (0.0351) (0.00783) Unemployed 0.00697 2 0.0204** (0.0437) (0.0104) Retired 0.0509 2 0.00348 (0.0416) (0.0103) At home 0.0730** 2 0.0156** (0.0310) (0.00748) Student 0.105* 2 0.0279** (0.0621) (0.0124) Asset ownership Tv 2 0.0237 2 0.0229 2 0.0570 (0.0233) (0.0255) (0.0487) Fridge 0.00802 0.0363* 2 0.0826** (0.0184) (0.0206) (0.0388) Computer 0.0413** 0.0215 0.0317 (0.0175) (0.0203) (0.0241) Wash 0.0573*** 0.127*** 0.00452 (0.0169) (0.0203) (0.0198) Car 0.0792*** 0.0921*** 0.0857*** (0.0110) (0.0171) (0.0132) Secondary house 0.0551*** 0.0120 0.0901*** (0.0137) (0.0199) (0.0182) Home owner 0.00913 0.00901 0.0137 (0.0122) (0.0166) (0.0176) Access to services Drinking water 2 0.0682*** 2 0.0529** 2 0.105** (0.0216) (0.0252) (0.0408) Hot water 2 .0625*** 0.0936*** 0.0470* (0.0202) (0.0292) (0.0253) (Continued ) 18 THE WORLD BANK ECONOMIC REVIEW TABLE 2. Continued (6) (7) Interaction with country wealth Probit Poor Rich sewerage system 2 0.0285* 2 0.0460** 2 0.00331 (0.0162) (0.0213) (0.0240) Country level var. (6) (7) GNI per capita 0.0532* 0.0154 (0.0295) (0.0175) Employment status interacted with education Public sect. employee  educ 2 0.0296*** (0.00948) Private sect. employee  educ 0.0212*** (0.00783) Unemployed  educ 2 0.0204** (0.0104) Retired  educ 2 0.00348 (0.0103) At home  educ 2 0.0156** (0.00748) Student  educ 2 0.0279** (0.0124) Country �xed effects Yes Yes Year �xed effects No No Observations 130,914 129,171 Robust standard errors in parentheses (clustered at the country level). Coef�cients signi�cant at 10%: *; 5%: **; 1%: ***. Variables coding: See Table 1. The reversal point is around 3.2 for poorer countries and 3.4 for richer ones, which corresponds to the fact that average education is also higher in the second group of countries (3.91 vs. 3.74; see Table A1). Some variation also appears for asset categories, a fact that can be related to the differences in asset distribution in the two groups of countries. For example, the washing machine variable is only signi�cant in the poorer countries subsample, consistently with the fact that it is indeed a luxury in these countries (average ownership is 24%), but much less in the group of richer ones (average ownership 70%), while secondary house is only signi�cant in the rich countries sample. Similarly, fridge ownership is less discriminating in richer countries (90% ownership) than in poorer ones (69%), and as a con- sequence, results from the �rst category differ from those in Table 1. On the other hand, car ownership remains a luxury across the region (average owner- ship 24 and 35% in poorer and richer countries respectively), and the results for this variable are consistent across both subsamples. Bonnet, Dubois, Martimort and Straub 19 As mentioned above, in these speci�cations, the validity of the results from country-level variables rests on the assumption that unobserved country-year characteristics are not systematically correlated with observed individual and aggregate aspects. Alternatively, one can estimate model (1) and then regress the resulting country-year effects on country-level variables. Figure 2 presents the distribution of the gct across countries and years. These are estimated from the model in column (1) of Table 1, where no country-level variables are introduced. They represent the country-year effects on satisfaction that cannot be explained by individual characteristics of respondents in Latinobarometro. There are variations across years within a given country but also between countries. Indeed, for most countries, Figure 2 con�rms the fact that average satisfaction has decreased between 1998 and 2008 and also show that discontent was the highest around 2003 and started to decrease in 2005. For some countries like Brazil or Ecuador this average country level satisfaction catches up with the late 90’s level during 2006-2008. Table 3 shows the regression of these country-year �xed effects on country level variables. The country-year effects on satisfaction that cannot be explained by the individual characteristics of the respondents are positively cor- related with lagged growth, and negatively with proceeds from privatization. The statistically signi�cant results for lagged GDP growth indicate that the economic cycle seems to be key in explaining residual country-year effects, F I G U R E 2. Unexplained country-year effects ( probit model) 20 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Regression of country-year effects on country level variables OLS (1) (2) GNI per capita 2 7.26e-05 0.000540 (0.0114) (0.0128) GDP growth -1 0.0981*** 0.0897*** (0.0313) (0.0334) Privat. proceeds (106) 2 8.095e*** 2 8.685e*** (1.843e þ 06) (2.037e þ 06) Corruption 0.0270 0.0239 (0.0454) (0.0516) Bureaucratic quality 0.0208 (0.0657) Democracy index 2 0.0148 (0.00951) Constant 0.0677 0.183 (0.172) (0.197) Observations 142 142 R-squared 0.162 0.172 Robust standard errors in parentheses. Coef�cients signi�cant at 10%: *; 5%: **; 1%: ***. Variables coding: See Table 1. while satisfaction seems higher in countries that privatized less. Other variables fail to be signi�cant. Pseudo-Panel Method and Results The disadvantage of the previous model is that it does not take into account unobserved individual preferences. A speci�c instance of that problem is the so-called “anchoring effect�, which implies that individuals may be using different satisfaction scales and provide different answers to characterize the same level of satisfaction.15 The lack of follow-up data on individuals prevents for example the implementation of panel data models where one could take into account unobserved individual �xed effects. However, one can construct a “pseudo panel� using the observed characteristics of respondents to try to over- come such issues and test the robustness of previous results (see Deaton, 1985, or Attanasio and Weber, 1995, for the use of such method also called “syn- thetic� panel). Let us de�ne an “average� individual representative of a set of characteristics Z such that we can de�ne K types based on these observed Z : i [ k if hðZi Þ ¼ k where h is a function mapping individual with characteristics Z into a type space. If the true model is yict ¼ X0ict b þ ui þ gct þ 1ict ; 15. See Bertrand and Mullainathan (2001) and Senik (2004). Bonnet, Dubois, Martimort and Straub 21 where ui is an unobserved individual �xed effect, then this model cannot be identi�ed because each household is observed only once. However, if ek 1 P we assume that ui ¼ uhðZi Þ ¼ u and de�ne ykct ¼ #fi=hðZi Þ¼kg i=hðZi Þ¼k yict , then 1 X ykct ¼ yict ; #fi=hðZi Þ ¼ kg i=hðZ Þ¼k i 1 X 1 X ¼ X0ict b þ ui þ gct #fi=hðZi Þ ¼ kg i=hðZ Þ¼k #fi=hðZi Þ ¼ kg i=hðZ Þ¼k i i 1 X þ 1ict #fi=hðZi Þ ¼ kg i=hðZ Þ¼k i 1 X e k þgct þ ¼ X0kct b þ u 1ict if X , Z; #fi=hðZi Þ ¼ kg i=hðZ Þ¼k i ¼ X0kct b e k þgct þ jkct ; þu P where jkct ¼ #fi=hð1 Zi Þ¼kg i=hðZi Þ¼k 1ict . Then, one can identify b using the regression e k þgct þ jkct ; ykct ¼ X0kct b þ u ð3Þ where ykct is the average response of type k individuals and ue k is the unob- served type k �xed effect. Table 4 presents the results of such estimation where the variables Z used to create the pseudo panel are the country, the region within the country, the age category and gender. It yields up to 2,640 pseudo individ- uals, for a total of up to 13,041 observations across the 9 rounds of survey and 17 countries. As a robustness check, we investigated alternative de�nitions of pseudo individuals, taking for example birthday cohorts instead of age groups (in addition to country, region, and gender). The results with this alternative de�nition, not shown here to save space, do not exhibit major differences. As a matter of facts, both ways of construct- ing the pseudo panel are feasible but with obviously different interpret- ations. In the case where we use age cohorts instead of birth cohorts, the interpretation is similar as the one that can be done with education cohorts (Deaton, 1985). But both de�nitions have some interesting interpretation. In the case where the pseudo panel is de�ned using birth year cohorts, it assumes that unobserved characteristics that affect satisfaction are speci�c to the set of people born during the same years (or that their deviation to the cohort-speci�c mean is uncorrelated with other observable variables) so that the time evolution of satisfaction is concomitant with their aging but 22 T A B L E 4 . Pseudo panel �xed effects (1) (2) (3) (4) (5) OLS OLS OLS OLS OLS Demographics Couple 2 0.0246 2 0.0236 2 0.00787 2 0.00783 2 0.0302 (0.0288) (0.0288) (0.0311) (0.0311) (0.0346) Education respondent 2 0.0620** 2 0.0562* 2 0.0845** 2 0.0855*** 2 0.118*** (0.0302) (0.0323) (0.0329) (0.0329) (0.0371) Education respondent (sq) 0.00797** 0.00829** 0.0102** 0.0104** 0.0139*** (0.00382) (0.00403) (0.00408) (0.00409) (0.00453) Employment status Public sect. employee 0.0348 0.108 0.0280 0.0303 0.0330 (0.0507) (0.121) (0.0530) (0.0530) (0.0621) Private sect. employee 0.0516 0.0495 0.0510 0.0519 0.000854 (0.0359) (0.0749) (0.0384) (0.0384) (0.0434) THE WORLD BANK ECONOMIC REVIEW Unemployed 2 0.0220 2 0.0593 2 0.0209 2 0.0247 2 0.0207 (0.0592) (0.119) (0.0596) (0.0596) (0.0668) Retired 2 0.0204 0.0390 2 0.0376 2 0.0378 2 0.0337 (0.0634) (0.122) (0.0670) (0.0670) (0.0843) At home 2 0.00450 0.0488 2 0.00387 2 0.00671 2 0.00639 (0.0310) (0.0656) (0.0337) (0.0337) (0.0385) Student 0.0523 0.220 0.0208 0.0155 2 0.103 (0.0676) (0.170) (0.0691) (0.0693) (0.0808) Asset ownership Tv 2 0.0320 2 0.0338 2 0.00874 2 0.00614 0.0135 (0.0397) (0.0398) (0.0427) (0.0427) (0.0488) Fridge 0.0565* 0.0559* 0.0383 0.0378 0.0404 (0.0334) (0.0334) (0.0356) (0.0356) (0.0421) Computer 0.0355 0.0329 0.0444 0.0437 0.0617 (0.0376) (0.0376) (0.0390) (0.0390) (0.0423) Wash 0.0145 0.0149 2 0.00782 2 0.00435 2 0.0321 (0.0340) (0.0340) (0.0364) (0.0364) (0.0382) Car 2 0.000776 2 0.00132 0.0199 0.0221 2 0.00400 (0.0331) (0.0331) (0.0347) (0.0347) (0.0386) Secondary house 0.104** 0.105** 0.0964** 0.0984** 0.0862* (0.0426) (0.0425) (0.0444) (0.0444) (0.0513) Home owner 0.0332 0.0334 0.0284 0.0274 0.0220 (0.0278) (0.0278) (0.0297) (0.0297) (0.0340) Access to services Drinking water 2 0.0377 2 0.0391 2 0.0508 2 0.0495 2 0.0544 (0.0342) (0.0341) (0.0383) (0.0383) (0.0443) Hot water 2 0.0282 2 0.0290 2 0.0433 2 0.0462 2 0.0178 (0.0285) (0.0286) (0.0301) (0.0301) (0.0324) sewerage system 2 0.0810*** 2 0.0807*** 2 0.0950*** 2 0.0930*** 2 0.0938*** (0.0258) (0.0258) (0.0275) (0.0275) (0.0304) Country level var. GNI per capita 0.0129* 0.0120* 0.0221*** (0.00693) (0.00696) (0.00756) GDP growth -1 0.00182 0.00183 2 0.00146 (0.00185) (0.00185) (0.00188) Privat. proceeds 3.400*** 4.932*** 3.731*** (1.180) (1.391) (1.219) Corruption 0.0191* 0.0146 2 0.00110 (0.0110) (0.0113) (0.0114) Bureaucratic quality 2 0.293*** (0.0694) Democracy index 2 0.000833 (0.000558) Opinions variables Better situation 2 0.106*** (0.0190) Future situation 2 0.0689*** (0.0188) Left/right 0.0114** (0.00465) Bonnet, Dubois, Martimort and Straub (Continued ) 23 TABLE 4. Continued 24 (1) (2) (3) (4) (5) OLS OLS OLS OLS OLS Law con�dence 2 0.110*** (0.0161) Trust 0.0843*** (0.0285) Democracy preference 2 0.0278 (0.0179) Employment status interacted with education Public sect. employee  educ 2 0.0191 (0.0265) Private sect. employee  educ 2 0.00154 (0.0192) THE WORLD BANK ECONOMIC REVIEW Unemployed  educ 0.0105 (0.0302) Retired  educ 2 0.0200 (0.0321) At home  educ 2 0.0189 (0.0189) Student  educ 2 0.0371 (0.0326) Year Dummies yes yes yes yes yes Observations 13,041 13,041 11,372 11,372 9,187 Number of identi 2,640 2,640 2,304 2,304 2,263 R-squared 0.062 0.062 0.058 0.059 0.108 Robust standard errors in parentheses (clustered at the country level). Coef�cients signi�cant at 10%: *; 5%: **; 1%: ***. Variables coding: See Table 1. Pseudo-panel de�ning variables: country, region, age, sex. Bonnet, Dubois, Martimort and Straub 25 not attributable to differences in life cycle experiences since they all have lived during the same years at each survey round. However, in the case of pseudo panels based on age, we assume that the unobservable character- istics that affect satisfaction are speci�c to the age group so that the time evolution of satisfaction of these age groups is concomitant with their change of cohort but not attributable to differences in aging. The two ways to construct pseudo panels are thus dif�cult to disentangle but our results do not change qualitatively, which seems to prove their robustness. Once the pseudo panel has been created, one can use �xed effects linear regression to estimate (3) and asymptotic consistency is obtained as the number of pseudo units goes to in�nity (Deaton, 1985). Year dummies are also included. Table 4 con�rms some of the previous insights. As for the effect of edu- cation, we note that the U-shaped relationship between satisfaction and the education level is preserved, meaning that the less satis�ed with privatizations are the people with medium education. Contrary to the individual data esti- mations, employment variables, as well as the interactions of these employment categories with the level of education, are not signi�cant and their inclusion does not modify the direct effects. The most likely explanation is that the spatial distribution of jobs implies a correlation with the regions used to de�ne the pseudo-panel. As a result, �xed effects pick up the variations along these dimensions.16 As for country-level variables, income per capita and privatization proceeds are now positive and signi�cant, as well as corruption in column 3. Again, sat- isfaction appears to increase more in years in which speci�c groups see their country have increased their income per capita and their proceeds from priva- tizing more than the average across countries. Finally, looking at assets owner- ship and services, we again observe that being rich in the sense of owning a secondary house implies more satisfaction. In terms of access to services, the result on people not having access to sewerage services being more satis�ed is also maintained. With this pseudo panel and �xed effects, we may now interpret the results on access to sew- erage services as being driven by changes within identifying groups. It means that groups that saw their average level of access increase more than the average across all groups are more dissatis�ed. If anything this result seems consistent with the previous interpretation that the (low-)middle class is more dissatis�ed. Finally, results on opinion variables are also broadly stable, with in particu- lar people with weaker preferences for democracy and lower levels of trust in others being more dissatis�ed. A �rst look at a range of questions included in the Latinobarometro survey shows that the evolution of opinions on the 16. As a matter of fact, a previous version of this paper (Bonnet et al., 2005) found signi�cant effects of employment variables using a pseudo-panel that did not include regions as a determinant. 26 THE WORLD BANK ECONOMIC REVIEW T A B L E 5 . Pseudo panel �xed effects Correlation Opinion variables with opinion on Privatization Current country situation 2 0.1371* Better country situation than before 2 0.1014* Future country situation 2 0.1007* Law con�dence 2 0.1333* Trust 0.0476* Preference for Democracy 2 0.0063* Left/right position 0.0563* * Pairwise correlation signi�cant at the 5% level bene�ts of privatizations is closely paralleled by the evolution of some other beliefs. Table 5 shows the correlation between the yearly country-level average opinions on privatization and other opinion variables.17 The evolution of these opinions in the period under study shows that at the time respondents in Latin America expressed growing negative perceptions of privatization, they also increasingly perceived the economic situation of their country to be bad, worse than 12 months ago, and they were also increasingly thinking this situation would worsen. As signaled in Section 3, note that when estimations similar to those dis- cussed previously are run with these alternative opinions as dependent vari- ables, a number of different results are found, meaning that these opinions, although correlated, have distinct informational contents. Overall, there seem to be a strong co-movement of opinion variables. This is especially true for what we will call “super�cial� opinions, i.e., those on short- term aspects, and a bit less so for “deep� beliefs like the overall level of trust in others, the preference for democracy or the situation on a left to right political spectrum. Finally, trust in the judicial system is to some extent a mix of deep beliefs and more super�cial opinions: although we would expect the level of trust in such institution to be to some extent beyond purely cyclical consider- ations, it is also conceivable that it may be subject to strong short-term fluctu- ations following for example some widely publicized scandal in a given country. The main problem is that the inclusion on the right-hand side of the esti- mations of additional opinion variables might induce an endogeneity bias to the extent that both these variables and the opinion on privatizations are corre- lated with some individual or group unobserved effects. Unobserved effects would only be controlled for by the country dummies, as well as the �xed effects in the pseudo panel setting, if they are time invariant. Year �xed effects 17. The �rst four lines in this table show the correlation between the rate of approval of privatizations and the percentage of respondents that think that the country situation is bad/worse than 12 months ago/likely to worsen and that the judicial system is not trustworthy. Hence, the negative correlations mean that more pessimistic respondents on these aspects are less happy with privatizations. Bonnet, Dubois, Martimort and Straub 27 may take care of some time varying unobserved effects, but only if these are common across countries and individuals. Any residual time varying individual unobserved effects would still induce a correlation between opinion variables and the error term. Moreover, this endogeneity bias might also affect the coef- �cients and standard errors of the other right-hand side variables included in the estimations, such as the demographics, if, as is very likely, these variables are correlated with the opinion variables through the unobserved individual or group effects. Results on these variables must therefore be taken with caution. I V. D I S C U S S I O N AND CONCLUSION We have performed a systematic empirical analysis of the determinants of public discontent with privatizations in Latin America, using survey data from Latinobarometro covering 17 countries over the period 1998-2008, comple- mented by country level data on macroeconomic, political, and institutional aspects as well as data on the extent of privatizations. The strong surge in dis- satisfaction in the region since the end of the 1990s appears to respond �rst to a mix of absolute and relative welfare effects. Speci�c categories that are likely to have suffered directly from privatizations, such as unemployed and public sector employees, do indeed express more dissatisfaction. As for relative effects, the fact that the extreme of the distribution in terms of income or education are less dissatis�ed is consistent with the middle class expressing concerns about an unequal distribution of ef�ciency gains among the population, as put forward in previous contributions on the subject. To summarize in more details the insights from both individual data and pseudo-panel estimations, we get the following picture with respect to traits and environmental features that fuel dissatisfaction with privatizations. Women, older individuals and people living in couple are more dissatis�ed. As for employment status, dissatisfaction is more important among public employ- ees, unemployed and home workers, while private sector employees, above an average level of education, and students, below an average education level, appear positively correlated with satisfaction. The �rst category is likely to capture the discontent from public employees who are under the threat of being laid off or of seeing their job characteristics modi�ed as the result of the privatization process.18 Indeed, a number of studies show that there were sub- stantial job losses in privatized �rms and, despite the fact that these cuts were generally small when compared to the total workforce and tended to be par- tially reversed in the medium run, they also mention serious workers’ concerns about the quality of their new jobs, including the obligation to work longer hours and a degradation of health and social security bene�ts. This is com- pounded by the well known fact that in many Latin American countries, public �rms have been used for patronage purpose by successive governments, with 18. Unfortunately, the surveys do not provide information on this aspect. 28 THE WORLD BANK ECONOMIC REVIEW the result that many public employees had relatively non-demanding jobs with bene�ts that largely exceeded what was available to the population at large. The dissatisfaction expressed by unemployed individuals could be related both to these job losses, in case they were among the victims, and to worries about the conditions of possible future employment. Finally, students may also be expressing preoccupations with the evolution of the labor market. In general, it seems likely that the change in jobs characteristics induced by the privatization of some big �rms is taken by these categories of individuals, that are or will soon be looking for a job, as a signal that the labor market and the reward structure has become more competitive. In other words, the evaluation of privatization in this case seems to be affected by people’s beliefs about what to expect from the economic situation rather than by actual welfare changes directly induced by the policy.19 Discontent is more pronounced among people with an intermediate level of education or with intermediate socioeconomic levels, who can be interpreted as being middle class individuals. One potential explanation is the one pro- posed by Martimort and Straub (2009), who argue that the middle class per- ceives itself as being the main loser in the distribution of ef�ciency gains, partly because of instances of corruption that have pushed up the price for public services. In terms of assets, it appears that ownership of what can be considered as “luxury� assets in a developing country context (computer, secondary house, and to a lesser extent car and washing machine) corresponds to higher satisfac- tion with privatization. So is access to hot water. At the other extreme, not having access to drinking water or sewerage systems, a proxy for being in the poorer part of the population, is also associated with greater satisfaction. Both facts can again be related to the inverse U-shaped effect in terms of education and wealth, with higher level of dissatisfaction in the middle of the distri- bution. The top part of the distribution, corresponds to people that may actu- ally have bene�ted from the change in the pattern of corruption mentioned above, which went from affecting rich taxpayers, through the soft budget con- straint of the State, to falling mostly on service consumers through regulated prices. Moreover, they may also have bene�ted from the elimination of cross-subsidies that followed the privatization of key services like telecommuni- cations and water. At the other end, very poor people, located in rural commu- nities or less developed urban areas previously unconnected to the networks, are likely to have gained access to electricity, telecommunication or water after the change in ownership.20 This may explain their more positive evaluation of the bene�ts of privatizations. Note that some asset ownership effects are 19. Another category of belief that could matter here is happiness. Unfortunately, we do not have exploitable information on that aspect. 20. This is consistent with the evidence from McKenzie and Mookherjee (2003) discussed above, showing that these categories often experienced substantial welfare gains. Bonnet, Dubois, Martimort and Straub 29 weakened by the introduction of �xed effects in the pseudo-panel estimations. Again, one can conjecture that ideological effects, likely to be stronger among middle class, urban groups, are now captured by �xed effects. At the aggregate level, dissatisfaction appears to strive in the context of poorer countries experiencing a dif�cult macroeconomic situation. As a matter of fact, the negative time trend between 1998 and 2008 (despite a slight rever- sal in 2005), seems to capture mainly the effect of low economic growth. A number of policy implications can be derived from these results. First of all, dealing with actual or perceived fairness issues in the implementation of privatization programs appears crucial. This means �rst that structuring the programs and their articulation with the institutional environment to minimize the risks and suspicions of corruption should be a priority for governments willing to convince their population of the opportunity of privatization. Additionally, more transparency in the way potential ef�ciency gains are allo- cated following the move to private operators is likely to alleviate negative evaluations by the public. The fact that dissatisfaction peaks among the middle class also calls for speci�c policies towards categories of households that often lose from the pri- vatization process, as they already had access to services at lower prices. Giving them a stake in these privatizations, as was done in some transition countries, may be a �rst step towards that goal. REFERENCES Andres, L., V. Foster, J.L. Guasch, and T. Haven. 2007. Private Sector Participation in Infrastructure in Latin America: What Really Happened, World Bank. Attanasio, O., and G. Weber. 1995. “Is consumption Growth Consistent with Intertemporal Optimization? Evidence from the Consumer Expenditure Survey.� Journal of Political Economy 103: 1121– 1157. Benabou, R., and J. Tirole. 2006. “Belief in a Just World and Redistributive Politics�. The Quarterly Journal of Economics 121: 699– 746. Bertrand, M., and S. Mullainathan. 2001. “Do People Mean What They Say? Implications for Subjective Survey Data�. American Economic Review 91: 67–72. Bonnet, C., P. Dubois, D. Martimort, and S. Straub. 2009. “Empirical Evidence on Satisfaction with Privatization in Latin America: Welfare Effects and Beliefs�. IDEI Working Paper 566. Toulouse: IDEI. Bortolotti, B., and D. Siniscalco. 2003. The Challenges of Privatization: An International Analysis. Oxford: Oxford University Press. Checchi, D., M. Florio, and J. Carrera. 2009. “Privatization Discontent and Utility Reform in Latin America�. The Journal of Development Studies 45(3): 333– 350. Chong, A., and F. Lo´ pez-de-Silanes. 2004. “Privatization in Latin America: What Does the Evidence Say?� Economı ´a 4: 37– 111. ´ , F. Postel-Vinay, C. Senik, and K. Van der Straeten. 2005. “Heterogeneity in Clark, A., F. Etile Reported Well-Being: Evidence from Twelve European Countries�. Economic Journal 115: 118– 132. 30 THE WORLD BANK ECONOMIC REVIEW Davis, F., O. Rossowski, T. Richardson, and S. Barnett. 2000. “Fiscal and Macroeconomic Impact of Privatization�. Occasional Paper No. 194. Washington D.C.: International Monetary Fund. Deaton, A. 1985. “Panel Data from Time Series of Cross-Sections�. Journal of Econometrics 30: 109– 126 Di Tella, R., and R. MacCulloch. 2009. “Why Doesn’t Capitalism Flow to Poor Countries?�. Brookings Papers on Economic Activity (spring): 285–321. ———. 2005. “Reform Fatigue: The Connection between Crime, Crisis and Pro-Market Beliefs�. mimeo. Di Tella, R., S. Galiani, and E. Schargrodsky. 2007. “The Formation of Beliefs: Evidence from the Allocation of Land Titles to Squatters�. Quarterly Journal of Economics 122: 209–241. ———. 2008. “Reality versus Propaganda in the Formation of Beliefs about Privatization�. NBER Working Paper Series 14483. Cambridge, Mass.: National Bureau of Economic Research. Earle, J.S., and S. Gehlbach. 2003. “A Spoonful of Sugar: Privatization and Popular Support for Reform in the Czech Republic�. Economics and Politics 15: 1– 32. Estache, A., L. Guasch, and L. Trujillo. 2003. “Price Caps, Ef�ciency Payoffs, and Infrastructure Contract Renegotiation in Latin America�. World Bank Policy Research Working Paper 3129. Washington D.C: World Bank. Galiani, S., P. Gertler, and E. Schargrodsky. 2005. “Water for Life: The Impact of Privatization of Water Services on Child Mortality�. Journal of Political Economy 113: 83 –120. Guasch, J.L., J.J. Laffont, and S. Straub. 2008. “Renegotiation of Concession Contracts in Latin America�. International Journal of Industrial Organization 26: 421–442. Hopkins, E., and T. Kornienko. 2004. “Running to Keep in the Same Place: Consumer Choice as a Game of Status�. American Economic Review 94: 1085–1107. Kahneman, D., and R. Thaler. 1991. “Economic Analysis and the Psychology of Utility: Applications to Compensation Policy�. American Economic Review 81: 341–346. Kikeri, S., and J. Nellis. 2002. “Privatization in Competitive Sectors, the Record to Date�. World bank Policy Research Working Paper 2860. Washington D.C: World Bank. ´ pez-de-Silanes. 1999. “The Bene�ts of Privatization: Evidence from Mexico�. La Porta, R., and F. Lo Quarterly Journal of Economics 114: 1193–1242. ´ pez-Calva, L., and J. Rosello Lo ´ n. 2002. “The Distributive Impact of Privatization: The Case of ´ ricas. Mexico�. IPD Working Paper 2003-3. Mexico: Universidad de las Ame Lora, E., and U. Panizza. 2002. “Structural Reforms in Latin America under Scrutiny�. IADB Working Paper 470. Washington D.C.: Inter-American Development Bank. Lora, E., and M. Olivera. 2005. “The Electoral Consequences of the Washington Consensus�. IADB Working Paper 530. Washington D.C.: Inter-American Development Bank. Martimort, D., and S. Straub. 2009. “Infrastructure Privatization and Changes in Corruption Patterns: The Roots of Public Discontent�. Journal of Development Economics 90: 69– 84. McKenzie, D., and D. Mookherjee. 2003. “Distributive Impact of Privatization in Latin America: An Overview of Evidence from Four Countries�. Economı´a 3: 161–218. Megginson, W., and J. Netter. 2000. “From State to Market: A Survey of Empirical Studies on Privatization�. Journal of Economic Literature 39: 321–389. Nellis, J., R. Menezes, and S. Lucas. 2004. “Privatization in Latin America, the Rapid Rise, Recent Fall, and Continuing Puzzle of a Contentious Economic Policy�. Center for Global Development Policy Brief 3(1). Washington D.C.: Center for Global Development. Nellis J., and M. Shirley, eds. 2005. Reality Check. The Distributional Impact of Privatization in Developing Countries. Washington D.C.: Center for Global Development. ˜ ez. 2005. “Why Are Latin Americans so Unhappy about Reforms?�. Journal of Panizza, U., and M. Yan Applied Economics 8: 1– 29. Piketty, T. 1995. “Social Mobility and Redistributive Politics�. Quarterly Journal of Economics 110: 551 –84. Bonnet, Dubois, Martimort and Straub 31 Ravallion, M., and M. Lokshin. 2001. “Identifying Welfare Effects from Subjective Questions�. Economica 68: 335– 357. ———. 2000. “Who Wants to Redistribute? The Tunnel Effect in 1990s Russia�. Journal of Public Economics 76: 87 –104. Senik, C. 2004. “When Information Dominates Comparison. Learning from Russian Subjective Panel Data�. Journal of Public Economics 88: 2099–2123. Shirley, M. 2005. “Why Is Sector Reform so Unpopular in Latin America?�. The Independent Review 10(2): 195–207. APPENDIX Data construction Sources of data: † Latinobarometro surveys 1998–2008 † Political risk variables (Bureaucracy quality, Corruption) from the International Country Risk Guide, 1984–2008. † World Bank Privatization Database, transactions by country, region or sector, by year 1988–2008 http://rru.worldbank.org/Privatization/ (last consulted, May 23rd, 2011.This site provides information on more than 10,000 privatization transactions of at least US$1 million in developing countries from 1988 to 2008. † World Development Indicators, World Bank, 1960–2008. † Democracy index from the Polity IV project (codes the authority charac- teristics of states in the world system for purposes of comparative, quan- titative analysis). The Democracy indicator is an additive eleven-point scale (0–10) taking into account the competitiveness of executive recruit- ment, the openness of executive recruitment, the constraints on chief executive, and the competitiveness of political participation. Descriptive statistics T A B L E A 1 . Individual characteristics and opinions All countries Poor countries Rich countries Individual variables Mean Median Mean Median Mean Median Individual characteristics Sex (0 ¼ woman, 1 ¼ man) 0.49 0 0.50 0 0.48 0 Age (years) 38.75 36 37.48 35 40.02 37 Couple (1 ¼ living in couple, 0 ¼ single) 0.57 1 0.57 1 0.57 1 Education level (1-9) 3.82 4 3.74 4 3.91 4 TV (0 ¼ no, 1 ¼ yes) 0.88 1 0.82 1 0.92 1 Fridge (0 ¼ no, 1 ¼ yes) 0.81 1 0.69 1 0.90 1 (Continued ) 32 THE WORLD BANK ECONOMIC REVIEW TABLE A1. Continued All countries Poor countries Rich countries Individual variables Mean Median Mean Median Mean Median Wash (0 ¼ no, 1 ¼ yes) 0.48 0 0.24 0 0.70 1 Car (0 ¼ no, 1 ¼ yes) 0.30 0 0.24 0 0.35 0 Secondary house (0 ¼ no, 1 ¼ yes) 0.12 0 0.11 0 0.12 0 Home owner (0 ¼ no, 1 ¼ yes) 0.74 1 0.72 1 0.74 1 Drink water (0 ¼ no, 1 ¼ yes) 0.90 1 0.86 1 0.92 1 Hot water (0 ¼ no, 1 ¼ yes) 0.43 0 0.25 0 0.59 1 Sewerage system (0 ¼ no, 1 ¼ yes) 0.75 1 0.69 1 0.79 1 Opinion variables Better situation (1 ¼ better, 2 ¼ same, 2.30 2 2.35 2 2.26 2 3 ¼ worse) Future situation (1 ¼ better, 2 ¼ same, 2.08 2 2.14 2 2.02 2 3 ¼ worse) Left Right (0-10 from left to right) 5.46 5 5.56 5 5.36 5 Law con�dence* 2.91 3 2.99 3 2.83 3 Trust (0 ¼ no, 1 ¼ yes) 0.19 0 0.18 0 0.19 0 Democracy preference (-1 ¼ no, 0 ¼ same, 0.44 1 0.39 1 0.49 1 1 ¼ yes) * (1 ¼ very high,2 ¼ high,3 ¼ low,4 ¼ very low) T A B L E A 2 Employment status and opinion on privatization Employment status Percentage Employment status ¼ 1 (self employed) 29.5 % Employment status ¼ 2 (public sector employee) 8.9% Employment status ¼ 3 (private sector wage laborer) 17.2 % Employment status ¼ 4 (temporarily unemployed) 6.8 % Employment status ¼ 5 (retired) 7.0 % Employment status ¼ 6 (at home) 21.5 % Employment status ¼ 7 (student) 9.1 % Whether privatization has been bene�cial agree strongly 8.64 % agree 26.74 % disagree 40.04 % disagree strongly 24.58 % Bonnet, Dubois, Martimort and Straub 33 T A B L E A 3 . Country level variables Country level variables Mean Median Std. Dev. Min Max Political risk variables Bureaucracy quality 1.87 2 0.673 0.16 3 Corruption 3.35 3 1.079 1 5.83 Other country level variables GNI per capita 5296.1 4885 2387.75 1640 12460 GDP growth (%) 3.31 3.56 3.69 2 11.03 17.32 Proceeds from privatizations (per year in 675.87 145.5 1355.31 0 9457 1000 $ US) Democracy index 7.505 8 1.82 2 10 Skills, Exports, and the Wages of Seven Million Latin American Workers Irene Brambilla, Rafael Dix-Carneiro, Daniel Lederman, and Guido Porto The returns to schooling and the skill premium are key parameters in various �elds and policy debates, including the literatures on globalization and inequality, inter- national migration, and technological change. This paper explores the skill premium and its correlation with exports in Latin America, thus linking the skill premium to the emerging literature on the structure of trade and development. Using data on employment and wages for over seven million workers from sixteen Latin American economies, the authors estimate national and industry-speci�c returns to schooling and skill premiums and study some of their determinants. The evidence suggests that both country and industry characteristics are important in explaining returns to schooling and skill premiums. The analyses also suggest that the incidence of exports within industries, the average income per capita within countries, and the relative abundance of skilled workers are related to the underlying industry and country characteristics that explain these parameters. In particular, sectoral exports are posi- tively correlated with the skill premium at the industry level, a result that supports recent trade models linking exports with wages and the demand for skills. JEL codes: F13, F14 In the �eld of international trade, the skilled-wage premium is a key parameter linking globalization with income distribution. Goldberg and Pavcnik (2007), for instance, highlight that increases in the returns to schooling might reflect trade-induced skill-biased technical change, a channel through which Irene Brambilla is a Professor at Universidad Nacional de La Plata, Argentina; her email address is irene.brambilla@econo.unlp.edu.ar. Rafael Dix-Carneiro is at Princeton University; his email address is rdc@princeton.edu. Daniel Lederman is a Senior Economist at DEC-TI, The World Bank; his email address is dlederman@worldbank.org. Guido Porto is a Professor at Universidad Nacional de La Plata; his email address is guido.porto@depeco.econo.unlp.edu.ar. The authors gratefully acknowledge the �nancial support from the World Bank’s Latin American and Caribbean Studies Program and a grant funded by the World Bank executed Multi-Donor Trust Fund on Trade. Invaluable insights and comments on previous versions of this paper were received from William F. Maloney, J. Humberto Lo´ pez, Augusto de la Torre, Pravin Krishna, Betty Sadoulet, and three anonymous referees. The opinions expressed herein do not represent the views of the World Bank, its Executive Directors, or the governments theyrepresent. All remaining errors are the authors’ responsibility. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 34– 60 doi:10.1093/wber/lhr020 Advance Access Publication July 6, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 34 Brambilla, Dix-Carneiro, Lederman and Porto 35 globalization has bene�ted skilled workers relative to unskilled workers. Also, Galiani and Porto (2010) explore how tariff reforms have affected the skill premium across time and sectors in Argentina. The skill premium also plays an important role in the literature on international migration and the brain drain (Beine, Docquier and Rapoport, 2001). A central concern is that the education of workers in developing countries might lead to emigration of skilled workers who seek higher wages in developed economies. Thus the issue of the so-called “brain drain� has permeated policy discussions about the developmental conse- quences of public education policies in poor countries. In spite of the central role played by the returns to schooling parameter or skill premiums in various literatures of importance for developing countries, there has been surprisingly little research about the relative roles played by industrial structure versus national characteristics in developing countries. If skill premiums vary systematically across industries, then industrial policies that favor one sector over another could have important consequences for closing the gap between the private and social returns to education, for redu- cing the scope of the brain drain due to emigration of highly educated workers, and for affecting the relationship between globalization and income inequality. Hence this paper can also be seen as a contribution to the literature on whether the industrial composition of exports matter for development as in Hausmann, Hwang, and Rodrik (2005). This article examines the returns to schooling and skill premiums in Latin America with two objectives. First, we document the patterns of returns to schooling and skill premiums in the region by estimating these parameters with data from eighty-eight household surveys for sixteen countries: Argentina, Brazil, Chile, Colombia, Costa Rica, the Dominican Republic, Guatemala, Ecuador, El Salvador, Honduras, Mexico, Nicaragua, Panama, Paraguay, Peru, and Uruguay. The data cover over seven million workers. Oursecond objective is to explore industry-speci�c returns to schooling and skill premiums in Latin America. Following the literature on industry wage differentials (Dickens and Katz, 1986; Dickens and Lang, 1988; Gibbons and Katz, 1992), we allow the skill premiums to vary across industries, as in Galiani and Porto (2010).1 Using the eighty eight household surveys, we estimate and document industry-speci�c skill premiums for sixty industries in each country. We then work with those estimates to study the empirical relationship between industry-speci�c returns to schooling and skill premiums and the level of sectoral exports. Our motivation lies in the work of Bernard and Jensen (1995, 1999), which documents the better performance of exporting �rms vis-a` -vis �rms that sell in domestic markets: exporters are larger, more 1. The existence of skill premiums at the industry level requires either some sort of labor immobility or compensating differentials due to working conditions or ef�ciency wages. In Galiani and Porto (2010), imperfect labor mobility is generated by union membership. On ef�ciency wages, see Krueger and Summers (1988). 36 THE WORLD BANK ECONOMIC REVIEW productive, and pay higher wages.2 In this paper, we extend this literature by investigating the correlation between exports and the returns to schooling and skill premiums at the industry level. Brambilla, Lederman, and Porto (2010) review theories linking exports and the skill premium based on skill-intensive activities associated with exports. These include marketing activities as well as quality upgrades (labeling, war- ranties, certi�cation) needed to export. Using �rm-level data from Argentina, the authors �nd support for such a link. Previous related research by Pavcnik et al. (2004) found a positive partial correlation between export exposure and skill premiums for university graduates inBrazil. In this paper, we generate additional supportive evidence for models of exports and skills. In cross- country, cross-industry regressions we �nd a positive and statistically signi�cant partial correlation between industry-speci�c returns to schooling and skill pre- miums, on the one hand, and sectoral exports on the other hand. This corre- lation, however, is not large in magnitude: doubling sectoral exports (a reasonable shock in our data) is associated with a 0.26 percentage point increase in themanufacturing-industry returns to education. With an average return of 7.9 percent in the region, this is equivalent to an increase of slightly over 3 percent. The skill premium would increase by 1.8 percentage points. With an average skill premium of around 62 percent, this is also equivalent to an increase in the average premium of around 3 percent. The related analytical issues might have important policy implications. Most countries in Latin America and the Caribbean currently pursue various export- promotion policies, including trade liberalization, export-processing zones, and export promotion agencies. One justi�cation for such policies might be the apparent existence of wage premiums for workers employed by �rms that sell a large share of their production abroad. If sectoral wage premiums are in fact related to foreign markets, then export-promotion policies could be welfare enhancing if the private returns to schooling are lower than the social returns to schooling. In other words, such policies could help narrow the gap between the private and social returns to schooling. More generally, industry-speci�c policies, including other forms of industrial policies, could help reduce the gap between the private and social returns to schooling. However, the existence of export driven industry-speci�c skill premiums or returns to schooling do not by themselves imply welfare gains from exporting, because they could simply reflect wage inequality. The rest of this paper is organized as follows. Section I reports several esti- mates of average skill premiums for the countries under investigation. To test their robustness, we discuss results from various model speci�cations that differ in terms of de�nitions of skilled workers, sub-samples of the data, and 2. For details, see Bernard and Wagner (1997), Isgut (2001), Bernard and Jensen (2004), Alvarez and Lopez (2005), De Loecker (2007), Schank, Schnabel, and Wagner (2007), Verhoogen (2008), Clerides, Lach, and Tybout (1998), Pavcnik (2002), and Park, Yang, Shi, and Jiang (2010). Brambilla, Dix-Carneiro, Lederman and Porto 37 econometric estimators. In addition, Section I assesses whether international differences in skill premiums are associated with relative endowments of skilled workers, heterogeneity in the composition of skilled workers, or heterogeneity in gender-speci�c skill premiums. Section II presents estimates of industry- speci�c skill premiums for 60 tradable and non-tradable sectors covered by the employment survey data, including 23 manufacturing sectors. After a brief analytical discussion of inter-industry wage differentials and the role of exports, we discuss the empirical analysis of exports as determinants of the skilled premium in manufacturing sectors. Section III concludes by summar- izing the main �ndings. I . E S T I M AT I O N OF NATIONA L SKILL PREMIUMS We start by estimating national wage premiums paid to skilled workers using household-level data from sixteen economies: Argentina, Brazil, Bolivia, Chile, Colombia, Costa Rica, Dominican Republic, Ecuador, El Salvador, Guatemala, Honduras, Mexico, Nicaragua, Panama, Paraguay, Peru and Uruguay. In our analysis, we include information on wages, skills, industry af�liation and characteristics of workers from 88 household surveys. Details of the household surveys, years of data and number of observations are found in Table 1. For each country we have between two (Nicaragua) and eight (Dominican Republic, El Salvador, Peru, Uruguay) years of data. We use surveys from 2000 to 2007, and the number of observations by country range from 60,000 in Nicaragua to 1,742,000 in Brazil, covering more than seven million workers.3 Table 2 displays descriptive statistics on education and skill levels of the workers. The �rst two columns show sharp differences in the average number of years of education and in the share of skilled workers (de�ned as individuals who hold a high school diploma) across countries. Average years of education are comparatively high in Argentina (10.73), Uruguay (9.68), Chile (9.10), Panama (8.97), Colombia (8.55), and Ecuador and the Dominican Republic (around 8). These countries also have the highest shares of skilled workers, ranging from 30 percent in the Dominican Republic to 52 percent in Argentina, although in Colombia the share is relatively lower at 20 percent. The lowest years of education are observed in Nicaragua, Guatemala and Honduras (5.31,5.70, and 5.92) but the lowest share of skilled workers are observed in Nicaragua and Brazil (9 and 15 percent). In the cases of Argentina and Uruguay, the relatively high averages are partly explained by survey design 3. The survey data come from the Socioeconomic Database for Latin America and the Caribbean (SEDLAC), a joint-project of CEDLAS at the Universidad Nacional de La Plata (Argentina) and the World Bank’s LAC poverty group (LCSPP). All variables in SEDLAC are constructed using consistent criteria across countries and years and identical programming routines. Note that our sample starts in 2000, because the information on workers’ industry af�liation (see Section II) became comprehensive and consistent only in recent years. Also, the surveys have become more homogeneous in coverage (for most countries we have yearly data) and this makes cross-country analysis more convincing. 38 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Description of Household Surveys Country Name of Survey Survey years Obs. Argentina Encuesta Permanente de Hogares (EPH-C) 2003, 2004, 2005, 2006 339,884 semestre II Brazil Pesquisa Nacional por Amostra de 2002, 2003, 2004, 2005, 1,742,448 Domicilios (PNAD) 2006, 2007 Chile Encuesta de Caracterizacio ´n 2000, 2003, 2006 778,698 Socioecono ´ mica Nacional (CASEN) Colombia Encuesta Continua de Hogares (ECH) 2001, 2003, 2004, 2006 436,111 Costa Rica Encuesta de Hogares de Propo ´ sitos 2001, 2002, 2003, 2004, 308,502 Mu´ ltiples (HPM) 2005, 2006, 2007 Dominican Encuesta Nacional de Fuerza de Trabajo 2000, 2001, 2002, 2003, 213,080 Rep. (ENFT) onda Octubre 2004, 2005, 2006, 2007 Ecuador Encuesta de Empleo, Desempleo y 2003, 2004, 2005, 2006, 397,296 Subempleo (ENEMDU) 2007 El Salvador Encuesta de Hogares de Propo ´ sitos 2000, 2001, 2002, 2003, 546,543 Mu´ ltiples (EHPM) 2004, 2005, 2006, 2007 Guatemala Encuesta Nacional de Empleo e Ingresos 2002, 2003, 2004 91,343 (ENEI) Honduras Encuesta Permanente de Hogares de 2001, 2003, 2004, 2005, 312,118 Propo ´ sitos Mu ´ ltiples (EPHPM) 2006 Mexico Encuesta Nacional de Ingresos y Gastos 2000, 2002, 2004, 2005, 384,168 de los Hogares (ENIGH) 2006 Nicaragua Encuesta Nacional de Hogares sobre 2001, 2005 59,424 Medicio ´ n de Nivel de Vida (EMNV) Panama Encuesta de Hogares (EH) 2001, 2002, 2003, 2004, 314,531 2005, 2006 Paraguay Encuesta Permanente de Hogares (EPH) 2002, 2003, 2004, 2005, 158,762 2006, 2007 Peru Encuesta Nacional de Hogares (ENAHO) 2000, 2001, 2002, 2003, 595,917 2004, 2005, 2006, 2007 Uruguay Encuesta Continua de Hogares (ECH) 2000, 2001, 2002, 2003, 657,911 2004, 2005, 2006, 2007 Table lists the surveys used in the estimation of the national-level and industry-speci�c skil premiums. because the surveys cover only urban households. In the other fourteen countries the surveys are representative of rural and urban populations. Columns 3 and 4 compare male and female workers. For some countries the share of skilled workers is higher among females than among males, most noticeably in Argentina, Brazil, Chile, the Dominican Republic, Honduras, Panama, Paraguay, and Uruguay. This difference ranges from 1 to over 6 percentage points. In contrast, in Colombia, El Salvador, Mexico, Peru and Guatemala the share of skilled workers is between 2 and 4 percentage points higher among males than females. It is also informative to look at skilled workers at a �ner level of disaggrega- tion, as workers of different educational levels are grouped together in the skilled category. Column 5 presents the share of highly-skilled workers Brambilla, Dix-Carneiro, Lederman and Porto 39 T A B L E 2 . Skill Endowments Share of skilled Share of highly-skilled workersa workersb Country Average years of education All Male Female All Male Female (1) (2) (3) (4) (5) (6) (7) Argentina 10.73 0.52 0.49 0.54 0.24 0.24 0.24 Brazil 7.37 0.15 0.13 0.17 0.25 0.29 0.22 Chile 9.10 0.40 0.39 0.41 0.24 0.25 0.24 Colombia 8.55 0.20 0.21 0.19 0.55 0.55 0.54 Costa Rica 7.68 0.18 0.18 0.18 0.34 0.35 0.32 Dominican Rep. 8.02 0.30 0.28 0.33 0.34 0.34 0.34 Ecuador 8.06 0.32 0.32 0.32 0.33 0.33 0.32 El Salvador 6.20 0.23 0.24 0.22 0.22 0.21 0.17 Guatemala 5.70 0.19 0.22 0.16 0.27 0.32 0.22 Honduras 5.92 0.19 0.19 0.20 0.30 0.37 0.24 Mexico 7.94 0.27 0.28 0.26 0.41 0.45 0.37 Nicaragua 5.31 0.09 0.09 0.09 0.46 0.49 0.42 Panama 8.97 0.37 0.34 0.40 0.31 0.28 0.34 Paraguay 7.45 0.25 0.25 0.26 0.23 0.23 0.23 Peru 7.98 0.23 0.24 0.21 0.45 0.45 0.46 Uruguay 9.68 0.33 0.30 0.35 0.35 0.32 0.37 (a): Share of workers with a high school diploma or more (skilled) in the total number of workers. (Semi-skilled þ Highly-skilled)/(Unskilled þ Semi-skilled þ Highly-skilled). Summay statistics for the skill premium. Based on estimates from a low age regression on a skill dummy (column (2) of Table 3). conditional on being skilled, that is, the share of workers with more than a high school diploma (individuals with tertiary education, some college experi- ence, college degree, and graduate degrees) in the total number of workers with at least a high school diploma. This statistic indicates the composition of skilled labor in each country. The differences across countries are again sharp, thus implying that the composition of the skilled labor force varies across countries. Countries with high shares of highly-skilled workers in the skilled group (41 to 55 percent) are Colombia, Peru, Mexico and Nicaragua. Notice, for instance, that because Nicaragua has the lowest skill share, the relatively few workers with degrees tend to reach a high educational attainment. Countries with low shares of highly-skilled workers are Argentina, Brazil, El Salvador, Paraguay, Argentina and Chile (22 to 25 percent). The participation of highly-skilled workers in the total labor force can be obtained by multiply- ing column 5 by column 2. To estimate the returns to skills by country, we pool data from all years and estimate Mincer-type regressions with the log (hourly) wage of each worker explained by individual worker characteristics. The main variable of interest is a binary variable that indicates whether the worker is skilled or 40 THE WORLD BANK ECONOMIC REVIEW unskilled. The equation to be estimated for each country takes the follow- ing standard form: ln wijt ¼ gSkijt þ x0ijt b þ dj þ dt þ 1ijt ; ð1Þ Subscript i denotes individuals, j the industry of employment, and t denotes years. The hourly wage is given by w. It is computed as the reported weekly wage divided by the number of hours worked per week. For robust- ness, we also estimate the model using the log of total wage income as the dependent variable. In our �rst speci�cation, we measure skills Sk with years of education and g is thus interpreted as the returns to schooling. In a second speci�cation, we de�ne skilled workers as those with a high school diploma or more. Thus, the binary variable Sk is equal to one if the individual has at least a high school diploma. In this case, g measures the skill premium, that is, the percentage difference in wages of skilled workers relative to unskilled workers. In both cases, we control for individ- ual characteristics in the vector x and for industry and year effects in the indicator variables dt and dj. The controls included in x are gender, age and age squared, marital status, whether the individual works full-time or part-time, a dummy for individuals in rural areas, and regional (within countries) dummies. The estimates from these equations are correlations from cross-sections of workers, which raises econometric issues that have been discussed at length in the labor literature (see, for example, Griliches 1977, Card 1999, and Krueger and Lindahl 2001). A key concern in this literature is that the estimated correlations capture the ability or talent of workers, which is correlated with both educational attainment and wages, which would yield upwardly biased estimates of the returns to schooling. On the other hand, because wages and educational attainment are reported by the surveyed workers, the estimates might suffer from attenuation bias due to random reporting errors.4 Therefore, the econometric results should be interpreted as reduced-form coef�cients measuring the average difference in wagesbetween skilled and unskilled workers, not as predictions of the wages that would be received by individual workers who enter the skilled- workers category. In a second speci�cation, we de�ne two groups of skilled workers: semi- skilled workers (those with a high school diploma) and highly-skilled workers (those with tertiary education, some college, a college degree, or a graduate education). In this case we include two binary variables, Sk 1 for 4. Krueger and Lindahl (2001, p. 1101) conclude in their literature review that there is surprisingly little evidence of ability bias in estimates of the returns toschooling. For our purposes, ability bias is not a serious concern because there is no reason to believe that the magnitude of the ability bias varies across countries. It may vary systematically across industries, which is the focus of sections II and III below. However, we do want to capture complementarities between unobserved worker ability and skills allocated across sectors. Brambilla, Dix-Carneiro, Lederman and Porto 41 the semi-skilled and Sk 2 for the highly-skilled, as shown in the following equation: ln wijt ¼ g1 Sk1 2 2 0 ijt þ g Skijt þ xijt b þ dj þ dt þ 1ijt ; ð2Þ The coef�cients g1 and g2 measure the wage premium for semi-skilled and highly-skilled workers. Both coef�cients are de�ned relative to unskilled workers. To estimate the returns to skills in equations (1) and (2), we restrict the sample to employed workers between 22 and 65 years of age. We drop employed workers who report a wage of zero. Results are in Table 3. In column 1, we report the estimates of the returns to schooling. This par- ameter ranges from a minimum of 5.8 percent in Peru to a maximum of 11.1 percent in Chile. The average return to schooling (unweighted) is 7.9 percent. In all cases, the returns to schooling are, as expected, highly statistically signi�- cant. It is worth noting that these estimates tend to be on average lower than existing estimates that are commonly used for international comparisons. For example, for our sample of countries, Psacharopoulus et al. (1996), Psacharopoulus and Patrinos (2004), Giovagnoli et al. (2005) and Sanroman (2006) present estimates that are on average 4.5 percentage points higher than our estimates. These differences are most likely due to the differences in the speci�cation of the wage equation, as this existing literature utilizes the basic Mincerian speci�cation with years of education, age and age-squared as the only regressors.5 Estimates of skill premiums based on equation (1) are presented in column 2 of Table 3. The coef�cients are interpreted as the percentage difference in wages between skilled (high school diploma) and unskilled workers. For example, in Ecuador the wage of an employed individual with a high school diploma is, on average and after controlling for observable worker character- istics and industry af�liation, 50.8 percent higher than the wage of an employed unskilled worker. Coef�cients range from 44 to91 percent. Brazil shows the highest skill premium, 91.2 percent, and Colombia follows with a premium of 87.6 percent. Countries with returns to skill over 60 percent are Nicaragua, Guatemala, Costa Rica, Honduras, and Mexico. In Chile, Paraguay, Ecuador, and Uruguay the skill premium is above 50 percent. In the remaining countries –Dominican Republic, Panama, Argentina, El Salvador, and Peru – the skill premium ranges from 49 to 44 percent. 5. The existing estimates cover mostly years from the late 1980s to the early 1990s, but Sanroman (2006) utilizes Uruguayan data from 2001-2005, and Giovagnoli et al. (2005) use Argentine data from 2002. On average, the existing studies cover samples with less education than our more recent data, but the difference (of 0.14 years of schooling) is too small to be the source of the differences in the estimates of the returns to schooling. Also, rising trade and skill-biased technical change would have pushed our estimates upward relative to the previous estimates. Thus most of the differences in the estimates are probably due to differences in the speci�cation of the wage equation. 42 T A B L E 3 . Skill Premium Average Premium Gender Differences Years of Skill Semi- Highly- Years of Skill HS Some College Country Education Premium skilled skilled Education Premium Semi-Skilled Highly-Skilled Element. Some HS Diploma College Degree THE WORLD BANK ECONOMIC REVIEW (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) Argentina 0.078*** 0.481*** 0.383*** 0.819*** 2 0.008*** 2 0.075*** 2 0.076*** 2 0.040*** 0.187*** 0.324*** 0.523*** 0.716*** 1.030*** (0.001) (0.005) (0.005) (0.008) (0.001) (0.010) (0.011) (0.014) (0.010) (0.010) (0.010) (0.011) (0.011) Brazil 0.097*** 0.912*** 0.823*** 1.008*** 2 0.002*** 0.128*** 0.106*** 0.128*** 0.263*** 0.308*** 0.532*** 0.926*** 1.396*** (0.000) (0.002) (0.003) (0.004) (0.000) (0.005) (0.005) (0.008) (0.003) (0.004) (0.002) (0.003) (0.003) Chile 0.111*** 0.574*** 0.373*** 1.102*** 2 0.001*** 0.014*** 2 0.023*** 0.053*** 0.151*** 0.247*** 0.489*** 0.827*** 1.317*** (0.000) (0.004) (0.004) (0.005) (0.001) (0.007) (0.007) (0.009) (0.006) (0.005) (0.005) (0.007) (0.006) Colombia 0.089*** 0.876*** 0.569*** 1.130*** 2 0.020*** 2 0.067*** 2 0.088*** 2 0.051*** 0.186*** 0.308*** 0.518*** 0.916*** 1.358*** (0.001) (0.007) (0.009) (0.008) (0.001) (0.012) (0.017) (0.014) (0.007) (0.007) (0.008) (0.012) (0.009) Costa Rica 0.086*** 0.700*** 0.541*** 0.977*** 2 0.005*** 0.029*** 2 0.014*** 0.069*** 0.135*** 0.285*** 0.487*** 0.941*** 1.353*** (0.001) (0.009) (0.010) (0.013) (0.002) (0.017) (0.020) (0.025) (0.011) (0.013) (0.014) (0.014) (0.022) Dominican 0.061*** 0.489*** 0.282*** 0.908*** 2 0.008*** 2 0.055*** 2 0.064*** 0.013*** 0.130*** 0.185*** 0.302*** 0.491*** 1.019*** Rep. (0.001) (0.007) (0.007) (0.009) (0.001) (0.012) (0.014) (0.016) (0.009) (0.009) (0.009) (0.011) (0.010) Ecuador 0.071*** 0.508*** 0.363*** 0.890*** 2 0.011*** 2 0.110*** 2 0.152*** 2 0.033*** 0.195*** 0.300*** 0.499*** 0.785*** 1.126*** (0.001) (0.006) (0.006) (0.008) (0.001) (0.010) (0.011) (0.014) (0.007) (0.009) (0.009) (0.010) (0.010) El Salvador 0.058*** 0.472*** 0.339*** 0.788*** 0.000*** 2 0.032*** 2 0.078*** 2 0.066*** 0.132*** 0.184*** 0.336*** 0.603*** 1.016*** (0.001) (0.006) (0.006) (0.009) (0.001) (0.010) (0.011) (0.009) (0.006) (0.011) (0.006) (0.009) (0.009) Guatemala 0.081*** 0.744*** 0.619*** 1.066*** 2 0.016*** 2 0.197*** 2 0.192*** 2 0.318*** 0.259*** 0.367*** 0.777*** 0.988*** 1.362*** (0.002) (0.020) (0.022) (0.032) (0.003) (0.036) (0.040) (0.063) (0.017) (0.023) (0.022) (0.037) (0.034) Summay 0.083*** 0.716*** 0.552*** 1.021*** 2 0.006*** 2 0.001*** 2 0.071*** 2 0.068*** 0.235*** 0.426*** 0.730*** 0.949*** 1.476*** statistics f Mexico 0.087*** 0.676*** 0.464*** 1.033*** 2 0.021*** 2 0.138*** 2 0.283*** 2 0.056*** 0.237*** 0.396*** 0.697*** 0.975*** 1.320*** (0.001) (0.012) (0.014) (0.016) (0.002) (0.022) (0.026) (0.029) (0.015) (0.015) (0.018) (0.023) (0.020) Nicaragua 0.065*** 0.735*** 0.447*** 1.023*** 0.000*** 0.155*** 0.024*** 0.221*** 0.129*** 0.278*** 0.413*** 0.613*** 1.174*** (0.002) (0.023) (0.030) (0.030) (0.003) (0.042) (0.057) (0.055) (0.021) (0.020) (0.026) (0.039) (0.031) Panama 0.078*** 0.487*** 0.340*** 0.889*** 2 0.013*** 2 0.124*** 2 0.136*** 0.017*** 0.194*** 0.327*** 0.526*** 0.823*** 1.234*** (0.001) (0.006) (0.006) (0.008) (0.001) (0.011) (0.012) (0.014) (0.009) (0.009) (0.010) (0.011) (0.011) Paraguay 0.077*** 0.528*** 0.432*** 0.926*** 2 0.001*** 2 0.069*** 2 0.072*** 0.009*** 0.165*** 0.355*** 0.545*** 0.731*** 1.107*** (0.001) (0.011) (0.012) (0.018) (0.002) (0.019) (0.021) (0.032) (0.012) (0.013) (0.017) (0.017) (0.019) Peru 0.058*** 0.442*** 0.270*** 0.737*** 2 0.009*** 2 0.137*** 2 0.153*** 2 0.130*** 0.165*** 0.258*** 0.331*** 0.506*** 0.800*** (0.001) (0.006) (0.006) (0.008) (0.001) (0.010) (0.012) (0.013) (0.007) (0.008) (0.007) (0.010) (0.008) Uruguay 0.086*** 0.539*** 0.411*** 0.834*** 0.004*** 0.037*** 0.024*** 0.112*** 0.186*** 0.416*** 0.658*** 0.812*** 1.196*** (0.000) (0.003) (0.004) (0.005) (0.001) (0.006) (0.007) (0.009) (0.006) (0.006) (0.007) (0.007) (0.007) Column (1): Log wage regression on years of education (returns to schooling); Columns (2): Log wage regression with skill dummy (returns to skill); Columns (3)-(4): Log wage regression with a semi-skilled and highly-skilled dummies. All resuts are relative to unskilled dummy (the omitted category). Columns (5)-(8) display differences in skill premium between males and females. Columns (9)-(13): Log wage regressions on �ve different educational attainment variables. Standard errors in brackets. All results are signi�cant at the 1 percent level (***). Brambilla, Dix-Carneiro, Lederman and Porto 43 44 THE WORLD BANK ECONOMIC REVIEW Columns 3 and 4 in Table 3 present results from equation (2), where the skill premium is split into the premium for semi-skilled workers and highly-skilled workers. Both premiums are relative to the unskilled category. Thus, in Costa Rica, semi-skilled workers earn on average 54.1 percent more than unskilled workers, and highly-skilled individuals earn 97.7 percent more than the unskilled. Across countries, the premium for semi-skilled workers ranges from 27 (in Peru) to 82.3 percent (in Brazil); the premium for highly-skilled workers ranges from 3.7 (in Peru) to over 110 percent (in Chile and Colombia). In general, countries with a high premium for the semi-skilled also exhibit a high premium for the highly-skilled. The correlation between the two measures is 0.76. We now turn to several robustness checks. First, the samples used to obtain the results described above include workers in all sectors of the economy and the estimates consequently reveal patterns of skill premiums at the national level. Because section II below is about the relationship between industry- speci�c skill premiums and exports, we also estimated the average skill premium restricting the sample to workers employed in manufacturing sectors only. Our estimates of skill premiums do not differ much from the baseline case where all workers are included in the regressions. In 10 of the 16 countries, the national skill premium is higher than the manufacturing skill premium; in 2 cases (Chile and Colombia), there are almost no differences in those pre- miums, an in four countries, the premium in manufacturing is actually higher. Among these countries, the maximum difference, in Peru, is of 1.8 percentage points (5.8 percent at the national level and 7.6 percent in the manufacturing sector). In Mexico, the difference is 1.2 percentage points in favor of manufac- turing. In all the other cases, the differences are relatively small. In fact, the correlation between the skill premium at the national level and the skill premium in the manufacturing sector is 0.88. To save space, we do not report these results in the paper, but are available in Table A1 in the on-line appendix.6 As additional robustness tests, we restricted the sample to full time workers and experimented with a median regression, which is theoretically less sensitive to outliers. Again, results are very close to the baseline speci�cation. We also explored regressions using the log of total monthly wage income as the depen- dent variable. The results, reported in Table A2 of the online appendix, remain robust. The correlation between both estimates is very high (0.995). This means that working with monthly wages or with hourly wages is not likely to affect our conclusions, either qualitatively or quantitatively. Our results uncover considerable differences in the returns to skill across countries. One obvious explanation for the differences in skill premiums could be factor endowments. Comparing the returns to skill presented in column 1 with the skill endowmentsin Table 2, column 2, we �nd a negative association 6. The link is http://sites.google.com/site/guidoportounlp/. Brambilla, Dix-Carneiro, Lederman and Porto 45 between the skill ratio and the skill premium. The correlation between the two variables is –0.63. Another plausible explanation for the estimated cross-country differences in the average skill premium is gender differences in returns to skill, which could vary across countries as a consequence of cultural attitudes and social norms related to gender. Gender differences in the returns to schooling could also be due to country differences in industrial structure, with some industries employ- ing relatively more (less) female workers with different skill levels. For example, export assembly operations (“maquilas�) are known to employ more women than men, and these industries tend to be located in economies that are close to the U.S. market. To explore this possibility, we allow the skill premium to vary by gender by adding an interaction term to the baseline regression: e Skijt ÃMijt þ x0ijt b þ dj þ dt þ 1ijt ; lnwijt ¼ gSkijt þ g ð3Þ where M is a binary variable that is equal to one for males (the gender dummy is separately included in x). The skill premium for females is given by g, while the premium for males is given by g þ g e, where ge represents the differential skill premium for males. In the case of two skill groups, the regression equation is ln wijt ¼ g1 Sk1 e 1 Sk1 ijt þ g 2 2 e 2 Sk2 ijt ÃMijt þ g Skijt þ g ijt ÃMijt ð4Þ þ x0ijt b þ dj þ dt þ 1ijt ; where g e 2 are the differential premiums for semi-skilled and highly-skilled e 1 and g males relative to females. Results for the differential premiums are displayed in columns 5 to 8 of Table 3.7 We begin in column 5 with gender differences in the returns to schooling (when Sk is measured with years of education). The male premium ranges from -2 percent (in Colombia and Mexico) to 0.4 percent in Uruguay. In column 6, we report gender differences in the skill premium. They range from negative 19.7 percent (in Guatemala) to positive 15.5 percent (in Nicaragua). Countries with a positive differential for males are Brazil, Nicaragua, Costa Rica, Chile and Uruguay. In almost all other countries, with the exception of a few results that are not statistically signi�cant, the male differential is negative and signi�cant, which implies that the gender wage gap is lower among skilled than among unskilled workers. For most countries, splitting skilled workers into semi-skilled and highly-skilled does not affect the direction of the gender difference in skill premiums, but there are signi�cant international differences in the gender-speci�c skill premiums. Because the pattern of these gender-speci�c premiums is somewhat erratic across countries, our results suggest that the cross-country differences in skill 7. In Table A3 of the online appendix, we report robustness results where we run the main speci�cation, equations (1), for a sample of only males. 46 THE WORLD BANK ECONOMIC REVIEW premiums are more likely due to differences in relative factor endowments than to gender differences. Additional support for this conclusion comes from a sim- plistic regression model with the national skill premium as the dependent vari- able (and a corresponding sample of sixteen observations) and these two explanatory variables. The results (not reported) show that only the ratio of skilled over unskilled workers is statistically signi�cant with a coef�cient esti- mate of –0.791 and a corresponding p-value for the null hypothesis of 0.025. The male-speci�c skill premium by country is positive, 0.292, but it is not stat- istically signi�cant. In fact, the estimate of the skill endowment variable changes only slightly, to –0.885, after the exclusion of the gender-speci�c premium. Another plausible explanation for the large differences in skill premiums across countries could be the composition of skill groups, because skilled workers are not homogeneous. In particular, the highly-skilled group includes individuals with tertiary education, some college, a college degree, and a post- graduate degree. Column 9 to 13 of Table 3 present the skill premiums of �ve groups: individuals who completed elementary school, individuals who did not �nish high school, high school graduates, individuals with some college or ter- tiary education, and college graduates. The results are markedly different across countries even for these arguably more homogeneous groups. Moreover, the average of the �ve coef�cients is highly correlated with the skill premium in that same country (the correlation is 0.72). Thus far, it seems that the skill endowments are our preferred country-level correlate of national skill pre- miums, but in subsequent exercises (reported in Table 6 and 7) we explore the role of the level of development, proxied by GDP per capita. II. INDUSTRY-SPECIFIC SKILL PREMIUMS After this detailed characterization of the national skill premium and its vari- ation across Latin American countries, in what follows we explore cross- country differences in skill premiums at the industry level. With perfect factor mobility (and leaving aside compensating differentials), wages should equalize across sectors and there should thus be an aggregate skill premium affecting all skilled workers in the labor market. With departures from that model, includ- ing imperfect mobility of skilled labor (but also of unskilled labor), wage equalization does not follow, and skill premiums at the industry level can vary in equilibrium. To investigate this possibility, we augmented our previous model to estimate skill premiums by sector. Speci�cally, we multiply the skill categories, using the different de�nitions described above, by dummy variables for each industry code at the 2-digit International Standard Industry Classi�cation (ISIC) Revision 3.8 The coef�cient on this interaction provides an 8. For those surveys that do not use ISIC Rev.3 to classify industries, concordance tables were utilized. See Table A8 of the online Appendix for a list of sectors. Brambilla, Dix-Carneiro, Lederman and Porto 47 estimate (relative to the industry of reference) of industry-speci�c skill premiums. At the 2-digit level, there are 60 sectors in the ISIC Revision 3 classi�cation. With a sample of 16 countries, we estimate approximately 960 industry-skill premiums (which are listed in Table A4, for the case of years of education, and Table A5, for thecase of the skilled dummy, in the on-line Appendix). There are signi�cant differences in the skill premiums, both across sectors for a given country and across countries for a given sector. Table 4 presents the distri- bution of industry-skill premiums within countries. Consistent with the esti- mates of the aggregate skill premiums (Table 3), there are wide differences in the average (and median) skill premium across countries that unsurprisingly mimic the patterns observed in Table 3. Figure 1 illustrates the notable dis- persion in the estimated skill premiums across industries within countries. In addition, there is considerable dispersion in the average skill premium across countries (for a given industry). For instance, the cross-country averages in the skill premium range from 1.12 in sector 99 (“Extra-territorial organization and bodies�) to 0.13 in sector 95 (“Private households with employed persons�).9 Exports and Industry-Speci�c Skill Premiums: Theory Skill premiums are affected by numerous factors, including demand and supply conditions, policies, and various shocks. Our interest in exports as a correlate of the skill premium is motivated by the literature on the wage premia paid by exporters. This literature, pioneered by Bernard and Jensen (1995, 1999) and later complemented by numerous researchers (see for instance the review in Bernand, Jensen, Redding, and Schott 2007), documents the better perform- ance of exporting �rms relative to non-exporting�rms in terms of employment, wages, and productivity. In this article, we explore a reduced-form analysis to assess sectoral exports as a determinant of industry skill premiums.10 Two leading theories explain this potential link between industry exports and skill premiums. One argues that the act of exporting requires activities that are skill-intensive, although the production of the good may require unskilled labor. Exporting �rms, and therefore industries with more exports in general, will thus demand higher skills and pay a higher skill premium.11 The alterna- tive theory argues that exporting is associated with higher pro�ts (because 9. We also investigated the dispersion of skill premiums (across sectors and countries) for the semi-skilled and highly-skilled categories. There is still signi�cant dispersion in the premiums. For the highly-skilled, for instance, the highest average premium is estimated for Chile (1.23) and the lowest for Uruguay (0.64). For the semi-skilled, the highest premium appears in Brazil (0.88) and the lowest in Peru (0.27) and Uruguay (0.24). 10. In a related paper, Brambilla, Lederman and Porto (2010) develop a model of exports and skills tested with �rm data from Argentina. 11. Exporters could either hire more skilled workers or provide on-the-job training. 48 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Skill Premium by Industry: Summary Statistics Mean Median Standard Deviation 10th Percentile 90th Percentile Obs All Countries 0.62 0.59 0.36 0.26 1.01 795 Argentina 0.46 0.44 0.24 0.17 0.74 54 Brazil 0.92 0.90 0.24 0.64 1.20 57 Chile 0.68 0.60 0.33 0.38 1.03 54 Colombia 0.84 0.83 0.29 0.53 1.23 57 Costa Rica 0.66 0.66 0.39 0.37 1.01 54 Dominican Rep. 0.53 0.51 0.30 0.19 0.81 50 Ecuador 0.56 0.53 0.27 0.26 0.93 57 El Salvador 0.41 0.47 0.58 0.15 0.73 46 Guatemala 0.69 0.65 0.42 0.12 1.14 36 Honduras 0.71 0.66 0.44 0.34 1.16 50 Mexico 0.62 0.58 0.28 0.27 1.03 47 Nicaragua 0.71 0.72 0.45 0.24 1.41 33 Panama 0.48 0.47 0.23 0.20 0.78 50 Paraguay 0.49 0.52 0.24 0.19 0.78 43 Peru 0.51 0.49 0.24 0.23 0.79 51 Uruguay 0.54 0.50 0.24 0.31 0.89 56 Summay statistics for the skill premium. Based on estimates from a log wage regression on a skill dummy (column (2) of Table 3. F I G U R E 1 . Skill Premium by Industry Graph displays skill premiums that are industry and country speci�c. The dash within each box corresponds to the median skill premium, the edges of the box correspond to the 75th and 25th percentile skill premiums, and the upper and lower dashes correspond to the next adjacent skill premiums. Outside values are excluded. more productive �rms self select into exports) and these higher pro�ts are shared with the workers via pro�t sharing rules. The theory focusing on the need to engage in skill-intensive activities in order to export a product is based on Brambilla, Lederman, and Porto (2010). Brambilla, Dix-Carneiro, Lederman and Porto 49 F I G U R E 2 . Industry Exports and Industry Skill Premium Skill Intensive Tasks Note: Equilibrium wages for skilled workers in industries with different levels of exports in a model of skill-intensive exports. The curve Ls (ws) is the supply of skilled labor. The curve Ld (ws) is the industry demand for skilled labor, which depends positively on the level of sectoral exports. For our present purposes, we assume that skilled labor is imperfectly mobile, as in Goldberg and Pavcnik (2005), Ferreira et al. (2008), and Galiani and Porto (2010). Unskilled workers are perfectly mobile across sectors and earn the economy-wide competitive wage, wu. While total labor supply in a given industry may be �xed due to labor speci�city, workers can be induced to supply more effort at higher offered wages. In Figure 2, for instance, the relationship between effective skilled labor supply in industry j and skilled wages ws is increasing (the function Ls (ws )). Exporting requires both the production of the physical units of the product and the provision of export services. These services include labeling, marketing, technical support, consumer support (webpage, email, warranty).12 Brambilla, Lederman, and Porto (2010) assume that these export services are skill-intensive activities because they require the effort Ls of highly skilled man- agers and engineers. It follows that the demand for the effort of skilled labor in industry j will depend on the level of exports of the industry.13 In Figure 2, we plot two such demand functions for two industries with different levels of exports, ExpH . ExpL ; the high-export industry has a higher demand for skilled workers. As Figure 2 shows, the high-export sector pays higher wages to their skilled workers. Since the wage offered to the unskilled workers is assumed to be the 12. In Manasse and Turrini (2001) and Verhoogen (2008), exporting requires quality upgrades. 13. The demand for unskilled labor may depend on exports. For illustration purposes, this is not really relevant in our discussion. See Brambilla, Lederman, and Porto (2010) for details. 50 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3 . Industry Exports and Industry Skill Premium Fair Wages and Pro�t Sharing Note: Equilibrium wages for skilled workers in industries with different levels of exports in a model of fair-wages. The curve f ( p) is the supply of effort of skilled workers. The curve f(wsj.) shows that pro�ts are decreasing in wages. For a given ws, pro�ts are increasing in the level of exports. same across industries (given by the competitive national market for unskilled labor), it follows that high-export sectors pay a higher skilled premium. An alternative theory is based on pro�t sharing mechanisms. In the trade lit- erature, pro�t sharing originates in a fair-wage hypothesis, as in Egger and Kreickemeier (2009) and Amiti and Davis (2008). In short, skilled workers demand a wage premium to exert the necessary effort because it is considered fair to share the pro�ts of the �rms. In consequence, while marginal �rms pay the competitive outside wage, more pro�table �rms pay increasingly higher wages. In Figure 3, this is represented by the fair-wage constraint ws ¼ fðpÞ, where fð�Þ is increasing in the level of pro�ts p Pro�ts, on the other hand, are a decreasing function of the wages offered to skilled workers. This is represented by the function p (ws) in Figure 3. In addition, following Melitz (2003), we assume that pro�ts are higher for expor- ters, and consequently the pro�t function p(ws) of high export sectors are higher, for a given level of wages, than in low export sectors. In equilibrium, high-export �rms offer higher wages ws to skilled workers. Together with com- petitive labor markets for unskilled labor with equilibrium wages wu and some degree of speci�city of skilled labor (as before), in the end the industry-speci�c skill premium is an increasing function of the level of sectoral exports. It is worth noting that the theories described above imply that exports either demand higher skills (observed and unobserved, thus including innate worker ability) or offer higher pro�ts, which can be shared with skilled workers. The empirical exercises that follow, however, should not be interpreted strictly as as Brambilla, Dix-Carneiro, Lederman and Porto 51 F I G U R E 4 . Industry Exports and Industry Skill Premium Cross-Latin American Correlation Note: Scatter plot and linear �t of the skill-premium (at 2-digit level) and the level of sectoral exports (relative to GDP). tests of exports as causing high skill premiums. This would be the case only if exports are strictly exogenous and industry-speci�c demand for skilled workers does not by itself cause exports. As will become apparent, it is somewhat com- forting that the effects of industry-speci�c exports appear correlated with skill premiums even after controlling for industry-speci�c effects. Still, the results must be interpreted with caution because it does not follow that skilled workers that move from an industry with low estimated premiums to another with higher premiums will receive higher wages. This is so because industries and exports may require speci�c skills that may not be transferable to other activities. Exports and the Industry-Speci�c Skill Premiums: Evidence Figure 4 summarizes our claim: The skill premium rises with exports (as a share of GDP) across Latin American manufacturing industries. The remainder of this section explores this statistical relationship. As a �rst step, we assessed the role of country and industry dummies by estimating three sets of regressions where industry-country skill premiums are explained by i) country dummies alone; ii) industry dummies alone; and iii) country and industry dummies. Country dummies alone account for 32 percent of the variance of the skill premium, industry dummies alone account for 37 percent, and both sets of dummies jointly explain around 66 percent. The dummies are always jointly statistically signi�cant. For the manufacturing sector, country dummies account for 51 percent and industry dummies for 19 percent. In contrast, 52 THE WORLD BANK ECONOMIC REVIEW country dummies account for 27 percent in the non-tradable sector, whereas industry dummies account for 48 percent. These �ndings suggest a stronger role of country variables for the manufacturing sector but a stronger role for industry variables in the non-tradable sector. To study whether sectoral exports are an important determinant of the industry-speci�c skill premiums, we estimated several versions of the following model:   export jc g jc ¼ a ln þ z0jc b þ fj þ fc þ m jc : ð5Þ gdpc The skill premium in industry j, country c ( gjc), is a function of the ratio of sector j exports in country c to GDP (exportjc /gdpc). The coef�cient of interest is a. The model can include industry effects fj, countryeffects, f c, as well as other country-sector characteristics (zjc). The model was estimated with weighted least squares, where the weights are the inverse of the standard errors of the sectoral skill premiums. This GLS strategy accounts for the fact that the industry-speci�c skill premiums are estimated (in equations (1) or (2), for instance). Note that we do not attach a causal interpretation to our estimates. In fact, our results have a reduced-form interpretation, namely to illustrate whether the data support any link between sectoral exports and the skill premiums. Table 5 presents the results. In Panel A), the return to schooling (the coef�- cient of years of education) is the dependent variable gjc; and in Panel B), the coef�cient of the dummy representing skilled workers is the dependent vari- able. Column 1 shows the estimate of the model when the skill premiums are regressed on a constant and the log of the ratio of exports over GDP. The esti- mate for a is positive and signi�cant, thus suggesting that the skill premium rises with exports. The estimates in column 1 imply that doubling a sector’s share of exports over GDP (or a change in the log of exports over GDP equal to one) is associated with an increase of 0.0025 in the return to schooling (Panel A) and with an increase of 0.033 in the skillpremium. The latter �nding suggests that the wage differential between skilled and unskilled workers rises by 3.3 percentage points. Notice that the simulated shock of a change of 1 in the log of exports over GDP is reasonable because the standard deviation of the variable in our sample is about 2.1. Thus this association is positive and signi�cant but it is not very large when compared to the average skill premium of 62 percent or its standard deviation of 36 percent (recall Table 4). In columns 2 to 5 of Table 5, we perform several robustness tests. Column 2 shows the results from the estimation of (5) with industry dummies. The inci- dence of industry exports remains signi�cant, with a similar magnitude as in column 1. Column 3 includes country dummies only, and the link between exports and the skill premium disappears. In Column 4, we include both sets of dummies and the partial correlation remains insigni�cant. Controlling for T A B L E 5 . Exports and the Industry-Skill Premium (1) (2) (3) (4) (5) (6) (7) (8) (9) A) Return to Schooling 0.00250*** 0.00342*** 0.00025 2 0.00027 0.00252*** 0.00309*** 0.00261** 0.00249** 0.00238** log Exports/GDP [0.00091] [0.00096] [0.00099] [0.00120] [0.00093] [0.00102] [0.00104] [0.00113] [0.00113] log GDP_pc 0.01943*** 0.01340*** 0.01439*** 0.02057*** 0.01968*** [0.00344] [0.00512] [0.00509] [0.00592] [0.00601] log Skilled/Unskilled 2 0.01131*** 2 0.00923*** 2 0.01247*** 2 0.00710** 2 0.00899** [0.00320] [0.00336] [0.00363] [0.00351] [0.00411] enrollment rate 0.00032* 0.00021 0.00021 0.00018 [0.00018] [0.00019] [0.00020] [0.00020] export constraints 0.00401 0.00228 0.00628 0.00501 [0.00425] [0.00429] [0.00454] [0.00476] doing business index 2 0.00010** 2 0.00005 [0.00005] [0.00005] average �rm size 2 0.09317** 2 0.07769* [0.03803] [0.04188] Country Dummies No No Yes Yes No No No No No Industry Dummies No Yes No Yes Yes Yes Yes Yes Yes Observations 287 287 287 287 287 287 287 261 261 Summay statistics for 0.026 0.419 0.273 0.608 0.485 0.493 0.502 0.506 0.507 the B) Skill Premium log Exports/GDP 0.03292*** 0.03330*** 0.02187*** 0.00375 0.00997 0.01707** 0.01817** 0.01997** 0.02098** [0.00798] [0.00968] [0.00687] [0.00913] [0.00714] [0.00788] [0.00808] [0.00883] [0.00891] log GDP_pc 0.17993*** 0.12134*** 0.11874*** 0.13265*** 0.13921*** [0.02762] [0.03959] [0.03985] [0.04683] [0.04741] log Skilled/Unskilled 2 0.42187*** 2 0.41095*** 2 0.40506*** 2 0.40962*** 2 0.39591*** [0.02687] [0.02791] [0.02948] [0.02947] [0.03314] Brambilla, Dix-Carneiro, Lederman and Porto enrollment rate 0.00290** 0.00317** 0.00262 0.00289* (Continued ) 53 54 TABLE 5. Continued (1) (2) (3) (4) (5) (6) (7) (8) (9) [0.00145] [0.00151] [0.00159] [0.00162] export constraints 2 0.02069 2 0.01805 2 0.01319 2 0.00466 [0.03309] [0.03340] [0.03606] [0.03727] doing business index 0.00022 0.00038 [0.00035] [0.00042] THE WORLD BANK ECONOMIC REVIEW average �rm size 0.02015 2 0.10155 [0.27825] [0.30904] Country Dummies No No Yes Yes No No No No No Industry Dummies No Yes No Yes Yes Yes Yes Yes Yes Observations 285 285 285 285 285 285 285 259 259 R-squared 0.057 0.256 0.562 0.722 0.619 0.626 0.626 0.624 0.625 Resuls from the second stage regression of the skill premium (and the returns to schooling) on sectoral exports (and other controls). Standard errors in parenthesis. Signi�cance at 1, 5 and 10 percent denoted by ***, ** and * Brambilla, Dix-Carneiro, Lederman and Porto 55 both country and industry dummies might be too restrictive, however. Country �xed effects explain about a third of the variation in skill premiums, and both country and industry dummies account for about 60 percent. This leaves little room for exports to explain the skill premium because much of the variation of the dependent variable is attenuated by the dummies. To learn more about the role of sectoral exports, we work with a more parsimonious version of equation (5) where instead of country dummies we control for country characteristics, namely the log of per capita GDP and the ratio of skilled (high school completed) over unskilled labor. These results are reported in column 5 of Table 5. Both per capita GDP and the skill compo- sition are statistically signi�cant determinants of the industry-country skill pre- miums (using both years of education or a skilled worker dummy) with the expected signs: disparities between skilled and unskilled wages rise with GDP per capita, and countries with a greater fraction (supply) of skilled workers pay lower skill premiums. The signi�cance of these variables supports their use in lieu of the country �xed effects. Also, the R 2 of the model remains high (at 0.485 in Panel A and 0.619 in Panel B), which is much higher than the R 2 from the model with country dummies. Note that, in these models (column 5), the coef�cient of exports as a fraction of GDP is positive and statistically signi�cant, and the estimates are of similar magnitude as those reported in columns 1 and 2. While these results suggest a positive partial correlation between exports and the skill premium, the fact that this correlation does not survive the inclusion of country �xed effects deserves further attention. What country characteristics that are correlated with exports and skill premiums might be lurking behind the country effects? One candidate provided by the literature is school coverage. As pointed out by Card (1991), countries with higher school coverage may also provide education of lower quality (or to students from rela- tively poor families) thus reducing the magnitude of estimated skill premiums. Another factor that may affect the results is the business climate, which may facilitate exports and allow �rms to pay higher wages for skill laborat the same time. Using data from the World Bank’s Doing Business database, we add two controls in our regressions, the lead time to export measured by the number of days and an index of the ease for doing business in each country. The corre- sponding results are reported in columns 6 and 7 of Table 5. Even after con- trolling for these country characteristics, the positive partial correlation between sectoral exports and the industry-country skill premium remains statistically signi�cant. This occursfor both dependent variables, returns to schooling (Panel A) and the skill premium (Panel B). Yet another factor that may contaminate our results is �rm size.14 If, for instance, labor turnover is an issue, larger �rms may need to pay higher wages (especially to skilled workers) to retain workers. Alternatively, if larger �rms 14. We thank a referee for pointing this out. 56 THE WORLD BANK ECONOMIC REVIEW offer more stable employment prospects, smaller �rms may need to pay higher premiums to attract (skilled) workers. In any case, failure to control for �rm size may bias our results. While some household surveys in Latin America include partial information of �rm characteristics, we use standardized data from UNIDO, compiled by Nicita and Olarreaga (2007), to construct a measure of the average size of �rms at the sector level for all countries in our sample (except Uruguay). The UNIDO data includes the number of establish- ments and the total number of employees in each industry-country, and we use the ratio of these two variables (aggregated at the 2-digit level) to control for �rm size. The results of the estimations that include this variable are in columns 8 and 9. We �nd, especially in Panel A, that �rm size is negatively cor- related with the skill premium (which indicates that the mechanisms outlined here could be of practical relevance). More importantly for our purposes, exports remain positive and statistically signi�cant. We performed two additional robustness tests. First, we estimated all model speci�cations reported in Table 5 but with monthly total wages as the depen- dent variable. Second, we estimated all the models on the skill premiums with male workers. Results arereported in Tables A6 and A7 of our online appendix, and sectoral exports are positive and statistically signi�cant.15 Thus far, the most important �nding in Table 5 (and in the robustness checks in the appendix) is that, after controlling for industry �xed effects, country characteristics such as per capita GDP, relative endowments, enroll- ment rates, export constraints, and the cost of doing business, the data still suggests a positive and statistically signi�cant partial correlation between exports and skill premiums. We �nish by studying other trade-related determi- nants of industry skill premiums. Unit values, which are proxies for product quality, might also be important. A model of the impact of quality upgrading on wage inequality (or increases in skill premiums) is developed and estimated by Verhoogen (2008). Product variety, measured by the dispersion of unit values within industries, might be a correlate of skill premiums. Perhaps �rms in sectors with wide scope for product differentiation can exercise monopoly power, charge higher mark-ups, and perhaps pass-on those pro�ts to their workers. Alternatively, product differentiation itself might require skilled labor. The calculation of unit values using data from the U.N. Comtrade database is not straightforward and inevitably brings measurement errors. We used three 15. It is worth recalling that this �nding does not necessarily mean that increasing exports, which raise skill premiums and the returns to schooling, will raise social welfare. Indeed, the results suggest that exports raise the wage gap between skilled and unskilled workers. A remaining issue is whether exports raise wage inequality above and beyond their effects on the skill premium. To explore this issue, we estimated auxiliary models with the ratio of wage-equation residuals of the 90th percentile over the bottom 10th percentile as the dependent variable, following the same structure of explanatory variables presented in Table 5.These results, available from the online appendix, suggest that exports do not robustly raise wage-residual dispersion. We are indebted to an anonymous referee for raising this issue related to wage inequality. Brambilla, Dix-Carneiro, Lederman and Porto 57 different measures for unit values to test the robustness of the partial corre- lation between exports and skill premiums. Some empirical issues need to be addressed, however. First, in Comtrade, recorded transactions for a single HS code appear with different quantity codes. To address this concern, for a given HS code, we pooled data from all countries and picked the quantity code that is reported most frequently. For the calculation of unit values, we only considered those transactions that were reported in the “most frequent quantity code,� to make sure that unit values for a given HS code are expressed in the same units across countries. Unit values were then aggregated to the ISIC Rev 3, 2-digit level by taking weighted averages (weights are given by the importance of a given HS code exports on total exports of the corresponding 2-digit ISIC industry). The indicator of the dispersion of unit values is the variance of unit values across HS codes within a country and 2-digit ISIC industry. Second, unit values are highly dispersed, and therefore we used the median unit values (without any weighting) as a second measure of unit values. The corresponding indicator of dispersion is still the variance of unit values. Third, to account for outliers we trimmed the top and bottom �ve percent of the observations on unit values. In turn, we calculated the weighted average within countries and 2-digit ISIC industries as in the �rst approach. The regression model is similar to equation (5). That is, we regress the skill premium in industry j and country c on the measures of unit values and the variance of unit values plus industry dummies and national characteristics instead of country dummies, namely the log of per capita GDP and the ratio of skilled to unskilled endowments.16 The main results are in Table 6. Each panel (A to C) in the table corresponds to one of thethree indicators of unit values. In columns 1-3, the dependent variables is the return to years of schooling, and in columns 4-6, it is the skill premium (with skills measured with a dummy variable). Our �rst conclusion is that neither unit values nor the dispersion of unit values explain the skill premium. While these results appear robust, it is always plausible that they are the consequence of noise in the unit values. For instance, in speci�cation in which we trim the top and bottom 5% of the unit values, the dispersion in unit value becomes signi�cant in some regressions. This result hints at the relevance of the scope for product differentiation. Nevertheless, the key �nding from Table 6 is that in all models that control for unit values, sectoral exports are still signi�cant in explaining skill pre- miums. Also, the magnitudes of the estimates are similar to those in Table 5. This robustness test supports the view that exports signi�cantly affect the premium paid for skills at the industry level. 16. These results are not reported for the sake of brevity. 58 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . Export Unit Values and the Skill Premium Years of Education Skilled Dummy (1) (2) (3) (4) (5) (6) PANEL A Log Unit value 0.0006 2 0.0001 0.0037 2 0.009 [0.0009] [0.001] [0.006] [0.01] log Var(Unit_value) 0.00031 0.0003 0.003 0.006 [0.0003] [0.0007] [0.003] [0.006] log Exports/GDP 0.0024** 0.0024** 0.0024** 0.017** 0.016* 0.016* [0.001] [0.001] [0.001] [0.008] [0.008] [0.008] Observations 287 287 287 285 285 285 R-squared 0.5 0.5 0.5 0.63 0.63 0.63 PANEL B Log Unit value 2 0.0002 2 0.0006 0.024 0.021 [0.002] [0.003] [0.023] [0.023] log Var(Unit_value) 0.0003 0.0003 0.0029 0.0025 [0.0003] [0.0003] [0.0029] [0.0029] log Exports/GDP 0.0026** 0.0024** 0.0024** 0.017** 0.016* 0.015* [0.001] [0.001] [0.001] [0.008] [0.008] [0.008] Observations 287 287 287 285 285 285 R-squared 0.5 0.5 0.5 0.628 0.628 0.629 Log Unit value 0.0011 2 0.0009 0.005 2 0.004 [0.0009] [0.001] [0.007] [0.014] log Var(Unit_value) 0.0007* 0.001 0.003 0.004 [0.0004] [0.0008] [0.003] [0.006] log Exports/GDP 0.0023** 0.0022** 0.0024** 0.017** 0.016** 0.016** [0.001] [0.001] [0.001] [0.008] [0.008] [0.008] Observations 287 287 287 285 285 285 R-squared 0.5 0.5 0.5 0.63 0.63 0.63 Panel (A): dispersion in unit values measured with the variance of unit values across Harmonized System codes within a country and 2-digit ISIC industry. Panel B): median of the unit values. Panel C): variance of unit values across HS codes, after trimming for ourliers. Standard errors in parenthesis. Signi�cance at 1, 5 and 10 percent denoted by ***, ** and * III. CONCLUDING REMARKS This paper studied the returns to schooling and the skill premium in Latin America and the Caribbean and its correlation with exports. We �rst estimated and described national skill premiums for over seven million workers from sixteen countries. Motivated by recent models featuring limited inter-industry factor mobility, we estimated industry-speci�c skill-premiums for sixty 2-digit ISIC sectors. Finally, we investigated reduced-form regressions linking these country-industry skill premiums with exports. An interesting and previously unknown �nding is that unobserved industry- and country-speci�c effects jointly explain over 60 percent of the observed var- iance in the skill premium. Each set of factors has about the same explanatory power for skill premiums in manufacturing sectors. It is thus not clear that Brambilla, Dix-Carneiro, Lederman and Porto 59 industrial policies would succeed anymore than industry-neutral national policies in changing the skill premium. In addition, sectoral exports are related to sectoral skill premiums: sectors with higher exports pay higher skilled pre- miums. This evidence supports recent trade theories linking exports to wages and to skills, as in Brambilla, Lederman and Porto (2010) and Verhoogen (2008), and it highlights the need for further research to understand the mech- anisms at work. However, the welfare implications of these results remain unclear, because export-driven skill premiums would raise national welfare only if there is a gap between the social and private returns to schooling. If not, raising skill premiums by raising exports would only result in an increase in wage inequality between skilled and unskilled workers. REFERENCES Alvarez, R., and R.A. Lo´ pez 2005. “Exporting and Performance: Evidence from Chilean Plants.� Canadian Journal of Economics 38: 1384– 1400. Amiti, M., and D. Davis 2008. “Trade, Firms, and Wages: Theory and Evidence.� National Bureau for Economic Research Working Paper No. 14106, Cambridge, Massachusetts. Beine, M., F. Docquier, and H. Rapoport 2001. “Brain Drain and Economic Growth: Theory and Evidence.� Journal of Development Economics 64(1): 275– 289. Bernard, A., and J. Jensen 1995. “Exporters, Jobs, and Wages in U.S. Manufacturing: 1976-1987.� Brooking Papers on Economic Activity: Microeconomics 1995: 67– 119. ——— 1999. “Exceptional Exporter Performance: Cause, Effect, or Both?� Journal of International Economics 47: 1 –25. Bernard, A., and J. Wagner 1997. “Exports and Success in German Manufacturing.� Weltwirtschaftliches Archiv 133: 134– 157. Bernard, A.B., and J.B. Jensen 2004. “Why Some Firms Export.� Review of Economics and Statistics 86: 561– 569. Bernard, A., B. Jensen, S. Redding, and P. Schott 2007. “Firms in International Trade.� Journal of Economic Perspectives 21(3): 105 –130. Brambilla, I., D. Lederman, and G. Porto 2010. “Exports, Export Destinations and Skills.� National Bureau for Economic Research Working Paper No. 15995, Cambridge, Massachusetts. Card, D. 1999. “The Causal Effect of Education on Earnings,� in A. Orley and D. Card, editors, Handbook of Labor Economics, Vol. 3, pp. 1801– 1863. Amsterdam: Elsevier Science B.V. Clerides, S., S. Lach, and J. Tybout 1998. “Is Learning by Exporting Important? Micro-Dynamic Evidence from Colombia, Mexico, and Morocco.� Quarterly Journal of Economics 108(3): 903–947. De Loecker, J. 2007. “Do Exports Generate Higher Productivity? Evidence from Slovenia.� Journal of International Economics 73: 69 –98. Dickens, W., and L. Katz 1987. “Inter-Industry Wage Differences and Industry Characteristics,� in K. Lang and J. Leonard, editors, Unemployment and the Structure of Labor Markets. New York and Oxford: Basil Blackwell: Basil Blackwell. Dickens, W., and K. Lang 1988. “Labor Market Segmentation and the Union Wage Premium.� Review of Economics and Statistics 70(3): 527–530. Egger, H, and U. Kreickemeier 2009. “Firm Heterogeneity And The Labor Market Effects of Trade Liberalization.� International Economic Review 50(1): 187–216. 60 THE WORLD BANK ECONOMIC REVIEW Ferreira, F., P. Leite, and M. Wai-Poi 2007. “Trade Liberalization, Employment Flows, and Wage Inequality in Brazil.� World Bank Policy Research Working Paper No. 4108, Washington, DC. Galiani, S., and G. Porto 2010. “Trends in Tariff Reforms and in the Structure of Wages.� The Review of Economics and Statistics 92(3): 482 –494. Gibbons, R., and L. Katz 1992. “Does Unmeasured Ability Explain Inter-Industry Wage Differentials?� Review of Economic Studies 59: 515 –35. Giovagnoli, P., A. Fizsbein, and H. Patrinos. 2005. “Estimating the Returns to Education in Argentina: 1992-2002.� World Bank Policy Research Working Paper No. 3715, Washington, DC. Goldberg, P., and N. Pavcnik 2005. “Trade, Wages, and the Political Economy of Trade Protection: Evidence from the Colombian Trade Reforms.� Journal of International Economics 66: 75– 105. ——— 2006. “Distributional Effects of Globalization in Developing Countries.� Journal of Economic Literature 45(1): 39 –82. Griliches, Z. 1977. “Estimating the Returns to Schooling: Some Econometric Problems.� Econometrica 45: 1– 22. Hausmann, R., J. Hwang, and D. Rodrik 2005. “What You Export Matters.� Journal of Economic Growth 12: 1– 25. Isgut, A. 2001. “What’s Different about Exporters? Evidence from Colombian Manufacturing.� Journal of Development Studies 37(5): 57 –82. Krueger, A.B., and M. Lindahl 2001. “Education for Growth: Why and For Whom?� Journal of Economic Literature 39(4) 1101– 36. Krueger, A.B., and L.H. Summers 1988. “Ef�ciency Wages and Inter-Industry Wage Structure.� Econometrica 56(2) 259–93. Manasse, P., and A. Turrini 2001. “Trade, Wages, and ‘Superstars’.� Journal of International Economics 54: 97– 117. Melitz, M. 2003. “The Impact of Trade on Intra-industry Reallocations and Aggregate Industry Productivity.� Econometrica 71(6): 1695– 1725. Nicita, A., and M. Olarreaga 2007. “Trade, Production, and Protections Database, 1976-2004.� World Bank Economic Review 21: 165–171. Park, A., D. Yang, X. Shi, and Y. Jiang 2010. “Exporting and Firm Performance: Chinese Exporters and the Asian Financial Crisis.� The Review of Economics and Statistics 92(4): 822–42. Pavcnik, N. 2002. “Trade Liberalization, Exit, and Productivity Improvements: Evidence from Chilean Plants.� Review of Economic Studies 69(1): 245–76. Pavcnik, N., A. Blom, P. Goldberg, and N. Schady 2004. “Trade Liberalization and Industry Wage Structure: Evidence from Brazil.� World Bank Economic Review 18(3): 319– 44. Sanroman, G. 2006. “Returns to Schooling in Uruguay.� Working Paper No. 1406, Departamento de Economia, Universidad de la Republica, Uruguay. Schank, T., C. Schnabel, and J. Wagner 2007. “Do Exporters Really Pay Higher Wages? First Evidence from German Linked Employer-Employee Data.� Journal of International Economics 72(1): 52– 74. Verhoogen, E. 2008. “Trade, Quality Upgrading, and Wage Inequality in the Mexican Manufacturing Sector.� Quarterly Journal of Economics 123(2): 489– 530. How to Deal with Covert Child Labor and Give Children an Effective Education, in a Poor Developing Country Alessandro Cigno* Because credit and insurance markets are imperfect and intrafamily transfers and how children use their time outside school hours are private information, the second-best policy makes school enrollment compulsory, forces overt child labor below its ef�cient level (if positive), and uses a combination of need- and merit-based grants, �nanced by earmarked taxes, to relax credit constraints, redistribute, and insure. Existing con- ditional cash transfer schemes can be made to approximate the second-best policy by incorporating these principles in some measure.child labor, education, uncertainty, moral hazard, optimal taxation. JEL codes: D82, H21, H31, I28, J24 Developing country governments and international development agencies have long been aware that human capital accumulation, more than physical capital accumulation, is the mainspring of economic and civil progress. But many chil- dren in poor developing countries fail to complete even primary education, and some do not go to school at all. The reasons are well known.1 Baland and Robinson (2000) demonstrate that child labor will be inef�ciently high if parents are either credit or bequest constrained.2 Evidence that parent inability to borrow discourages education and encourages child labor is reported by a host of researchers, including Jacoby (1994) and Fuwa and others (2009). Loury (1981) and Pouliot (2006) demonstrate that parent inability to insure against the risk of a low return causes education investment to be inef�ciently low and child labor to be inef�ciently high, even when credit is not rationed * Alessandro Cigno (cigno@uni�.it) is a professor of economics at the University of Florence. Work on this article was completed while the author was visiting the Institute of Economic Research at Hitotsubashi University. Comments by three anonymous referees and editorial advice by Alain de Janvry are gratefully acknowledged. 1. For a systematic exposition, see Cigno and Rosati (2005). 2. Cigno (1993, 2006) shows, however, that the problem is alleviated when a set of self-enforcing, renegotiation-proof family rules oblige working-age family members to support their young children and elderly parents. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 61 – 77 doi:10.1093/wber/lhr038 Advance Access Publication July 26, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 61 62 THE WORLD BANK ECONOMIC REVIEW and bequests are interior.3 Ram and Schultz (1979) and Jacoby and Skou�as (1997) �nd evidence that parent inability to insure against the risk of a low return discourages education investment in developing countries. Parent incomes may also be uncertain. Evidence in Beegle, Dehejia, and Gatti (2006), that parents respond to a negative income shock by making their children work more, suggests that households cannot insure against that kind of risk either. Fitzsimons (2007) reports, however, that parents respond in this way to a downturn not in their own income, but in village aggregate income, suggesting that idiosyncratic income shocks are neutralized by informal insurance arrange- ments at the local level.4 Because idiosyncratic shocks even out on average, governments face less risk than do individual households. Partly because of this lower risk, governments also have easier access to international money markets than do most citizens. Thus in imperfect domestic credit and insurance markets there is an ef�ciency argument for governments to lend to and insure parents of school-age children. Given evidence of diminishing absolute risk aversion in an education context (Kodde 1986, for example) and given an unequal distribution of parent wealth, there is also an equity argument. Ef�ciency-enhancing policies are politically easier to implement when they do not involve redistribution, but it is dif�cult to see how redistribution could be avoided. Even if the government could �nance the education of poor children entirely by borrowing against their future tax payments, insuring families against the risk of a low return to edu- cation would still imply redistribution from rich to poor school leavers. Similarly, insuring families against the risk of a downturn in parent income would involve redistribution from rich parents to poor parents.5 The tension between equity and ef�ciency will be minimal for a small-scale project, especially if �nanced largely by international aid,6 but not for a large-scale one. It is also dif�cult to imagine that any project, large or small, could be sup- ported by the international community forever. Information asymmetries give rise to another set of problems. In developing countries many children work, but much of what they do is invisible to the government. A small fraction of this covert child labor involves physically damaging or morally degrading activities—the “worst forms� of child labor— which national governments are committed by international treaty to eradicate. 3. According to Levhari and Weiss (1974), the return to education is uncertain because a child’s learning ability is fully revealed only after the education investment is carried out. For evidence of that, see Belzil and Hansen (2002). According to Razin (1976), the uncertainty concerns the rental price of the human capital accumulated through education. In developing countries the uncertainty also concerns the length of time for which the future adult will be able to enjoy the bene�t. 4. Evidence of such arrangements in a developing country is reported by Besley (1995) and Townsend (1994), among others. 5. See Johnson (1987). 6. It will not even arise when education is privately �nanced by migrant remittances. Dessy and Rambeloma (2010), Epstein and Kahana (2008), and Hanson and Woodruffs (2003), report evidence that such remittances reduce child labor in the families left behind. Cigno 63 But most covert child labor consists of activities conducted for and under the direct supervision of the child’s parents (such as helping in the home, working on the family farm, and contributing to the family business).7 While compara- tively harmless in themselves, these activities conflict with education and thus have an opportunity cost in terms of forgone future earnings. The government might want to regulate them, but it cannot because they are private infor- mation; this creates a moral hazard problem. Similar considerations apply from an education standpoint. Scholastic perform- ance depends not only on how much time children spend attending school, but also on how much time they spend doing homework and on how alert and well rested they are when doing both. A child who falls asleep during lessons and does not �nd the time or is too tired to give homework the necessary attention will have poorer results in school than will a child of the same learning ability who comes to school well rested and with homework conscientiously done. Because school enrollment and school attendance are common knowledge but much of what chil- dren do outside school is private information, another moral hazard problem exists. And a similar problem arises from the fact that intrafamily transfers are private information, because the government cannot be sure that a public subsidy intended for children does not end up as extra consumption for the parents. This article describes the second-best policy and compares it with two benchmarks—a low one represented by a laissez-faire policy and a high one rep- resented by the �rst-best policy—in a situation where parents can neither borrow nor insure, where parents are better informed than the government about their children’s time allocation, and where the government does not observe intrafam- ily transfers. The analysis assumes that the expected return to education is posi- tive.8 Fasih (2008) reports evidence of high returns to education, especially in low- and middle-income countries. The same study reports that these returns are lower for poor children than for rich children. That may be a sign that the poor can afford only poor-quality schools,9 but the line of reasoning in this article suggests another (not necessarily rival) explanation: that children in poor families study less, or less effectively, per year of school enrollment or day of school attendance than do children in rich families. The worst forms of child labor raise moral issues that transcend the materialistic calculations underlying this article10 and are thus omitted from the formal analysis, but the analysis argues that the proposed policy reduces child labor in all its forms. The policy optimization has an optimal taxation, or principal-agent, format.11 This type of analysis does not appear to have been attempted before 7. See Cigno and Rosati (2005). 8. For evidence of a causal effect of education on future earnings, see Card (1999) and Oreopoulos (2006). 9. On the subject, see Alderman, Orazem, and Paterno (2001). 10. See Dessy and Pallage (2005) for a strictly economic analysis. 11. For a survey of the ways in which the approach can be used in a family policy context, see Cigno (2011). For an application in higher education, see Cigno and Luporini (2009). 64 THE WORLD BANK ECONOMIC REVIEW in the context of a poor developing country. Given the context, “school age� is taken to mean primary school age. Assuming that children in that age range are under parent control, parents, rather than the children themselves, are taken to be the agents and are modeled as risk-averse, expected-utility maxi- mizers.12 Because the implications of an education externality are well under- stood and allowing for it would merely reinforce the argument for public intervention, the analysis excludes it (but �nds that the policy itself gives rise to a �scal externality). Because the argument for having the policy �nanced out of general tax revenue is weak in the absence of an education externality, and assuming that international aid cannot go on forever, the constraint that the policy must be self-�nancing is imposed. Section I lays down the technical assumptions and characterizes parent decisions. Section II examines the laissez- faire equilibrium. Section III derives the �rst- and second-best policies. And section IV discusses actual policy practice (including conditional cash transfer schemes) in the light of the theoretical �ndings. I. TECHNICAL ASSUMPTIONS AND PARENT DECISIONS There are a large number of families, i ¼ 1, 2 . . . n, each consisting of a couple with a given number of children, the same for every family and normalized to unity. The assumption that all parents have the same number of children is less than realistic, but the normative implications of departing from it have been examined in depth elsewhere13 and do not impinge on the points at issue here. Learning ability is randomly distributed across children and imperfectly obser- vable until the education investment is carried out. Parent income varies exogenously across families and is observable by the government. Later in the analysis income will be allowed to be either uncertain or private information. There are two periods, t ¼ 1, 2. Children are alive in both, parents only in the �rst. For brevity, the child in the ith family is referred to as i. Ex post, i’s utility will be Ui ¼ u(ci1) þ u(ci2), where cit denotes i’s consumption in period t. 12. That is not the only possible representation of individual behavior in the face of uncertainty. In prospect theory (Kahneman 2003) individuals are assumed to be risk averse in the domain of gains and risk lovers in that of losses. Although this alternative approach has some empirical justi�cation, it is not followed here for two reasons. First, in a situation where most people live slightly above the subsistence level, little risk-loving behavior is likely to be observed. Second, the policymaker may not approve of such behavior and may consequently maximize an objective function that is not a mere aggregation of the individual ones. Kanbur, Pirttila ¨ , and Tuomala (2008) show that if the government corrects for what it considers an aberrant behavior, the solution to an optimal taxation problem with moral hazard may have the same properties as if the agents were risk-averse, expected-utility maximizers. 13. If the number of children is exogenous and parent income or work effort are private information, the optimal income tax rate is zero for the family with the highest income. If the number of children is exogenous but varies across families as in Cremer, Dellis, and Pestieau (2003), the optimal policy will redistribute in favor of families with more children. Neither of these properties necessarily applies, however, if the number of children is endogenous, as in Cigno (2001) and Balestrino, Cigno, and Pettini (2002): the �rst one (no distortion at the top) because children’s visibility makes mimicking much harder, the second one (children reduce tax liability) because children yield utility. Cigno 65 Assuming descending altruism, the ex post utility of i’s parents may be written as Vi ¼ v(ai) þ bUi, 0 , b , 1, where ai denotes parent consumption and b is a measure of altruism. The functions u(.) and v(.) are assumed to be increasing 0 0 and concave, with u (0) ¼ v (0) ¼ 1. In an uncertain environment concavity implies risk aversion. Assuming that marginal utility becomes very large as consumption approaches zero implies that subsistence consumption is normal- ized to zero. Because utility does not depend on time allocation, this implies that leisure is not a good and that work does not yield direct disutility. This may be justi�ed by saying that, with the worst forms of child labor out of the picture and consumption likely to be low, the marginal utility of income is likely to be higher than that of leisure. In period 1 a child may or may not be enrolled at school. Enrollment has a �xed cost p, equal to the average cost of tuition.14 The previous section argued that effective education time is increasing in school attendance, homework, and rest time and decreasing in work time. If i is enrolled, effective education time will be positive (or there would be no point in paying p). To simplify, effective education time is measured as the amount of time that the child does not work. If i is not enrolled, effective education time will be zero.15 Child labor may be overt or covert. Overt child labor consists of work done for an employer other than the child’s parents and carries a wage. Covert child labor involves either participating in a family-run, income-generating activity, such as farming or retailing, or replacing the child’s parents in performing household chores such as cooking, cleaning, fetching water, and gathering fuel.16 Although neither form of covert child labor carries a wage, the former produces income directly, and the latter indirectly by allowing the child’s parents to spend more time raising income. Let ei denote i’s effective education, and let Li denote i’s overt labor. Normalizing a child’s time endowment to unity, i’s covert child labor is then (1 – ei – Li). Because children can contribute to the production of family income, parent income in period 1 is de�ned as family income if child labor in both its forms are zero. Let yi denote parent income in family i. The income generated by overt child labor is Li w1, where w1 is the child wage rate, and income gener- ated by covert child labor is z(1 – ei – Li), where z(.) is a revenue function, 0 increasing and concave, with z(0) ¼ 0 and z (0) ¼ 1. By de�nition of revenue 14. Tuition fees are usually per year (or shorter period such as the semester) of school enrollment, so the total a family spends for a child’s tuition reflects the number of years for which the child is enrolled at school but not the number of days for which the child actually attends school or the number of hours during which the child studies at home, in each of those years. This lumpiness of tuition fees is accounted for by treating p as a constant. 15. In many developing countries a substantial minority of school-age children is reported as neither working nor studying. This can be explained without introducing leisure by allowing for the existence of �xed costs of access to school and work; see Cigno and Rosati (2005). 16. See Cigno and Rosati (2005) for an analysis of the incidence of these activities and for the effects of making fetching water and gathering fuel unnecessary by providing homes with electricity and running water. 66 THE WORLD BANK ECONOMIC REVIEW function, z(1 – ei – Li) is the maximum amount of income that the family can produce with (1 – ei – Li) units of i’s time by optimally allocating this time between direct participation in income-raising activities conducted by i’s parents and replacement of i’s parents in performing household chores. Concavity reflects diminishing marginal rates of technical substitution between adult and child work. Assuming that marginal revenue gets very large as covert child labor gets very small is realistic in a poverty context like the present one and ensures that such labor will never be zero. In period 2 i will earn w2 þ xi, where w2 denotes the income of an unskilled adult and xi denotes the individ- ual skill premium. If i does not enroll at school, xi will be zero. In period 1, i receives a transfer, mi, from i’s parents,17 and another, gi, from the govern- ment. In period 2 i will make a transfer to the government, ui. All these trans- fers can be positive, negative, or zero. Parents make their decisions in period 1, after the government has announced its policy. Anticipating a result to be obtained in section III, gi is taken to be a function of yi, and ui to be a function of xi. While w1, w2, and gi are certain, xi and consequently ui are uncertain. Because ei must be chosen in period 1, edu- cation is a risky investment. The supplementary assumption (to be relaxed later) is made that xi is independent and identically distributed over the closed interval [0, x¯ ] [ R þ with density f(.jei) conditional on ei and f(.j0) ¼ 0. To simplify the notation, xi is used to measure both the �nal school result and the skill premium.18 The cumulative distribution of xi, F(xijei), associated with a higher ei, �rst-order stochastically dominates the one associated with a lower ei, ð1Þ Fei ðxi jei Þ 0: In other words, the more i studies and the less i works, the more of a chance i has of getting good marks and thus of attracting a high skill premium. For each ei there will be values of xi such that equation (1) holds as an inequality. The standard convexity of distribution function assumption, that F(x ijei) is convex in fei ð:jei Þ ei, and monotone likelihood ratio assumption, that f ð:jei Þ is increasing in xi,19 are made, which allow the �rst-order approach to be adopted. 17. One might be tempted to simplify the analysis by taking the utility aggregation problem as solved and viewing Vi as a family welfare function. This would allow intrafamily transfers to be left out and all costs and revenues to be treated as pertaining to the family as a whole, but doing so would be misleading, because, as Baland and Robinson (2000) show, transfers from parents to children may be inef�ciently low. 18. Using one random variable with density conditional on study time to represent the school result and another with density conditional on the school result to represent the skill premium would make no substantive difference to the results so long as both are independent and identically distributed and the skill premium is not conditional on some decision variable. 19. This property might not hold if xi depended on systemic factors and (1 – ei) depended on employment opportunities. In the present context, however, it seems reasonable to assume that there is nothing to stop wi from falling low enough to clear the (overt) child labor market and that there will also be plenty of opportunities for covert child labor. Cigno 67 If i enrolls for school and overt child labor is not regulated by the govern- ment, i’s parents will choose the (ei, Li, mi) that maximizes  �  EðVi Þ ; vi þ b ui1 þ xi ui2 f i dxi , where vi ; v( yi þ zi – mi), zi ; z(1 – Li – ei), ui1 ; u(mi þ w1Li þ gi – p}), ui2 ; u(w2 þ xi – ui), and f i ; f(xijei), subject to ð2Þ ei ! 0; ð3Þ Li ! 0 and ð4Þ 1 À ei À Li ! 0: Because equation (4) will never be binding for the restrictions imposed on the revenue function, the �rst-order conditions are ð ð5Þ 0 0 À v i zi þ b ui2 feii dxi þ ji ¼ 0 for ei xi ð6Þ À v0i z0i þ ci þ bu0i1 w1 ¼ 0 for Li and ð7Þ À v0i þ bu0i1 ¼ 0 for mi where ji is the Lagrange multiplier of equation (2) and ci that of equation (3). If i does not enroll, ei cannot be positive. Again assuming that overt child labor is free to vary, i’s parents will then choose (Li, mi) to maximize V(Li, mi) ; v( yi þ z(1 – Li) – mi) þ b [ u(mi þ w1 Li) þ u(w2) ], subject to equation (3) – equation (4). The solution will satisfy equation (6) – equation (7) for p ; ei ; 0. If Li is regulated by the government, equation (6) need not hold. Irrespective of whether i is enrolled and Li is regulated, it is clear from equation (7) that mi is decreasing in gi. In other words, public transfers crowd out private transfers. II. LAISSEZ-FAIRE EQUILIBRIUM Under laissez-faire school enrollment is not compulsory, overt child labor is free to vary, and g i ; u i ; 0. The payoff of enrolling i at school is ð8Þ pS ðyi ; pÞ ; max EðVi Þ; subject to equation ð2Þ À equation ð4Þ ðLi ;ei ;mi Þ and the payoff of not enrolling is ð9Þ pW ðyi Þ ; max V ðLi ; mi Þ; subject to equation ð2Þ À equation ð4Þ: ðLi ;mi Þ 68 THE WORLD BANK ECONOMIC REVIEW The child will enroll if p S ( yi, p) is at least as large as p W ( yi). There is then a y, de�ned by pS ð~ threshold value of yi, ~ y; pÞ ¼ pW ð~yÞ, below which i will not be y is the same for every i, because the expected return to education is enrolled. ~ the same for all of them, so if any children are not enrolled, it will be those whose parents have a low income. This result differs from the one in Ranjan (2001), where children’s learning ability is assumed to be directly observable ex ante and the threshold is consequently lower for parents of high-ability chil- dren than for parents of low-ability children. Given that ji will be zero if ei is positive, equation (5) implies ð ð10Þ 0 0 either ei ¼ 0 or vi zi ¼ b ui2 feii dxi : xi Therefore, either ei is zero and i is not enrolled or ei is positive and increasing in yi. Taken together with equation (7) and given that ci will be zero if Li is positive, equation (6) similarly implies ð11Þ either Li ¼ 0 or z0 ð1 À Li À ei Þ ¼ w1 : Therefore, Li is either zero, or positive and increasing in w1. It is then clear that overt child labor is the same in all families.20 What differs is effective edu- cation and total (overt plus covert) child labor. Proposition 1. Under laissez-faire children from very poor families are not enrolled at school; children from less poor families are enrolled, but their effective education increases with parent income; and overt child labor is either zero or increasing in the child wage rate. The second part of this proposition provides a possible explanation for the empirical �nding that poor children get a smaller increase in their future income in return for an extra year of school enrollment or an extra day of school attendance than do rich children, because it says that poor children receive less effective education during that extra year or day than do rich children. III. FIRST- AND SECOND-BEST POLICIES The government’s preferences are represented by the Benthamite social welfare function, X n ð12Þ SW ¼ EðVi Þ: i¼1 Because there are many parents and children and because risks are assumed to be uncorrelated, the government does not face any uncertainty about its tax 20. It would vary across families if the z() function did (for example, if the return to covert child labor were higher in a farming family that owns land than in one that does not). Cigno 69 revenue and thus has easier access to international credit than individual citi- zens. The usual “small country� assumption—that the real interest rate is con- stant and normalized to zero—is made. Because the expected return to education is the same for every i, and assuming it is positive, the government will then make school enrollment compulsory. Because the optimization can determine only relative tax rates, the tax on w2 is normalized to zero, and the socially optimal values of gi and ui are investigated. Because the government does not face budget uncertainty, it will choose (ei, Li, mi, gi, ui), for i ¼ 1, 2, . . . n, to maximize equation (12), subject to the budget constraint, Xn  ð  ð13Þ gi À ui f i dxi ¼ 0; i¼1 xi and equation (2) – equation (4). If (ei, mi) is private information, the maximi- zation will also be subject to incentive-compatibility constraints. Because E(Vi) is concave in (e1, Li, mi), SW will be concave in it too. For the independent and identically distributed assumption, the optimal ( gi, ui) can depend only on (ei, mi, xi, yi) and not on any (ej, mj, xj, yj ) for j = i. First-Best Policy Under the �rst-best policy the government prescribes (ei, Li, mi) and designs personalized lump-sum transfers, ( gi, ui), for each i. Because there are no incentive-compatibility constraints, and denoting the Lagrange multiplier of equation (13) as l, the �rst-order conditions are equation (6) for Li, equation (7) for mi, ð ð14Þ À v0i z0i þ ðbui2 þ lui Þfeii dxi þ ji ¼ 0 for ei ; xi ð15Þ bu0i1 À l ¼ 0 for gi ; and, at each possible realization of xi, À � ð16Þ À bu0i2 À l f i ¼ 0 for ui : Because equation (11) must still hold, it is clear that the �rst-best Li is the same for every i, Li ¼ L FB, and not necessarily zero. The �rst-order condition on ei is not the same as under laissez-faire because it takes account � of the expected marginal bene�t of tax revenue for society as a whole, l xi ui feii dxi . Given this �scal externality, ei will be larger than under laissez-faire for every i. In view of equations (7), (15), and (16), it is also clear that ai ¼ a FB, ci1 ¼ ci2 ¼ c FB, and mi ¼ m FB. Because this implies that parent income is equalized across families and children are ex ante identical, the �rst-best level of ei is the same for every i, ei ¼ e FB. 70 THE WORLD BANK ECONOMIC REVIEW Proposition 2. Under the �rst-best policy the government uses lump-sum taxes and subsidies to achieve perfect equity, perfect consumption smoothing, and full insurance; all school-age children allocate their time in the same way; overt child labor is either zero or increasing in the child wage rate; each school-age child receives more effective education than under laissez-faire. The last part of this proposition implies that the laissez-faire level of effective education is inef�ciently low. Second-Best Policy Under the second-best policy, (ei, mi) is private information. According to the logic of optimal taxation, the government will then make school enrollment compulsory, �x Li, and influence parent decisions by announcing how gi and ui will be related to the information available in the relevant period. Because gi is payable in period 1, it can depend only on yi. Because ui is payable in period 2, it can also depend on xi. If it seems odd that a benevolent government might actually oblige children to do a certain amount of paid work, think of the second-best value of Li as a legal maximum. Because of the potential moral hazard problem, the maximization of equation (12) is subject not only to equation (2) – equation (4) and equation (13), but also to the incentive- compatibility constraints represented by equations (5) and (7). Let wi denote the Lagrange multiplier of equation (5) and mi that of equation (7). The �rst- order conditions are ð ð ! À 0 �2 À v0i z0i þ ðbui2 þ lui Þfeii dxi þ ji þ wi v0i z00 i þ v 00 i z i þb u f i i 2 ei ei dxi ð17Þ xi xi þ mi v00 0 i zi ¼ 0 for ei, h À 0 �2 i  0 à ð18Þ À v0i z0i þ bu0i1 w1 þ ci þ wi v0i z00 i þ v00 i zi þ mi v00 00 i zi þ bui1 w1 ¼ 0 for Li,  00 à ð19Þ À v0i þ bu0i1 þ wi v00 0 00 i zi þ mi vi þ bui1 ¼ 0 for mi, ð20Þ bu0i1 À l þ mi bu00 i1 ¼ 0 for gi and, at each possible realization of xi, À � ð21Þ À bu0i2 À l f i À wi bu0i2 feii ¼ 0 for ui. l Using equation (7), equation (20) can be rewritten as 1 þ mi ri ¼ v0 , where i u00 ri ; Àui01 is the Arrow-Pratt measure of absolute risk aversion. As long as ri is i1 Cigno 71 nonincreasing in i’s income,21 and given that v0i is decreasing in yi, ð22Þ gi ¼ gðyi Þ; g0 , 0: f fei Condition (21) may be similarly rewritten as 1 þ wi feii ¼ bu l 0 . Because fi is i2 increasing in xi, and u0i2 in ui, ð23Þ ui ¼ uðxi Þ; u0 , 0: Because there is nothing to prevent gi from falling below zero for some yi, gi can be interpreted as the difference between an education grant equal to p 22 and an earmarked tax increasing in parent income. Similarly, because there is nothing to stop ui from being negative for some xi, ui can be interpreted as the difference between another earmarked tax, equal to p, and another education grant, this time increasing in the school result. Having established that g(.) and u (.) are decreasing functions, it is clear that the policy redistributes from the rich to the poor and insures parents and chil- dren against the risk of a low return to effective education. Comparing equation (20) with equation (15), and equation (21) with equation (16), however, it is also clear that the policy does not go as far as the �rst-best policy does. The reason is that redistribution has an ef�ciency cost, because the government cannot use personalized lump-sum transfers as under the �rst-best policy. What happens to (ei, Li, mi)? Compare equation (17) with equations (4) and (5) to see that ei is lower than under the �rst-best policy. In particular, because ( yi þ gi) is not the same for all i, ei increases with yi, as (albeit more slowly than) under laissez-faire. Because gi can be negative, it cannot be ruled out that ei will be lower than � under laissez-faire for some i. Because the govern- i ment can borrow against xi ui f dxi , however, gi will be negative only if yi is very high. Because the government is also insuring parents against the risk of a low return to effective education, it is thus unlikely that ei will be lower than under laissez-faire for any i. Comparing equation (18) with equation (6) also shows that Li will be no higher than under either the �rst-best policy or laissez- faire. The intuition is that if w1 is high enough for the ef�cient Li to be posi- tive, imposing a ceiling on overt child labor will distort the allocation of i’s total working time between overt and covert labor and thus make work as a whole less attractive than education. Comparing equation (19) with equation (7) shows that mi will be lower than under either the �rst-best policy or laissez-faire. Proposition 3. Under the second-best policy school enrollment is compulsory, and all school-age children, with the possible but unlikely exception of those from very rich families, receive more effective education than under laissez-fare; the government uses a net subsidy decreasing in 21. For evidence, see Johnson (1987). 22. Recall that subsistence consumption is normalized to zero. 72 THE WORLD BANK ECONOMIC REVIEW parent income, and a net tax decreasing in the individual skill premium, to redistribute and insure, but stops short of perfect equity, full insurance, and perfect consumption smoothing; if the ef�cient level of overt child labor is positive, the government sets a limit, lower than the ef�- cient level, on the amount of paid work a child can legally do. The implications of relaxing some of the assumptions made so far can be intuited without formal analysis. Suppose that the returns to education invest- ment have an aggregate as well as an idiosyncratic component. Because aggre- gate risks cannot be insured against by redistributing within cohorts, the government must use its ability to borrow and lend on the international credit market to redistribute not only within, but also between, cohorts. A similar argument applies to parent incomes. If the shocks to parent income are purely idiosyncratic, the policy prescription remains qualitatively the same, because redistributing from rich parents to poor parents will insure families against the risk of a downturn in that income. The prescription also remains the same when the shocks have an area component, because the policy redistributes not only within, but also between, areas. If the shocks have a countrywide com- ponent, however, the government must use its ability to borrow and lend on the international credit market to redistribute not only within, but also between, cohorts (as in the case where the aggregate shocks concern the return to education investment). I V. P O L I C Y P R A C T I C E IN THE LIGHT OF THEORETICAL FINDINGS Under laissez-faire, if credit and insurance markets are imperfect or contracts between parents and young children are unenforceable, effective education is inef�ciently low. If parent income is below a certain threshold, the child will not enroll at school. Above that threshold, the child will enroll, but children from poor families will receive less effective education than children from rich families. This prediction is consistent with evidence in Ram and Schultz (1979), Jacoby (1994), Jacoby and Skou�as (1997), Belzil and Hansen (2002), and Fuwa and others (2009) that inability to borrow and insure reduces edu- cation investment. It is consistent also with evidence, surveyed in Fasih (2008), that the return to measurable education inputs such as school enrollment or attendance is positive and particularly large in low- to middle-income countries but lower for poor children than for rich children. Because the amount of effec- tive education that children receive in a year of school enrollment or day of school attendance is lower if they come from a family with low parent income than if they come from one with high parent income, the return to enrollment or attendance will in fact understate the return to effective education of poor children relative to that of rich children. This explanation does not conflict with other possible explanations, such as that poor children have access to poor quality schools only. The optimal (�rst- or second-best) policy relaxes the credit constraint on education investment by giving parents an advance on the expected return and Cigno 73 provides insurance against the risk of a low return by redistributing from lucky to unlucky school leavers. Because it redistributes from rich parents to poor parents, it will also reduce inequality and, if parent income is uncertain, provide insurance against the risk of a downturn in that income. The �rst-best policy uses personalized lump-sum transfers to achieve perfect equity, full insurance, and perfect consumption smoothing. Because children are ex ante identical, all parents enroll their children at school and give each child the same ef�cient amount of effective education. The second-best policy also redis- tributes and insures. Because it cannot use personalized lump-sum transfers, however, it stops short of perfect equity, full insurance, and perfect consumption-smoothing. It also raises effective education above the laissez- faire level for most children, but not to the ef�cient level. Although the worst forms of child labor are outside the scope of this analysis, a policy that encourages effective education will discourage all forms of covert child labor, including the worst ones. Under the second-best policy school enrollment is compulsory (whereas under the �rst-best policy, it does not need to be compulsory because it is in the interest of all parents to send their children to school). If the child wage rate is high enough for the ef�cient level of overt child labor to be positive, the government will also impose a legal ceiling, lower than the ef�cient level, on such labor. This distorts the mix of overt and covert child labor and thus makes child labor as a whole less attractive relative to education. Furthermore, the government makes a transfer decreasing in parent income to every school child and exacts a transfer decreasing in the individual education result from every school leaver. The �rst transfer can be interpreted as the difference between a need-based education grant, covering maintenance and tuition, and an earmarked tax increasing in parent income. Similarly, the second transfer can be interpreted as the difference between an earmarked tax, equal to (the capitalized value of ) the need-based education grant, and a merit-based edu- cation grant increasing in the school result. The �rst transfer may be negative for school-age children from families with high parent income. If the expected return to effective education is high enough, however, this transfer may be positive for anyone. The second transfer may be negative for school leavers with a high education result. In the model the �rst transfer occurs at the begin- ning of the education process, and the second at the end. In practice, however, the government could deliver the need-based grant and collect the tax on parent income in installments over the education period. Similarly, it could deliver the merit-based grant in installments over the education period, as partial results become available, and collect the tax on school leavers, again in installments, as the individual skill premia gradually unfold. The analysis here is tailored for a poor developing country; it may be inter- esting to compare the results with those of a model tailored for a rich devel- oped economy. Hanushek, Leung, and Yilmaz (2003) use a calibrated general equilibrium model to assess the welfare effects of a range of policy instruments, 74 THE WORLD BANK ECONOMIC REVIEW including need- and merit-based education grants, under the assumptions that child labor is out of the question and that parents are rich enough to be risk neutral (or, equivalently, that there is a well developed insurance market). In such a world, education subsidies generally perform worse than other forms of redistribution, and a merit-based education grant can be justi�ed only in the presence of an education externality (while the current study �nds that it is optimal anyway). These differences highlight the importance of the stage of development in designing education policy. In most countries primary school enrollment is compulsory, and work at a very young age is forbidden (though enforcement is not always effective). In poor developing countries education is subsidized only through the price of school enrollment, if at all. Is that better than nothing? That question is best answered in two steps. First, starting from laissez-faire, would compulsory school enrollment raise social welfare? The answer is no, because it would oblige all parents, including those who would not let their children study anyway, to bear the tuition cost. Forbidding child labor instead, or on top, of that would also reduce welfare, because the ban would apply only to overt child labor, thereby distorting time allocation. Second, given compulsory enrollment, and with or without a ban on child labor, would a price subsidy raise welfare? If the subsidy is �nanced by a poll tax, the policy will affect welfare to the extent that the number of children varies across families. If all families had the same number of children, the policy would have no effect, because the parents would be taking a lump-sum subsidy with one hand and giving it back with the other. If the number varies exogenously across families, the policy will affect welfare, but the effect will be positive only if the marginal utility of income is higher in families with many children than in families with few children (in other words, if income and fertility are not positively corre- lated). If fertility is endogenous, the policy could actually reduce welfare, because it will trigger a substitution of quantity for quality of children; see Cigno (1986). If the subsidy is �nanced by a tax increasing in parent income, the net transfer schedule will look almost like g(.), though not exactly the same because a price subsidy cannot be larger than the price, and may thus be insuf- �cient for a second-best policy. In any case, a second-best policy would also require some form of insurance against the risk of a low return to effective edu- cation. In other words, the u(.) schedule is needed too. Finally, considerable attention has been given to schemes that effectively pay children to attend school, such as Mexico’s Programa de Educacio ´ n, Salud y Alimentacio ´ n.23 Skou�as and Parker (2001) �nd evidence that such schemes encourage school attendance and discourage child labor. If the nonobservable determinants of effective education were positively correlated with the observa- ble ones, one could be con�dent that offering transfers conditional on observa- ble determinants would encourage nonobservable determinants. But there is 23. For a comprehensive exposition, see Fiszbein, Schady, and Ferreira (2009). Cigno 75 evidence that the correlation is actually negative. Ravallion and Woodon (2000) report that the increase in school attendance elicited by an enrollment subsidy is four to eight times larger than the corresponding reduction in child labor. Consistent with this �nding, Fuwa and others (2009) estimate that a credit constraint reduces average school attendance by 60 percent, but raises child labor by double that percentage. Why? The answer given by the current study’s model is that paying a child to attend school triggers a substitution away from not only labor but also homework and rest. This has an ef�ciency cost and may actually reduce effective education time. The model further shows that paying a child to attend school will crowd out parent transfers ( parents will give their children less money or take more money away from them). In the light of these theoretical results and empirical �ndings, cash transfers should be made conditional not only on the child attending school, but also on the child doing no more than a certain amount of overt labor (less than the ef�- cient amount, if that is positive). Furthermore, cash transfers to children in the scheme should be increasing in education results and decreasing in parent income. Such corrections would improve the scheme but would not be enough for a second-best policy, because the parents would still get no insurance against the risk of a low-skill premium, let alone against the risk of a negative shock to their own income. Of course, the distance from the second best will be even greater if parent income is private information or if overt child labor is not overt after all, because it will then be impossible to make cash transfers conditional on either or both of these variables. The optimal taxation approach adopted in this study gives new insights into how best to discourage labor at a very young age and provide all children an effective education, in poor developing countries. One such insight is that sub- sidizing school attendance without rewarding school attainment at the same time is not optimal and may even be counterproductive. Another is that, in a second-best perspective, it is optimal to force overt child labor below its ef�- cient level, if this is positive, despite the fact that (indeed, precisely because) covert child labor cannot be similarly regulated. The analysis lends support to the notion that school enrollment should be made compulsory but not necess- arily to the notion that overt child labor should be banned. These results should help in designing a universal education system and in improving partial forms of public intervention such as conditional cash transfers. REFERENCES Alderman, H., P. F. Orazem, and E. M. Paterno. 2001. “School Quality, School Cost, and the Public/ Private School Choices of Low-Income Households in Pakistan.� Journal of Human Resources 36: 304–26. Baland, J. M., and J. A. Robinson. 2000. “Is Child Labor Inef�cient?� Journal of Political Economy 108: 663–79. 76 THE WORLD BANK ECONOMIC REVIEW Balestrino, A., A. Cigno, and A. Pettini. 2002. “Endogenous Fertility and the Design of Family Taxation.� International Tax and Public Finance 9: 175– 93. Beegle, K., R. H. Dehejia, and R. Gatti. 2006. “Child Labor and Agricultural Shocks.� Journal of Development Economics 81: 80 –96. Belzil, C., and J. Hansen (2002). “Unobserved Ability and the Return to Schooling.� Econometrica 70: 2075– 91. Besley, T. 1995. “Non-Market Institutions for Credit and Risk-Sharing in Low-Income Countries.� Journal of Economic Perspectives 9: 115 –27. Card, D. 1999. “The Causal Effect of Education on Earnings.� In O. Ashenfelter, and D. Card, eds., Handbook of Labor Economics. Vol. 3. Amsterdam: North Holland. Cigno, A. 1986. “Fertility and the Tax-Bene�t System: A Reconsideration of the Theory of Family Taxation.� Economic Journal 96: 1035–51. ———. 1993. “Intergenerational Transfers Without Altruism: Family, Market and State.� European Journal of Political Economy 9: 505– 18. ———. 2001. “Comparative Advantage, Observability, and the Optimal Tax Treatment of Families With Children.� International Tax and Public Finance 8: 455–70. ———. 2006. “A Constitutional Theory of the Family.� Journal of Population Economics 19: 259– 83. ———. 2011. “Agency in Family Policy: A Survey.� CESifo Economic Studies 57: 305 –31. Cigno, A., and A. Luporini. 2009. “Scholarships or Student Loans? Subsidizing Higher Education in the Presence of Moral Hazard.� Journal of Public Economic Theory 11: 55 –87. Cigno, A., and F. C. Rosati. 2005. The Economics of Child Labour. New York and Oxford: Oxford University Press. Cremer, H., A. Dellis, and P. Pestieau. 2003. “Family Size and Optimal Income Taxation.� Journal of Population Economics 16: 37 –54. Dessy, S. E., and S. Pallage. 2005. “A Theory of the Worst Forms of Child Labor.� Journal of Development Economics 115: 68– 87. Dessy, S. E., and T. Rambeloma. 2010. “Is Temporary Emigration of Unskilled Workers a Solution to the Child Labor Problem?� Cahiers de Recherche 1037. Centre Interuniversitaire sur le Risque, les Politiques E´ conomiques et l’Emploi, Montre ´ al, Canada. Epstein, G. S., and N. Kahana. 2008. “Child Labor and Temporary Emigration.� Economics Letters 99: 545–48. Fasih, T. 2008. Linking Education Policy to Labor Market Outcomes. Washington, DC: World Bank. Fiszbein, A., N. Schady, and F. H. G. Ferreira. 2009. Conditional Cash Transfers: Reducing Present and Future Poverty. Washington, DC: World Bank. Fitzsimons, E. 2007. “The Effects of Risk on Education in Indonesia.� Economic Development and Cultural Change 56: 1 – 25. Fuwa, N., S. Ito, K. Kubo, T. Kurosaki, and Y. Sawada. 2009. “How Does Credit Access Affect Children’s Time Allocation? Evidence from Rural India.� Discussion Paper 183. Japan External Trade Organization, Institute of Developing Economies, Chiba, Japan. Hanson, G. H., and C. Woodruff. 2003. “Emigration and Education Attainment.� University of California– San Diego. Hanushek, E. A., C. K. Y. Leung, and K. Y. I. Yilmaz. 2003. “Redistribution through Education and Other Transfer Mechanisms.� Journal of Monetary Economics 50: 1719–50. Jacoby, H. 1994. “Borrowing Constraints and Progress Through School: Evidence from Peru.� Review of Economics and Statistics LXXVI: 151–60. Jacoby, H., and E. Skou�as. 1997. “Risk, Financial Markets and Human Capital in Developing Countries.� Review of Economic Studies 64: 311– 35. Johnson, W. R. 1987. “Income Redistribution as Human Capital Insurance.� Journal of Human Resources 22: 269– 80. Cigno 77 Kahneman, D. 2003. “Maps of Bounded Rationality: Psychology for Behavioral Economics.� American Economic Review 93: 1449– 75. ¨ , and M. Tuomala. 2008. “Moral Hazard, Income Taxation and Prospect Kanbur, R., J. Pirttila Theory.� Scandinavian Journal of Economics 110: 321 –37. Kodde, A. D. 1986. “Uncertainty and the Demand for Education.� Review of Economics and Statistics 68: 460– 67. Levhari, D., and Y. Weiss. 1974. “The Effect of Risk on the Investment in Human Capital.� American Economic Review 64: 950– 63. Loury, G. 1981. “Intergenerational Transfers and the Distribution of Earnings.� Econometrica 49: 843–67. Oreopoulos, P. 2006. “Average Treatment Effects of Education when Compulsory School Laws Really Matter.� American Economic Review 96: 152– 75. Pouliot, W. 2006. “Introducing Uncertainty in Baland and Robinson’s Model of Child Labour.� Journal of Development Economics 79: 264– 72. Ram, R., and T. Schultz. 1979. “Life Span, Health, Savings and Productivity.� Economic Development and Cultural Change 13: 399–421. Ranjan, P. 2001. “Credit Constraints and the Phenomenon of Child Labor.� Journal of Development Economics 64: 81 –102. Ravallion, M., and Q. Woodon. 2000. “Does Child Labour Displace Schooling? Evidence on Behavioural Responses to an Enrolment Subsidy.� Economic Journal 110: c158–c175. Razin, A. 1976. “Lifetime Uncertainty, Human Capital and Physical Capital.� Economic Enquiry 14: 439–48. Skou�as, E., and S. Parker. 2001. “Conditional Cash Transfers and the Impact on Child Work and Schooling. Evidence from Mexico’s Progresa Program.� Economia 2: 45–96. Townsend, R. 1994. “Risk and Insurance in Village India.� Econometrica 62: 539– 92. Resource Windfalls and Emerging Market Sovereign Bond Spreads: The Role of Political Institutions ¨ ckner* Rabah Arezki and Markus Bru We examine the effect that revenue windfalls from international commodity price booms have on sovereign bond spreads using panel data for 38 emerging market econ- omies during the period 1997-2007. Our main �nding is that commodity price booms lead to a signi�cant reduction in the sovereign bond spread in democracies, but to a signi�cant increase in the spread in autocracies. To explain our �nding we show that, consistent with the political economy literature on the resource curse, revenue wind- falls from international commodity price booms signi�cantly increased real per capita GDP growth in democracies, while in autocracies GDP per capita growth decreased. JEL codes: C33, D73, D74, D72, H21. I. INTRODUCTION Some researchers have argued that international commodity price booms may spawn an over-accumulation of external debt in commodity exporting countries that increases the risk of external debt default (e.g. Krueger, 1987; Berg and Sachs, 1988).1 We examine this hypothesis empirically by analyzing how the spread on sovereign bonds reacted in these countries to the booms and slumps of the export-relevant commodity prices. Changes in the spread on sovereign bonds reflect changes in investors’ beliefs of the risk that a country * International Monetary Fund (Arezki, corresponding author) and University of Adelaide (Bruckner). Contact e-mails: rarezki@imf.org; markus.bruckner@adelaide.edu.au. We thank three anonymous referees, the editor Elisabeth Sadoulet, and members of the editorial board for helpful comments and suggestions. We are grateful to Amine Mati for providing us with his dataset on sovereign bond spreads and to Daniel Lederman for providing us with his dataset on export diversi�cation. The views in this paper are those of the authors alone and do not necessarily represent those of the IMF or IMF policy. All remaining errors are our own. Bru ¨ ckner gratefully acknowledges the �nancial support of the Spanish Ministry of Science and Technology provided by CICYTECO2008- 04997. 1. The recent concern that Dubai may default on its external debt is an example par excellence that higher commodity prices may be associated with a higher risk of external debt default. Further examples are, among others, Russia and Nigeria. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 78– 99 doi:10.1093/wber/lhr015 Advance Access Publication May 18, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 78 ¨ ckner Arezki and Bru 79 defaults on its external debt. An increase in the spread on sovereign bonds is in turn a cost for the bond issuing country that may trigger in a self-ful�lling way the default on its external debt. Both for investors and policy makers, it is therefore important to have knowledge about how international commodity price shocks, which induce large upturns and downturns in foreign currency revenues in emerging market economies, affect the spread on sovereign bonds. We �nd that increases in international commodity prices for exported com- modity goods are associated with a signi�cant reduction in sovereign bond spreads on average. However, the reduction in the spread on sovereign bonds is particularly large in countries with sound democratic institutions and strong political checks and balances. In autocratic regimes and countries where the political rule is characterized by weak checks and balances, windfalls from international commodity prices lead to a signi�cant increase in the spread on sovereign bonds. The heterogeneous response of sovereign bond spreads to international com- modity price shocks sheds new light on the resource curse literature, that has argued for the importance of political institutions in determining whether windfalls from natural resources are a curse or a blessing for the economic development of resource exporting countries (e.g. Mehlum et al., 2006; Robinson et al., 2006).2 We provide further evidence in this direction by showing that, consistent with the political economy model developed in Mehlum et al. (2006), international commodity price booms signi�cantly increased real per capita GDP growth in countries with sound democratic insti- tutions. In countries with autocratic institutions, revenue windfalls from inter- national commodity price booms led to a signi�cant decrease in output growth. Hence, while our empirical results are consistent with general equili- brium models that predict a countercyclical relationship between sovereign bond spreads and the business cycle in emerging market economies (e.g. Arellano, 2008), our results highlight the importance of political economy factors in shaping the relationship between commodity price shocks and sover- eign bond spreads in these countries. The remainder of our paper is organized as follows. Section II describes the data. Section III discusses the estimation strategy. Section IV presents the main results. Section V concludes. I I . D ATA COMMODITY REVENUE WINDFALLS. We construct a country-speci�c international commodity export price index that captures revenue windfalls from 2. See also Van der Ploeg (2010) for a review and overview of the resource curse literature. 80 THE WORLD BANK ECONOMIC REVIEW international commodity prices as: Y ComPIi;t ¼ ComPricec;tui;c c [C where ComPricec,t is the international price of commodity c in year t, and ui,c is the average (time-invariant) value of exports of commodity c in the GDP of country i.3 We obtain data on annual international commodity prices from UNCTAD Commodity Statistics and our data on the value of commodity exports are from the NBER-United Nations Trade Database. The commodities included in our index are aluminum, beef, coffee, cocoa, copper, cotton, gold, iron, maize, oil, rice, rubber, sugar, tea, tobacco, wheat, and wood. In case there were multiple prices listed for the same commodity we used a simple average of all the relevant prices. We note that even though some of the countries in our sample are net resource importers (in sum, across all commodities) our commodity export price index captures that there may still be some commodities for which the country is an exporter. For example, according to Lederman and Maloney (2008) Egypt is a net natural resource importer. However, Egypt also exports a signi�cant amount of crude oil. When the international price of oil increases Egypt experi- ences a positive revenue windfall, and this is captured by our export price index. On the other hand, when the international prices of other commodities increase Egypt experiences a negative terms of trade shock but not necessarily a negative revenue shock (which depends among other things on the structure of ad valorem import duties). We therefore follow the resource curse literature (e.g. Sachs and Warner, 1995, 2001) and focus on a gross export price index as our measure for resource windfalls. As a robustness check we will present estimates that are restricted to the sample of countries that are net natural resource exporters. SOVEREIGN BOND SPREADS. Our data on the spread on sovereign bonds are from the Emerging Markets Bond Index Global (EMBI Global). The bond spreads are measured against a comparable US government bond and are period averages for the whole year. POLITICAL INSTITUTIONS. Our two main measures of political institutions are the average (time-invariant) Polity2 score from the Polity IV database (Marshall and Jaggers, 2009) and the average (time-invariant) checks and balance score from the Database of Political Institutions (Beck et al., 2001). The Polity2 score is based on the constraints placed on the chief executive, the competitiveness of political participation, and the openness and competitiveness of executive recruitment. The Polity2 score ranges from 2 10 to þ 10, with higher values 3. This functional form of the commodity export price index follows common practice in the literature. See for example Collier and Goderis (2007) and the references cited therein. ¨ ckner Arezki and Bru 81 indicating stronger democratic institutions. The checks and balance score is based on the number of veto players in the political system, their respective party af�liations, and the electoral rules. The checks and balance score ranges between 1 to 6, with higher values indicating stronger checks and balances. Following Persson and Tabellini (2003, 2006) and the Polity IV project we also construct an autocracy indicator variable that takes on the value of unity in countries with negative (average) Polity2 scores. The main purpose of this auto- cracy indicator variable is to facilitate the interpretation of the results from the regression analysis. Note that we use countries’ average polity and checks and balance scores because we want to capture long-run and thus more fundamental differences in countries’ political institutions. Countries’ political institutions are also highly persistent as about three-fourths of the countries in our sample did not experience changes in their political institutions score. OTHER CONTROL VARIABLES. Data on real per capita GDP are from the Penn World Table, version 6.3 (Heston et al., 2009). Data on corruption are from Political Risk Service (2010). Data on ethnic fractionalization are from Alesina et al. (2003). Data on the Her�ndahl index of export diversi�cation are from Lederman and Xu (2010). Data on the Gini coef�cient are from the World Development Indicators (2010). Data on British colonial origin, French colonial origin, and historical settler mortality are from Acemoglu et al. (2001). Descriptive statistics of these variables are provided in Data Appendix Table 1. A list of countries included in the sample is provided in Data Appendix Table 2. I I I . E S T I M AT I O N S T R AT E GY To examine the effects that revenue windfalls from international commodity price booms have on sovereign bond spreads, we estimate the following econo- metric model: DlogðSpreadi ; tÞ ¼ ai þ bt þ hDlogðComPIi;t Þ þ ui;t where ai are country �xed effects and bt are year �xed effects. ui,t is an error term that is clustered at the country level. As a baseline regression, we estimate the average marginal effect h that commodity price booms have on sovereign bond spreads. We then examine how this marginal effect varies as a function of countries’ political institutions by estimating: DlogðSpreadi;t Þ ¼ ai þ bt þ cDlogðComPIi;t Þ þ dDlogðComPIi;t Þ Ã Poli þ ei;t where Poli is a measure of cross-country differences in political institutions. In order for the estimate on the parameter c to reflect the average marginal effect we compute Poli for the Polity2 score as the Polity2 score of country i minus the Polity2 sample average. Formally: Poli ¼ Polity2i - Avg.(Polity2). We do the same for the checks and balance score. This rescaling does not affect the 82 THE WORLD BANK ECONOMIC REVIEW parameter estimate d but it is useful for interpretation purposes as it ensures that the parameter estimate c reflects the average marginal effect (i.e. the effect for the “average� country). Note that our measures of political institutions Poli are time-invariant and therefore we do not need to control for them in the �xed effects regression (the reason is that the direct effect of these variables on the sovereign bond spread is already accounted for by the country �xed effects ai). We estimate both static and dynamic panel data models. For the dynamic panel data model we report system-GMM estimates (Blundell and Bond, 1998) as the presence of country �xed effects leads the �xed effects estimator to produce inconsistent estimates.4 We address the important issue of political institutions being correlated with other cross-sectional variables that could possibly affect the relationship between commodity price booms and sovereign bond spreads by including additional interaction terms in the regression. In particular, we include in all regressions an additional interaction term between DComPI and cross-country differences in GDP per capita. In addition, we use instrumental variables tech- niques to further address endogeneity biases. In particular, we build on the seminal work of Acemoglu et al. (2001) and instrument the political insti- tutions interaction term Pol*DComPI with the interaction between DComPI and indicator variables for colonial origin and historical settler mortality. We test the validity of these instrumental variables using the Hansen test. I V. M A I N R E S U L T S Table 1, column (1) presents our estimates of the average marginal effect that resource windfalls from international commodity price booms have on sover- eign bond spreads in the largest possible sample of 38 emerging market econ- omies during the period 1997-2007. The main �nding is that commodity windfalls lead on average to a signi�cant reduction in sovereign bond spreads. Panel A presents panel data estimates that control for country �xed effects and Panel B presents panel data estimates that control in addition to the country �xed effects for year �xed effects. The panel data estimates reported in column (1) imply that an increase in the commodity export price index of size 1 stan- dard deviation would signi�cantly reduce the spread on sovereign bonds on average by over 0.1 standard deviations. Column (2) of Table 1 shows that the marginal effect of international com- modity price booms on the spread on sovereign bonds signi�cantly varies across countries as a function of cross-country differences in political 4. In the system-GMM estimation we use the �rst and second lags as instruments for the lagged dependent variable to reduce the concern that too many moment conditions are used (for further discussion on this issue see e.g. Roodman, 2009). We note that the dynamic panel data bias associated with the �xed effects estimator is bounded of order T 21, where T is the time-series dimension of the panel (see Nickell, 1981). For comparison purposes we also report estimates from the �xed effects estimator. ¨ ckner Arezki and Bru 83 T A B L E 1 . Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds (Static Panel Regression) DSpread Panel A: Controlling for Country Fixed Effects (1) (2) (3) (4) LS LS LS LS DComPI 2 10.950*** 2 7.417** 2 29.694*** 2 8.072*** ( 2 3.03) ( 2 2.26) ( 2 4.27) ( 2 2.83) DComPI* 2 2.610*** Avg. Polity2 Score ( 2 2.81) DComPI* 55.815*** Autocracy Indicator (4.15) DComPI* 2 16.939*** Avg. Checks & Balance Score ( 2 3.35) DComPI* 0.001** 0.004*** 0.002*** Avg. GDP Per Capita (1.98) (4.21) ( 2 2.63) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects No No No No Observations 291 291 291 291 Panel B: Controlling for Country and Year Fixed Effects (1) (2) (3) (4) LS LS LS LS DComPI 2 6.127* 2 1.644 2 20.727*** 2 3.108 ( 2 1.72) ( 2 0.37) ( 2 3.46) ( 2 0.74) DComPI* 2 2.121** Avg. Polity2 Score ( 2 2.33) DComPI* 45.676*** Autocracy Indicator (3.57) DComPI* 2 11.420** Avg. Checks & Balance Score ( 2 2.17) DComPI* 0.002** 0.004*** 0.002** Avg. GDP Per Capita (2.13) (3.88) (2.09) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations 291 291 291 291 Note: The method of estimation is least squares. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The dependent variable is the log- change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this vari- able is constructed). The cross-section (average time-series) dimension of the panel is 38 (7.7). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. institutions. The estimated interaction effect between revenue windfalls from international commodity price booms and the Polity2 score is negative and statistically signi�cant at the 5% level. The point estimate on the interaction term implies that at the sample maximum Polity2 score (democracies), an increase in the commodity export price index of size 1 standard deviation 84 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds Note: The left-hand side �gure shows the relationship between changes in countries’ commodity export price index and the spread on their sovereign bonds for countries that had on average a strictly positive Polity2 score. The right-hand side �gure shows the relationship between changes in countries’ commodity export price index and the spread on their sovereign bonds for countries that had on average a negative Polity2 score. would signi�cantly reduce the spread on sovereign bonds by over 0.3 standard deviations. On the other hand, at the sample minimum Polity2 score (autocra- cies), a shock of similar magnitude would be associated with a signi�cant increase in the spread on sovereign bonds by 0.2 standard deviations. Column (3) of Table 1 shows that we obtain similar heterogeneity in the marginal effect of international commodity price booms on sovereign bond spreads when we discretize the Polity2 score into an autocracy indicator vari- able that is unity for negative Polity2 scores and zero otherwise. The signi�cant positive coef�cient on the autocracy interaction term implies that in autocracies revenue windfalls from commodity price booms signi�cantly increased the spread on sovereign bonds, while in democracies sovereign bond spreads sig- ni�cantly decreased. Figure 1 illustrates this nonlinear relationship graphically. We show in column (4) of Table 1 as a robustness check on our measure of political institutions, that windfalls from international commodity price booms signi�cantly decreased sovereign bond spreads in countries with strong checks and balances, while in countries with weak checks and balances the sovereign bond spreads signi�cantly increased.5 Table 2 shows that our results are robust to controlling for lagged changes in the sovereign bond spread. Columns (1) to (3) present the least squares 5. We document in Appendix Table 1 that the results in Table 1 are robust to outliers. In particular, we report in columns (1)-(3) of Appendix Table 1 median (quantile) estimates, and in columns (4)-(6) least-squares estimates that exclude observations which fall in the top/bottom 1 percentile of the distribution of the change in the commodity export price index. T A B L E 2 . Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds (Dynamic Panel Regression) DSpread (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 4.123 2 16.369*** 1.984 2 0.032 2 14.994*** 2 2.224 (0.85) ( 2 2.76) (0.41) ( 2 0.01) ( 2 2.78) ( 2 0.56) DComPI* 2 2.305** 2 1.685** Avg. Polity2 Score ( 2 2.34) ( 2 2.43) DComPI* 49.324*** 33.101*** Autocracy Indicator (3.37) (2.68) DComPI* 2 10.407** 2 8.086** Avg. Checks & Balance Score ( 2 2.02) ( 2 2.17) DComPI* 0.003*** 0.005*** 0.002** 0.002*** 0.003*** 0.001* Avg. GDP Per Capita (2.87) (4.00) (2.56) (2.89) (1.91) (2.53) L.DSpread 0.183*** 0.182*** 0.180*** 0.241*** 0.231*** 0.232*** (3.73) (3.65) (3.58) (5.06) (5.22) (4.84) Hansen J, p-value . . . 0.232 0.220 0.259 AR(2) test, p-value . . . 0.125 0.151 0.134 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 253 253 253 253 253 253 Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The Arezki and Bru dependent variable is the log-change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the ¨ ckner panel is 37 (6.8). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. 85 86 THE WORLD BANK ECONOMIC REVIEW estimates and columns (4) to (6) present the system-GMM estimates. The dynamic panel data estimates reveal a signi�cant positive autocorrelation in the log-change of the sovereign bond spreads. Importantly, they show that the interaction between changes in the commodity export price index and political institutions remains statistically signi�cant at the 5% level when we allow for dynamics in the dependent variable. So far we only controlled in our regressions for an interaction term between changes in the commodity export price index and cross-country differences in GDP per capita. The GDP per capita interaction control is important because there exists a large literature that has argued for a positive effect of cross- country per capita income differences on political institutions (see for example Barro, 1999, or Przeworski et al., 2000). To demonstrate that the interaction between political institutions and commodity price windfalls is robust to additional interaction controls we report in Table 3 estimates when controlling for an interaction between changes in the commodity export price index and ethnic fractionalization, an interaction between changes in the commodity export price index and the Gini coef�cient, an interaction between changes in the commodity export price index and a Her�ndahl index of export diversi�ca- tion, and an interaction between changes in the commodity export price index and an indicator variable that is unity if the country is a net natural resource importer. Some of these additional interaction controls are indeed statistically signi�cant. But nevertheless, the inclusion of these additional interaction con- trols on the right-hand side of the estimating equation continues to produce a signi�cant interaction effect between commodity price booms and political institutions. Table 4 shows that we obtain similar results to our baseline estimates if we restrict the sample to the natural resource net-exporting countries. The natural resource net-exporting countries are strongly affected by the booms and slumps in the international commodity prices. It is thus reassuring from the standpoint of identi�cation that in this restricted sample our results continue to hold. We can go even further and examine the relationship between commodity price windfalls, political institutions and sovereign bond spreads using instru- mental variables techniques that correct for possible endogeneity bias of the estimated interaction effect. Building on the seminal work by Acemoglu et al. (2001), we use historical settler mortality data and indicator variables of countries’ colonial origin as instrumental variables for political institutions. Table 5 reports our two-stage least squares estimates where the political insti- tutions interaction term is instrumented by the interaction between changes in the commodity export price index and the Acemoglu et al. instruments for institutions. The main result is that the political institutions interaction con- tinues to be signi�cant in the instrumental variables regression. Also, with the exception of the autocracy interaction term the Hausman test does not indicate a signi�cant difference between the least squares and instrumental variables estimates. We also note that the quality of the instrumental variables is good as T A B L E 3 . Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds (Robustness to Additional Interaction Control Variables) DSpread (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 2 16.203** 2 24.292*** 2 19.010** 2 17.839* 2 24.920** 2 20.853** ( 2 2.06) ( 2 2.70) ( 2 2.00) ( 2 1.92) ( 2 2.24) ( 2 2.16) DComPI* 2 2.572*** 2 2.305** Avg. Polity2 Score ( 2 3.74) ( 2 2.56) DComPI* 31.557*** 27.595** Autocracy Indicator (2.97) (2.35) DComPI* 2 8.884* 2 9.517** Avg. Checks & Balance Score ( 2 1.87) ( 2 2.10) DComPI* 0.007*** 0.006*** 0.006*** 0.006*** 0.006*** 0.005*** Avg. GDP Per Capita (5.95) (5.81) (4.67) (5.11) (4.82) (4.48) DComPI* 12.794 9.712 22.785 2 10.773 7.279 20.443 Ethnic Fractionalization (0.86) (0.59) (1.25) (0.63) (0.43) (1.02) DComPI* 2 1.755*** 2 1.479** 2 2.264*** 2 1.497*** 2 1.280** 2 1.885*** Avg. Gini Coef�cient ( 2 3.40) ( 2 2.22) ( 2 4.02) ( 2 2.93) ( 2 2.08) ( 2 4.41) DComPI* 56.488*** 35.402** 42.894*** 50.247*** 32.609** 38.955*** Avg. Export Diversi�cation (3.28) (2.24) (2.65) (3.60) (2.24) (2.87) DComPI* 0.631 2.646 1.482 1.079 2.362 0.286 Nat. Res. Importer Indicator (0.02) (0.08) (0.04) (0.05) (0.11) (0.01) L.DSpread 0.194*** 0.193*** 0.195*** 0.246*** 0.245*** 0.246*** (3.49) (3.56) (3.58) (3.73) (3.79) (3.75) Hansen J, p-value . . . 0.376 0.367 0.377 AR(2) test, p-value . . . 0.192 0.190 0.197 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 247 247 247 247 247 247 Arezki and Bru Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The ¨ ckner dependent variable is the log-change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the panel is 35 (7.1). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. 87 88 T A B L E 4 . Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds (Robustness to Restricting the Sample to Natural Resource Exporting Countries) DSpread (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 10.729* 2 10.365 7.876 5.774 2 13.272** 3.766*** (1.64) ( 2 1.42) (1.23) (1.04) ( 2 2.32) (3.69) DComPI* 2 2.274*** 2 2.039** Avg. Polity2 Score ( 2 3.31) ( 2 2.15) DComPI* 45.230*** 41.569*** Autocracy Indicator (3.02) (3.00) DComPI* 2 9.329 2 10.762** THE WORLD BANK ECONOMIC REVIEW Avg. Checks & Balance Score ( 2 2.07) ( 2 2.23) DComPI* 0.003*** 0.005*** 0.002*** 0.002** 0.004*** 0.002** Avg. GDP Per Capita (3.11) (3.78) (2.83) (2.38) (3.26) (2.11) L.DSpread 0.205*** 0.217*** 0.202*** 0.198*** 0.206*** 0.189*** (3.00) (3.38) (2.90) (3.24) (3.58) (3.00) Hansen J, p-value . . . 0.281 0.359 0.301 AR(2) test, p-value . . . 1.000 1.000 0.999 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 125 125 125 125 125 125 Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The dependent variable is the log-change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the panel is 17 (7.4). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. T A B L E 5 . Commodity Windfalls, Political Institutions, and the Spread on Sovereign Bonds (Robustness to Instrumental Variables Estimation) (1) (2) (3) (4) (5) (6) 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS Panel A: Second Stage (Dependent Variable is DSpread) DComPI 2 8.497 2 23.636*** 2 16.827* 2 10.780* 2 26.276*** 2 20.367** ( 2 1.01) ( 2 2.92) ( 2 1.66) ( 2 1.64) ( 2 3.54) ( 2 2.43) DComPI 2 3.019*** 2 3.141*** * Avg. Polity2 Score ( 2 5.54) ( 2 6.60) DComPI* 46.731*** 48.514*** Autocracy Indicator (4.67) (5.38) DComPI* 2 14.079*** 2 15.197*** Avg. Checks & Balance Score ( 2 3.50) ( 2 4.19) DComPI* 0.006*** 0.006*** 0.005*** 0.007*** 0.008*** 0.007*** Avg. GDP Per Capita (5.80) (6.55) (4.42) (8.71) (9.22) (7.36) DComPI* 2.511 4.035 24.284 25.202* 26.442* 49.703** Ethnic Fractionalization (0.21) (0.32) (1.44) (1.66) (1.68) (2.52) DComPI* 2 0.877 2 0.162 2 1.321 2 1.626** 2 0.877 2 2.049*** Avg. Gini Coef�cient ( 2 1.19) ( 2 0.18) ( 2 1.87) ( 2 2.10) ( 2 0.98) ( 2 2.73) DComPI* 47.443** 18.220 30.884 43.143** 12.781 25.692 Avg. Export Diversi�cation (2.36) (0.80) (1.50) (2.24) (0.58) (1.31) L.DSpread 0.233*** 0.232*** 0.233*** (3.41) (3.41) (3.40) Hansen J, p-value 0.336 0.467 0.319 0.221 0.399 0.218 Hausman test, p-value 0.776 0.028 0.967 0.724 0.083 0.645 Panel B: First Stage (Dependent Variable is DComPI*Polity Variable) DComPI* 2 4.184*** 0.407** 2 0.870*** 2 4.139*** 0.411** 2 0.859*** Log Settler Mortality ( 2 4.94) (2.11) ( 2 8.58) ( 2 4.76) (2.11) ( 2 8.15) Arezki and Bru DComPI* 2 4.081*** 0.078 2 0.570*** 2 4.169*** 0.070 2 0.590*** British Colony ( 2 3.56) (0.30) ( 2 4.30) ( 2 3.52) (0.26) ( 2 4.29) ¨ ckner (Continued ) 89 90 TABLE 5. Continued (Robustness to Instrumental Variables Estimation) DComPI* 2 7.171*** 0.335* 2 1.812*** 2 7.275*** 0.362* 2 1.814*** French Colony ( 2 8.34) (1.77) ( 2 18.25) ( 2 8.63) (1.88) ( 2 18.82) Country Fixed Effects Yes Yes Yes Yes Yes Yes THE WORLD BANK ECONOMIC REVIEW Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 148 148 148 128 128 128 Note: The method of estimation is two-stage least squares. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. Panel A shows the second-stage estimates and Panel B shows the �rst-stage estimates. The dependent variable in Panel A is the log-change in the spread on sovereign bonds. The dependent variable in Panel B, columns (1) and (4) is the interaction between DComPI and countries’ average Polity2 score; in columns (2) and (5) of Panel B the dependent variable is the interaction between DComPI and countries’ autocracy indicator; in columns (3) and (6) of Panel B the dependent variable is the interaction between DComPI and countries’ average checks and balance score. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension in columns (1)-(3) of the panel is 19 (7.8); columns (4)-(6) 18 (7.1). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. ¨ ckner Arezki and Bru 91 the �rst-stage F-statistic easily exceeds the Stock and Yogo (2005) critical values for instruments to be declared weak and the Hansen test does not reject that the instruments are uncorrelated with the second-stage error term. As an intermediate step to explain the heterogeneity in the marginal effect that international commodity price booms have on sovereign bond spreads, we report in Table 6 the effect that international commodity price booms have on countries’ real per capita GDP growth. We �nd that higher international prices for exported commodity goods are associated with a signi�cant increase in real per capita GDP growth in democracies. But in countries with deeply autocratic regimes, windfalls from international commodity prices are associated with a signi�cant decrease in real per capita GDP growth. Taking for example the esti- mates in column (5) of Table 6, a one standard deviation increase in the export price index growth rate was associated with a signi�cant increase in real per capita GDP growth in the democracy sample by about 0.29 standard deviations while in the autocracy sample it was associated with a signi�cant reduction in GDP per capita growth by about 0.16 standard deviations. Similarly, columns (4) and (6) show that the marginal effect of commodity price booms on GDP per capita growth is signi�cantly increasing in countries’ Polity2 and checks and balances scores. So much so, that at sample maximum Polity2 and checks and balances scores a commodity windfall was associated with a signi�cant increase in GDP per capita growth while at sample minimum Polity2 and checks and balances scores a commodity windfall was associated with a signi�- cant decrease in GDP per capita growth. The estimates in Table 6 therefore show that while in countries with strong political institutions a plausibly exogenous windfall from international commodity price booms was associated with a signi�cant increase in GDP per capita growth, in countries with weak political institutions it was associated with a signi�cant decrease. The political economy model developed in Mehlum et al. (2006) can provide an explanation for this heterogeneous response in real per capita GDP growth: in countries with grabber friendly political institutions, revenue wind- falls from international commodity price booms increase rent-seeking activity and lead to a crowding out of production activity. Democratic institutions, in particular, stronger checks and balances constrain politicians in their policy space. Relative to an autocratic regime, politicians are also held more accounta- ble to the public. Hence, in a more democratic regime the expected returns to rent-seeking activities are lower. This in turn means that production activity will remain strong in the democratic regime despite the high rents that are rea- lized in the commodity exporting sector when international commodity prices are booming. In the autocratic regime, on the other hand, where there are rela- tively high gains from specializing in grabbing activities, production activity will be crowded out in the presence of a revenue windfall. Thus, revenue wind- falls from international commodity prices may be associated with lower per capita GDP growth in more autocratic regimes. 92 T A B L E 6 . Commodity Windfalls, Political Institutions, and Economic Growth DGDP (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 0.164 2.890*** 0.610 0.470 2.100*** 0.732 (0.24) (3.61) (0.83) (0.96) (5.66) (1.40) DComPI* 0.375*** 0.219*** Avg. Polity2 Score (2.65) (2.85) DComPI* 2 5.623*** 2 3.328*** Autocracy Indicator ( 2 3.94) ( 2 4.40) DComPI* 1.417** 0.948*** THE WORLD BANK ECONOMIC REVIEW Avg. Checks & Balance Score (2.25) (2.71) DComPI* 2 0.001* 2 0.001*** 2 0.001 2 0.001** 2 0.001*** 2 0.001* Avg. GDP Per Capita ( 2 1.69) ( 2 3.28) ( 2 1.28) ( 2 2.24) ( 2 3.36) ( 2 1.76) L.DGDP 0.020 0.014 0.017 0.172 0.172 0.170 (0.30) (0.21) (0.24) (1.60) (1.60) (1.16) Hansen J, p-value . . . 0.815 0.833 0.822 AR(2) test, p-value . . . 0.887 0.877 0.968 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 253 253 253 253 253 253 Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The dependent variable is the log-change in real GDP per capita. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the panel is 37 (6.8). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. ¨ ckner Arezki and Bru 93 Table 7 provides further evidence on this political economy channel by doc- umenting that political institutions played a key role in shaping the relationship between commodity windfalls and corruption. The signi�cant positive auto- cracy interaction term in the corruption equation implies that in autocracies commodity windfalls are associated with a signi�cant increase in corruption. On the other hand, in democracies and countries with strong checks and bal- ances commodity windfalls did not lead to a signi�cant increase in corruption. This result is consistent with the political economy literature that has high- lighted the importance of political institutions in shaping political leaders’ incentive constraints and thus economic outcomes (e.g. North, 1990; Acemoglu et al., 2001). The growth results in Table 6 are in line with the political economy model developed in Mehlum et al. (2006). However, an open and conceptually inter- esting question is whether beyond their effect on GDP per capita growth com- modity price booms exhibit signi�cant effects on sovereign bond spreads. The business-cycle literature on the link between GDP per capita growth and sover- eign bond spreads has argued for a countercyclical average relationship between economic growth and sovereign bond spreads (see e.g. Neumeyer and Perri, 2005; Aguiar and Gopinath, 2006; or Arellano, 2008). Given this litera- ture which does not emphasize the role of political institutions but instead argues for a countercyclical relationship between economic growth and sover- eign bond spreads in an environment where �nancial markets are incomplete, it is interesting to explore whether beyond their effects on economic growth the interaction between commodity price booms and political institutions still matters for sovereign bond spreads. To explore the above issue Table 8 reports estimates of the effects that com- modity price booms have on sovereign bond spreads when GDP per capita growth is included as a right-hand- side regressor in the sovereign bond spreads estimating equation. Because we condition in this regression on GDP per capita growth the estimates should be interpreted as capturing the effects that com- modity price booms (and the interaction between commodity price booms and political institutions) have on sovereign bond spreads beyond the effects that these variables have on GDP per capita growth. We report in Table 8 both least squares and system-GMM estimation. To address possible reverse effects of changes in the sovereign bond spreads on GDP per capita growth we instru- ment GDP per capita growth with the lagged �rst differences. The main result in Table 8 is that, conditional on GDP per capita growth, the interaction effect between commodity price booms and political institutions are quantitatively smaller, but still statistically signi�cant for the majority of the speci�cations. Hence, while the effect on aggregate output is clearly of �rst-order importance, we �nd that commodity price booms and political institutions exhibit additional effects that go beyond aggregate output. This result highlights the importance of political institutions in shaping the relationship between resource windfalls and the spreads on sovereign bonds; it is also consistent with our 94 T A B L E 7 . Commodity Windfalls, Political Institutions, and Corruption Corruption (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 2 15.920 2 32.482 2 17.700 2 7.077 2 20.706 2 5.011 ( 2 0.93) ( 2 1.57) ( 2 1.02) ( 2 0.38) ( 2 0.87) ( 2 0.26) DComPI* 2 3.593** 2 2.671* Avg. Polity2 Score ( 2 2.35) ( 2 1.88) DComPI* 59.695*** 45.972** Autocracy Indicator (2.94) (2.07) DComPI* 2 16.864** 2 9.179 Avg. Checks & Balance Score ( 2 2.19) ( 2 1.07) THE WORLD BANK ECONOMIC REVIEW DComPI* 0.005** 0.008*** 0.005*** 0.003 0.005* 0.002 Avg. GDP Per Capita (2.32) (2.97) (2.64) (1.55) (1.92) (1.15) L.Corruption 0.439*** 0.437*** 0.441*** 0.515*** 0.512*** 0.518*** (6.04) (6.14) (6.06) (4.35) (4.30) (4.39) Hansen J, p-value . . . 0.833 0.789 0.837 AR(2) test, p-value . . . 0.366 0.440 0.331 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 242 242 242 242 242 242 Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The dependent variable is the corruption score from Political Risk Service. The corruption score is rescaled so that higher values indicate more political cor- ruption. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the panel is 35 (6.9). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. T A B L E 8 . Commodity Windfalls, Political Institutions, and Sovereign Spread (Effect Beyond GDP Per Capita Growth) DSpread (1) (2) (3) (4) (5) (6) LS LS LS GMM GMM GMM DComPI 5.789* 2 5.272 5.390 3.450 2 8.975 2.184 (1.74) ( 2 0.70) (1.61) (1.10) ( 2 1.51) (0.64) DComPI* 2 0.777 2 1.327* Avg. Polity2 Score ( 2 0.96) ( 2 1.67) DComPI* 27.461* 28.059** Autocracy Indicator (1.80) (2.22) DComPI* 2 4.673 2 7.751* Avg. Checks & Balance Score ( 2 1.02) ( 2 1.72) DComPI* 0.002*** 0.004*** 0.002*** 0.001*** 0.003*** 0.002** Avg. GDP Per Capita (3.66) (2.95) (3.26) (2.99) (2.94) (2.48) DGDP 2 4.073*** 2 3.892*** 2 4.089*** 2 3.471*** 2 3.430*** 2 3.509*** ( 2 3.91) ( 2 3.80) ( 2 3.98) ( 2 3.68) ( 2 3.71) ( 2 3.69) L.DSpread 0.148*** 0.148*** 0.146*** 0.209*** 0.205*** 0.201*** (2.82) (2.89) (2.76) (4.56) (4.59) (4.34) Hansen J, p-value (DGDP) . . . 0.199 0.184 0.194 Hansen J, p-value (L.DSpread) . . . 0.197 0.199 0.227 Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Observations 253 253 253 253 253 253 Note: The method of estimation in columns (1)-(3) is least squares; columns (4)-(6) system-GMM (Blundell and Bond, 1998) with two-step Windmeijer (2005) small sample correction. t-values (in brackets) are based on Huber robust standard errors that are clustered at the country level. The Arezki and Bru dependent variable is the log-change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the ¨ ckner panel is 37 (6.8). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. 95 96 THE WORLD BANK ECONOMIC REVIEW �nding that political institutions signi�cantly affect the relationship between resource windfalls and corruption. V. C O N C L U S I O N We investigated in this paper the effects that international commodity price booms have on sovereign bond spreads using panel data for 38 emerging market economies during the period 1997-2007. Our main �nding is that revenue windfalls from international commodity price booms lead to a signi�- cant reduction in sovereign bond spreads in emerging market economies with sound democratic institutions. In countries with more autocratic institutions revenue windfalls lead on the other hand to a signi�cant increase in the sover- eign bond spreads. To explain this heterogeneity in the marginal effect that international com- modity price booms have on sovereign bond spreads, we showed that revenue windfalls from international commodity price booms lead to a signi�cant increase in real per capita GDP growth in countries with sound democratic institutions. In countries with deeply autocratic regimes, revenue windfalls lead to a decrease in real per capita GDP growth. Our empirical results are consist- ent therefore with general equilibrium models that predict a countercyclical relationship between sovereign bond spreads and the business cycle in debtor countries (e.g. Arellano, 2008). However, our empirical results also highlight the importance of political economy factors in shaping the relationship between commodity price booms and sovereign bond spreads. Further research, in particular, theoretical contributions along the lines of Cuadra and Saprinza (2008) may therefore be of interest in advancing our understanding of the relationship between revenue windfalls from international commodity price booms, economic growth, and the spread on sovereign bonds in emerging market economies. We conclude on a cautious note that our empirical analysis is based on a relatively short time period. Ideally, an empirical analysis of the effects of com- modity price booms on sovereign bond spreads should include also the 70s and 80s. Manzano and Rigobon (2007) argued that the commodity boom of the 70s led many of the developing (in particular, Latin American countries) to overborrow. When commodity prices collapsed in the 80s, these countries had large debt to GDP ratios and were unable to service their debt, leading to a debt crisis. There exist, unfortunately, no panel data on sovereign bond spreads for the 70s and 80s. This means that we are unable to cover in our analysis the 70s and 80s. We thus end on a note that interestingly, and in line with our results, many of the developing countries were much less democratic in the 70s and 80s than they are today. ¨ ckner Arezki and Bru 97 REFERENCES Acemoglu, D., S. Johnson, and J. Robinson (2001). “The Colonial Origins of Comparative Development: An Empirical Investigation.� American Economic Review 91: 1369– 1401. Aguiar, M., and G. Gopinath (2006). “Defautable Debt, Interest Rates, and the Current Account.� Journal of International Economics 69: 64– 83. Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg (2003). “Fractionalization.� Journal of Economic Growth 8: 155–194. Arellano, C. (2008). “Default Risk and Income Fluctuations in Emerging Markets.� American Economic Review 98: 690– 712. Barro, R. (1999). “Determinants of Democracy.� Journal of Political Economy 107: 158– 183. Beck, T., G. Clarke, A. Groff, P. Keefer, and P. Walsh (2001). “New Tools in Comparative Political Economy: The Database of Political Institutions.� World Bank Economic Review 15: 165 –176. Berg, A., and J. Sachs (1988). “The Debt Crisis: Structural Explanations of Country Performance.� Journal of Development Economics 29: 271– 306. Blundell, R., and S. Bond (1998). “Initial Conditions and Moment Restrictions in Dynamic Panel Data Models.� Journal of Econometrics 87: 115– 43. Collier, P., and B. Goderis (2007). “Commodity Prices, Growth, and the Natural Resource Curse: Reconciling a Conundrum.� CSAE Working Paper 2007–15. Cuadra, G., and H. Sapriza (2008). “Sovereign Default, Interest Rates, and Political Uncertainty in Emerging Markets.� Journal of International Economics 76: 78 –88. Heston, A., R. Summers, and B. Aten (2009). “Penn World Table Version 6.3�, Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania, August 2009. Krueger, A. (1987). “Origins of the Developing Countries’ Debt Crisis, 1970 to 1982.� Journal of Development Economics 27: 141–163. Lederman, D., and W. Maloney (2008). “In Search of the Missing Research Curse.� World Bank Policy Research Paper 4766. Lederman, D., and L. Xu (2010). “Commodities and Structural Volatility.� World Bank, mimeo. Manzano, O., and R. Rigobon (2007). “Resource Curse or Debt Overhang?� in Natural Resources and Development: Neither Curse nor Destiny. Editors: D. Lederman, and W. Maloney, Stanford Press. Marshall, M., and K. Jaggers (2009). Polity IV Project: Dataset Users’ Manual. Center for Global Policy, George Mason University (www.cidcm.umd.edu/polity). [Polity IV Data Computer File, Version p4v2004. College Park, MD: Center for International Development and Conflict Management, University of Maryland.] Mehlum, H., K. Moene, and R. Torvik (2006). “Institutions and the Resource Curse.� Economic Journal 116: 1 – 20. Nickell, S. (1981). “Biases in Dynamic Models with Fixed Effects.� Econometrica 49: 1417– 1426. Neumeyer, P., and F. Perri (2005). “Business Cycles in Emerging Economies: The Role of Interest Rates.� Journal of Monetary Economics (2): 345–380. North, D. (1990). Institutions, Institutional Change, and Economic Performance. New York: Cambridge University Press. Persson, T., and G. Tabellini (2003). The Economic Effects of Constitutions. MIT Press, Cambridge. ——— (2006). “Democracy and Development. The Devil in Detail.� American Economic Review Papers and Proceedings 96 (2): 319– 324. Political Risk Service (2010). International Country Risk Guide. Online Database. Przeworski, A., M. Alvarez, J. Cheibub, and F. Limongi (2000). Democracy and Development: Political Institutions and the Well-Being of the World, 1950-1990. Cambridge, UK: Cambridge University Press. 98 THE WORLD BANK ECONOMIC REVIEW Robinson, J., R. Torvik, and T. Verdier (2006). “Political Foundations of the Resource Curse.� Journal of Development Economics 2: 447–468. Roodman, D. (2009). “A Note on the Theme of Too Many Instruments.� Oxford Bulletin of Economics and Statistics 71: 135– 158. Sachs, J., and A. Warner (1995). “Natural Resource Abundance and Economic Growth.� NBER Working Paper 5398. ——— (2001). “The Curse of Natural Resources.� European Economic Review 45: 827–838. Stock, J., and M. Yogo (2005). “Testing for Weak Instruments in Linear IV Regression.� In Identi�cation and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg, ed. D. Andrews, and J. Stock, 80–108. Cambridge: Cambridge University Press. Van der Ploeg, F. (2010). “Natural resources: curse or blessing?� Journal of Economic Literature, forthcoming. WDI (2010). World Development Indicators. Online Database. Windmeijer, F. (2005). “A Finite Sample Correction for the Variance of Linear Ef�cient Two-Step GMM.� Journal of Econometrics 126: 25– 51. A p p e n d i x Ta b l e 1 . Robustness to Outliers DSpread (1) (2) (3) (4) (5) (6) Median Median Median Excluding Excluding Excluding Regression Regression Regression Max/Min Max/Min 1% Max/Min 1% 1% DComPI 2 0.478 2 16.023*** 2 0.310 2 5.858 2 30.213*** 2 7.860 ( 2 0.12) ( 2 2.91) ( 2 0.07) ( 2 0.77) ( 2 4.28) ( 2 1.19) DComPI* 2 1.879** 2 3.371** Avg. Polity2 ( 2 2.39) ( 2 2.42) Score DComPI* 29.437** 55.051*** Autocracy (2.42) (4.26) Indicator DComPI* 2 7.734 2 16.651** Avg. Checks ( 2 1.43) ( 2 2.54) & Balance Score DComPI* 0.002** 0.003** 0.002* 0.001 0.004*** 0.001 Avg. GDP Per (2.08) (2.40) (1.75) (0.12) (3.13) (1.18) Capita Country Fixed Yes Yes Yes Yes Yes Yes Effects Year Fixed Yes Yes Yes Yes Yes Yes Effects Observations 291 291 291 284 284 284 Note: The method of estimation in columns (1)-(3) is maximum likelihood; columns (4)-(6) least-squares. The least-squares regressions in columns (4)-(6) exclude observations where the change in the commodity export price index is in the top/bottom 1 percentile. The dependent variable is the log-change in the spread on sovereign bonds. DComPI stands for the log-change in the international commodity export price index (see page 3 in the paper for a detailed explanation of how this variable is constructed). The cross-section (average time-series) dimension of the panel is 38 (7.7). *Signi�cantly different from zero at the 10 percent signi�cance level, ** 5 percent signi�cance level, *** 1 percent signi�cance level. ¨ ckner Arezki and Bru 99 Data Appendix Table 1. Descriptive Statistics Mean Std. Dev. Min Max Obs. DLog Sovereign Bond Spread (DSpread) 2 0.11 0.39 2 2.02 1.32 291 DLog Export Price Index (DComPI) 0.002 0.006 2 0.02 0.04 291 Polity2 Score 4.98 5.35 27 10 291 Checks and Balance Score 3.20 1.43 1 6 291 GDP Per Capita 9189 5085 1236 21331 291 Ethnic Fractionalization 0.42 0.23 0.002 0.85 289 Export Concentration 0.11 0.19 0.006 0.98 282 Gini 43.16 9.16 27 60.4 291 Corruption 2.41 0.96 1 5 278 Settler Mortality 206.5 486.6 17.7 2004 148 Data Appendix Table 2. List of Countries Country Observations Spread Polity2 GDP GINI Ethnic Frac Algeria 4 748.88 23 5432 0.35 0.34 Argentina 10 2135.98 7.9 12956 0.5 0.26 Brazil 10 684.82 8 8666 0.58 0.54 Bulgaria 10 420.05 8.7 7303 0.3 0.4 Chile 8 132.52 9.2 15765 0.55 0.19 China 10 102.81 27 5209 0.42 0.15 Colombia 10 446.4 7 6919 0.58 0.6 Croatia 9 2288.7 0.7 11209 0.29 0.82 Cuba 7 305.35 27 7706 0.27 0.37 Dominican Republic 6 539.44 8 8194 0.51 0.43 Ecuador 10 1271.83 6.6 5351 0.56 0.66 Egypt 6 195.49 2 4.5 5102 0.32 0.18 El Salvador 5 259.21 7 5325 0.51 0.2 Greece 2 89.99 10 19117 0.34 0.16 Hungary 8 69.66 10 14881 0.27 0.15 Indonesia 3 249.39 8 4944 0.39 0.74 Korea, Republic of 7 255.87 8 18806 0.32 0 Lebanon 9 400.77 7 7679 0.6 0.13 Malaysia 10 197.84 3 14952 0.43 0.59 Mexico 10 315.8 7.6 10226 0.49 0.54 Morocco 9 379.89 26 4855 0.4 0.48 Nigeria 10 908.19 3.5 1664 0.45 0.85 Pakistan 6 492.48 2 3.8 3112 0.31 0.71 Panama 10 346.38 9 7464 0.55 0.55 Peru 10 434.95 7 5339 0.51 0.66 Philippines 10 414.73 8 3918 0.45 0.24 Poland 10 155.88 9.6 11568 0.33 0.12 Russia 10 972.75 5.2 9718 0.39 0.25 South Africa 10 234.17 9 9223 0.35 0.75 Thailand 9 170.31 7.4 7713 0.43 0.63 Tunisia 5 148.6 24 9034 0.41 0.04 Turkey 10 488.33 7 6569 0.42 0.32 Ukraine 7 677.71 6.2 7696 0.3 0.47 Uruguay 6 508.43 10 10962 0.45 0.25 Venezuela 10 715.42 6.2 10689 0.48 0.5 Vietnam 2 158.72 27 3492 0.38 0.24 When Should We Worry about Inflation? Raphael Espinoza, Hyginus Leon, and Ananthakrishnan Prasad At what level should inflation be a concern? From a growth perspective, high and rising levels of inflation as in 2006– 2008 raise concerns that inflation, if uncontained, could undermine growth. On the other hand, higher levels of inflation could create more space for using monetary policy to reduce nominal and real interest rates during �nancial crises. A nonlinear growth regression for 165 countries over 1960 –2007 shows that for developing countries, inflation above 10 percent quickly hurts growth. For advanced economies, there is no speci�c threshold: in the medium term, higher inflation hurts growth for any initial level of inflation, suggesting that there is a real cost to maintaining higher inflation as a buffer. JEL codes: E31, O40 High output growth and low inflation are among the central objectives of macroeconomic policy. But can they coexist? Or is there a tradeoff between lowering inflation and achieving higher growth? At the operational level, there is a recognition that the growth-inflation relationship depends on the level of inflation. At some low levels, inflation may be positively correlated with growth, by “greasing the wheels� of the economy or as a signal of overheating, but at higher levels inflation is likely to be harmful to growth. In the academic literature, this relationship has been translated into threshold models, which suggest that when inflation exceeds a certain threshold level, higher inflation becomes immediately costly for growth, a result that would call for radical policy changes once inflation exceeds the threshold. This structural break in the relationship between inflation and growth is assumed to occur instantaneously, since the preferred model in the literature is a piecewise linear model (the threshold regression, or TAR, model). The policy implications of the �nding that inflation is quickly penalizing for growth call for further analysis of how quickly inflation becomes costly. For Raphael Espinoza (corresponding author; respinoza@imf.org) is an economist in the Research Department at the International Monetary Fund (IMF). Hyginus Leon (hleon@imf.org) is an advisor in the IMF Western Hemisphere Department and IMF senior resident representative in Jamaica. Ananthakrishnan Prasad (aprasad@imf.org) is the IMF mission chief for Qatar. The views expressed in this article are those of the authors and do not necessarily represent those of the IMF or IMF policy. The authors thank Chris Papageorgiou, Abdel Senhadji, participants at an IMF Institute seminar, and three anonymous referees for very useful comments. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 100– 127 doi:10.1093/wber/lhr043 Advance Access Publication November 4, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 100 Espinoza, Leon and Prasad 101 developing countries, the focus has been on the level of inflation beyond which growth falters. When inflation rose rapidly as a global phenomenon in mid-2007, there were concerns that it was reaching high levels that, if uncon- tained, would undermine growth by raising inflationary expectations. Consequently, a policy priority for many countries was to head off inflationary pressures while preserving growth. For advanced economies, which have much lower inflation thresholds, the focus was on maintaining very low levels of inflation to sustain macroeconomic stability. Citing the twin risks of deflation and near zero nominal interest rates, a recent study challenged this view, arguing that higher targets for inflation (and resulting higher nominal interest rates) would leave more room to reduce nominal and real interest rates during crises and allow for stronger monetary stimulus (Blanchard, Dell’Ariccia, and Mauro 2010). Motivated by this experience and by the renewed discussion of the optimal level of inflation, this article uses a logistic smooth transition regression model (LSTR; see Tera ¨ svirta 1994, 1998, and van Dijk, Tera¨ svirta, and Franses 2002) to revisit the relationship between inflation and growth. It augments the litera- ture in several ways: estimating the speed of transition from one regime (low effect of inflation on growth) to another (high effect on growth), using boot- strap techniques to validate the precision of the estimates and Monte Carlo experiments to investigate the performance of LSTR in estimating thresholds and speeds of transition, distinguishing among advanced economies, develop- ing economies, and oil exporters, and extending the estimation period and increasing the number of countries. The estimated threshold is about 10 percent for most of the country groups (except for the advanced economies). This rate is in line with many of the esti- mates in the literature, although several papers have suggested higher thresholds (around 20 percent for developing economies). The speed of transition is fairly high, which implies that inflation is harmful to growth soon after it exceeds the threshold. Monte Carlo experiments reveal that although the LSTR estimator provides some useful additional degrees of freedom in the estimation, the dis- tinction between the LSTR and a more standard TAR model is not crucial because of the rapid speed of transition from one regime to another. The accuracy of the estimates is validated using bootstrapping techniques—providing, to the best of our knowledge, new robustness checks in the inflation-growth literature. These techniques also help explain the range of estimates found in the literature to be reflective of the natural vari- ation inherent in the cross-country data as well as of the estimation tech- nique. Although the bootstrapping exercises suggest that the value of the threshold is robust to outliers, the range of bootstrap estimates is almost as large as the range of estimates in the literature. For instance, for developing countries, the estimated inflation threshold lies between 7 percent and 13 percent for around 90 percent of the bootstrapped sample. Furthermore, the results indicate that inflation is more costly for oil exporters than for the 102 THE WORLD BANK ECONOMIC REVIEW other country groups if the dependent variable is nonoil GDP growth (as opposed to total GDP growth). Section I discusses previous inflationary episodes, as context for the volatility of inflation surrounding the recent global recession, and summarizes the policy measures taken around the world to �ght overheating in 2007 and 2008. Section II reviews the related academic literature. Section III identi�es a threshold model for the relationship between inflation and growth, and pre- sents the results and robustness checks to the model. Section IV presents impli- cations of the �ndings. I. TH E BACKDROP Since the late 1960s, expansionary �scal policies and accommodative monetary policies have contributed to a strong cyclical upswing in the global economy, creating supply-demand imbalances in many nonfuel primary commodities (table 1). Capacity constraints were already putting upward pressure on wages and prices, but the oil price shock of 1973 boosted an inflation surge, with inflation reaching double digits in many industrial countries (including Japan, the United Kingdom, and the United States). Developing countries (and oil importers) were affected even more. Meanwhile, falling real interest rates during the 1970s gave another boost to aggregate demand, further stoking inflation. In the 1980s and 1990s, during the period dubbed the Great Moderation, inflation declined steadily, especially among advanced economies. Although the interpretation is not de�nitive, econ- omists have attributed part of the improvements in the macroeconomic environment to stronger monetary policy (Bernanke 2004). The 2006–08 rapid increase in headline inflation was driven mostly by food and energy prices. Broadly speaking, the global economy faced three types of underlying inflationary impulses. First, a number of countries—including several in emerging Europe—faced a combination of strong capital inflows, rapid credit growth, tightening labor markets, and widening current account de�cits, all evidence of overheating. Second, many commodity-exporting countries, including members of the Gulf Cooperation Council and the Russian T A B L E 1 . Inflation in Historical Perspective 1900– 1930– 1950– 1961– 1971– 1981– 1991– 1996– Country group 13 39 60 70 80 90 95 2000 Advanced 1.5 0.2 4.3 4.0 10.8 8.1 3.9 2.0 economies Selected emerging 1.2 1.6 15.2 18.3 29.8 139.7 94.4 23.4 economies Source: IMF 2002. Espinoza, Leon and Prasad 103 Federation, experienced rapidly rising export earnings, which boosted aggre- gate demand and asset prices. Third, commodity and food prices surged, increasing inflation across the globe. With energy and food contributing signi�- cantly more to inflation in emerging markets and low-income countries, inflation rose more sharply in those countries. For example, food items accounted for 30 percent of the consumption basket in developing Asia (com- pared with 15 percent in Organisation for Economic Co-operation and Development countries) and about 70 percent of inflation in 2007. Headline inflation rose to 8.6 percent in May 2008 (year on year) compared with 3.5 percent in advanced economies (Figure 1). With mixed results, countries used a combination of monetary, trade, and �scal measures to counter inflationary pressures, depending on whether countries were commodity exporters (especially fuel) or importers. Some gov- ernments resorted to price and quantity controls on imports and exports, using administered prices, subsidies, and buffer stocks. Countries with flexible exchange rate regimes used a combination of monetary and �scal policies. Countries with pegged or relatively �xed exchange rate regimes depended more on �scal policy, efforts to increase productive capacity, and prudential measures to rein in inflation. In 2009, fuel prices plummeted some 40 percent and food prices declined almost 15 percent. With global demand squeezed, world inflation of 2.5 percent was less than half its 2008 peak of 6 percent. By the end of 2010, the average annual inflation rate over the period 2005–10 had shrunk to around 7.5 percent for food and fuel, but accommodating monetary policies around the world and sustained energy and food inflation in 2010 have rekindled worries of inflationary risks (and growth concerns) in regions that recovered fastest from the world recession. F I G U R E 1. Price indices (2005 ¼ 100) Source: IMF 2010. 104 THE WORLD BANK ECONOMIC REVIEW II. A REVIEW OF THE L I T E R AT U R E Research on the inflation-growth nexus has addressed three key questions: Is there a robust negative relationship between inflation and growth? Is there an optimal level of inflation, and what de�nition of inflation should be targeted? Is there a kink in the relationship, so that the relationship is positive at very low levels of inflation (perhaps due to Phillips curve effects) and negative at higher levels, and does the kink occur at similar levels in industrial and devel- oping economies? The Keynesian view of a static tradeoff between inflation and growth domi- nated the literature on inflation until the stagflation episode in the 1970s elev- ated Friedman’s view that the tradeoff could exist only in the short term. In the long run, the Phillips curve is vertical because agents form rational expectations on inflation. Neo-Keynesian models have synthesized these short- and long- term effects using rational expectation models with sticky prices in economies with monopolistic competition. In the long run, activity depends only on factors of production, but in the short run inflation surprises can increase pro- duction because inflation affects relative prices for companies that cannot reset prices immediately. The long-run supply curve with respect to inflation may even slope back- wards because of the effect of inflation on uncertainty: high inflation increases macroeconomic uncertainty because it is unclear to the private sector whether policymakers will want to face the costs of disinflation (Ball 1992). Uncertainty, in turn, reduces investment and growth because there is an option value to waiting when the investment decision is irreversible (McDonald and Siegel 1986). Changes in relative prices also create distortions in the production structure, and as a result the welfare effect of inflation is negative. In most dynamic stochastic general equilibrium (DSGE) models, optimal inflation is zero or lower (Schmitt-Grohe and Uribe 2004). However, optimal inflation is positive if nominal wages cannot be decreased, since inflation then allows for downward adjustments in real wages when needed (Akerlof, Dickens, and Perry 1996). Indeed, central banks in many advanced countries adopted (explicitly or implicitly) inflation targets of about 2 percent. The debate on the optimal level of inflation was revived following the proposal by Blanchard, Dell’Ariccia, and Mauro (2010) that central banks increase their inflation targets from 2 percent to 4 percent. This proposal arose from the observation that central banks in many countries with low inflation and low interest rates had been unable to reduce nominal (and thus real) interest rates as much as needed to �ght the Great Recession. Raising the inflation target from 2 percent to 4 percent was seen as reducing the probability of hitting the zero-bound of the interest rate (Reifschneider and Williams 2000; Billi and Kahn 2008). From a policy perspective, it was assumed that central banks could also affect expected inflation rates and therefore reduce real rates without moving nominal rates (Eggertsson and Woodford 2003). However, Espinoza, Leon and Prasad 105 such a policy would require committing to higher inflation in the future, and this proved dif�cult in the 2009 crisis as central bankers worried about the effect on the credibility of the monetary authority. With food and fuel prices driving recent inflationary episodes, the debate also extended to the appropriate de�nition of inflation (core or headline). In DSGE models, inflation is costly because of the relative price distortions created when prices are sticky. But food and fuel prices are less sticky than the other components in the CPI and therefore contribute less to these distortions. Consequently, targeting core inflation may be optimal (Aoki 2001). However, this argument ignores the effects that food and fuel price inflation might have in models with other frictions. In contrast, Anand and Prasad (2010) argue that food price shocks affect inflation expectations because households cannot hedge against these shocks and cannot factor food prices into wage bargaining. As a result, targeting headline inflation, with some weight given to the output gap, is optimal. This argument could be especially important for developing countries, where food accounts for a large fraction of household expenditure and �nancial constraints ( preventing hedging) are tighter. Cata ˜ o and Chang (2010) offer a similar argument and show that targeting headline inflation is optimal when food price volatility is high. A starting point for answering the empirical question on the kink in the inflation-growth relationship is identifying the threshold beyond which inflation has a negative effect on growth. Ideally, inflation thresholds should be esti- mated for each country, incorporating country-speci�c characteristics. Nevertheless, studies have relied mostly on panel techniques, because the relationship between inflation and growth is likely to be stronger at low fre- quencies, and the data rarely cover more than 40 years. However, measuring the threshold level of inflation in a cross-country framework runs the risk that extreme values will influence the results, since samples typically include countries with inflation as low as 1 percent and as high as 200 percent. Empirical studies have found a signi�cant statistical relationship between inflation and growth, even after controlling for �scal performance, wars, droughts, population growth, openness, and even human and physical capital and allowing for simultaneity bias. Among studies using a time series approach, Mubarik (2005) analyzed data on Pakistan for 1973–2005 and found that although inflation below 5 percent has a positive impact on econ- omic growth, inflation above 9 percent depresses growth. Rangarajan (1997), working with Indian data, suggested a range of 5 percent to 7 percent, a result con�rmed by Samantaraya and Prasad (2001), who estimated the threshold at 6.5 percent. Studies based on panel estimates have produced mixed results. Based on a cross-country regression of 101 countries over 1960–89, Fischer (1993) found that high inflation retards the growth of output by reducing investment and the productivity growth rate. Using different panels, Barro (1995, 2001) and Bruno and Easterly (1996) noticed that the negative effect of inflation on growth was 106 THE WORLD BANK ECONOMIC REVIEW signi�cant only when high-inflation episodes were included in the sample. Using annual data for 87 countries over 1970–90, Sarel (1996) found evidence of a structural break at an 8 percent inflation rate. Inflation and growth are positively correlated below 8 percent but negatively correlated above that, suggesting that ignoring this nonlinearity would signi�cantly underestimate the impact of inflation on growth. Ghosh and Philips (1998) found that although inflation and growth are posi- tively correlated at very low inflation rates (of 2 –3 percent a year), the relation- ship is reversed at higher rates. Furthermore, the relationship is convex, so that the decline in growth associated with an increase in inflation from 10 percent to 20 percent is much larger than that associated with an increase from 40 percent to 50 percent. Khan and Senhadji (2001) reexamined this result and found a sig- ni�cant threshold effect that differed for industrial and developing countries. Their study, based on a panel of 140 countries over 1960–98, established a threshold above which inflation signi�cantly slows growth of 1 –3 percent for industrial countries and 7–11 percent for developing countries. Using a panel of 138 countries, Drukker, Gomis-Porqueras, and Hernandez-Verme (2005) con- �rmed the existence of a threshold, but they estimated its level to be higher, at about 19 percent. This threshold level, higher than usually estimated, was also found by Pollin and Zhu (2006). A much lower threshold was obtained by Burdekin and others (2004), who estimated a panel model on 72 countries using annual data and, after allowing for multiple thresholds, found that inflation is costly for developing countries when it is higher than 3 percent. They also ident- i�ed a second break at 50 percent, above which marginal growth costs decreases by 25 percent. I I I . I N F L AT I O N AND GROWTH THRESHOLD LEVEL This section presents the results of the LSTR model and of Monte Carlo exper- iments analyzing the difference between the LSTR model and a TAR model. The LSTR Model The Khan and Senhadji (2001) model is extended by estimating the link between inflation and GDP growth using a panel of 165 countries over 1960– 2007 and allowing both for a smooth transition model and a convex relation- ship above (and below) the threshold. The explanatory variable is a logarithmic function of inflation. In line with the growth literature, data were averaged over �ve-year periods1 to smooth business cycle fluctuations, and controls were included for the other determinants of growth (see below). The following 1. The time dimension of the data is therefore nine periods of nonoverlapping �ve-year averages. This prevents any moving-average dynamics that would be generated by using annual data of �ve-year moving averages. Espinoza, Leon and Prasad 107 F I G U R E 2. Explanatory Variable f( p), as a Function of p Source: Authors’ calculations based on model described in text. LSTR model was estimated: 8 > > Dyit ¼ ai þblow Wlow ð f ðpit Þ À f ðcÃÞÞ þ bhigh Whigh ð f ðpit Þ À f ðcà ÞÞ þ Q:Xit þ 1it > > & > > lnð1þpit Þ if pit ! 0 > > f ðpit Þ ¼ > > Àlnð1Àpit Þ if pit , 0 > < Wlow ¼ 1 À Whigh > > 1 : > > Whigh ¼ > > à f ðpit Þ À f ðc Þ > > > > 1 þ expðÀgÃ Þ > : s and s is the standard deviation of f ðpit Þ ð1Þ where y is the logarithm of GDP and X a vector of control variables. The func- tion f( p), where p is annual inflation expressed in percentage points, is used to model the effect of inflation on growth (�gure 2). The choice of a logarithmic function rather than a linear function such as f( p) ¼ p enables capturing the fact that multiplicative shocks in inflation have similar effects on growth for any initial level of inflation. For instance, in a linear model, an increase in inflation from 10 percent to 20 percent would have the same effect on growth as an increase from 50 percent to 60 percent. In a multiplicative (logarithmic) model, an increase from 10 percent to 20 percent inflation would have the same effect as an increase from 50 percent to 100 percent.2 2. A constant was added in the logarithmic function to smooth the distribution of f( p) around zero. The distribution for f( p) in the data was almost symmetric (skewness at 0.9). 108 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3. Weight on bhigh (threshold ¼ 10, standard deviation ¼ 1) Source: Authors’ calculations based on model described in text. The explanatory variable f( p) enters the model through two regressors— low W ( f( p)-f(c*)) and Whigh ( f( p)-f(c*))—to capture the possibility that a different b should be used for different levels of inflation. The parameter c* can be interpreted as the level of inflation above which the parameters of the upper regime dominate and below which the parameters of the lower regime dominate. In particular, † If inflation is very low, the effect of inflation on growth is represented by blow only (since Whigh ¼ 0—see �gure 3—and therefore Wlow ¼ 1). † If inflation is very high, the effect of inflation on growth will be rep- resented mainly by bhigh (since Whigh ¼ 1 when inflation is high). More generally, when inflation is above c* the impact on growth is nearer bhigh, and when inflation is below c* the impact on growth is nearer blow. When inflation is in the neighborhood of c*, the actual effect of inflation on growth is given by a weighted average of blow and bhigh, where the weights are given by Whigh, as plotted in �gure 3, and Wlow ¼ 1-Whigh. When inflation is equal to c*, Whigh ¼ Wlow ¼ 1 2. The parameter g * captures the speed of transition from one regime to another. So that the value of g * can be compared across models, it is divided by the standard deviation of f( pit). Even if inflation is above the threshold (say inflation is 13 percent and the estimated threshold is 10 percent), the effect of inflation on growth may not be strongly negative when g * is low (for example, g * ¼ 3 in �gure 3), since it will be captured by 0.33 blow þ 0.67 bhigh.3 If g * ¼ 15, the effect is instead 0.03* blow þ 0.97* bhigh. 3. All the calculations and �gure 3 are computed with standard deviation ¼ 1 in line with the overall sample statistics. Espinoza, Leon and Prasad 109 The estimation of g * is important in the LSTR model: the policy impli- cations of a high g * would be that inflation pressures have to be tackled as a priority because inflation becomes immediately costly. The standard threshold model used in the literature assumes that g * ! 1 and, therefore, is unable to determine how quickly to worry about high inflation. When c* and g * are known, the model collapses to a linear model and can be estimated by ordinary least squares (OLS). The model is therefore estimated by searching for the c* and g * that minimize the concentrated sum of the squared errors (maximize the �t) of the OLS model. It is important, however, to test that equation (1) is an appropriate speci�cation of the data and in par- ticular that the nonlinearity in the relationship between inflation and growth can be captured by exactly two regimes.4 If there is only one regime (no nonli- nearities), the transition parameters are not identi�ed, and if there are more than two regimes, the model is misspeci�ed. The tests proposed by Gonza ´ lez, Tera¨ svirta, and van Dijk (2005) for panel models with heteroskedastic errors are applied to test for the existence of nonlinearities and determine the number of regimes.5 The literature surveyed in section II has not tackled this issue as only two-regime threshold models were assumed and estimated, and heteroske- dasticity was not taken into account. The technical details on the estimation of the LSTR model and the speci�cation tests are in the appendix. The LSTR model is estimated using several control variables as determinants of growth: the ratio of investment to GDP, population growth, initial GDP, the rate of change in terms of trade, and variability in terms of trade. Since invest- ment is included as an explanatory variable, the total effect of inflation on growth could be biased by the correlation between inflation and investment. To check for this, a model without investment is estimated, and the effect of inflation on investment and indirectly on growth is also attributed to inflation. Further, because the model controls for population growth rather than employ- ment growth (owing to data availability), the effect of inflation on labor force participation and structural unemployment is implicitly included in the com- puted costs of inflation. Time dummy variables control for world business cycles, and country dummy variables capture country-speci�c characteristics. Results The results for the baseline speci�cation are presented in table 2. The model is estimated for all countries in the data set (columns 1 –3) and for several country groups as de�ned by the International Monetary Fund (see the statistical appendix in IMF 2008), including advanced economies (columns 4– 6), developing economies (columns 7–9), a group of oil producers (columns 4. In the typical smooth transition autoregressive model, it is common to follow a modeling process of testing for linearity against logistic and exponential nonlinearity, searching for the best-�t transition lag, and undertaking misspeci�cation tests for remaining nonlinearities and variance reduction (Leon and Najarian 2005). 5. The authors are grateful to an anonymous referee for suggesting these speci�cation tests. T A B L E 2 . Ordinary Least Squares, Threshold Regression, and Logistic Smooth Transition Regression Results 110 All countries Advanced economies (1) (2) (3) (4) (5) (6) Logistic Logistic Ordinary smooth Ordinary smooth least Threshold transition least Threshold transition Variable squares regression regression squares regression regression Investment/GDP 0.0691** 0.0640** 0.0637** 0.0947*** 0.0919*** 0.0923*** [0.0285] [0.0280] [0.0280] [0.0287] [0.0287] [0.0286] Population growth 0.497*** 0.503*** 0.502*** 0.581** 0.519** 0.520** [0.148] [0.148] [0.148] [0.224] [0.226] [0.226] Initial GDP – 0.0215*** –0.0207*** –0.0206*** –0.0374*** – 0.0366*** – 0.0366*** [0.00436] [0.00433] [0.00433] [0.00527] [0.00531] [0.00531] Terms of trade growth 0.0573*** 0.0531*** 0.0530*** 0.0613* 0.0569 0.0568 [0.0182] [0.0178] [0.0178] [0.0352] [0.0349] [0.0350] Standard deviation of the terms of trade 4.61e-05*** 4.41e-05*** 4.41e-05*** –0.000153 – 0.000135 – 0.000138 THE WORLD BANK ECONOMIC REVIEW [1.67e-05] [1.68e-05] [1.68e-05] [0.000365] [0.000357] [0.000357] f(inflation) – 0.00452*** –0.00634*** [0.00103] [0.00224] f(inflation) 2 f(T) if inflation , T 0.00255 0.00775 [0.00184] [0.00622] f(inflation) 2 f(T) if inflation . T –0.00988*** – 0.00795*** [0.00137] [0.00235] Wlow*( f(inflation) 2 f(c*)) 0.00198 0.00777 [0.00173] [0.00621] Whigh*( f(inflation) 2 f(c*)) –0.0102*** – 0.00791*** [0.00140] [0.00234] Observations 1,304 1,304 1,304 259 259 259 R-squared 0.412 0.426 0.426 0.749 0.755 0.755 TAR or LSTR parameter c* 8 9 1 1 LSTR parameter g * 6 13 LM test (GTD 2005) H0: linear model 37.36 17.36 p-value nonlinearity 0.00 0.30 LM test for remaining nonlinearities 12.33 7.41 p-value remaining nonlinearity 0.65 0.95 T A B L E 2 . continued Developing countries Oil producers Oil producers nonoil GDP (7) (8) (9) (10) (11) (12) (13) (14) (15) Logistic Logistic Logistic Ordinary smooth Ordinary smooth Ordinary smooth least Threshold transition least Threshold transition least Threshold transition Variable squares regression regression squares regression regression squares regression regression Investment/GDP 0.0651** 0.0616** 0.0615** 0.212*** 0.202*** 0.208*** 0.333*** 0.333*** 0.334*** [0.0286] [0.0287] [0.0286] [0.0776] [0.0751] [0.0747] [0.109] [0.101] [0.101] Population growth 0.497*** 0.502*** 0.504*** 0.768*** 0.811*** 0.796*** 0.397 – 0.177 –0.193 [0.158] [0.158] [0.158] [0.194] [0.195] [0.196] [0.827] [0.892] [0.891] Initial GDP – 0.0165*** – 0.0163*** – 0.0164*** –0.0230* –0.0230* – 0.0243* 0.204*** 0.206*** 0.206*** [0.00529] [0.00525] [0.00524] [0.0135] [0.0132] [0.0133] [0.0712] [0.0686] [0.0685] Terms of trade growth 0.0550*** 0.0519*** 0.0517*** 0.0757* 0.0646* 0.0637 0.156 0.122 0.123 [0.0190] [0.0188] [0.0187] [0.0400] [0.0383] [0.0386] [0.119] [0.122] [0.122] Standard deviation of the terms of trade 4.70e-05*** 4.55e-05*** 4.53e-05*** 0.000468 0.00043 0.000474 0.000139** 0.000119* 0.000120* [1.62e-05] [1.65e-05] [1.65e-05] [0.000349] [0.000348] [0.000349] [6.08e-05] [6.07e-05] [6.07e-05] – 0.00475*** –0.000825 – 0.00661 [0.00111] [0.00343] [0.0117] f(inflation) 2 f(T) if inflation , T 0.000629 0.00707 0.0284** [0.00190] [0.00559] [0.0129] f(inflation) 2 f(T) if inflation . T – 0.0102*** –0.0193** – 0.0826*** [0.00164] [0.00763] [0.0255] Wlow*( f(inflation) 2 f(c*)) 0.000777 – 0.00582* 0.0280** [0.00190] [0.00303] [0.0129] high W *( f(inflation) 2 f(c*)) – 0.00987*** – 0.135*** –0.0818*** [0.00157] [0.0475] [0.0252] Observations 1,042 1,042 1,042 200 200 200 253 253 253 R-squared 0.386 0.396 0.396 0.499 0.517 0.519 0.643 0.67 0.671 TAR or LSTR parameter c* 11 10 10 74 13 13 LSTR parameter g * 6 1 11 LM test (GTD 2005) H0: linear model 30.15 18.73 20.48 p-value nonlinearity 0.01 0.18 0.12 LM test for remaining nonlinearities 11.94 7.68 2.58 p-value remaining nonlinearity 0.61 0.91 1.00 *** Signi�cant at p , .01; ** signi�cant at p , .05; * signi�cant at p , .1. Espinoza, Leon and Prasad LSTR is logistic smooth transition regression. LM is linear model. GTD 2005 is Gonza ´ lez, Tera ¨ svirta and van Dijk (2005). TAR is threshold regression. Note: Numbers in brackets are robust standard errors. Coef�cients for time and country dummy variables are not shown. 111 Source: Authors’ calculations based on data described in text. 112 THE WORLD BANK ECONOMIC REVIEW 10–12), and all oil producers for which nonoil GDP data were available (columns 13–15). The results identify a threshold inflation of about 9 percent for the entire sample of 165 countries. The coef�cient bhigh for the variable Whigh ( f( p) – f(c*)) is signi�cant. (The standard errors are heteroskedastic-robust, conditioned on the value of the LSTR or TAR par- ameters in the estimation. Bootstrapping results are discussed below.) When inflation is above 9 percent, a doubling of inflation decreases GDP by 0.7 per- centage point a year (table 3). The order of magnitude of the estimate is in line with that of Khan and Senhadji (2001), who use a discrete threshold model and the same control variables. The full sample estimate in the current study is driven by the developing country data and produces almost identical results to those for the full panel. For the advanced economies group, the nonlinearity test does not reject the null hypothesis of linearity (against the logistic nonlinear alternative), and the linear speci�cation cannot be ruled out. Indeed, the threshold obtained is the lowest allowed by the estimation algorithm,6 and the coef�cient in the OLS model is similar to that of the LSTR or TAR for inflation above the transition parameter. In any case, the difference between OLS and a nonlinear relation- ship with a threshold at 1 percent is not statistically or economically meaning- ful. Inflation was above 1 percent (for a �ve-year average) for 95 percent of observations for advanced economies over the sample, and aiming at inflation rates below 1 percent would pose serious deflation risks. Nevertheless, the model and the nonlinearity test do provide a result for advanced economies: higher inflation hurts growth in the medium term, for any reasonable initial level of inflation. For the group of oil-exporting countries, the nonlinearity test does not reject the linearity assumption in this model, which implies that the threshold par- ameter is not identi�ed (this is also what the bimodal bootstrapping distri- bution suggests; see below). In particular, since GDP numbers may be contaminated by the importance of oil production in real GDP, the model is also estimated using nonoil GDP data for all countries for which the data are available. The results become meaningful, and the effect of inflation is strong: the threshold is 13 percent, and a doubling of inflation from higher than 13 percent lowers real nonoil GDP 2.7 percent a year. Although the data are not available to investigate the cause of this �nding, it could be either because the nonprimary commodity sector is more sensitive to inflation (since production in the primary sector tends to depend on other elements such as international prices and production capacity)—in which case the �nding would be relevant to a varied range of countries—or because the structure differs for the 6. For the advanced economies group, there are only 12 observations with inflation below 1 percent, which means that the estimate may not be robust. To check for robustness, the search was narrowed to inflation above 2 percent, and 2 percent was also found to minimize the concentrated sum of squares for the TAR and 3 percent for the LSTR (with a very low speed of transition). The nonlinearity test again did not reject the linearity hypothesis. Espinoza, Leon and Prasad 113 T A B L E 3 . Effect of Inflation on Growth All All Effect on All countries (if countries (if Advanced Developing nonoil countries g ¼ 1) g ¼ 15) economies countries real GDP (a) (b) (c) (d) (e) (f) Effect on annual GDP growth Threshold 9 9 9 1 10 13 blow 0.002 0.002 0.002 0.008 0.001 0.028 bhigh –0.01 – 0.01 – 0.01 – 0.008 – 0.01 – 0.082 g 6 1 15 13 6 11 Standard 1.08 1.08 1.08 0.84 1.12 0.98 deviation Initial inflation 1 New 1 0.00 0.00 0.00 0.00 0.00 0.00 inflation 5 0.25 0.10 0.22 – 0.88 0.13 3.08 8 0.35 0.01 0.32 – 1.20 0.21 4.25 9 0.32 – 0.03 0.32 – 1.29 0.20 4.59 10 0.27 – 0.07 0.25 – 1.36 0.17 4.94 12 0.12 – 0.16 0.07 – 1.50 0.06 5.49 15 2 0.11 – 0.28 – 0.15 – 1.66 – 0.16 4.62 20 2 0.41 2 0.48 – 0.42 – 1.88 – 0.45 2.17 30 2 0.81 2 0.81 – 0.81 – 2.19 – 0.86 – 1.07 45 2 1.20 2 1.20 – 1.20 – 2.51 – 1.26 – 4.31 Initial inflation 5 New 1 2 0.25 2 0.10 – 0.22 0.88 – 0.13 – 3.08 inflation 5 0.00 0.00 0.00 0.00 0.00 0.00 8 0.09 2 0.09 0.10 – 0.32 0.07 1.17 9 0.07 2 0.13 0.10 – 0.41 0.07 1.51 10 0.02 2 0.17 0.03 – 0.48 0.04 1.86 12 2 0.13 2 0.26 – 0.15 – 0.62 – 0.08 2.41 15 –0.36 – 0.38 – 0.37 – 0.78 – 0.29 1.54 20 –0.66 – 0.58 – 0.64 – 1.00 – 0.59 – 0.91 30 –1.06 – 0.91 – 1.03 – 1.31 – 1.00 – 4.15 45 –1.46 – 1.30 – 1.42 – 1.63 – 1.39 – 7.38 Initial inflation 10 New 1 –0.27 0.07 – 0.25 1.36 – 0.17 – 4.94 inflation 5 –0.02 0.17 – 0.03 0.48 – 0.04 – 1.86 8 0.08 0.08 0.07 0.16 0.04 – 0.69 9 0.05 0.04 0.07 0.08 0.03 – 0.35 10 0.00 0.00 0.00 0.00 0.00 0.00 12 –0.15 – 0.08 – 0.18 – 0.13 – 0.11 0.55 15 –0.38 – 0.21 – 0.40 – 0.30 – 0.33 – 0.32 20 –0.67 – 0.40 – 0.67 – 0.52 – 0.63 – 2.77 30 –1.08 – 0.74 – 1.06 – 0.83 – 1.03 – 6.01 45 –1.47 – 1.13 – 1.45 – 1.14 – 1.43 – 9.25 (Continued ) 114 THE WORLD BANK ECONOMIC REVIEW TABLE 3. Continued All All Effect on All countries (if countries (if Advanced Developing nonoil countries g ¼ 1) g ¼ 15) economies countries real GDP (a) (b) (c) (d) (e) (f) Initial inflation 15 New 1 0.11 0.28 0.15 1.66 0.16 – 4.62 inflation 5 0.36 0.38 0.37 0.78 0.29 – 1.54 8 0.46 0.29 0.47 0.46 0.36 – 0.38 9 0.43 0.25 0.47 0.38 0.36 – 0.03 10 0.38 0.21 0.40 0.30 0.33 0.32 12 0.23 0.12 0.21 0.17 0.21 0.87 15 0.00 0.00 0.00 0.00 0.00 0.00 20 – 0.30 – 0.20 – 0.27 – 0.22 – 0.30 – 2.45 30 – 0.70 – 0.53 – 0.66 – 0.53 – 0.71 – 5.69 45 – 1.09 – 0.92 – 1.06 – 0.84 – 1.10 – 8.93 Source: Authors’ calculations based on data described in text. economies for which data are available (countries in Africa, the Middle East, and Central Asia). The LSTR model provides fairly high estimates of g *. This implies that when starting from a level of inflation of 10 percent, an increase in inflation is more costly than would be case with a much smaller speed of transition (compare columns a and b in table 3).7 Overall, the LSTR models con�rm that inflation tends to be costly quickly, which means that inflation has to be tackled sooner rather than later. Since this result suggests that the discrete TAR model might also be a good—albeit simpli�ed—model of the relationship between inflation and growth, TAR models are also estimated as a check, and the results are comparable. The heteroskedastic-robust Lagrange multiplier (LM) tests for remaining nonlinearities also suggest that there is no need for additional nonlinearities in equation 1. Statistical Inference through Bootstrapping Statistical inference was performed using bootstrapping (500 iterations) and taking into account the panel structure of the data (each random selection cor- responds to a panel, not to an individual observation). The results are pre- sented for all country groups (�gures 4 –6), but since the bootstrapping exercise is computed under the hypothesis of nonlinearity, it is unlikely to be valid for advanced economies and for oil producers when real total GDP is used. 7. The exception is for the group of oil producers, for which the LSTR is clearly not a correct speci�cation. Espinoza, Leon and Prasad 115 F I G U R E 4. Bootstrapping for Advanced Economies and Developing Economies Source: Authors’ calculations based on data described in text. The bootstrapping exercises suggest that the results are robust for advanced economies, for developing economies, and for oil producers when nonoil real GDP is used. The distribution of the thresholds is concentrated around the base estimates. For developing economies, the estimated inflation threshold lies 116 THE WORLD BANK ECONOMIC REVIEW F I G U R E 5. Bootstrapping for Oil Producing Developing Economies Source: Authors’ calculations based on data described in text. between 7 percent and 13 percent for more than 90 percent of the boot- strapped sample. For advanced economies, although the mode of the distribution is at 1 percent, the inflation threshold is found to be lower than Espinoza, Leon and Prasad 117 F I G U R E 6. Lagged Inflation and Lagged Investment (all countries) Source: Authors’ calculations based on data described in text. 5 percent in only 43 percent of the bootstrapped samples and lower than 18 percent in 75 percent of the samples. When total GDP is used as the dependent variable for oil exporters, both the TAR and the LSTR estimations are not robust (there seem to be two poten- tial “modes� for the distribution of the threshold), likely because the LSTR model is not identi�ed when there are no nonlinearities or because there are 118 THE WORLD BANK ECONOMIC REVIEW several outliers. This result con�rms that the estimated LSTR model in table 2, column 12, is misleading.8 When nonoil real GDP (for which linearity was not clearly rejected) is used, the threshold estimate of around 13 percent is robust, roughly in line with the one found for the developing countries group. The bootstrapping exercise con�rms that the effect of higher inflation on nonoil growth is stronger for oil producing developing countries than for the other developing countries. Overall, it is noteworthy that although the coef�cients linking growth and inflation at low levels of inflation are not consistently signi�cantly different from zero, the relationship between growth and inflation at high levels of inflation is almost always negative. This con�rms the importance of nonlineari- ties and the negative relationship between high inflation and growth. Robustness Checks Several other speci�cations were estimated to test for the robustness of these results (table 4). The main lesson of the robustness checks is that when the Gonza ¨ svirta, and van Dijk (2005) test rejects nonlinearity, the esti- ´ lez, Tera mates are robust to different speci�cations. This is the case for the model esti- mated on the whole sample and for the model estimated on developing countries. However, when the test does not reject linearity (for advanced econ- omies and oil producers when total GDP is used), estimates of the transition parameters are not as robust. The �rst robustness check consists of dropping the investment share of GDP variable, to assess the indirect effect of inflation on growth through investment. The transition parameters and slope coef�cients are mostly unchanged for the groups for which the linearity assumption was rejected. For advanced econom- ies, the 1 percent threshold estimated in the baseline is now estimated at 6 percent, but the Gonza ¨ svirta, and van Dijk (2005) test is again ´ lez, Tera unable to reject the linearity hypothesis. A second check estimates the model with the square root of inflation instead of the logarithm of inflation (see table 4).9 The LSTR transition parameters c* are lower by 1 or 2 percentage points in this square-root speci�cation, but the difference is within the error bands suggested by the bootstrapping exercise. The estimates, at 11 percent and with g * ¼ 6, are also very similar for the whole sample and for developing countries when only the subsample of post-1990 data are selected for all countries. Similarly, the cost of inflation for growth is also of the same order of magnitude. For the advanced economies, the model seems misspeci�ed, as the linearity assumption is again not rejected. 8. As a result, the very low g is not robust either—there is a very strong negative correlation between g and c. 9. Twenty periods of very high inflation (with rates exceeding 150 percent a year on average for �ve years) were dropped as they strongly affected that estimation. T A B L E 4 . Robustness Checks All countries Advanced economies Developing countries Oil producers Oil producers nonoil GDP (1) (2) (3) (4) (5) (5) (7) (8) (9) (10) Logistic Logistic Logistic Logistic Logistic smooth smooth smooth Ordinary smooth Ordinary smooth Ordinary least transgression Ordinary least transgression Ordinary least transgression least transgression least transgression Robustness check and indicator squares regression squares regression squares regression squares regression squares regression 1. Robustness check model without investment/GDP f(inflation) -0.00433*** -0.00580** -0.00469*** -0.0037 -0.0112 [-4.483] [-2.473] [-4.525] [-0.937] [-0.938] low W *( f(inflation) 2 f(c*)) 0.00343** 0.00176 0.0021 0.00956 0.0236* [2.211] [0.628] [1.243] [1.638] [1.768] Whigh*( f(inflation) 2 f(c*)) -0.0113*** -0.0107*** -0.0112*** -0.0264*** -0.0855*** [-7.936] [-3.852] [-6.953] [-3.437] [-3.752] LSTR parameter c* 9 6 10 9 13 LSTR parameter g * 6 11 7 8 10 LM-test (GTD 2005) H0: linear model 39.46 17.11 26.63 16.80 20.48 p-value nonlinearity 0.00 0.25 0.01 0.21 0.08 LM-test for remaining nonlinearities 16.89 9.80 15.89 37.58 2.68 p-value remaining nonlinearity 0.26 0.78 0.26 0.00 1.00 p 2. Robustness check f( p) ¼ p if p . 0; p f( p) ¼ 2 2 p if p , 0 f(inflation) -0.00221*** -0.00215* -0.00231*** -0.00186 -0.0161** [-3.286] [-1.720] [-3.142] [-0.974] [-2.070] Wlow*( f(inflation) 2 f(c*)) 0.00399** 0.0259** 0.00254 0.0123** 0.0253* [2.089] [2.023] [1.278] [2.192] [1.730] Whigh*( f(inflation) 2 f(c*)) -0.00380*** -0.00243* -0.00372*** -0.00823*** -0.0305*** [-4.860] [-1.926] [-4.324] [-3.028] [-3.102] LSTR parameter c* 7 1 8 8 12 LSTR parameter g * 15 10 6 10 11 LM-test (GTD 2005) H0: linear model 24.98 16.63 20.39 15.07 19.41 p-value nonlinearity 0.05 0.34 0.12 0.37 0.15 LM-test for remaining nonlinearities 6.35 10.21 6.19 5.78 5.10 p-value remaining nonlinearity 0.97 0.81 0.96 0.97 0.98 Espinoza, Leon and Prasad (Continued ) 119 TABLE 4. Continued All countries Advanced economies Developing countries Oil producers Oil producers nonoil GDP 120 (1) (2) (3) (4) (5) (5) (7) (8) (9) (10) Logistic Logistic Logistic Logistic Logistic smooth smooth smooth Ordinary smooth Ordinary smooth Ordinary least transgression Ordinary least transgression Ordinary least transgression least transgression least transgression Robustness check and indicator squares regression squares regression squares regression squares regression squares regression 3. Robustness check: post 1990 estimation f(inflation) -0.00930*** -0.01 -0.00904*** 0.00 -0.02 [-4.222] [-1.529] [-3.727] [0.356] [-1.085] low W *( f(inflation) 2 f(c*)) 0.00 0.00 0.00 0.01 0.00 [-0.118] [-0.394] [0.0292] [1.636] [-0.245] Whigh*( f(inflation) 2 f(c*)) -0.0140*** -0.0206*** -0.0134*** -0.0999* -0.251*** [-4.406] [-4.077] [-4.008] [-1.853] [-6.396] Observations 616 616 116 116 500 500 93 93 168 168 LSTR parameter c* 11 14 11 48 27 LSTR parameter g * 5 15 5 14 1 LM-test (GTD 2005) H0: linear model 26.16 10.75 21.37 16.75 7.78 p-value nonlinearity 0.00 0.29 0.01 0.05 0.46 LM-test for remaining nonlinearities 15.81 4.14 15.09 3.80 4.23 THE WORLD BANK ECONOMIC REVIEW p-value remaining nonlinearity 0.07 0.90 0.09 0.92 0.84 4. Robustness check: 3-year average model with lagged inflation and lagged investment/GDP f(inflation) 0.00103 -0.00648*** 0.00135 -2.26E-05 -0.0013 [1.087] [-3.133] [1.333] [-0.00676] [-0.168] Wlow*( f(inflation)-f(c*)) 0.00387*** -0.00618*** 0.00370*** 0.00768 0.00824 [2.795] [-3.216] [2.760] [1.543] [1.025] high W *( f(inflation) 2 f(c*)) -0.00390** 0.0179** -0.00423** -0.0192* -0.0715** [-2.355] [1.974] [-2.202] [-1.875] [-2.276] LSTR parameter c* 12 100 17 10 26 LSTR parameter g * 15 1 6 15 15 LM-test (GTD 2005) H0: linear model 25.09 21.94 16.73 14.26 11.85 p-value nonlinearity 0.123 0.235 0.542 0.712 0.809 LM-test for remaining nonlinearities 7.097 58.93 6.293 12.33 7.436 p-value remaining nonlinearity 0.989 3.05E-06 0.995 0.83 0.977 *** Signi�cant at p , .01; ** signi�cant at p , .05; * signi�cant at p , .1. LSTR is logistic smooth transition regression. LM is linear model. GTD 2005 is Gonza ´ lez, Tera ¨ svirta and van Dijk (2005). TAR is threshold regression. Note: Numbers in brackets are robust t-statistics. Source: Authors’ calculations based on data described in text. Espinoza, Leon and Prasad 121 All the models are identi�ed using the assumption that over the medium term (the variables are expressed as nonoverlapping �ve-year averages) the causality runs from inflation to growth, a common assumption in the growth-inflation literature (the OLS assumption). To check for the robustness of the results to the endogeneity of inflation, growth is regressed on lagged inflation (which is unlikely to be caused by future growth), as opposed to con- temporaneous inflation. The lagged investment to GDP ratio was also used for the same reason. The dataset was generated as nonoverlapping three year averages for this exercise, to keep the relationship meaningful. The threshold estimates were again similar for the whole sample and slightly higher (at 17 percent) for developing countries, con�rming that this estimation is robust to potential endogeneity of inflation. However, the Gonza ¨ svirta, and van Dijk (2005) test does not reject ´ lez, Tera linearity for the subgroups of developing countries and of oil producers: at the three-year frequency and using past inflation, the nonlinearity is not strong enough to reject the linear model. For the full sample, the linearity assumption is marginally rejected at the 15 percent level, and the inflation threshold is about 12 percent, with a very high speed of transition. In addition, the effect of high inflation is weaker using lagged inflation than using contemporaneous inflation. The bootstrapping exercise (see �gure 6) con�rms that the LSTR par- ameters are well estimated, with the mode of the distribution being at 11 percent for the threshold inflation. The coef�cient on inflation is negative in more than 95 percent of the bootstrapping iterations, and the mode of this dis- tribution is –0.006. The TAR model yields similar threshold and coef�cient distributions. TAR or LSTR: A Monte Carlo Analysis Monte Carlo experiments were run to analyze the difference between the LSTR model with a reasonably low g * ( g* ¼ 6) and one with an in�nite g * (a TAR model). The Monte Carlo simulations consist of generating arti�cial data from the low-g * LSTR model (equation 1), using the actual estimates of the parameters of the model ( g *, c*, and the standard error of the model) for the different groups, and then estimating a TAR model. The reverse experiment is also performed: arti�cial data are generated using the TAR model parameters and employed to estimate a LSTR model. The results show that the LSTR estimation procedure is good at capturing the value of the threshold, as well as the negative sign of the bhigh coef�cient (�gure 7). However, the blow coef�cient is not statistically different from zero. And the LSTR often estimates g to be lower than its true value: for a TAR, the estimate should be g * ! 1, which is approximated by g* ! 15, but this happens only 25 percent of the time. For the converse exercise (when a LSTR model is simulated with a low g * and estimated as a TAR), the TAR model 122 THE WORLD BANK ECONOMIC REVIEW F I G U R E 7. Monte Carlo Simulations (Data Generating Process as estimated for developing economies) Source: Authors’ calculations based on simulations described in text. does not signi�cantly bias the threshold estimate. It also yields a signi�cantly negative coef�cient for bhigh. Of course, the TAR cannot be used to evaluate the size of g*. Espinoza, Leon and Prasad 123 I V. C O N C L U S I O N Motivated by the global inflation episode of 2007 and recent discussions of the optimal level of inflation that followed heightened risks of deflation and of a liquidity trap, this article revisits the inflation-growth nexus. A smooth tran- sition model is used to investigate the speed at which inflation beyond a threshold becomes harmful to growth, an important consideration in the policy response to rising inflation as the world economy recovers. For a panel of 165 countries over 1960–2007, estimates �nd that for developing countries inflation above a threshold of about 10 percent quickly becomes harmful to growth, suggesting the need for a prompt policy response to inflation at or above that threshold. For the advanced economies, there is no evidence that a speci�c threshold effect is at play: any level of inflation hurts growth. For oil exporting countries, the estimates are less robust, but the threshold is again estimated to be about 10 percent. The effect of higher inflation for oil produ- cers is also stronger than for other developing countries. The impact on growth of core and headline inflation could not be distin- guished because of a lack of data. This is an important area for further research. To the extent that core inflation is better anticipated and less volatile than headline or the noncore component of inflation, core inflation is likely to have less impact on agents’ price-setting behavior and on long-term growth. Also, with policymakers’ reactions more predictable for core inflation than for headline inflation, core inflation would likely create less macroeconomic uncer- tainty (Ball 1992). On the other hand, food and fuel inflation is likely to affect growth more strongly and at lower levels than is core inflation. Policymakers’ reactions to food and fuel inflation are less predictable—food and fuel inflation is generally more volatile, hurts the poor disproportionately, and affects subsi- dies and government balances. Since the food and fuel components of inflation are more exogenous to policy, especially in developing countries, the best route to reducing the costs of inflation may well be to increase the flexibility of markets and develop insurance mechanisms. Another potential area for research is explaining the difference in thresholds and speed of transitions among the advanced, developing, and oil producing economies using factors such as measures of �nancial friction and �nancial development, production structure, and absorptive capacity. APPENDIX The model described in equation (1) is estimated using the standard nonlinear least squares method (see Chan 1993). If g * and c* were known, the model would collapse to a simpler formulation: Dyit ¼ a0 i þ d1 g1 ðpit Þ þ d2 g2 ðpit Þ þ Q:Xit þ 1it ðA1Þ 124 THE WORLD BANK ECONOMIC REVIEW where d1 and d2 are coef�cient to be estimated, and g1( p) ¼ Wlow( p)[ f( p)– f(c*)] and g2( p) ¼ Whigh ( p)[ f( p)– f(c*)]. This model would be estimated using OLS (with country dummy variables). Chan (1993) and Gonza ¨ svirta ´ lez, Tera and van Dijk (2005) suggested therefore that g * and c* be estimated by mini- mizing the concentrated sum of squared errors: Xh   ðcà ;gÃ Þ ¼ Argminðc;gÞ[f1;���;100gÂf1;���;15g Dyit À a ^ low Wlow ðc; gÞ ^i þb i;t à � ð f ðpit Þ À f ðc ÞÞ þ b ^ high high W ^ :Xit ފ2 Þ ðc; gÞ � ð f ðpit Þ À f ðcÞÞ þ Q ðA2Þ where a ˆ high ;Q ˆ low ; b ˆ i; b ^ are themselves estimated by OLS at each iteration of the nonlinear optimization. The solution (c*, g *) is found through a simple grid search algorithm that scrolls through all potential vectors (c*, g *) of natural numbers in f1, . . . ,100g  f1, . . . ,15g. The search for g * stops at 15 as the transition functions for g . 14 are in practice indistinguishable. Periods of severe recession—when losses were higher than 5 percent of GDP per year over �ve years—are excluded from the data. These episodes corre- spond to the breakup of the Soviet Union (Azerbaijan, Belarus, Kazakhstan, and the Russian Federation during 1990–94), 1990–94 in Angola, and 1960– 64 in Argentina and Taiwan, China. Gonza ´ lez, Tera ¨ svirta, and van Dijk (2005) propose several speci�cation tests 10 of the model. The �rst is a test for the existence of nonlinearities—a test that g * ¼ 0 (in which case the model collapses to a linear model as Whigh ¼ Wlow). The test is nonstandard because under the null hypothesis H0: g * ¼ 0, the par- ameters c* and bhigh are not identi�ed. The problem of hypothesis testing in this context was �rst studied by Davies (1977) and applied by Hansen (1999) in the panel context. The solution proposed by Gonza ¨ svirta, and van ´ lez, Tera Dijk, following Luukkonen, Saikkonen, and Tera ¨ svirta (1988), is to replace Whigh by its Taylor series approximation around g * ¼ 0. The following short presentation is taken from Gonza ¨ svirta, and van Dijk, which contains ´ lez, Tera additional details. After a Taylor series approximation, the model becomes Dyit ¼ a0 i þ bà 0 Ã0 à 0 xit þ b1 xit f ðpit Þ þ 1it ðA3Þ where xit ¼ ð f ðpit Þ; Xit Þ and the vector parameter b*1 is a multiple of g *. 0 Testing g * ¼ 0 is then equivalent to testing that b* 1 ¼ 0. Under the null hypothesis, 1* it ¼ 1it, which implies that the Taylor series approximation does 10. Equations (1) and (2) in Gonza ´ lez, Tera ¨ svirta, and van Dijk (2005), with m ¼ 1 and only one transition variable, are equivalent to the formulation here, with Whigh(.) matching their g(.) and bhigh – blow corresponding to their b1. The Gonza ´ lez, Tera ¨ svirta, and van Dijk (2005) formulation is used for the calculations of the speci�cation tests described here. Espinoza, Leon and Prasad 125 not affect the asymptotic distribution theory. A LM test, robust for heteroske- dasticity, can then be computed as ~0^ ^ HACÀ1 W ~ S 10 0 W LMx ¼ ^ 10 ðA4Þ where ^10 is the vector of residuals obtained under the null hypothesis (the ~ is the vector xit f( pit) in equation 3 (after the residuals of the linear model), W �xed effects demeaning transformation) and S ^ HAC is the heteroskedastic-robust estimator of the covariance matrix: S ~ 0X ^ HAC ¼ ½ÀW ~ ðX~ 0X~ ÞÀ1 : Ik ŠD^ ½ÀW ~ 0X ~ 0X ~ ðX ~ ÞÀ1 : Ik Š 0 ; with X ðA5Þ ^¼ ~ i Š0 ½^ ~ i; W ½X 10 10 0 ~ ~ D i^ i Š½Xi ; W i Š i where X ~ i is the vector xit for country i in equation (A3) (after the �xed effects demeaning transformation), and k is the number of explanatory variables in the linear model. Under the null hypothesis LMx is asymptotically distributed as x2 (k). The test of remaining heterogeneity proposed by Gonza ¨ svirta, and ´ lez, Tera van Dijk (2005) was also computed. The test is based on the same principles as the test described above. The null hypothesis is now that the LSTR model with one transition is appropriate, and the LM statistics tests whether an additional element, bà 0 2 xit f ðpit Þ, can be added to the one-transition LSTR model. The details of the test are presented in section 3.3.2 of Gonza ¨ svirta, and ´ lez, Tera van Dijk (2005). REFERENCES Akerlof, G., W. Dickens, and G. Perry. 1996. “The Macroeconomics of Low Inflation.� Brookings Papers on Economic Activity 27: 1 –76. Anand, R., and E.S. Prasad. 2010. “Optimal Price Indices for Targeting Inflation under Incomplete Markets.� IMF Working Paper 10/200. International Monetary Fund, Washington, D.C. Aoki, K. 2001. “Optimal Monetary Policy Responses to Relative-Price Changes.� Journal of Monetary Economics 48: 55 –80. Ball, L. 1992. “Why does High Inflation Raise Inflation Uncertainty.� Journal of Monetary Economics 29: 371– 388. Barro, R. 1995. “Inflation and Growth.� NBER Working Paper W5326. Cambridge MA: National Bureau of Economic Research. ———. 2001. “Inflation and Growth.� In J. Rabin and G. L. Stevens, eds., Handbook of Monetary and Fiscal Policy. New York: Marcel Dekker. Bernanke, B. 2004. “The Great Moderation.� Remarks at the meetings of the Eastern Economic Association, February 20, Washington, D.C. Billi, R., and G. Kahn. 2008. “What Is the Optimal Inflation Rate?� Federal Reserve Bank of Kansas City, Economic Review Second Quarter: 5 –28. Blanchard, O., G. Dell’Ariccia, and P. Mauro. 2010. “Rethinking Macroeconomic Policy.� IMF Staff Position Note SPN/10/03. International Monetary Fund, Washington, D.C. 126 THE WORLD BANK ECONOMIC REVIEW Bruno, M., and W. Easterly. 1996. “Inflation and Growth: In Search of a Stable Relationship.� The Federal Reserve Bank of Saint Louis Review 78: 139– 146. Burdekin, R.C.K, A.T. Denzau, M.W. Keil, T. Sitthiyot, and T.D. Willett. 2004. “When Does Inflation Hurt Economic Growth? Different Nonlinearities for Different Economies.� Journal of Macroeconomics 26: 519–532. Catao, L., and R. Chang. 2010. “World Food Prices and Monetary Policy.� IMF Working Paper 10/ 161. International Monetary Fund, Washington, D.C. Chan, K.S. 1993. “Consistency and Limiting Distribution of the Least Squares Estimator of a Threshold Autoregressive Model.� The Annals of Statistics 21: 520–533. Davies, R. 1977. “Hypothesis Testing When a Nuisance Parameter Is Present Only under the Alternative.� Biometrika 64: 247– 254. Drukker, D., P. Gomis-Porqueras, and P. Hernandez-Verme. 2005. “Threshold Effects in the Relationship between Inflation and Growth: A New Panel Data Approach.� Texas A&M University, Department of Economics. Processed. Eggertsson, G., and M. Woodford. 2003. “The Zero Bound on Interest Rates and Optimal Monetary Policy.� Brookings Papers on Economic Activity 1: 139 –211. Fischer, S. 1993. “The Role of Macroeconomic Factors in Economic Growth.� Journal of Monetary Economics 32: 485–512. Ghosh, A., and S. Phillips. 1998. “Warning: Inflation May Be Harmful to Your Growth.� IMF Staff Papers 45: 672–710. ´ lez, A., T. Tera Gonza ¨ svirta, and D. van Dijk. 2005. “Panel Smooth Transition Regression Models.� SSE/EFI Working Paper Series in Economics and Finance 604. Stockholm School of Economics, Stockholm, Sweden. Hansen, B. 1999. “Threshold Effects in Non-dynamic Panels: Estimation, Testing, and Inference.� Journal of Econometrics 93: 345– 368. International Monetary Fund. 2002. World Economic Outlook. May. Washington, D.C.: International Monetary Fund. ———. 2008. World Economic Outlook. September. Washington, D.C.: International Monetary Fund. ———. 2010. World Economic Outlook. September. Washington, D.C.: World Economic and Financial Surveys. Khan, M. S., and A. S. Senhadji. 2001. “Threshold Effects in the Relationship between Inflation and Growth.� IMF Staff Papers 48: 1 –21. Leon, H. L., and S. Najarian. 2005. “Modeling Nonlinear Dynamics in Real Exchange Rates.� International Journal of Finance and Economics 10: 15–39. Luukkonen, R., P. Saikkonen, and T. Tera¨ svirta. 1988. “Testing Linearity against Smooth Transition Autoregressive Models.� Biometrika 75: 491– 499. McDonald, R., and D. Siegel. 1986. “The Value of Waiting to Invest.� Quarterly Journal of Economics 101: 707– 727. Mubarik, Y. A. 2005. “Inflation and Growth: An Estimate of the Threshold Level of Inflation in Pakistan.� State Bank of Pakistan Research Bulletin 1: 35 –44. Pollin, R., and A. Zhu. 2006. “Inflation and Economic Growth: A Cross-country Nonlinear Analysis.� Journal of Post Keynesian Economics 28: 593– 614. Rangarajan, C. 1997. “Dimensions of Monetary Policy.� Thirteenth Anantharamakrishnan Memorial Lecture, February 7, Chennai, India. Reifschneider, D., and J. Williams. 2000. “Three Lessons for Monetary Policy in a Low-inflation Era.� Journal of Money, Credit, and Banking 32: 936–966. Samantaraya, A., and A. Prasad. 2001. “Growth and Inflation in India: Detecting the Threshold.� Asian Economic Review 3: 414– 420. Sarel, M. 1996. “Nonlinear Effects of Inflation on Economic Growth.� IMF Staff Papers 43: 199– 215. Espinoza, Leon and Prasad 127 Schmitt-Grohe, S., and M. Uribe. 2004. “Optimal Fiscal and Monetary Policy under Sticky Prices.� Journal of Economic Theory 114: 198– 230. ¨ svirta, T. 1994. “Speci�cation, Estimation, and Evaluation of Smooth Transition Autoregressive Tera Models.� Journal of the American Statistical Association 89: 208–218. ———. 1998. “Modelling Economic Relationships with Smooth Transition Regressions.� In A. Ullah and D. E. Giles, eds., Handbook of Applied Economic Statistics. New York: Marcel Dekker. ¨ svirta, and P. H. Franses. 2002. “Smooth Transition Autoregressive Models—A Van Dijk, D., T. Tera Survey of Recent Developments.� Econometric Reviews 21: 1–47. Is Economic Volatility Detrimental to Global Sustainability? Yongfu Huang In a dynamic panel data model allowing for error cross-section dependence, output volatility is found to impede sustainable development. Through a �nancial develop- ment channel (liquidity liability ratio), output volatility exerts a signi�cant effect on depletion of natural resources, a key component of sustainability. Low-income countries, low energy-intensity countries, and low trade-share countries tend to be especially vulnerable to macroeconomic volatility and shocks. The �ndings highlight the interaction between global �nancial markets and the wider economy as a key factor influencing sustainable development, with important implications for macroeco- nomic and environmental policies in an integrated global green economy. JEL codes: E32, O11, O16 The world economy grew at an average annual rate of 2.7 percent in the 1990s and then performed exceptionally well over 2000–08, growing at an average annual rate of 3.2 percent a year (World Bank 2010b). Following these long economic boom periods, the world economy, especially �nancial markets, experienced a period of uncertainty, volatility, and severe crises. At the same time, climate change, in large part due to human activities (IPCC 2001), began to emerge clearly as the greatest global challenge of our age. Between 1981 and 2005, some 60 percent of the world’s ecosystems became degraded or were being unsustainably exploited (Barbier 2009). This suggests an important question: Could economic volatility lead countries down an unsustainable path? Against a background of multiple crises—climate, fuel, food—the global �nancial crisis of 2007–09 has caused enormous damage to the world Yongfu Huang (yh279@cam.ac.uk) is a Tyndall Research Fellow at the University of Cambridge. The author is grateful to Philip Arestis, Unai Pascual, Vasilis Sara�dis, the journal editor, and three anonymous referees for constructive comments and suggestions. He also thanks two Tyndall Centre referees and seminar participants at the Department of Land Economy of Cambridge University for helpful comments and discussions. A supplemental appendix to this article is available at http://wber. oxfordjournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 128– 146 doi:10.1093/wber/lhr042 Advance Access Publication October 27, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 128 Huang 129 economy, resulting in the most severe global recession in generations. The �nancial crisis spread rapidly around the globe. Nearly all stock markets experienced bursts of volatility. Lin (2009, p. 2) points out that the current economic downturn is “possibly turning a short-run macroeconomic adjust- ment into a long-term development problem.� But empirical evidence on the impact of economic volatility on long-run sustainable development remains sparse. This article looks at whether economic volatility has a damaging effect on sustainable development, as measured by changes in genuine savings or adjusted net savings (net national savings minus natural resource depletion). Research on the link between volatility and growth indicates a signi�cant impact of volatility on savings through growth. To determine whether the link between volatility and sustainability reflects more than the link between vola- tility and savings, this article goes a step further to examine whether output volatility affects natural resource depletion. A dynamic panel data study was conducted with data for 128 countries over 1979–2008. To address the effects of global shocks in recent decades—driven by the increases in international trade and private capital flows—that could cause error dependence across countries, the study considers a common factor structure in the error term. More speci�cally, it applies the system of general- ized method of moments (GMM), adjusted to allow for error cross-section dependence (Sara�dis and others 2009). The study �nds that output volatility exerts a strong negative impact on genuine savings. The impact is strongest in low-income countries, low energy-intensive countries, and low trade-share countries. In addition to the signi�cant impact of volatility on savings suggested in the literature, it �nds that the negative effect of output volatility on genuine savings is due to the positive impact of output volatility on natural resource depletion through a �nancial development channel (liquidity liability ratio). The �ndings highlight the role of the interaction between global �nancial markets and the wider economy in promoting global sustainability. This article contributes to the literature in several dimensions. First, it explores the effects of output volatility on both genuine savings and the depletion of natural resources. Second, it allows for the possibility of error cross-section dependence and tries to correct for reverse causality and unob- served country-speci�c effects. Third, it considers macroeconomic and environ- mental policies in an integrated global green economy in which development strategies and programs take account of the state of natural resources (energy, forests, minerals, soils, freshwater, and �sheries) on which future growth depends. Section I reviews the literature. Section II describes the data and outlines the methodology of GMM estimation without and with cross-section dependence. Section III details the empirical results. Section IV presents some implications of the �ndings. 130 THE WORLD BANK ECONOMIC REVIEW I. THE THEORETICAL AND EMPIRICAL RESEARCH ON V O L AT I L I T Y, G ROW T H, AND SU STAIN ABI LIT Y Volatility measures the variation or movement in an economic variable such as the growth rate, usually by a standard deviation of the variable over some period. Volatility has declined in general over recent decades, but as an inde- pendent research area, it has moved from a second-order concern to “a central position in development economics� (Aizenman and Pinto 2005, p. 2). There is considerable research on the link between economic volatility and long-run growth, along with considerable controversy.1 Since high growth performance does not necessarily lead to high levels of development, policymakers have increasingly emphasized sustainable development as the primary national objective, especially for developing countries. Sustainable development—a different kind of growth that preserves the environment—is about economic growth together with environmental protection, with the two reinforcing each other.2 It is thus important to understand the impact of economic volatility on sustainable development. The research references the widely used sustainability indicator of genuine savings. Based on standard national accounting conventions, genuine savings considers depletion of natural resources, pollution damage, and investments in human capital. It is the true savings rate in an economy in terms of creating and maintaining total net wealth, which is inclusive of manufactured capital, human capital, and natural capital. In recent years, considerable research has tried to explain differences in genuine savings rates across countries, identifying several variables, such as per capita gross national income, per capita GDP growth rate, age dependency, and urbanization (Dietz and others 2007). Research indicates a signi�cant impact, positive or negative, of volatility on savings through growth; however, the impact of volatility on natural resources depletion, a key element of genuine savings, is not yet well understood. It is still not clear whether the impact of volatility on genuine savings is due to the impact of volatility on natural resources depletion. If output volatility does play a role in natural resources depletion, what are the channels? Potentially, volatility could accelerate natural resources depletion and environmental degradation through �nancial markets, trade, or investment; 1. One line of theoretical research �nds that volatility is positively related to growth (Sandmo 1970; Ghosh and Ostry 1997; Canton 2002). Another line of research supports a negative impact of volatility on growth (Kharroubi 2007; Aysan 2007). Empirically, while Kormendi and Meguire (1985) and Grier and Tullock (1989), among others, �nd evidence that output volatility promotes growth, most cross country studies suggest that economic volatility hurts long-run growth (Ramey and Ramey 1995; Hnatkovska and Loayza 2005; Loayza and others 2007; Koren and Tenreyro 2007). 2. According to the World Commission on Environment and Development (1987, ch. 2), sustainable development is “a development path that meets the needs of the present without compromising the ability of future generations to meet their own needs.� It has typically been regarded as having three dimensions or pillars—environmental, social and economic sustainability—which are not mutually exclusive and can be mutually reinforcing (United Nations 2005). Huang 131 however, these channels have not been carefully examined.3 For �nancial markets, generally speaking, global economic volatility is likely to increase credit risk and uncertainty. The loss of con�dence means that investors are likely to refuse to provide funds to banks, while the banks tend to ration bor- rowers by providing less credit than requested or restricting loan maturity.4 The credit crunch reduces investment in energy sectors and low-carbon projects and discourages switching to low-carbon and renewable energy technologies that conserve natural resources and are less environmentally destructive but that are typically more costly than conventional technologies (OECD/IEA 2009). With companies unable to fund capital-intensive offset projects in either regulated or voluntary carbon markets, the credit crunch slows development of global carbon markets, which are designed to achieve the dual objectives of sustainable development and emission reductions.5 The credit crunch also leads to severe funding cuts for environmental agencies and initiatives to tackle environmental problems such as deforestation (Kittiprapas 2002; Kasa and Næss, 2005). Finally, the curtailed availability of credit and insurance also reduces trade and investment flows in the natural resources sectors (Committee of Ten 2009). The economic crisis and volatility could also affect natural resources depletion through unemployment and weaker enforcement of environmental law. Dauvergn (1999) shows that the lost jobs and falling income associated with �nancial crisis could trigger greater natural resources extraction as people in poor and densely populated countries revert to ecological commons such as �sheries and forests to secure basic subsistence. With detailed survey data, Gaveau and others (2009) �nd that the 1997–98 crisis caused a reversal in law enforcement efforts in Indonesia, resulting in substantial losses of “protected� forests and biodiversity. This study aims to contribute to this emerging line of research. I I. DATA AND ME T H O D O LO GY This study examines whether economic volatility has signi�cant impacts on sustainable development, controlling for per capita GDP growth, per capita gross national income (GNI), and the age-dependency ratio (ratio of 3. Several transmission channels for the effect of volatility on growth have been studied. Wolf (2005) discusses factor accumulation, domestic �nance, trade, capital mobility, and political institutions. In addition, Ramey and Ramey (1995) emphasize the level of investment, while Aysan (2007) supports the productivity of investment rather than its level. 4. Huang (2011, p. 45) provides evidence that output volatility (and volatility of the black market premium) has a negative impact on the total size of the �nancial system (banking sector plus stock market) in the developing world. 5. Hamilton and others (2010) indicate that, due to the �nancial crisis, global voluntary carbon markets transactions declined 26 percent and their total value dropped 47 percent in 2009 compared with 2008. 132 THE WORLD BANK ECONOMIC REVIEW dependents to working-age population).6 Appendix Table S1 in the supplemen- tal appendix (available at http://wber.oxfordjournals.org/) describes the variables and lists their sources. The Data The �rst dependent variable is “genuine savings,� or adjusted net saving excluding particulate emission damage,7 (GENSAV) as a percent of GNI. (The data for all variables, unless otherwise noted, are from World Bank 2010b.) The regressions use three-year averages over 1979–2008 of log(1 þ GENSAV/ 100). The second dependent variable is “resource depletion� (DEPLETION), or the value of net national savings plus education expenditure, minus adjusted net saving and carbon dioxide damage. The regressions use three-year averages over 1979–2008 of log(1 þ DEPLETION/100). This analysis focuses mainly on “output volatility� (VGR), de�ned as the standard deviation for three-year intervals over 1979–2008 of log(1 þ GR/ 100). The analysis controls for “per capita GDP growth� (GR), “per capita GNI� (GNIPC), and “age-dependency ratio� (AGE). The analysis uses three-year averages over 1979–2008 of log(1 þ GR/100), log(1 þ GNIPC/100), and log(1 þ AGE/100). Two potential channels of transmission are investigated: “liquidity liability ratio� (LLY) and “investment� (KI). LLY, the �nancial channel, measures the liquid liabilities of banks and nonbank �nancial intermediaries (currency plus demand and interest-bearing liabilities) as a share of GDP. It reflects the size of �nancial intermediaries relative to the economy, including the central bank, deposit money banks, and other �nancial institutions. The analysis uses log(100  LLY). KI, the investment channel, is the percentage share of invest- ment in real GDP per capita (RGDPL). The regressions use log(RGDPL) and log(KI). The data are from Heston, Summers, and Aten (2009). The sample contains data for 128 nontransition economies over 1979–2008 (listed in table S2 in the supplemental appendix), with non-overlapping three- year periods, so that each country has a maximum of 10 observations. Countries with fewer than 10 annual observations are excluded. The 6. Although the life expectancy ratio and the urbanization rate were also considered as controlling variables, and energy use per capita and �nal energy intensity as potential channels, no evidence was found for them. Data on life expectancy at birth (total years), urban population (percent of total), and energy use per capita (kilograms of oil equivalent per capita) are from World Bank (2010b). Data on �nal energy intensity are from Enerdata (2010). 7. The adjusted net savings series is equal to net national savings (gross savings less depreciation of produced assets) plus investment in human capital (education expenditure) minus resource depletion (energy depletion, mineral depletion, net forest depletion) and environmental degradation (carbon dioxide). Since there are more missing values when adjusted net savings includes particulate emission damage, this analysis uses adjusted net savings excluding particulate emission damage. Huang 133 regressions include dummy variables for 31 low-income countries, 63 low- and middle-income countries, 65 countries whose average �nal energy intensity of GDP at purchasing power parity over the period is below the median (“lower energy intensity countries�), and 70 countries whose average trade share as a percent of GDP over the period is below the median (“lower trade-share countries�). Income classi�cations are from the World Bank (2011). Data on �nal energy intensity of GDP are from Enerdata (2010). Methodology This analysis of the impact of economic volatility on sustainable development under globalization employs the GMM estimator adjusted to allow for error cross-section dependence, recently proposed by Sara�dis, Yamagata, and Robertson (2009) for a linear dynamic panel model. In recent decades, cross-country dependence has become an important phenomenon in an increasingly globalized world where common factors— macroeconomic shocks, common technological shocks, and environmental shocks—can cause strong interactions in the world economy. The following model allows for error cross-section dependence: DEPVARit ¼gi þ aDEPVARi;tÀ1 þ b1 VGRi;tÀ1 þ b2 GRi;tÀ1 þ b3 GNIPCi;tÀ1 þ b4 AGEit þ l0i ft þ vit ð1Þ i ¼ 1; 2; � � � ; 128; t ¼ 2; � � � 10 where DEPVAR denotes the dependent variable, GENSAV or DEPLETION, and g is individual effects. The autoregressive coef�cient a is assumed to lie inside the unit circle, j aj , 1, to ensure model stability. The coef�cients b1 to b4 reflect the existence and direction of any speci�c effect on GENSAV or DEPLETION. The term ft is a (r  1) vector of unobserved time-varying common factors assumed to be nonstochastic and bounded, and li is a vector of factor loadings assumed to be independent and identically distributed, such 0 that li ft ¼ li1 ft1 þ li2 ft2 þ . . . þ lir ftr (here r is the number of common factors).8 The error term vit is the transitory disturbance term, which is assumed to be independently distributed with zero mean and �nite variance. It is also assumed to be uncorrelated with individual effects and common factors, but correlations are possible between either individual effects or common factors ( ft and subsequent shocks) and the regressors. The explanatory variables VGRi,t21, GRi,t21, and GNIPCi,t21 are predeter- mined with respect to vit and so may be correlated with vi,t21 and earlier shocks, but they are uncorrelated with vit and subsequent shocks. The 8. Bai (2009) suggests an interactive effects model that includes the interaction between factors, ft, and factor loadings, li, which is more general than an additive effects model, the traditional one-way 0 0 0 model, or a two-way �xed effects model. Taking r ¼ 2 gives ft ¼ (1 ht) , li ¼ (ai 1), and li ft ¼ ai þ ht, where ai is the individual effect and ht is the time effect. 134 THE WORLD BANK ECONOMIC REVIEW assumption of predetermination for the explanatory variables other than AGEit rules out the possibility of reverse causality or joint determination. The �rst-differences of equation (1) are DDEPVARit ¼ aDDEPVARi;tÀ1 þ b1 DVGRi;tÀ1 þ b2 DGRi;tÀ1 þ b3 DGNIPCi;tÀ1 þ b4 DAGEit þ l0i Dft þ Dvit ð2Þ i ¼ 1; 2; � � � ; 128; t ¼ 3; � � � 10 where DDEPVARit ¼ DEPVARit 2 DEPVARi,t21, which applies to DAGEit DVGRi,t21 ¼ VGRi,t21 2 VGRi,t22, which also applies to DGRi,t21 and D GNIPCi,t21 D ft ¼ ft 2 ft21 and D vit ¼ vit 2 vi,t21. When common factors are assumed to have an identical effect on each cross- section unit, several methods have been proposed using �rst differencing to eliminate the individual effects when estimating dynamic panel data models with a short time dimension. Arellano and Bond (1991) propose the �rst- differenced GMM estimator (DIF-GMM).9 For simplicity, yit denotes DEPVARit and xit is a vector of independent variables (VGRit, GRit, GNIPCit, AGEit). The moment conditions for errors in differences on which the DIF-GMM estimator in this application is based can be written as " ! # yt i À2 E ðDvit Þ ¼ 0 xitÀ2 ð3Þ t ¼ 3; � � � 10 0 0 where yti 22 ¼ ( yi1, yi2, . . . yi,t22) and xti 22 ¼ (xi1, xi2, . . . xi,t22) . The DIF-GMM estimator has been found to suffer from the weak- instruments problem associated with highly persistent data. To address this issue, Arellano and Bover (1995) and Blundell and Bond (1998) developed a system GMM estimator (SYS-GMM) by considering a mean stationarity assumption on initial conditions. In addition to the moments for errors in differences described above, the SYS-GMM estimator is also based on the additional moments for errors in levels as follows:10   ! Dyi;tÀ1 E ðgi þ vit Þ ¼ 0 Dxi;tÀ1 ð4Þ t ¼ 3; � � � 10: However, in reality, common factors typically have a differential effect across 9. DIF-GMM uses all lagged values of the dependent variable and independent variables from t – 2 and earlier as suitable instruments for the differenced values of the original regressors; for example, D DEPVARi,t21, D VGRi,t21, D GRi,t21, D GNIPCi,t21, and D AGEit in this context. 10. The additional mean stationarity condition of ( yit, xit) enables use of the lagged �rst differences of the series ( yit, xit) dated t – 1 as instruments for the untransformed equations in levels. Huang 135 cross-sectional units, causing heterogeneous error cross-section dependence. Sara�dis and Robertson (2009) show that the standard DIF-GMM and SYS-GMM estimators are not consistent in the presence of heterogeneous error cross-section dependence; the standard instruments these estimators rely on for lagged values of the dependent variable, in levels or �rst differences, are invalid. Under the assumption of heterogeneous error cross-section dependence, Sara�dis, Yamagata, and Robertson (2009) suggest a consistent �rst-differenced GMM estimator (DIF-GMM-C) and a consistent system GMM estimator (SYS-GMM-C). These two GMM estimators rely on partial instruments consist- ing of the regressors. More speci�cally, based on partial moment condition (5) below, the DIF-GMM-C estimator is consistent under the assumption of hetero- geneous error cross-section dependence. This also applies to the SYS-GMM-C estimator, which is based on partial moment conditions (5) and (6):  À2 à E xti ðDvit Þ ¼ 0 ð5Þ t ¼ 3; � � � 10:  à E Dxi;tÀ1 ðgi þ vit Þ ¼ 0: ð6Þ Sara�dis, Yamagata, and Robertson (2009) have proposed a new test for detect- ing error cross-section dependence in a linear dynamic panel model (CSD test). Under the null hypothesis of homogeneous error cross-section dependence, the CSD test enables examination of whether any error cross-section dependence remains after time dummy variables are included.11 The �nite sample simulation-based results in Sara�dis, Yamagata, and Robertson show the good performance of the CSD test, especially for the version based on a system GMM estimator. I II. EM P I R I CA L RE S U LTS This section presents the econometric evidence on the impact of output vola- tility on genuine savings and its key element, natural resources depletion, and examines the channels through which volatility affects natural resources depletion. A �nite sample correction is made to the two-step covariance matrix using the method developed by Windmeijer (2005) for both the �rst-differenced 11. The CSD test is the Sargan’s difference test based on either the two-step �rst-differenced GMM estimator or two-step system GMM estimator. The Sargan’s difference test statistic based on the two-step �rst-differenced GMM estimator is the difference between the Sargan statistic for DIF-GMM with a standard set of moment conditions (condition 3) and the Sargan statistic for DIF-GMM-C using a restricted set of moment conditions (condition 5). The Sargan’s difference test statistic based on the two-step system GMM estimator is the difference between the Sargan statistic for SYS-GMM with a standard set of moment conditions (conditions 3 and 4) and the Sargan statistic for SYS-GMM-C using a restricted set of moment conditions (conditions 5 and 6). 136 THE WORLD BANK ECONOMIC REVIEW GMM estimators and the system GMM estimators under either assumption. This analysis uses only lagged values of yit and xit from t – 2 to t – 4 as instruments.12 For any GMM estimators, three speci�cation tests are conducted to address consistency. The �rst two are serial correlation tests, which test the null hypothesis of no �rst-order serial correlation (M1) and no second-order serial correlation (M2) in the residuals in the �rst-differenced equation. Given that the errors in levels are serially uncorrelated, a signi�cant �rst-order serial corre- lation would be expected, but an insigni�cant second-order correlation would be expected in the �rst-differenced residuals. The third speci�cation test is a Sargan test of overidentifying restrictions, which examines the overall validity of the instruments by comparing the moment conditions with their sample ana- logue. For SYS-GMM and SYS-GMM-C, an additional test is carried out. The difference Sargan (Diff-Sargan) test examines the null hypothesis that the lagged differences of the explanatory variables are uncorrelated with the errors in the levels equations, as in Blundell and Bond (1998).13 Once the model is well-speci�ed, various tests are used to further examine its properties. A Granger causality test is conducted to examine whether genuine savings or natural resources depletion is Granger-caused by economic volatility.14 The point estimates for the long-run effect are also reported with the standard errors approximated using the delta method. However, since the long-run effect is calculated as a nonlinear function of the model parameters, this effect may be imprecisely estimated. To test for unobserved individual effects, a heterogeneity test for the null hypothesis of no heterogeneity is conducted.15 Finally, the new CSD tests based on the �rst-differenced GMM estimators and system GMM estimators are carried out, as explained above. The Impacts of Volatility on Sustainability GENUINE SAVINGS. Did output volatility lead to economic unsustainability during 1979–2008? This section �rst reports the evidence for the whole sample of 128 countries and then for the model that includes the interaction terms between output volatility and the dummy variables for low-income countries, lower energy-intensity countries, or lower trade-share countries to distinguish the volatility effects on genuine savings across country groups. 12. To avoid the possible over�tting bias associated with the full Arellano and Bond (1991) instrument set, Bowsher (2002) proposes selectively reducing the number of moment conditions for each �rst-differenced equation. 13. The statistic, called an incremental Sargan test statistic, is the difference between the Sargan statistics for DIF-GMM (or DIF-GMM-C) and the Sargan statistic for SYS-GMM (or SYS-GMM-C). 14. The Sargan statistics for the unrestricted and the restricted models, with the same moment conditions, are compared using an incremental Sargan test statistic, which is asymptotically distributed as x2g, where g is the number of restrictions. 15. The test also uses an incremental Sargan test statistic, the difference between the Sargan statistic for DIF-GMM (or DIF-GMM-C) and the Sargan statistic for SYS-GMM (or SYS-GMM-C), where the lagged levels are used as instruments in the levels equations. Huang 137 T A B L E 1 . Output Volatility and Genuine Savings, 1979–2008 (Dependent Variable: Genuine Savings) Cross-sectional independence Cross-sectional dependence Variable or test DIF-GMM SYS-GMM DIF-GMM-C SYS-GMM-C Lag 1 genuine savings 0.207 0.584 – 0.525 0.013 (0.406) (0.000)*** (0.358) (0.983) Lag 1 output volatility – 0.175 – 0.252 0.286 – 0.191 (0.360) (0.010)** (0.557) (0.650) Lag 1 per capita GDP growth 0.061 – 0.038 0.511 0.043 (0.715) (0.753) (0.462) (0.890) Lag 1 per capita GNI – 0.012 – 0.021 0.030 – 0.043 (0.765) (0.311) (0.779) (0.291) Age dependency – 0.024 – 0.208 – 0.172 – 0.415 (0.837) (0.022)** (0.318) (0.033)** M1 serial correlation test ( p-value) 0.10 0.01 0.96 0.52 M2 serial correlation test ( p-value) 0.56 0.88 0.22 0.65 Sargan test ( p-value) 0.33 0.84 0.26 0.32 Diff-Sargan test ( p-value) 1.00 0.46 Granger causality test ( p-value) 0.23 0.03 0.34 0.10 Long-run effect – 0.22 – 0.61 0.19 – 0.19 (0.29) (0.28) (0.25) (0.54) Heterogeneity test ( p-value) 0.00 0.01 CSD test ( p-value) 0.48 1.00 Number of observations 896 1,022 896 1,022 * Signi�cant at the 10 percent level; ** signi�cant at the 5 percent level; *** signi�cant at the 1 percent level. Note: Numbers in parentheses are the robust p-values. Data are for 128 nontransition countries over 1979– 2008. See text for variables de�nitions and data sources and for test descriptions. Source: Author’s analysis based on data described in text. The CSD tests based on �rst-differenced GMM estimators and system GMM estimators fail to reject the null hypothesis of homogeneous error cross-section dependence (table 1), suggesting that the assumption of cross-sectional inde- pendence is appropriate for this context. The focus is naturally on the DIF-GMM and SYS-GMM estimates. The speci�cation tests indicate that the two models are well speci�ed. The hypothesis of no �rst-order serial corre- lation can be rejected but not the hypothesis of no second-order serial corre- lation. The Sargan tests do not signal that the instruments used in the models are invalid. The Diff-Sargan tests for SYS-GMM cannot reject the null hypoth- esis of the additional moment conditions being valid, implying that SYS-GMM is a more reliable estimator than the DIF-GMM in this context. Results for the SYS-GMM estimator suggest that the impact of output volatility on genuine savings is statistically signi�cant and negative. The effect of the age-dependency ratio is also negative at the 5 percent signi�cance level. The Granger Causality test con�rms the negative effect of volatility, rejecting 138 THE WORLD BANK ECONOMIC REVIEW the null hypothesis at a 3 percent signi�cance level. The long-run effect is greater than the short-run effect. Under certain circumstances, the CSD test might lack power because it is based on the overidentifying restrictions test statistic, so the results should be interpreted cautiously. The following section examining transmission channels provides evidence of such cross-section dependence. In principle, the �rst-differenced GMM and system GMM estimates impose homogeneity on all slope coef�cients, under either the cross-sectional indepen- dence or cross-sectional dependence assumption. One concern with GMM esti- mates is that these parameters may be heterogeneous across countries, a contention that is con�rmed by the heterogeneity test, which clearly rejects the null hypothesis. This concern is addressed by separately including the interaction term between output volatility and the dummy variables for low-income countries, lower energy-intensity countries, and lower trade-share countries (table 2). For these regressions, the speci�cation tests (M1, M2, and Sargan) indicate that all six models are well-speci�ed. Diff-Sargan tests further show that SYS-GMM is a more reliable estimator than DIF-GMM for this case. The SYS-GMM estimates con�rm the strong negative impact from output volatility to genuine savings. The volatility effect is smaller for lower energy-intensity countries than for the whole sample and for other country groups. Thus, there is evidence that output volatility is indeed an impediment to global sustainability as measured by genuine savings, especially for low-income countries, lower energy-intensity countries, and lower trade-share countries. This �nding points to possible damaging consequences of output volatility for the economy as a whole, consistent with what happened during the global �nancial crisis of 2007–09. A global credit crunch triggered a sustained period of stress and instability in global �nancial markets and the worst global recession in generations. However, the negative impact of volatility on genuine savings (net national savings minus natural resources depletion) is likely driven by the negative impact of volatility on growth and savings, as widely examined in the litera- ture. The question then is whether output volatility depletes natural resources or causes environmental degradation as it impedes growth and savings. DEPLETION OF NATURAL RESOURCES. Table 3 reports evidence on resource depletion for the entire sample of 128 countries for the model with no inter- action terms. Based on the �rst-differenced GMM estimator, the CSD test clearly rejects the null hypothesis of homogeneous cross-sectional dependence. According to Sara�dis and Robertson (2009), the standard DIF-GMM estima- tor is not consistent in the presence of heterogeneous error cross-section depen- dence. As expected, Sargan’s test for DIF-GMM rejects the null hypothesis while Sargan’s test for DIF-GMM-C fails to reject the null hypothesis and the CSD test rejects the null hypothesis. However, for DIF-GMM-C, the serial correlation test M1 cannot reject the null hypothesis, suggesting that DIF-GMM-C is not reliable. T A B L E 2 . Output Volatility and Genuine Savings, 1979–2008, with Interaction Terms (Dependent Variable: Genuine Savings) Lower energy-intensity dummy Lower trade-share dummy Low-income dummy variablea variableb variablec Variable or test DIF-GMM SYS-GMM DIF-GMM SYS-GMM DIF-GMM SYS-GMM Lag 1 genuine savings 0.240 0.646 0.217 0.606 0.164 0.580 (0.287) (0.000)*** (0.351) (0.000)*** (0.421) (0.000)*** Lag 1 output volatility  dummy variable – 0.214 –0.270 –0.009 – 0.147 – 0.183 – 0.245 (0.423) (0.108) (0.932) (0.049)** (0.399) (0.060)* Lag 1 per capita GDP growth 0.134 0.030 0.132 0.018 0.080 – 0.021 (0.310) (0.802) (0.400) (0.885) (0.632) (0.863) Lag 1 per capita GNI 0.006 –0.007 –0.006 – 0.024 – 0.001 – 0.024 (0.874) (0.684) (0.898) (0.246) (0.977) (0.250) Age dependency – 0.055 –0.136 –0.024 – 0.212 – 0.072 – 0.216 (0.582) (0.093)* (0.841) (0.022)** (0.537) (0.014)** M1 serial correlation test ( p-value) 0.08 0.01 0.09 0.01 0.09 0.01 M2 serial correlation test ( p-value) 0.53 0.95 0.51 0.93 0.42 0.91 Sargan test ( p-value) 0.44 0.83 0.38 0.80 0.16 0.70 Diff-Sargan test ( p-value) 1.00 1.00 1.00 Granger causality test ( p-value) 0.31 0.13 1.00 0.34 0.36 0.09 Long-run effect – 0.28 –0.76 –0.01 – 0.37 – 0.22 – 0.58 (0.36) (0.51) (0.14) (0.24) (0.28) (0.33) Heterogeneity test ( p-value) 0.01 0.03 0.04 CSD test ( p-value) 0.60 1.00 0.65 1.00 0.49 1.00 Number of observations 896 1,022 896 1,022 896 1,022 * Signi�cant at the 10 percent level; ** signi�cant at the 5 percent level; *** signi�cant at the 1 percent level. Note: Numbers in parentheses are the robust p-values. See text for variables de�nitions and data sources and for test descriptions. a. Includes 31 low-income countries. b. Includes 65 lower energy-intensity countries whose average �nal energy intensity of GDP over 1979– 2008 are below the median value. Huang c. Includes 70 lower trade-share countries whose average trade shares (as a percent of GDP) over 1979– 2008 are below the median value. Source: Author’s analysis based on data described in text. 139 140 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Output Volatility and Resource Depletion, 1979–2008 (Dependent Variable: Resource Depletion) Cross-sectional independence Cross-sectional dependence Variable or test DIF-GMM SYS-GMM DIF-GMM-C SYS-GMM-C Lag 1 Resource Depletion – 0.023 – 0.418 0.058 – 0.077 (0.915) (0.006)*** (0.496) (0.623) Lag 1 Output Volatility 0.051 0.195 0.044 0.082 (0.470) (0.055)* (0.541) (0.253) Lag 1 per capita GDP Growth 0.038 0.102 0.018 0.024 (0.804) (0.343) (0.882) (0.761) Lag 1 per capita GNI 0.002 – 0.046 0.002 – 0.029 (0.940) (0.008)*** (0.926) (0.095)* Age dependency – 0.026 – 0.272 – 0.054 – 0.104 (0.654) (0.001)*** (0.342) (0.196) M1 serial correlation test ( p-value) 0.70 0.09 0.62 0.91 M2 serial correlation test ( p-value) 0.50 0.88 0.18 0.88 Sargan test ( p-value) 0.01 0.59 0.12 0.57 Diff-Sargan test ( p-value) 1.00 1.00 Granger causality test ( p-value) 0.31 0.17 0.17 0.11 Long-run effect 0.05 0.14 0.05 0.08 (0.06) (0.07) (0.07) (0.06) Heterogeneity test ( p-value) 0.00 0.00 CSD test ( p-value) 0.01 0.46 Number of observations 895 1,021 895 1,021 * Signi�cant at the 10 percent level; ** signi�cant at the 5 percent level; *** signi�cant at the 1 percent level. Note: Numbers in parentheses are the robust p-values. Data are for 128 nontransition countries over 1979–2008. Resource depletion is measured by three-year averages of the sum of energy depletion, mineral depletion, and net forest depletion, as de�ned in the text. See text for variables de�nitions and data sources and for test descriptions. Source: Author’s analysis based on data described in text. Based on system GMM estimators, the CSD test cannot reject the null hypothesis, suggesting that SYS-GMM rather than SYS-GMM-C is the appro- priate estimator for this context. For SYS-GMM, serial correlation tests M1 and M2 indicate that the null hypothesis of no �rst-order serial correlation can be rejected, but that the null hypothesis of no second-order serial correlation cannot be rejected. The Sargan tests cannot reject the null hypothesis that the instruments in the model are valid. The Diff-Sargan test cannot reject the null hypothesis that the additional moment conditions are valid, supporting SYS-GMM as the more reliable estimator for this context. The SYS-GMM esti- mates provide strong evidence of a positive effect of output volatility on natural resources depletion in the global economy. They also provide evidence that per capita GNI and the age-dependency ratio, both signi�cant at the 1 percent level, are important variables. Huang 141 T A B L E 4 . Output Volatility and Resource Depletion, 1979–2008, with Interaction Terms (Dependent Variable: Resource Depletion) Lower income dummy Lower energy-intensity variablea dummy variableb Variable or test DIF-GMM SYS-GMM DIF-GMM SYS-GMM Lag 1 resource depletion – 0.031 – 0.412 – 0.058 – 0.433 (0.823) (0.012)** (0.791) (0.010)*** Lag 1 output volatility  dummy variable 0.030 0.124 0.034 0.179 (0.408) (0.062)* (0.618) (0.052)* Lag 1 per capita GDP growth 0.043 0.066 0.047 0.053 (0.581) (0.522) (0.724) (0.582) Lag 1 per capita GNI 0.007 – 0.048 0.005 – 0.050 (0.761) (0.009)*** (0.791) (0.008)*** Age dependency – 0.031 – 0.288 – 0.036 – 0.280 (0.601) (0.001)*** (0.531) (0.001)*** M1 serial correlation test (p-value) 0.67 0.13 0.83 0.11 M2 serial correlation test (p-value) 0.46 0.78 0.54 0.63 Sargan test (p-value) 0.02 0.53 0.01 0.42 Diff-Sargan test (p-value) 1.00 1.00 Granger causality test (p-value) 0.14 0.60 0.55 0.40 Long-run effect 0.03 0.09 0.03 0.13 (0.03) (0.04) (0.06) (0.06) Heterogeneity test (p-value) 0.00 0.00 CSD test (p-value) 0.01 0.30 0.03 0.25 Number of observations 895 1,021 895 1,021 * Signi�cant at the 10 percent level; ** signi�cant at the 5 percent level; *** signi�cant at the 1 percent level. Note: Numbers in parentheses are the robust p values. See text for variables de�nitions and data sources and for test descriptions. a. Includes 31 low-income countries and 63 middle income countries. b. Includes 65 lower energy-intensity countries whose average �nal energy intensity of GDP over 1979– 2008 are below the median value. Source: Author’s analysis based on data described in text. Table 4 looks at whether these �ndings are robust to the inclusion of inter- action terms between output volatility and the lower income countries dummy variable or the lower energy-intensity dummy variable. The speci�cation tests (M1, M2, and Sargan) indicate that the SYS-GMM and SYS-GMM-C models are well-speci�ed. The Diff-Sargan test further shows that SYS-GMM is a more reliable estimator than the DIF-GMM for this case. The SYS-GMM estimates provide evidence for a strong positive impact from output volatility to natural resources depletion. Per capita GNI and the age-dependency ratio also have a statistically signi�cant negative impact on natural resources depletion. The effect of volatility on natural resources depletion is smaller for the lower income countries than for the whole sample and for the lower energy-intensity countries. 142 THE WORLD BANK ECONOMIC REVIEW In summary, output volatility has damaging effects on genuine savings, likely due to the positive impact on natural resources depletion in addition to the negative impact of output volatility on net national savings suggested in the literature. The signi�cant effects are greater in low-income countries, lower energy-intensity countries, and lower trade-share countries, suggesting that these countries are especially vulnerable to macroeconomic shocks. The results are robust to the use of alternative estimation methods and are not due to unobserved heterogeneity or reverse causality. The Channels What are the underlying channels through which volatility affects sustainable development? This section focuses on natural resources depletion as the depen- dent variable. Table 5 presents evidence on whether output volatility affects natural resources depletion through either �nancial development or investment share. The �nancial development channel considered is the liquidity liability ratio. Based on �rst-differenced GMM and system GMM estimators, the CSD tests clearly reject the null hypothesis of homogeneous cross-sectional dependence, suggesting that DIF-GMM-C and SYS-GMM-C are consistent estimators, unlike DIF-GMM and SYS-GMM. For DIF-GMM-C, serial correlation tests M1 and M2 indicate that the null hypothesis of no �rst-order serial correlation can be rejected but the hypothesis of no second-order serial correlation cannot. The Sargan tests cannot reject the null hypothesis that the instruments in the model are valid. For SYS-GMM-C, M1 cannot reject the null hypothesis. Thus the analysis focuses on the DIF-GMM-C estimates. The DIF-GMM-C estimates clearly indicate that output volatility is no longer signi�cant while the liquidity liability ratio is signi�cant. This result is supported by the Granger causality test for the effect of output volatility, which cannot reject the null hypothesis. Apparently, the liquidity liability ratio picks up the effect of output volatility on natural resources depletion and is the channel through which output volatility induces natural resources depletion, thereby impairing sustainability. What about the investment ratio channel? With the inclusion of the invest- ment ratio, the DIF-GMM, DIF-GMM-C, and SYS-GMM models are no longer well-speci�ed. The speci�cations tests are in general satisfactory for the SYS-GMM estimator. Based on �rst-differenced GMM and system GMM esti- mators, the CSD test clearly rejects the null hypothesis, further suggesting that SYS-GMM is a consistent estimator. Investment share is insigni�cant, while output volatility continues to be signi�cant. There is no evidence that the investment ratio picks up the effect of output volatility on natural resources depletion, but this does not necessarily rule out the possibility that the invest- ment ratio is a channel for output volatility to affect sustainability because several of the models are not well speci�ed when the investment ratio is added. Further research is needed. T A B L E 5 . Channels through Which Output Volatility Affects Resource Depletion, 1979–2008 (Dependent Variable: Resource Depletion) Financial development channel Investment channel Variable or test DIF-GMM SYS-GMM DIF-GMM-C SYS-GMM-C DIF-GMM SYS-GMM DIF-GMM-C SYS-GMM-C Lag 1 resource depletion 0.106 – 0.386 0.191 –0.401 –0.050 –0.397 –0.051 –0.091 (0.159) (0.021)** (0.104) (0.171) (0.585) 0.032)** (0.581) (0.448) Lag 1 output volatility –0.012 0.212 – 0.017 0.194 0.068 0.208 0.120 0.101 (0.862) (0.024)** (0.819) (0.037)** (0.098)* (0.031)** (0.003)*** (0.054)* Lag 1 liquidity liability ratio 0.024 0.006 0.031 0.004 (0.099)* (0.827) (0.055)* (0.927) Lag 1 investment –0.004 –0.005 0.023 –0.009 (0.816) (0.816) (0.217) (0.471) Lag 1 per capita GDP growth –0.030 0.105 – 0.040 0.097 0.053 0.124 0.121 0.046 (0.718) (0.324) (0.764) (0.291) (0.372) (0.288) (0.096)* (0.346) Lag 1 per capita GNI –0.011 – 0.072 – 0.018 –0.059 –0.002 –0.045 0.006 –0.031 (0.620) (0.002)*** (0.610) (0.064)* (0.933) (0.014)** (0.806) (0.073)* Age dependency –0.077 – 0.289 – 0.085 –0.234 0.005 –0.229 –0.046 –0.120 (0.195) (0.001)*** (0.167) (0.067)* (0.912) (0.001)*** (0.332) (0.096)* M1 serial correlation test ( p-value) 0.11 0.18 0.05 0.26 0.78 0.15 0.91 0.85 M2 serial correlation test ( p-value) 0.25 0.75 0.15 0.72 0.74 0.95 0.29 0.85 Sargan test ( p-value) 0.08 0.46 0.78 0.80 0.02 0.48 0.14 0.62 Diff-Sargan test ( p-value) 1.00 0.57 1.00 1.00 Granger causality test ( p-value) 0.18 0.02 0.22 0.10 0.08 0.18 0.01 0.01 Long-run effect –0.01 0.15 – 0.02 0.14 0.06 0.15 0.11 0.09 (0.08) (0.07) (0.09) (0.07) (0.04) (0.06) (0.04) (0.05) Heterogeneity test ( p-value) 0.00 0.00 0.01 0.00 CSD test ( p-value) 0.00 0.07 0.02 0.21 Number of observations 773 890 773 890 895 1,021 895 1,021 * Signi�cant at the 10 percent level; ** signi�cant at the 5 percent level; *** signi�cant at the 1 percent level. Note: Numbers in parentheses are the robust p-values. Data are for 128 nontransition countries over 1979– 2008. The �nancial development channel Huang (liquidity liability ratio) and the investment channel (investment share of GDP) are examined separately. See text for variables de�nitions and data sources and for test descriptions. Source: Author’s analysis based on data described in text. 143 144 THE WORLD BANK ECONOMIC REVIEW Final energy intensity and energy consumption per capita were also exam- ined as potential transmission channels. Neither was found to be signi�cant. This is another interesting area for further research. I V. S U M M A R Y AND CONCLUSIONS In sum, the interaction between global �nancial markets and the wider economy signi�cantly influences sustainable development. Through this �nan- cial development channel, speci�cally the liquidity liability ratio, output vola- tility affects natural resources depletion, a key component of genuine savings. This analysis sheds light on the interaction between the �nancial crisis and economic downturns during the 2007–09 global �nancial crisis. After the col- lapse of the U.S. subprime mortgage market in 2007 and the failure of Lehman Brothers in 2008, the global economic downturn caused by the credit crunch led to rising credit risk and uncertainty in the �nancial system. This intensi�ed the already sharp drop in global demand, with adverse implications for sustain- able growth and sustainable exploitation of natural resources. There is also evidence of cross-country dependence through the �nancial development channel. Macroeconomic volatility or �nancial crisis in one country tends to spread rapidly to other �nancial markets, perhaps reflecting the cross-border �nancial links typically associated with �nancial markets. Lower income countries, lower energy-intensity countries, and lower trade- share countries are in general more vulnerable to macroeconomic volatility or shocks. Increasing the resilience of these countries to �nancial crisis and econ- omic volatility should help to build a sustainable global economy. Understanding the nature of volatility and how to manage its consequences should be of considerable interest, especially to developing countries. However, the ability of these countries to tackle volatility might be constrained by an underdeveloped �nancial sector, weak institutions, and other political economy considerations. Empirical research continues to show that underdeve- loped �nancial sectors and weak institutions in developing countries can amplify the adverse effects of volatility on long-run growth and sustainable development and result in persistent long-run development problems. Governments should aim to liberalize the �nancial sector and strengthen capacities to mobilize and manage �nancial resources while ensuring adequate regulation and supervision and deli- vering public services more effectively. In addition, government actions to expand energy-saving development and strengthen macroeconomic fundamentals could support sustainable development. Internationally, supportive frameworks are needed to facilitate �nancial development, energy savings, climate change adaptation and mitigation, and a low carbon economy. Dedicated resources for development should be made available for vulnerable countries and people around the world to enable countries to cope with economic volatility and crises and improve safety nets and basic services such as health and education. Huang 145 REFERENCES Aizenman, Joshua, and Brian Pinto. 2005. “Overview.� In Joshua Aizenmann and Brian Pinto, eds., Managing Economic Volatility and Crises. Cambridge: Cambridge University Press. Arellano, Manuel, and Stephen Bond. 1991. “Some Tests of Speci�cation for Panel Data: Monte Carlo Evidence and an Application to Employment Equations.� Review of Economic Studies 58: 277–97. Arellano, Manuel, and Olympia Bover. 1995. “Another Look at the Instrumental-variable Estimation of Error-components Models.� Journal of Econometrics 68: 29– 51. Aysan, Ahmet F. 2007. “The Effects of Volatility on Growth and Financial Development through Capital Market Imperfections.� METU Studies in Development 34(1): 1–18. Bai, Jushan. 2009. “Panel Data Models with Interactive Fixed Effects.� Econometrica 77(4): 1229–79. Barbier, Edward B. 2009. “A Global Green New Deal.� Report prepared for the Green Economy Initiative and the Division of Technology, Industry and Economics of the UN Environment Programme. Department of Economics and Finance, University of Wyoming, Laramie. Blundell, Richard, and Stephen Bond. 1998. “Initial Conditions and Moment Restrictions in Dynamic Panel Data Models.� Journal of Econometrics 87: 115–43. Bowsher, Clive G. 2002. “On Testing Over-identifying Restrictions in Dynamic Panel Data Models.� Economics Letters 77: 211 –20. Canton, Erik. 2002. “Business Cycles in a Two-sector Model of Endogenous Growth.� Economic Theory 19: 477– 92. Committee of Ten. 2009. “Impact of the Crisis on African Economies–Sustaining Growth and Poverty Reduction, African Perspectives and Recommendations to the G20.� Report to the British Prime Minister, Gordon Brown, March 17. African Development Bank, the African Union and the United Nations Economic Commission for Africa, Abidjan, Co ˆ te d’Ivoire. Dauvergn, Peter. 1999. “The Environmental Implications of Asia’s 1997 Financial Crisis.� IDS Bulletin 30(3): 31–42. Dietz, Simon, Eric Neumayer, and Indra de Soysa. 2007. “Corruption, the Resource Curse and Genuine Saving.� Environment and Development Economics 12: 33– 53. Enerdata. 2010. Global Energy Market Data. Retrieved December 2010 from www.enerdata.fr. Gaveau, David L. A., Matthew Linkie, Patrice Levangd, and Nigel Leader-Williams. 2009. “Three Decades of Deforestation in Southwest Sumatra: Effects of Coffee Prices, Law Enforcement, and Rural Poverty.� Biological Conservation 142(3): 597 –605. Ghosh, Atish R., and Jonathan D. Ostry. 1997. “Macroeconomic Uncertainty, Precautionary Saving, and the Current Account.� Journal of Monetary Economics 40(1): 121 –39. Grier, Kevin B., and Gordon Tullock. 1989. “An Empirical Analysis of Cross-national Economic Growth, 1951–1980.� Journal of Monetary Economics 24: 259– 76. Hamilton, Katherine, Milo Sjardin, Molly Peters-Stanley, and Thomas Marcello. 2010. “Building Bridges: State of the Voluntary Carbon Markets 2010.� Washington, D.C. and New York: Ecosystem Marketplace and Bloomberg New Energy Finance. Heston, Alan, Robert Summers, and Bettina Aten. 2009. Penn World Table, version 6.3. Center for International Comparisons, University of Pennsylvania, Philadelphia. Hnatkovska, Viktoria, and Norman Loayza. 2005. “Volatility and Growth.� In Joshua Aizenmann and Brian Pinto, eds., Managing Economic Volatility and Crises. Cambridge: Cambridge University Press. Huang, Yongfu. 2010. Determinants of Financial Development. Basingstoke, U.K.: Palgrave Macmillan Press. Intergovernmental Panel on Climate Change (IPCC). 2001. Climate Change 2001. Synthesis Report. Cambridge: Cambridge University Press. Kasa, Sjur, and Lars Otto Næss. 2005. “Financial Crisis and State-NGO Relations: The Case of Brazilian Amazonia, 1998– 2000.� Society & Natural Resources 18(9): 791–804. 146 THE WORLD BANK ECONOMIC REVIEW Kharroubi, Enisse. 2007. “Crises, Volatility, and Growth.� World Bank Economic Review 21(3): 439 –60. Kittiprapas, Sauwalak. 2002. “Social Impacts of Financial and Economic Crisis in Thailand.� Regional Project on the Social Impact of the Asian Financial Crisis. East Asian Development Bank: Manila. ´ s, and Silvana Tenreyro. 2007. “Volatility and Development.� Quarterly Journal of Koren, Miklo Economics 122(1): 243–87. Kormendi, Roger, and Philip Meguire. 1985. “Macroeconomic Determinants of Growth: Cross-country Evidence.� Journal of Monetary Economics 16(2): 141–63. Lin, J. Yifu. 2009. “Learning from the Past to Reinvest the Future.� Address at the Annual Bank Conference on Development Economics, June 22 –24, Seoul. Loayza, Norman V., Romain Rancie ´ n, and Jaume Ventura. 2007. “Macroeconomic ` re, Luis Serve Volatility and Welfare in Developing Countries: An Introduction.� World Bank Economic Review 21(3): 343– 57. OECD/IEA (Organisation for Economic Co-operation and Development/International Energy Agency). 2009. “The Impact of the Financial and Economic Crisis on Energy Investment.� Background paper for the G-8 Energy Ministers’ Meeting, May 24 –25, L’Aquila, Italy. Ramey, Garey, and Valerie A. Ramey. 1995. “Cross Country Evidence on the Link between Volatility and Growth.� American Economic Review 85(5): 1138– 51. Sandmo, Agnar. 1970. “The Effect of Uncertainty on Saving Decisions.� The Review of Economic Studies 37(3): 353–60. Sara�dis, Vaslis, and Donald Robertson. 2009. “On the Impact of Error Cross-sectional Dependence in Short Dynamic Panel Estimation.� Econometrics Journal 12(1): 62 –81. Sara�dis, Vaslis, Takashi Yamagata, and Donald Robertson. 2009. “A Test of Cross Section Dependence for a Linear Dynamic Panel Model with Regressors.� Journal of Econometrics 148(2): 149 –61. United Nations General Assembly. 2005. “2005 World Summit Outcome.� Final document. A/60/L.1. New York. Windmeijer, Frank. 2005. “A Finite Sample Correction for the Variance of Linear Ef�cient Two-step GMM Estimators.� Journal of Econometrics 126(1):25– 51. Wolf, Holger. 2005. “Volatility: De�nitions and Consequences.� In Joshua Aizenmann and Brian Pinto, eds., Managing Economic Volatility Crises. Cambridge: Cambridge University Press. World Bank. 2010a. Financial Development and Structure Database. Washington, D.C. ———. 2010b. World Development Indicators 2010 Database. Washington, D.C. ———. 2011. World Bank List of Economies. Retrieved January 2011 from http://search.worldbank. org/all?qterm=list%20of%20economies. World Commission on Environment and Development. 1987. Our Common Future. Oxford: Oxford University Press. The Discriminatory Nature of Speci�c Tariffs Sohini Chowdhury This paper quanti�es the distortions from speci�c tariffs levied by the EU. The most- favored-nation (MFN) speci�c tariffs levied by the EU translate into higher tariff bar- riers for poor countries exporting low price goods. We show that for poor countries, these higher tariff barriers from speci�c tariffs offset the gains from preferential tariffs. We apply a two-stage analysis to show that the speci�c tariffs levied by the EU on its agricultural imports wash away more than half of the welfare bene�ts enjoyed by the Sub-Saharan African countries from EU preferential tariffs. Our results provide the �rst quantitative estimate of the distortions associated with speci�c tariffs. JEL codes: F13, F14, O12, Q17 The most-favored-nation (MFN) clause, a key feature of the WTO agreement, requires that each WTO member treat its trading partners equally by levying a common MFN tariff on every imported tariff line. But while a MFN speci�c tariff (e.g. $2 per ton) conforms to this non-discrimination principle in paper, it violates this principle in spirit. In fact, speci�c tariffs discriminate against imports from poor countries. We can identify two channels for this. First, poor countries tend to have lower export prices compared to rich countries (Schott, 2004);1 this means that when poor and rich exporters face the same level of speci�c tariff, poor exporters face a higher ad valorem equivalent (AVE).2 Sohini Chowdhury did this analysis towards her doctoral dissertation at the Krannert School of Management, Purdue University (403 W State St, West Lafayette, IN 47907). She is currently an economist at Moody’s Analytics Inc. (121 N Walnut Street, Suite 500, West Chester, PA 19380). She is grateful to Chong Xiang, Thomas Hertel, David Hummels and Luca Salvatici for excellent guidance, and to Badri Narayanan Gopalakrishnan for useful comments. Her phone number is þ 1 (410) 402- 4462, and her email address is sohini.chowdhury@gmail.com. 1. Schott (2004) analyzed US manufacturing imports for the year 2001 and found a signi�cant positive correlation between the per capita GDP of exporters and the unit value of their exports. He concluded that rich countries utilize their comparative advantage in skill and capital to supply higher quality and higher priced varieties. Abd-El-Rahman (1991) shows that unit value differences reveal quality differences, even at the detailed ten-digit HS level. Fontagne ´ et al. (1997) and Greenaway et al. (2001) show that vertical specialization is among the most salient features of trade between European countries. 2. The AVE of a speci�c tariff expresses the speci�c tariff on a commodity as a percentage of the price of the commodity. This means that the same speci�c tariff of $2/ton on rice translates into a higher AVE, say 50%, for rice imported from Bangladesh and a lower tariff equivalent, say 10%, for rice imported from Japan. This effect is seen with any per-unit speci�c price rise, arising either from an unit transport cost (Alchian and Allen, 1964; Hummels and Skiba, 2004) or from an import quota (Falvey, 1979; Boorstein and Feenstra, 1987). THE WORLD BANK ECONOMIC REVIEW, VOL. 26, NO. 1, pp. 147 –163 doi:10.1093/wber/lhr036 Advance Access Publication July 12, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 147 148 THE WORLD BANK ECONOMIC REVIEW Second, speci�c tariffs are concentrated on agricultural commodities3 which form the bulk of poor country’s exports (Gibson et al., 2001;4 Hoekman et al., 2002). Our initial data analysis shows that in 2004, nearly half (48%) of the global agricultural imports were sourced from poor countries. While the discri- minatory effects of speci�c tariffs have been discussed in the literature (Gibson et al., 2001; Von Kirchbach and Mondher, 2003; Boue ¨ t et al., 2004), there has been no quantitative estimate of the welfare loss faced by poor countries from such tariffs. In this paper we provide an estimate of the additional welfare loss from speci�c tariffs, vis-a ` -vis ad valorem tariffs, accruing to poor countries. Additionally, to get a better sense of the size of this loss, we express the loss as a percentage of the welfare gain enjoyed by poor countries from the tariff pre- ferences granted to them by rich countries. Article XXIV of the GATT/WTO together with the Enabling Clause,5 permit exemptions to the MFN obligation and create a legal basis for the formation of Preferential Trade Agreements. These PTAs can be bilateral/reciprocal tariff preferences like Free Trade Agreements or Customs Unions, or unilateral/non-reciprocal tariff preferences extended from developed to developing countries like the Generalized System of Preferences (GSP), and its extensions like the Everything but Arms (EBA)6 initiative by the EU and the African Growth and Opportunities Act (AGOA)7 by the US. Results show that more than half the welfare gains enjoyed by the Sub-Saharan8 countries (SSA) from the preferences granted to them by the EU are taken away by the speci�c tariffs levied by the EU. In other words, while on one hand the EU grants preferences to poor countries through lower tariff rates, on the other hand it takes away the bene�ts of these preferences through speci�c tariffs. To arrive at our results, we use a partial equilibrium model considering only the preferential and speci�c tariffs levied by the EU on its agricultural imports from Sub-Saharan Africa. We follow a two-stage procedure, sequentially elimi- nating preferential tariffs and speci�c tariffs by the EU. In Stage 1, we eliminate 3. Our analysis in Section I shows that speci�c tariffs were concentrated on agricultural imports whether analyzed in terms of the number of tariff lines, the tariff rates, or the AVEs. 4. Gibson et al. (2001) states that approximately 44 % of agricultural tariff-lines in the US and EU are speci�ed in non-ad valorem terms. The reasons include the increased protection that a non-ad valorem tax provides against large drops in import prices and the lack of transparency associated with these rates, which helps conceal the level of protection. 5. Decision on Differential and More Favourable Treatment, Reciprocity, and Fuller Participation of Developing Countries, GATT Document L/4903, 28 November 1979, BISD 26S/203 6. This initiative came into effect in 2001. Under this, the EU grants duty free access to imports of all products that originate in LDCs with the exception of arms and munitions. 7. This act came into effect in 2000. Under this the US extended preferences to 37 African countries, providing duty free access to agricultural commodities. 8. EU includes the 25 member countries of the European Union in 2004. SSA includes 50 countries (all of Africa – South Africa – a few small dependencies – the six North African countries of West Sahara, Morocco, Algeria, Libya, Tunisia and Egypt). Chowdhury 149 preferential tariffs by moving from the preferential tariff schedule to the MFN tariff schedule. In Stage 2, we eliminate speci�c tariffs by moving from the MFN tariff schedule to the mean ad valorem equivalent tariff schedule, where the mean is taken over all exporters for each importer-commodity pair. We then compare the welfare losses from each stage. There is a large literature that analyzes the welfare gains/losses from trade agreements.9 In comparison, our paper disentangles the effects of speci�c tariffs and trade agreements and compares the welfare loss from speci�c tariffs to the welfare gains from preferential tariffs. There is also a large literature that calculates the distortions from tariffs.10 In comparison, our paper calculates the additional distortion from speci�c tariffs vis-a ` -vis ad valorem tariffs. We de�ne the mean ad valorem tariff equivalent as a benchmark11 to calculate the additional loss from speci�c tariffs. The rest of our paper is organized as follows: We present some preliminary data analysis in Section I and discuss the general research methodology in Section II. Sections III and IV describe the model and data, and the results are analyzed in Section V. Section VI presents some robustness checks and Section VII concludes. I . P R E L I M I N A R Y D ATA A N A LY S I S We use trade and tariff data for the year 2004 from the Comtrade and MAcMapHS6v2.02 databases respectively. The data sources are discussed in details in Section IV. Figure 1 shows that globally, speci�c tariffs are concen- trated in the trade of agricultural commodities. 11% of all bilateral tariff lines in agriculture face speci�c tariffs, compared to only 1.6% of the bilateral tariff lines in non-agriculture and 2.4% of all bilateral tariff lines. In terms of the number of tariff lines, 65% of all the bilateral tariff lines facing speci�c tariffs are agricultural commodities. In terms of tariff rates, 67% of the value of all bilateral speci�c tariff rates (in $ per ton) are on agricultural commodities. Based on these data analyses, the rest of our paper will focus on speci�c tariffs on agricultural imports alone. The data identi�es Iceland, Vanatua, Switzerland and Norway as countries with high speci�c tariffs on their agricul- tural imports; and Kenya, Vietnam, Kazakhstan and Uzbekistan as countries facing high ad valorem equivalents of speci�c tariffs on their agricultural 9. These studies are either in a partial equilibrium framework using some variant of the gravity equation (Baier et al., 2007; Carrere, 2006) or in a general equilibrium framework (Boue¨ t and Laborde, 2009; Anderson et al., 2006, Francois et al., 2005) 10. Feenstra (1995), Anderson (1998), Kee et al. (2009), Irwin (2010), Lloyd et al. (2010) to name a few. 11. Boorstein and Feenstra (1987) do something similar when they calculate the ‘excess cost’ of the US steel import quotas during 1969– 74. They de�ne as the benchmark the ad valorem tariff which has the same effect on aggregate import price as the quota. 150 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. AVEs of Speci�c Tariffs by Commodity Notes: The AVEs corresponding to each HS2 is obtained by taking a simple average over all tariff lines for all importers, within each HS2. HS Chapter: 06 is ‘live trees & plants’; HS Chapter: 24 is ‘tobacco’; HS Chapter: 01-24 are agricultural commodities. Source: Authors’ analysis based on data sources discussed in the text. F I G U R E 2. AVEs of Speci�c Tariffs in Agriculture by Exporter Notes: Kenya (KEN), Kazakhstan (KGZ), Uzbekistan (UZB), Vietnam (VNM), Papua New Guinea (PNG), Mongolia (MNG). Source: Authors’ analysis based on data sources discussed in the text. exports. Figure 2 shows that poor exporters, on average, face higher ad valorem equivalents of speci�c tariffs. We check this claim further by testing, similar to Schott (2004), the relation- ship between an exporter’s per capita GDP and the unit values of its Chowdhury 151 agricultural exports. We estimate the following regression pooling across all exporters ( j ) and all importers (i), for all agricultural goods (k) (HS Chapter , ¼ 24), with HS6 commodity �xed effects   GDP ln uvijk ¼ aik þ b ln þjik ð1Þ L j where, aik is the importer-by-commodity �xed effect, uvijk is the unit value of good k imported into i from j, (GDP/L)j is the per capita GDP of exporter j and jik is the error term. For unit values, we use the Exporter Reference Group Unit Values (ERGUVs), which are computed as the weighted median unit value of worldwide exports from an exporter’s reference group (Boue ¨ t et al., 2004).12 Our estimation results, presented in Table 1, suggest that on average, doubling the per capita GDP of exporters increases the ERGUVs of their exports by 7.7%. The signi�cant positive correlation between the exporter’s unit value and per capita GDP corroborates the claim that on average, the ad valorem equiva- lents of MFN speci�c tariffs are higher for poor country exporters. It is this feature of speci�c tariffs that bias them against poor exporters. I I . R E S E A R C H M E T H O D O LO G Y While speci�c tariffs increase the effective tariff rates faced by poor exporters, preferential tariffs lower these rates by granting poor exporters tax concessions. So the welfare losses faced by poor exporters from speci�c tariffs can be ana- lyzed only if we control for the welfare gains from the preferential tariffs they face.13 We use a two-stage experiment to disentangle these two effects. In Stage 1, we eliminate preferential tariffs by removing all preferences granted by rich importers in agriculture. This is achieved by moving from the rich impor- ter’s preferential tariff rates (which we denote by tp) to their MFN tariff rates. 12. Boue¨ t et al. (2004) assign each reporting country to a reference group of similar countries using a hierarchical clustering analysis based on GDP per capita (in terms of PPP) and trade openness. They label the �ve resulting groups as: (1) richest countries; (2) high openness, middle income countries; (3) low openness, middle income countries; (4) high openness, low income countries; (5) low openness, low income countries. The full set of countries and reference groups is provided in Appendix A1 of their paper. They calculate ERGUVs using "weighted" medians assuming that each UV repeats as many times as the underlying trade flow contains dollars. For robustness, the UVs are computed based on three-year-average trade flows (across the 2000– 2002 period). Outliers are �ltered by truncating to the top or bottom limit, any ratio of ERGUVs to the world median unit value which fall outside the bracket [1/3; 3]. A sequential procedure is used to �ll missing values for reference groups: any blank is substituted by the value of the closest reference group. 13. This is especially relevant given the recent proliferation of PTAs. In 2008, there were over 350 PTAs (Bhagwati, 2008) suggesting that the rich countries now grant MFN tariffs to only a handful of countries. The EU applies its MFN tariffs to only six countries – Australia, New Zealand, Canada, Japan, Taiwan and the US – with all other countries enjoying more favorable tariffs. This has prompted Bhagwati to suggest that the MFN tariff should be more appropriately renamed as the LFN (Least Favored Nation) tariff (Bhagwati, 2008) 152 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Regression of Unit Values on Exporter Per Capita GDP ln(p1) ln(p2) ln(GDP/L) 0.077** 0.121 (0.002) (0.061) N 229640 229640 Source: Authors’ analysis based on data sources discussed in the text. p1 ¼ ERGUV; p2 ¼ bilateral unit value. **p , 0.01 T A B L E 2 . An Illustration of the 2-Stage Procedure Ad Valorem Equivalents of Swiss import tariffs on Raw Cane Sugar Stage 1 Stage 2 Exporter tp ! tm ! tc Guatemala 0.00 2.28 0.33 Japan 0.10 0.10 0.33 Source: Authors’ analysis based on data sources discussed in the text. Now while the MFN tariff rates are the same for all exporters corresponding to each rich importer-commodity pair, their ad valorem equivalents (which we denote by tm) are not, since these MFN rates also include speci�c tariffs. In Stage 2 therefore, we convert the speci�c tariffs levied by rich importers on agriculture to ad valorem tariffs. This is achieved by moving from the ad valorem equivalents of MFN tariff rates (tm) to the mean ad valorem equiva- lents across all exporters (which we denote by tc). The counterfactual mean ad valorem tariff tc is truly MFN since it is uniform across exporters for each importer-commodity pair. We construct tc for each importer-commodity pair according to Equation 18, as the average ad valorem equivalent faced by all exporters, weighted by the product of the price elasticity of import demand and free-trade imports, following Leamer (1974). We clarify the intuition behind tp, tm and tc in Table 2. The table shows the ad valorem equivalents of tariffs levied by Switzerland on its import of raw cane sugar from Japan and Guatemala. tp denotes the actual tariffs levied and includes both preferential and speci�c tariffs. Due to the preferential tariff bene�ts enjoyed by the poor country Guatemala, it faces a lower (in this case, a zero) tp than Japan. tm is the ad valorem equivalent of the MFN tariff rates levied by Switzerland and does not include preferential tariffs. Although the MFN tariff rates are identical across exporters, their ad valorem equivalents (tm) are not, due to the presence of speci�c tariffs. In particular, due to the lower price of Guatemala exports, Guatemala faces a higher tm than Japan. To even out the non-MFN effects of speci�c tariffs across Guatemala and Japan, we construct the counterfactual tariff tc as the mean ad valorem tariff equivalent across all exporters. Having Chowdhury 153 been stripped off the effects of speci�c tariffs, tc is the same across all expor- ters. Once we have the three tariff schedules tp, tm and tc, we compare the welfare losses corresponding to each. We run both stages of the experiment under a partial equilibrium model which considers only the EU’s tariff policy towards Sub-Saharan Africa. III. MODEL We consider a simple world with the EU as the only importer. All countries exporting to the EU are classi�ed as either SSA (Sub-Saharan Africa) or non-SSA. We consider only the trade in agricultural commodities. We select the particular combination of the EU and SSA because our data suggests that on average, EU agricultural imports from SSA face high MFN speci�c tariffs (31.3%) and are signi�cantly higher ($9.67 billion) than the imports of other rich countries from SSA. Moreover, almost every EU agricultural tariff line imported from SSA bene�ts from some tariff preference (only 17 out of the 516 tariff lines face no preference). These make the EU and SSA perfect candi- dates for our model. We use the model to compute the proportion of the bene�ts enjoyed by SSA from EU preferential tariffs that are taken away by EU speci�c tariffs. Consider the import of commodity k from country j into the EU given by q jk ¼ adjk pÀ jk sk ð2Þ Consider the export of commodity k from country j to the EU given by q jk ¼ asjk pS jk k ð3Þ where, j [ fall exportersg is the exporter index, k [ fagricultural tariff linesg is the commodity index, sk is the own price elasticity of import demand of good k in the EU, and Sk is the own price elasticity of export supply of good k from the rest of the world as faced by the EU. adjk and asjk are shift parameters corresponding to the import demand and export supply curves respectively. From the Comtrade dataset, we obtain data on bilateral c.i.f prices ( psjk) and import quantitities (q0jk) for the EU. All other prices and quantities are expressed in terms of these values and elasticities. We illustrate our model in Figure 3, showing the EU’s import demand curve and exporter j’s export supply curve for commodity k. p* is the free-trade world price and q* is the corresponding import quantity. An import tariff t levied by the EU introduces a wedge between the price paid by the domestic consumers and the price received by the foreign supplier by raising the domestic price to pd and lowering the world price to ps where pdjk ¼ psjk ð1 þ t jk Þ ð4Þ 154 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3. Welfare Loss Source: Authors’ analysis based on model discussed in the text. As Figure 3 shows, the distribution of the tariff incidence between the EU and the exporter is determined by the import demand and export supply elasticities sk and Sk. The post-tariff import quantity is q0. The EU faces an allocative ef�- ciency loss from the import tariff due to the higher domestic price and lower import, and a terms of trade gain due to the lower price it pays for its import. So the net welfare loss faced by the EU from the import tariff t is equal to the allocative ef�ciency loss it faces net of the terms of trade gain. SSA faces an allocative ef�ciency loss due to lower exports, and a terms of trade loss (equiv- alent to the terms of trade gain faced by the EU) due to the lower price it receives on its export. So the net welfare loss faced by SSA from the import tariff t is equal to the sum of its allocative ef�ciency loss and terms of trade loss. It is important to note that all losses are expressed with respect to the free trade i.e. zero tariff scenario. The welfare loss in EU and an exporting country j, from a tariff t levied by the EU on commodity k is derived as follows: From Figure 3, ignoring the subscripts, we have S q0 ¼ ad pÀ s d ¼ as ps ð5Þ From (4) and (5), we have Às ad pÀ s s ð1 þ t Þ ¼ as pS s ð6Þ Chowdhury 155 Rearranging (6), we have ad ¼ ð1 þ tÞs pS s þs ð7Þ as From Figure 3 we also have qà ¼ ad pÃÀs ¼ as pÃS ð8Þ From (8), we have ad ¼ pÃðSþsÞ ð9Þ as Combining (7) and (9), we have ð1 þ tÞs pS s þs ¼ pÃðSþsÞ ð10Þ Rearranging (10), we have s p à ¼ ps ð1 þ tÞsþS ð11Þ From (11), (8) and (5), we have sS qà ¼ q0 ð1 þ tÞsþS ð12Þ From Figure 3, AE loss in the EU à q � ¼ pðqd Þdq À ðq à À q0 Þp à q0 �à  q q Às 1 ¼ ad dq À ðqà À q0 Þp à q0  1 q à 1  1Às  ¼ ðad Þs q 1Às1 Àðq à À q0 Þp à q0 1 h sÀ1 sÀ1 i s ¼ sÀ ð a Þs qà s À q s À ðqà À q0 Þpà 1 d 0 Substituting the values of q* and p* from (12) and (11) respectively, we have 156 THE WORLD BANK ECONOMIC REVIEW AE loss in the EU s 1  sÀ1 h SðsÀ1Þ i ¼ ðq0 psd Þs q s 0 ð 1 þ t ÞsðsþSÞ À 1 sÀ1 h sS i s À q0 ð1 þ tÞsþS À 1 ps ð1 þ tÞsþS s h SðsÀ1Þ i ¼ ps ð1 þ tÞq0 ð1 þ tÞsðsþSÞ À 1 ð13Þ sÀ1 sð1þSÞ s À q0 ps ð1 þ tÞ sþS þ q0 ps ð1 þ tÞsþS ! 1 sð1þSÞ s s ¼ ps q0 ð1 þ tÞ sþS À ð1 þ tÞ þ ð1 þ tÞsþS sÀ1 sÀ1 From Figure 3, AE loss in the exporter j � qà ¼ ðq à À q0 Þp à À pðqs Þdq q0  1 � q à qs S ¼ ðq à À q0 Þp à À q0 dq as S h Sþ1 Sþ1 i À1 ¼ ðq à À q0 Þp à À as S qà S À q0S Sþ1 Substituting the values of q* and p* from (12) and (11) respectively, we have AE loss in exporter j   S sS s À1 ¼ q0 ð1 þ tÞsþS À 1 ps ð1 þ tÞsþS À as S Sþ1 h Sþ1 sð1þSÞ Sþ1 i q0S ð1 þ tÞ sþS À q0S ! S s 1 sð1þSÞ ¼ ps q0 À ð1 þ tÞsþS þ ð1 þ tÞ sþS Sþ1 Sþ1 ð14Þ TOT gain in the EU/loss in exporter j ¼ ð pà À ps Þq0 Substituting the value of p* from (11), we get Chowdhury 157 T A B L E 3 : Welfare Change Under Special Scenarios Scenario Result When tjk ¼ 0, for all j, k TOT loss and allocative ef�ciency loss for the EU is 0 TOT loss and allocative ef�ciency loss for each exporter is 0 Welfare loss for the EU and all exporters is 0 Although welfare loss is 0, welfare is not optimum for the EU since there can be welfare gains from positive TOT gains When Sk ¼ 1 for all k, indicating a flat The TOT gains for the EU and TOT losses for each export supply curve faced by the EU exporter is 0 When sk ¼ 1, for all k, indicating a flat Allocative ef�ciency loss for the EU is 0 import demand curve faced by the EU Source: Authors’ analysis based on data sources discussed in the text. TOT gain in the EU/loss in exporter j h s i ¼ ps q0 ð1 þ tÞsþS À 1 ð15Þ Finally we have, Net welfare loss in EU ! 1 sk ð1þSk Þ sk sk ¼ Sj Sk q0jk psjk ð1 þ t jk Þ sk þSk À ð1 þ t jk Þ þ ð1 þ t jk Þsk þSk sk À 1 sk À 1 AE LOSS ð16Þ h sk i ÀSj Sk q0jk psjk ð1 þ t jk Þsk þSk À 1 TOT LOSS Net welfare loss in exporter j ! Sk 1 sk ðSk þ1Þ sk ¼ Sk q0jk psjk þ ð1 þ t jk Þ Sk þsk À ð1 þ t jk Þsk þSk ðSkþ1 Þ ðSk þ 1Þ AE LOSS ð17Þ P h sk i þ q0jk psjk ð1 þ t jk Þsk þSk À 1 k TOT LOSS Then applying the available data to (16) and (17) we calculate the dollar values of the welfare loss corresponding to the three different tariff schedules tp, tm and tc. Using (16) and (17), we can also derive the allocative ef�ciency loss and terms of trade loss corresponding to the special scenarios shown in Table 3. 158 THE WORLD BANK ECONOMIC REVIEW I V. D ATA Data on EU bilateral c.i.f.14 prices ( ps) and the corresponding imports (q0) at the six-digit Harmonized System are obtained from the Comtrade database for the year 2004. Data on the ad valorem and speci�c components of EU MFN and preferential tariff rates for for the year 2004 are obtained from the MAcMap-HS6v2.02 database (Boumellassa et al., 2009). We construct the ad valorem equivalents of speci�c tariffs by dividing the speci�c tariffs by the Exporter’s Reference Group Unit Values. We denote by tp, the sum of the pre- ferential ad valorem tariff and the ad valorem equivalent of the preferential speci�c tariff. We denote by tm, the sum of the MFN ad valorem tariff and the ad valorem equivalent of the MFN speci�c tariff. Finally, we construct the mean ad valorem tariff equivalent tc for each importer-commodity pair as the average tm over all exporters, weighted by the product of the price elastici- ties of import demand and free-trade imports (Leamer, 1974) as: Sj sik q0ijk ð1 þ tmijk ÞÀsik tmijk tcik ¼ ð18Þ Sj sik q0ijk ð1 þ tmijk ÞÀsik where i [ fEUg is the importer index and all other notations follow the de�- nitions in Section III. The differences in the welfare loss corresponding to tp and tm show the effect of eliminating EU preferential tariffs on SSA agricul- tural imports. This is Stage 1 of our experiment. The differences in the welfare loss corresponding to tm and tc show the effect of converting EU speci�c tariffs on SSA agricultural imports to ad valorem tariffs. This is Stage 2 of our experiment. To avoid aggregation bias, we perform all tariff changes at the individual country and tariff line level. The results are then aggregated for pres- entation and analysis. Agricultural imports from SSA ($9.67 billion) constitute less than 10% of total EU imports. For agricultural exports to the EU, SSA faces a lower prefer- ential tariff (4.7%) when compared to non-SSA (9.3%). This is to be expected, since the EU extends preferential market access to agricultural imports from SSA. But SSA faces a higher average ad valorem equivalent of MFN tariffs (14.7%) relative to non-SSA (11.8%); consistent with the hypothesis that exports from SSA have lower unit values on average when compared to exports from non-SSA. To compute the welfare loss, we also need data on the price elasticities of import demand (s) and export supply (S) for the EU. We obtain the former from Kee et al. (2008), and the latter from Broda et al. (2008) which contains estimates of the commodity-speci�c export supply elasticities for the US. We apply these elasticities to the EU, assuming that the EU faces identical elasticities. 14. Cost, insurance and freight (the price on which import tariff is applied) Chowdhury 159 T A B L E 4 . Welfare Change. (US$ millions) Stage 1 Stage 2 AE TOT Welfare AE TOT Welfare Region Change Change Change Change Change Change EU 2 578 1237 659 389 2 113 276 SSA 2 719 2 1237 1956 522 729 1251 Non-SSA 0 0 0 2 722 2 616 2 1338 Source: Authors’ analysis based on data sources discussed in the text. V. R E S U LT S We present the welfare changes from Stages 1 and 2 in Table 4. Stage 1 gener- ates an allocative ef�ciency loss in the EU and SSA equal to $578 million and $719 million respectively, and a terms of trade gain in the EU matched by an equivalent terms of trade loss in SSA of $1237 million. Non-SSA faces no tariff change and therefore no welfare change. We attribute the higher allocative ef�- ciency loss in SSA to the observation that the EU faces a relatively steep export supply curve compared to the import demand curve. The average export supply elasticity on agricultural goods faced by the EU is 1.92 and the corre- sponding import demand elasticity is -3.21. These relative elasticities also explain the strong terms of trade effect relative to the allocative ef�ciency effect.15 In Stage 2, SSA gains $522 million and $729 million respectively in terms of allocative ef�ciency and terms of trade due to lower tariff rates. The EU faces an allocative ef�ciency gain of $389 million since it now levies a more uniform tariff across its exporters while maintaining the same trade- weighted average tariff rate.16 But the EU also faces a terms of trade loss from levying lower tariffs on SSA imports. This loss is partly compensated by the higher tariffs on non-SSA imports resulting in a net terms of trade loss of $113 million. The higher tariffs on non-SSA imports generate allocative ef�ciency and terms of trade losses for non-SSA equivalent to $722 million and $616 million respectively. So the uniform tariffs levied in Stage 2 essentially switch the destinies of SSA and non-SSA. The combined results from Stages 1 and 2 suggest that 64% of the welfare gains enjoyed by SSA from the preferential 15. This is consistent with the terms-of-trade argument (Bagwell et al., 2006) which shows that the terms of trade changes resulting from tariff cuts are signi�cantly large and play an important role in a country’s decision to join a trade agreement. According to this argument, trade agreements are useful to governments as a means of helping them escape from a terms-of-trade-driven Prisoners’ Dilemma. Broda et al. (2008) provides an empirical support of this argument by showing that the tariff choices of non-WTO countries reflect their ability to manipulate their terms of trade. 16. For a small importer, an increase in tariff dispersion increases the trade restrictiveness and welfare loss. (Anderson et al., 2007; Feenstra, 1995; Irwin, 2010). In our large importer case, it increases the allocative ef�ciency loss. This explains why trade agreements seek to lower both tariff distortions and tariff peaks. 160 THE WORLD BANK ECONOMIC REVIEW tariffs on its agricultural exports to EU are washed away by the speci�c tariffs levied by the EU. The welfare loss to SSA from EU speci�c tariffs is equivalent to 12.7% of SSA agriculture exports to the EU. This �nding highlights the dis- criminatory nature of speci�c tariffs. It suggests that we can add speci�c tariffs to the list of factors thought to be eroding away the effectiveness of non- reciprocal/unilateral trade agreements meant to expand trade and promote development in poor countries. This widely debated list includes other ‘cul- prits’ like limited product coverage, proliferation of regional trade agreements, attached side conditions and Rules of Origin.17 This model does not account for substitution across imported and domestic commodities and across different imported varieties, and also ignores all inter- sectoral linkages and income effects. To capture these effects, we run the two stages of the experiment in a computable general equilibrium Global Trade Analysis Project (GTAP) model (Hertel, 1997), analyzing simultaneously the trade policy of multiple importers and exporters. The results, to be presented in a separate working paper, suggest that 72% of the preferential tariff bene�ts enjoyed by poor country exporters are taken away by the speci�c tariffs levied by the rich country importers. But the merits of the general equilibrium model come at a cost. The GTAP model assumes a Constant Elasticity of Substitution (CES) demand function into which all goods enter symmetrically. This is not compatible with our scenario of different quality goods being exported from different countries at different prices. Moreover, implementing the general equilibrium model requires aggregating countries into regions and aggregating products into commodity groups. Due to multiple tariff lines within each aggregate commodity group, such aggregation may hide compositional effects where the aggregate tariff faced by an exporter is determined by the composition of its exports. VI. ROBUSTNESS CHECKS We also run our experiment using an alternative benchmark tariff schedule tc’, which is the importer-commodity version of the Uniform Tariff Equivalent in 17. The GSP was designed to promote the development of poorer countries based on an infant industry argument (UNCTAD, 1964). But recent studies discussed in Lima ˜ o et al. (2006) show that the GSP and other unilateral preferences provided by the EU and United States often have side conditions attached that are valued by the preference-granting country and potentially costly to the recipient. A FAO Trade Policy Brief (ftp://ftp.fao.org/docrep/fao/008/j5424e/j5424e01.pdf) argues that limited product coverage and constraints on preference utilization, together with supply side problems, have prevented the full use such of trade preferences by the majority of preference-receiving countries. Also, the proliferation of bilateral trade agreements has resulted in the erosion of unilateral preference bene�ts. As an example, they state the relative loss of preferences by the Caribbean Basin Initiative bene�ciaries when NAFTA was created. This explains the �nding that although close to 80 least developed and small island developing states bene�t from non-reciprocal preferences, they account collectively for less than 2 percent of world agricultural exports. Meeting rules of origin prove costly for developing countries (administrative costs, constrained supply of intermediate imports etc.) discussed in Krishna (2005), Krishna and Krueger (1995), Falvey et al. (1998, 2002). Chowdhury 161 Kee et al. (2009), which in turn is derived from the Trade Restrictiveness Index proposed by Anderson (1998). While the Uniform Tariff Equivalent is the uniform tariff across all commodities which leaves welfare for the importer unchanged, our benchmark tariff tc’ is the uniform tariff across all exporters which leaves welfare for each importer-commodity pair unchanged. Mathematically, " 2 #1 S s q j ik 0ijk tm ijk 2 tc0ik ¼ ð19Þ Sj sik q0ijk where i [ fEUg is the importer index and all other notations follow the de�- nitions in Section III. Using tc’ as our benchmark tariff does not change our results qualitatively and suggests that 67% of the welfare bene�ts enjoyed by Sub-Saharan Africa from EU preferential tariffs are taken away by the speci�c tariffs levied by the EU. We also run our experiment using an alternative measure of unit value. Our baseline results use the Exporter’s Reference Group Unit Value (ERGUV) to compute the ad valorem equivalents of speci�c tariffs. An alternative is the bilateral unit value constructed by dividing the c.i.f import value by the corre- sponding import quantity. Again, while this does not change our results quali- tatively, it makes them stronger. This is because the bilateral unit value has a higher dispersion than the ERGUV by construction, which generates larger within-commodity cross-exporter dispersion and a larger distortionary effect from speci�c tariffs. But the cost of using bilateral unit value is that it is more prone to measurement errors (Boue ¨ t et al., 2004). VII. CONCLUDING REMARKS Our paper is the �rst at quantifying the costs from speci�c tariffs accruing to poor countries. The MFN speci�c tariffs levied by rich WTO member countries translate into higher ad valorem equivalents for the exporters of low price goods, who are primarily low-income countries. This suggests that the prefer- ential tariff bene�ts enjoyed by poor countries might be offset by the losses from speci�c tariff. We calculate the losses from speci�c tariffs and compare them to the gains from preferential tariffs. We show that the speci�c tariffs levied by the EU on its agricultural imports wash away more than half of the welfare gains enjoyed by the Sub-Saharan African countries from EU preferen- tial tariffs. REFERENCES Abdel-Rahman, K. 1991. “Firms’ Competitive and National Comparative Advantages as Joint Determinants of Trade Composition. Weltwirtschaftliches Archiv�, vol. 127, 83– 97. 162 THE WORLD BANK ECONOMIC REVIEW Alchian, A., and William R. Allen 1964. University Economics, Belmont, CA: Wadsworth Publishing Company Anderson, J., and P. Neary 2007. “Welfare versus market access: the implications of tariff structure for tariff reform�, Journal of International Economics, 71(1), 187– 205. Anderson, J. 1998. “Trade Restrictiveness Benchmarks�, The Economic Journal, 108(449), 1111– 1125 Anderson, K., W. Martin, and D. van der Mensbrugghe. 2006. “Doha Merchandise Trade Reform: What’s at Stake for Developing Countries?� World Bank Economic Review 20(2), 169–95 Armington, P.S. 1969. “A Theory of Demand for Products Distinguished by Place of Production�, IMF Staff Papers 16(2), 179– 201. Arndt, C. 1996. “An Introduction to Systematic Sensitivity Analysis via Gaussian Quadrature Method�, GTAP Technical Paper No. 2, Center for Global Trade Analysis, Purdue University Bagwell, K., and R. Staiger 2006. “What Do Trade Negotiators Negotiate About? Empirical Evidence from the World Trade Organization�, NBER WP12727 Baier, S.L., and J.H. Bergstrand 2007. “Do Free Trade Agreements Actually Increase Members’ Trade?�, Journal of International Economics, 71(1), 72– 95 Bhagwati, J. 2008. “Termites in the Trading System: How Preferential Agreements Undermine Free Trade�, Oxford University Press Boorstein, R., and R. Feenstra 1987. “Quality upgrading and its welfare costs in US steel imports 1969-74�, NBER working paper # W2452 Boue¨ t, A., and D. Laborde 2009. “The Potential Cost of a Failed Doha Round�, IFPRI Discussion Paper, # 00886. ¨ t, A., Y. Decreux, L. Fontagne, J. Se Boue ´ bastien, and D. Laborde 2004. “A Consistent, Advalorem Equivalent measure of Applied Protection across the world: The MAcMap-HS6 database�, CEPII, working Paper no. 2004–22 Boumellassa, H., D. Laborde, and C. Mitaritonna 2009. “A consistent picture of the protection across the world in 2004: MAcMapHS6 version 2�, IFPRI Discussion Paper ˜ o, and D.E. Weinstein 2008. “Optimal Tariffs and Market Power: The Evidence�, Broda, C., N. Lima American Economic Review, 98(5), 2032– 2065 Broda, C., and D.E. Weinstein 2006. “Globalization and the Gains from Variety�, The Quarterly Journal of Economics, 121(2), 541–585. Carrere, C. 2006. “Revisiting the effects of regional trade agreements on trade flows with proper speci�- cation of the gravity model�, European Economic Review, 50(2), 223–247 Falvey, R., and G. Reed. 2002. “Rules of Origin as Commercial Policy Instruments�, International Economic Review, 43(2), pp. 393– 407. ———. 1998. “Economic Effects of Rules of Origin�, Weltwirtschaftliches Archiv, 134(2), 209– 229. Falvey, Robert. 1979. “The Composition of Trade within Import-Restricted Product Categories�, Journal of Political Economy, 87(5), 1105–1114. Feenstra, Robert. 1995. “Estimating the Effects of Trade Policy�, NBER WP# 5051 ´ , L., and M. Freudenberg 1997. “Intra-Industry Trade: Methodological Issues Reconsidered�, Fontagne CEPII Working Paper, 97 –01. Francois, J., F. Van Meijl, and F. Van Tongeren 2005. “Trade Liberalization in the Doha Development Round�, Economic Policy, 20(42) Gibson, P., J. Wainio, D.M. Whitley, and M. Bohman 2001. “Pro�les of Tariffs in Global Agricultural markets� USDA, Agricultural Economic Report 796 Greenaway, D., and J. Torstensson 2001. “Economic Geography, Comparative Advantage and Trade within Industries: Evidence from the OECD�, Journal of Economic Integration, 15(2), 60– 80. Harrison, J.W., and K.R. Pearson 1996. “Computing Solutions for large General Equilibrium Models using GEMPACK�, Computational Economics Chowdhury 163 Hertel, T.W., D. Hummels, M. Ivanic, and R. Keeney 2007. “How Con�dent can we be of CGE-based assessments of Free Trade Analysis?�, Economic Modeling 24(4), 611– 635 Hertel, T.W. 1997. “Global Trade Analysis: Modeling and Applications�, Cambridge University Press Hoekman, B., A. Nicita, and M. Olarreaga. 2006. “Estimating the Effects of Global Trade Reform.� In B. Hoekman, and M. Olarreaga, eds., Global Trade Liberalization and Poor Countries: Poverty Impacts and Policy Implications. Washington D.C.: Brookings Institution Press, and Paris: Institut de Sciences Politiques. Hoekman, B., F. Ng, and M. Olarreaga 2002. “Eliminating Excessive Tariffs on Exports of Least Developed Countries�, World Bank Economic Review’, 16(1), 1– 21 Horridge, J.M., and D. Laborde 2008. “TASTE: a program to adapt detailed trade and tariff data to GTAP-related purposes�, GTAP Technical Paper, Centre for Global Trade Analysis, Purdue University Huff, K., and T. Hertel 1996. “Decomposing Welfare Changes in GTAP,� GTAP Technical Paper No. 5, Center for Global Trade Analysis, Purdue University, West Lafayette, IN. (revised 2001). Hummels, D., and A. Skiba 2004. “Shipping the Good Apples Out: An Empirical Con�rmation of the Alchian-Allen Hypothesis�, Journal of Political Economy, 112(6), 1384–1402. Irwin, D. 2010. “Trade Restrictiveness and Deadweight Losses from US Tariffs, 1859–1961�, American Economic Journal: Economic Policy 2(3): 111– 133 Kee, H.L., A. Nicita, and M. Olarreaga 2009. “Estimating Trade Restrictiveness Indices�, Economic Journal 119(534), 172– 99. ———. 2008. “Import Demand Elasticities and Trade Distortions�, The Review of Economics and Statistics, 90(4), 666 –682. Krishna, K. 2005. “Understanding Rules of Origins�, NBER Working Paper 11150. Krishna, K., and A. Kruger. 1995. “Implementing Free Trade Areas: Rules of Origin and Hidden pro- tection� in A. Deardorff, J. Levinsohn, and R. Stern (eds), New Directions in Trade Theory. Ann Arbor, MI: University of Michigan Press. Leamer, E. 1974. ‘Nominal Tariff Averages with Estimated Weights’, Southern Economic Journal, 41(1), 34–46. ˜ o, N., and M. Olarreaga 2006. ‘Trade Preferences to Small Developing Countries and the Welfare Lima Costs of Lost Multilateral Liberalization’, World Bank Economic Review, 20(2), 217– 240. Lloyd, P.J., J.L. Croser, and K. Anderson. 2010. “Global Distortions to Agricultural Markets: New Indicators of Trade and Welfare Impacts, 1960 to 2007�, Review of Development Economics 14(2), 141–60. McDougall, R.A. 2002. “A New Regional Household Demand System for GTAP�, GTAP Technical Paper, vol. 20. Global Trade Analysis Project, Purdue University Narayanan G., and Terrie L. Walmsley, (Ed.). 2008. “Global Trade, Assistance, and Production: The GTAP 7 Data Base�, Center for Global Trade Analysis, Purdue University. Schott, Peter. 2004. “Across-Product versus Within-Product Specialization in International Trade�, QJE, 119(2), 647–678. UNCTAD. 1964. “Towards a New Trade Policy for Development,� Report by the Secretary General of UNCTAD, United Nations, New York. Von Kirchbach, F., and M. Mondher 2003. “Market Access Barriers: A Growing Issue for Developing Country Exporters�, International Trade Forum, Issue 2/2003. Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW • Conditional Cash Transfers and HIV/AIDS Prevention: Unconditionally Promising? Hans-Peter Kohler and Rebecca Thornton • Just Rewards? Local Politics and Public Resource Allocation in South India Timothy Besley, Rohini Pande, and Vijayendra Rao • An Axiomatic Approach to the Measurement of Corruption: Theory and Applications James E. Foster, Andrew W. Horowitz, and Fabio Méndez • How Much of Observed Economic Mobility is Measurement Error? IV Methods to Reduce Measurement Error Bias, with an Application to Vietnam Paul Glewwe • Inequality of Opportunity in Egypt Nadia Belhaj Hassine • Can Global De-Carbonization Inhibit Developing Country Industrialization? Aaditya Mattoo, Arvind Subramanian, Dominique van der Mensbrugghe, and Jianwu He • Trade Liberalization and Investment: Firm-level Evidence from Mexico ˘ lu Ivan T. Kandilov and Aslı Leblebiciog THE WORLD BANK 1818 H Street, NW Washington, DC 20433, USA World Wide Web: http://www.worldbank.org/ E-mail: wber@worldbank.org