30182 THE WORLD BANK ECONOMIC REVIEW Volume 18 2004 - Number 2 Do Macroeconomic Crises Always Slow Human Capital Accumulation? Norbert R. Schady The Distribution of Income Shocks during Crises: An Application of Quantile Analysis to Mexico, 1992-95 William F. Maloney, Wendy V. Cunningham, and Mariano Bosch 5,, Agricultural Tariffs or Subsidies: Which Are More Important for Developing Economies? Bernard Hoekman, Francis Ng, and Marcelo Olarreaga The Earnings Effects of Multilateral Trade Liberalization: Implications for Poverty Thomas W. Hertel, Maros Ivanic, Paul V. Preckel, and John A. L. Cranfield A Asymmetries in the Union Wage Premium in Ghana Niels-Hugo Blunch and Dorte Verner A DEVELOPMENT DATABASE Governance Matters III: Governance Indicators for 1996, 1998, 2000, and 2002 Daniel Kaufmann, Aart Kraay, Massimo Mastruzzi www.wber.oupiournals.org OXFORD ISSN 0258-6770 THE WORLD BANK ECONOMIC REVIEW EDITOR L. Alan Winters, World Bank EDITORIAL BOARD Abhijit Banerjee, Massachusetts Institute of Elizabeth M. King, World Bank Technology, USA Justin Yifu Lin, China Centerfor Economic Kaushik Basu, Cornell U.. ' , USA Research, Peking University, China Tim Besley, London School of Econonmics, UK Mustapha Kamel Nabli, World Bank Anne Case, Princeton University, USA Juan Pablo Nicolini, Universidad di Tella, Fran,ois Bourguignon, World Bank Argentinia Stijn Claessens, World Bank Howard Pack, University of Pennsylvania, Paul Collier, World Bank USA Augustin Kwasi Fosu, African Economic Jean-Philippe Platteau, Facultes Universitaires Research Council, Kenya Notre-Dame de la Paix, Belgium Mark Gersovitz, The Johns Hopkins Boris Pleskovic, World Bank U. ..'. USA Martin Ravallion, World Bank Jan Willem Gunning, Free University, Mark R. Rosenzweig, Harvard Amsterdam, The Netherlands University, USA Jeffrey S. Hammer, World Bank Joseph E. Stiglitz, Columbia University, USA Karla Hoff, World Bank Moshe Syrquin, University of Miami, USA Ravi Kanbur, Cornell University, USA Vinod Thomas, World Bank Following Alan Winters' appointment as Director of the Research Group at the World Bank, Professor Jaime de Alelo of the University of Geneva will become Editor of the Review. He will be responsible for volume 19 onwards. The World Bank Economic Review is a professional journal for the dissemination of World Bank-sponsored and outside research that may inform policy analyses and choices. It is directed to an international readership among economists and social scientists in government, business, and international agencies, as well as in universities and development research institutions. The Review emphasizes policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but not necessarily proficient in advanced mathematical or econometric techniques. Articles will illustrate how professional research can shed light on policy choices. Inconsistency with Bank policy will not be grounds for rejection of an article. Articles will be drawn from work conducted by World Bank staff and consultants and from papers submitted by outside researchers. Before being accepted for publication, all articles will be reviewed by two referees who are not members of the Bank's staff and one World Bank staff member. Articles must also be recommended by a member of the Editorial Board. Non-Bank contributors are requested to submit a proposal of not more than two pages in length to the Editor or a member of the Editorial Board before sending in their paper. Comments or brief notes responding to Review articles are welcome and will be considered for publication to the extent that space permits. Please direct all editorial correspondence to the Editor, The World Bank Economic Review, The World Bank, 1818 H Street, Washington, DC 20433, USA, or wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 18 * 2004 - Number 2 Do Macroeconomic Crises Always Slow Human Capital Accumulation? 131 Norbert R. Schady The Distribution of Income Shocks during Crises: An Application of Quantile Analysis to Mexico, 1992-95 155 William F. Maloney, Wendy V. Cunninghamn, and Mariano Bosch Agricultural Tariffs or Subsidies: Which Are More Important for Developing Economies? 175 Bernard Hoekmant, Francis Ng, and Marcelo Olarreaga The Earnings Effects of Multilateral Trade Liberalization: Implications for Poverty 205 Thomas W. Hertel, Maros Ivanic, Paul V Preckel, andJohn A. L. Cranfield Asymmetries in the UJnion Wage Premium in Ghana 237 Niels-Hugo Blunch and Dorte Verner A DEVELOPMENT DATABASE Governance Matters III: Governance Indicators for 1996, 1998, 2000, and 2002 253 Daniel Kaufmazzn, Aart Kraay, and Massinio Mastruzzi 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 18, 3 Issues, 2004): Academic libraries-Print edition and site-wide online access: US$113/179, Print edition only: USS107/£75, Site-wide online access only: US$101/71; Corporate-Print edition and site-wide online access: US$137/494, Print edition only: USS130489, Site- wide online access only: US$123/484; Personal-Print edition and individual online access: USS42432. Please note: £ Sterling rates apply in Europe, USS elsewhere. There may be other subscription rates available; for a complete listing, please visit www.wbro.oupjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary subscriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as firm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be considered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. BACK ISSUES: The current year and two previous years' issues are available from Oxford University Press. Previous volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORNIATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, Oxford OX2 6DP, UK. E-mail: jnls.cust.serv@oupjournals.org. Tel: +44 (0)1865 353907. Fax: + 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@oupjournals.org. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, 1-1-17-5F, Mukogaoka, Bunkyo-ku, Tokyo, 113-0023, Japan. E-mail: okudaoup@po.iijnet.orjp. Tel: (03) 3813 1461. Fax: (03) 3818 1522. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published 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. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 1818 H Street, NW, Washington, D.C. 20433, USA. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oupjournals.org/permissions. ADVERTISING: Inquiries about advertising should be sent to Helen Pearson, Oxford Journals Advertising, PO Box 347, Abingdon OX14 1GJ, UK. E-mail: helen@oxfordads.com. Tel/Fax: +44 (0)1235 201904. 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 WORI.D BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/TIIE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. PAPER USED: The World Bank Economic Review is printed on acid-free paper that meets the minimum requirements of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPY RIGHT ©) The International Bank for Reconstruction and Development/THE WORLD BANK 2004 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 WiP 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Do Macroeconomic Crises Always Slow Human Capital Accumulation? Norbert R. Schady The impact of macroeconomic crises on the investments made by parents in the human capital of their children is a question of considerable policy importance. Analysis of the effects of the profound 1988-92 macroeconomic crisis in Peru on the schooling and employment decisions of school-age children in urban areas finds no effect on atten- dance rates but a significant decline in the fraction of children who are both employed and attend school. It also finds significantly higher mean educational attainment for children exposed to the crisis than for those who were not. These findings may be related: Children who are not employed have more time available and may therefore put more effort into school. Normally Simon and I would have gone to work after high school, but jobs weren't to be had anyway, and the public college was full of students in our condition, because of the unemployment. -Saul Bellow, The Adventures of Augie March How do macroeconomic crises affect the schooling and employment decisions of children and their parents? Do such crises invariably slow human capital accumulation, especially among the poor, thus transmitting poverty across generations? Policymakers in the developing world, as well as such international organizations as the United Nations Children's Fund and the World Bank, often worry that households that are unable to smooth consumption during a crisis may cut back on expenditures on the education, health, and nutrition of their children (World Bank 2001). This is plausible but not self-evident. In the standard neoclassical model of human capital investment individuals acquire schooling until the (expected) marginal benefit of an additional year of Norbert R. Schady is Senior Economist, Development Research Group at the World Bank; his e-mail address is nschady@worldbank.org. The author would like to thank Jean Boivin and Olivier Deschenes for many useful conversations and work on an earlier draft of this article and, for their comments, Harold Alderman, Francois Bourguignon, Francisco Ferreira, Cesar Guadalupe, Gillette Hall, Hanan Jacoby, Jose Rodriguez, Jaime Saavedra, Carolina Sanchez-Paramo, Miguel Urquiola, three anonymous referees, participants at the Latin American and Caribbean Economic Association Conference in Montevideo, Uruguay, on October 18-20, 2001, and participants at the Inter-American Development Bank conference on "Crises and Disasters: Measurement and Mitigation of their Human Costs" in Washington, D.C., on November 13-14, 2001. The author also thanks Gilberto Moncada at Peru's National Institute for Statistics and Informatics, Jaime Saavedra at the Development Analysis Group in Peru, and Moises Ventocilla at Cuanto for help with the data. THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2. Oc The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi: 10.1 093/wber/lhhO36 18:131-154 131 132 THE WORLD BANK ECONOMIC REVIEW, VOL. i8, NO. 2 education equals the marginal cost (Rosen 1977; Willis 1986). The marginal benefit of one more year of schooling is the resulting increase in the discounted expected stream of earnings and the marginal cost is the forgone income and such direct private costs as tuition and transportation. Borrowing constraints may be an additional cost of attending school by preventing consumption smoothing, especially by poor households (Becker 1964; Jacoby 1994 on Peru). In general, an adverse macroeconomic shock will depress current employ- ment and wage prospects, so the opportunity cost of attending school will fall. Holding everything else the same, this should lead to increased investments in human capital. But a shock could also make borrowing constraints more bind- ing and thus reduce the total amount of schooling chosen. When macroeconomic shocks are persistent, they may also depress expected lifetime earnings, thus affecting the marginal benefit from schooling. If the lifetime earnings of all individuals are reduced by the same percentage, regard- less of their schooling, then the marginal benefit associated with an additional year of schooling will be lower (by this same percentage). But crises need not have a uniform across-the-board effect on expected earnings. For example, if the expected lifetime earnings for individuals with less education are disproportio- nately affected by the crisis, the marginal benefit to schooling could rise. Insofar as macroeconomic crises change the marginal costs of schooling, they may also affect the timing and intensity of schooling-in particular, the extent to which children combine schooling with part-time employment. Finally, the effect of a crisis on the wages and employment prospects of adults in a household may also have an effect on the schooling and employment decisions of children. Thus the effect of a macroeconomic crisis on schooling is ambiguous in theory. Children (or their parents) may choose more or less schooling, they may antici- pate or postpone further schooling, and they may expend more or less effort in school. The total effect of a crisis on schooling will depend on the relative magnitude of the changes in the marginal costs and benefits from education, as well as on the cross-price elasticity of child employment and adult wages. Given such uncertain predictions, careful, country-specific empirical work is needed. Goldin (1999) finds a large increase in secondary school enrollment rates during the Great Depression in the United States, especially in the states hit hardest by unemployment. In Latin America, De Ferranti and others (2000) suggest that by and large enrollment decisions are unaffected by macroeconomic crises, especially moderate ones. More recent work on Argentina confirms that the crisis of the late 1990s and early 2000s did not change overall enrollment levels, but it may have had negative effects on the quality of schooling because of high associated rates of teacher absenteeism (Espafia and others 2002). In Indonesia the deep financial crisis of 1998 appears to have had only very small effects on schooling outcomes (Strauss and others 2004; Thomas and others 2004). In rural India, Jacoby and Skoufias (1997) find that school attendance is responsive to seasonal fluctuations in income, although this does not appear to result in large losses of human capital, whereas Jensen (2000) shows large negative effects of rainfall shocks on school Schady 133 enrollment in C6te d'Ivoire. In sum, the effect of income shocks on schooling seems to vary considerably by country (as well as possibly by the nature of the crisis), with the largest effects being found in the poorest countries. This article analyzes the effects of the 1988-92 economic crisis in Peru on school attendance, employment, and years of schooling completed. The analysis finds no evidence that school attendance fell during the crisis, either overall or for specific population groups, such as the poor. Overall attendance was stable, and the propor- tion of children who combined school with work declined significantly during the crisis. The number of grades completed for a given age was higher for children exposed to the crisis and increased with the number of years of crisis exposure. I. THE PERUVIAN SETTING The Peruvian case is well suited to a study of the impact of macroeconomic crises on schooling outcomes for several reasons. First, the crisis in Peru was particularly deep and prolonged. National accounts data suggest that gross domestic product (GDP) per capita fell almost 30 percent between 1988 and 1992. The depth and duration of the crisis could have severely stretched the ability of households to protect investments in schooling. Second, reliable household data are available. Three Living Standards Measurement Study (LSNIS) surveys cover the period before (1985/86), during (1991), and after (1997) the crisis. These surveys provide detailed information on household characteristics, such as demographic structure, income, consumption, access to credit, parental education, and child outcomes, including educational attainment and employment. Economic Developments, 1985-97 Peru followed an erratic economic course in the 1980s and 1990s. Between 1985 and 1990 the government of Alan Garcia attempted to stimulate the economy with a heterodox stabilization program, relying on reduced foreign debt payments, a price freeze, and economic stimulus through wage increases, job creation pro- grams, and increased investments in education and health. These policies encour- aged high growth rates in 1986 and 1987, but the inflationary pressures and budget deficits quickly proved unsustainable (Glewwe and Hall 1994). The coun- try slid into a deep recession and hyperinflation in 1988. By the end of Garcia's term in 1990, the economy was in a state of near collapse: GDP per capita had fallen 10.5 percent in 1988, 13.4 percent in 1989, and 6.9 percent in 1990, and inflation had soared to 667 percent, 3,399 percent, and 7,482 percent (figure 1). Along with the economic crisis came a deterioration in public safety. By 1990 there was virtual 1. Other studies have considered the impact of macroeconomic volatility-although not macro- economic crises per se-on schooling outcomes. Using cross-country regressions, Flug and others (1998) report that income and employment volatility have negative effects on secondary school enroll- ment in Latin America. Behrman and others (2000) suggest that the low (or negative) growth rates in the 1980s in lIatin America set back the rate of growth of schooling attainment in the region. 134 THE WORLD BANK ECONONMIC REVIEW, VOL. 1 8, NO. 2 FIGURE 1. Real per capita GDP and Inflation in Peru, 1980-99 0 0 o~~~~~~~~~~~A .Lo 0 0 C~ U)~~~~~~~~~~~~~~~~~~~~~~~~U o / ) U o 0)( o- Inlto ID e aia(94sls L0O_ 0 r R t on iC~) Source: World Bank data.~~~~~~~~~~~~~~~~~~~~~~~~e cIvi war between the armed forces and the insurgents of the Shining Path and T6pac Amaru Revolutionary Movement in some areas of the country. The administration of Alberto Fujimori, who took office in 1990, opted for more orthodox economic remedies. Reforms included elimination of controls on prices, interest rates, and foreign exchange transactions; tariff reductions; labor market deregulation; and a far-reaching program of privatization. Growth picked up after 1992, inflation fell to 74 percent in 1992 and to less than 12 percent by 1995 (see figure 1), and poverty declined. Like the crisis, the recovery appears to have been far reaching, affecting all regions and most households (World Bank 1995, 1999a).2 Meanwhile, order was gradually reestablished in the countryside, in particular after the 1992 capture of Abimael Guzman, leader of the Shining Path. Education in Peru, 1985-97 Compared with many of its Latin American neighbors and other countries with similar income levels, Peru has attained impressively high enrollment rates in both primary and secondary school. Analysis of the 1997 LSlS survey suggests that 97.5 percent of 6-11-year-olds, 86.3 percent of 12-17-year-olds, and 2. Strict comparisons of poverty measures are not possible because of differences in the coverage of the surveys (see section 11). Schady 135 34.9 percent of 18-25-year-olds were attending some educational institution.3 These figures are particularly impressive because of the low levels of public spending on education. In 1997 total public spending on education in Peru was 3 percent of GDP (2.4 percent net of spending on pensions), compared with the Latin American average of 4.5 percent (World Bank 1999b). In contrast, private expenditures on education in Peru, at about 2 percent of GDP, are high and well above the average of 1.3 percent for Organisation for Economic Co-operation and Develoment member countries. As in other countries, private expenditures vary by education levels and income quintile. In 1997 households spent about three times as much to send an 18-25-year-old to school as to send a 6-11-year-old. Differences across income quintiles are starker. On average, house- holds in the richest income quintile spent about eight times as much for children ages 6-11 attending school as did households in the poorest income quintile, five times as much for children 12-17, and four times as much for children 18-25.4 II. DATA AND SAMPLE MEANS The main sources of data for this analysis are the 1985/86, 1991, and 1997 Peru LSMS surveys. These surveys contain information on household characteristics, school attendance, grade attainment, and labor force participation. The questions in all three surveys are generally comparable,5 but coverage of the surveys is not comparable across years. Although the 1985/86 and 1997 surveys visited households in each one of Peru's seven "natural regions"-Lima, the urban and rural areas of the coast, sierra (or highlands), and selva (or jungle)-the 1991 survey sampled households only in Lima, the urban coast, and the urban and rural areas of the sierra. There are also reasons for concern about the 1991 sample for the rural sierra. Mean income appears to be higher in 1991 than in 1997, and parents' mean years of education approximately doubles between 1985/86 and 1991 and stays constant thereafter (table 1). It is likely that inadequate funding for the 1991 survey or the dangerous conditions caused by widespread terrorism in the rural 3. Author's calculations based on the 1997 LSMS survey tapes. These numbers cover all seven regions in Peru, including rural areas and the urban areas of the selva and so do not correspond to the numbers in tables 2 and 3, which cover only Lima and the urban areas of the coast and sierra. Figures for school attendance for 6-14-year-olds can also be calculated from the 1996 Peru Demographic and Health Survey (Filmer 1999). These tend to be quite similar, except for 6-year-olds, whose attendance rates in the Demographic and Health Survey (61.3 percent) appear to be much lower than those from the LSMS surveys, likely a result of differences in the wording of the question. 4. Still, by international standards, these differences in expenditure patterns across income quintiles are modest. As a comparison, in the Philippines households in the richest quintile spend 20 times as much as households in the poorest quintile per student in primary school, 11 times as much per student in secondary school, and 11 times as much per student at the tertiary level (Schady 2001, p. 34). 5. There are some exceptions. For example, there are differences across surveys in the recall periods and in the level of disaggregation in the questions concerning household expenditures oii education. TABLE 1. Sample Means, by Region and Year Urban sample, excluding urban selva Lima only Urban coast only Urban sierra only Rural sierra only Variable 1985/86 1991 1997 1985/86 1991 1997 1985/86 1991 1997 1985/86 1991 1997 1985/86 1991 1997 Proportion attending school 0.90 0.97 0.92 0.92 0.98 0.94 0.91 0.99 0.93 0.84 0.94 0.88 0.42 0.69 0.37 only (ages 6-11) Proportion employed 0.(( 0.00 0.00 0.0( 0.00 0.00 0.00 0.00 0.00 0.00 0.(0 0.01 0.01 0.00 0.02 only (ages 6-11) Proportion neither in school 0.00 0.00 0.02 0.00 0.01 0.02 0.00 0.00 0.02 0.00 0.00 0.02 0.01 0.00 0.01 nor employed (ages 6-11) Proportion both in school 0.10 0.02 0.06 0.08 0.01 0.04 0.09 0.01 0.05 0.16 0.06 0.09 0.56 0.30 0.60 and employed (ages 6-11) Proportion attending 0.63 0.82 0.70 0.66 0.81 0.73 0.63 0.87 0.69 0.57 0.75 0.65 0.16 0.36 0.17 school only (12-17) Proportion emploved 0.04 0.03 0.06 0.03 0.04 0.06 0.05 0.03 0.07 0.02 0.01 0.05 0.21 0.07 0.20 Q.- only (ages 12-17) Proportion neither in school 0.02 0.04 0.03 0.01 (1.03 0.04 0.03 0.04 0.04 0.02 0.04 0.02 0.02 0.03 0.02 nor employed (ages 12-17) Proportion both in school and 0.31 0.12 0.20 (0.28 (1.12 0.17 0.30 0.05 0.20 0.39 0.19 0.28 0.61 0.54 0.61 employed (ages 12-17) Log per capita income 7.55 7.27 7.55 7.74 7.37 7.66 7.33 7.21 7.41 7.44 7.13 7.51 7.13 6.82 6.54 Proportion with access 0.22 0.17 0.41 0.25 0.15 0.39 0.18 0.16 0.42 0.26 0.21 0.42 0.10 0.11 0.20 to credit Proportion rural NA NA NA NA NA NA NA NA NA NA NA NA 1.00 1.(0 1.00 Mean age, all household 23.68 25.26 27.39 24.40 25.72 27.40 22.96 25.16 27.75 23.01 24.47 26.87 21.65 23.36 23.23 members Mean years of education 7.24 8.03 8.20 7.83 8.(06 8.34 6.44 7.13 7.47 6.95 9.17 8.94 3.12 5.14 5.00 of father Mean years of education 5.33 6.17 6.58 5.76 6.05 6.74 4.93 5.57 5.80 4.87 7.14 7.3(0 1.35 3.16 2.88 of mother Mean household size 7.81 7.24 6.96 7.68 7.42 7.08 7.82 7.47 7.05 8.09 6.60 6.54 7.36 6.47 6.89 Note: Weighted means, with the weights given by the expansion factors in the surveys. NA, not applicable. Souirce: Author's calculations based on t.s\is surveys for 1985/86, 1991, and 1997. Scbady 137 sierra in the early 1990s prevented enumerators from visiting outlying, poorer, less well-educated households. The sample for the analysis is therefore limited to Lima and the urban areas of the coast and sierra. In 1997 these three regions jointly accounted for about 58.1 percent of the population. Regressions were also run with samples that included the rural sierra. In all these results, which are available on request, the impact of the crisis was larger in absolute value than the impacts reported for the urban sample. Conzstructioni of Variables The analysis considers the impact of the 1988-92 crisis on three outcomes: school attendance, mean number of grades completed, and employment. Atten- dance is based on answers to two questions in the surveys. All household members age 6 and older are asked xvhether they are "currently attending school or studying something." Those who answer "no" are then asked whether they "attended school or studied something in the last 12 months."6 The variable constructed for attendance takes a value of one for all individuals who answer "yes" to either of these questions. Because a large share of households in the 1985/86 survey were questioned during the summer vacation months of December-March, attendance rates based on only the first question would be unreasonably low. The 1991 and 1997 surveys were both conducted between September and November.' A variable was also constructed for the total num- ber of grades a child has completed. The measure of employment is based on questions asked of all household members ages six and older. These questions first ask respondents whether they worked "as an employee for a business, corporation, government, a boss, or another individual" and second whether they worked "for themselves [a cuenlta propial or as an unpaid family member," or "on the [family] farm." The reference period for both questions is the past week. Follow-up questions with longer reference periods (12 months) are asked of those who answered that they did not work in the last week. All those who answer affirmatively to any of the four questions are considered to have been employed.8 6. This measure of attendance does not take into account the number of times children actually go to school-a shortcoming of the data. Using the term enrollment would also be somewhat inaccurate, however, because children may be formally enrolled in school but not attend. The Spanish word used in the survey questions (asistiralcolegio) is closer to attendan.ce than it is to cnrollbnent (the translation for enrollment would be estar iniscrito ; inatriculado eni el colegio). See Paxson and Schadv (2002) for a similar construction of variables. 7. Insofar as differences in the timing of the surveys make comn parisons of school attendance (or employment-see later discussion) across years more problematic, even after taking into account both attendance questions in the I SNiS surveys, this should not affect comparisons between 1991 and 1997. 8. In the 1985/86 and 1991 surveys workinig 'on the farm" and working "for themselves, or as an unpaid family member" are asked separatelv, whereas in the 1997 survey both options are listed as part of the same question. Household chores are meant to be excluded from these categories and are asked about in the following questions. 138 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. 2 The Peruvian think-tank the Development Analysis Group (GRADE) has calcu- lated comparable consumption and income aggregates for the 1985/86, 1991, and 1997 LSMS surveys as well as for a 1994 LSMS survey. It used 1997 price deflators to adjust for price differences across natural regions, and the national consumer price index to deflate consumption and income over time. GRADE'S income aggre- gates are used for the estimations in this study, with the exception of the 1994 survey because GRADE'S aggregates for that year suggest that mean income was lower in 1994 than in 1991, whereas the national accounts show a clear improve- ment during this period. It is therefore unclear whether 1994 should be treated as a crisis year.9 The 1994 data are included in some (unreported) regressions, however, and the results suggest that there are no significant differences in school attendance and employment patterns between 1991 and 1994. Some of the results reported here use a measure of crisis "exposure" as an explanatory variable. Years of exposure are determined by the number of years a child was between the ages of 6 and 17 during 1988-92. All school-age children surveyed in 1985/86 were unexposed to the crisis, and all school-age children surveyed in 1991 were exposed to the crisis (with the extent of expo- sure determined by their age). For school-age children surveyed in 1997 exposure depends on age; those ages 11-17 were exposed to the crisis during their school- age years, whereas those ages 6-10 were not. Overall, then, the range of this exposure variable is between zero for children who were never of school age during the crisis and five for children who were of school age during every year between 1988 and 1992 (see appendix table Al).'0 Sample Means The 1988-92 crisis affected all regions and all income quintiles (tables 1 and 2). Income dropped across the board, and credit became harder to come by. Only 17 percent of households report having had access to credit in 1991 compared with 22 percent in 1985/86 and 41 percent in 1997. Tables 1 and 2 also show expected secular trends in parental education (rising), age (rising), and household size (falling). Children (or their parents) make decisions about schooling and employment jointly and sort themselves into one of four mutually exclusive categories: attend school only, work only, neither attend school nor work, and both attend school and work. The means for these four categories are shown by survey year for the full urban sample (excluding the urban selva, which was not included in the 1991 survey) and for each geographic stratum separately (see table 1) and by gender and by poorest and richest income quintiles for the full urban sample (see table 2). Results on attendance and employment patterns are presented separately for children ages 6-11, corresponding roughly to primary school age, and ages 12-17, 9. See Deaton (2003) and Ravallion (2003) for possible explanations of such divergences between the survey and national accounts figures. 10. For the age calculations all children surveyed in 1985/86 are taken to have been surveyed in 1985, given that the majority of households were surveyed in 1985 rather than 1986. TABLE 2. Sample Means, by Year, Gender, and Income Quintile, Urban Sample, Excluding Urban Selva Boys Girls Poorest income quintile Richest income quintile Variable 1985/86 1991 1997 1985/86 1991 1997 1985/86 1991 1997 1985/86 1991 1997 Proportion attending school 0.90 0.98 0.92 0.90 0.97 0.92 0.89 0.96 0.91 0.90 0.97 0.96 only (ages 6-11) Proportion employed 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 only (ages 6-11) Proportion neither in school 0.00 0.00 0.01 0.00 0.00 0.03 0.00 0.00 0.04 0.01 0.01 0.01 nor employed (ages 6-11) Proportion both in school 0.10 0.02 0.06 0.10 0.03 0.06 0.10 0.04 0.05 0.09 0.02 0.03 and employed (ages 6-11) Proportion attending 0.59 0.82 0.67 0.67 0.82 0.73 0.60 0.83 0.68 0.72 0.84 0.77 school only (ages 12-17) Proportion employed 0.03 0.03 0.07 0.04 0.03 0.05 0.05 0.03 0.06 0.02 0.02 0.02 only (ages 12-17) Proportion neither in school 0.01 0.04 0.03 0.03 0.03 0.04 0.03 0.03 0.04 0.02 0.02 0.03 nor employed (ages 12-17) Proportion both in school 0.37 0.12 0.23 0.26 0.12 0.18 0.32 0.12 0.23 0.23 0.12 0.17 and employed (ages 12-17) Log per capita income 7.56 7.27 7.56 7.54 7.26 7.55 6.36 6.28 6.62 8.84 8.37 8.71 Proportion with access 0.22 0.16 0.41 0.23 0.17 0.40 0.11 0.18 0.37 0.32 0.22 0.41 to credit Mean age, all household 23.33 24.76 27.11 24.01 25.73 27.66 20.96 23.63 24.88 27.45 28.15 30.90 members Mean years of education 7,24 7.92 8.18 7.25 8.14 8.22 5.25 6.74 6.88 9.24 10.00 11.10 of father Mean years of education 5.28 6.16 6.55 5.37 6.18 6.62 3.68 5.24 5.27 7.46 8.47 9.79 of mother Mean household size 7.80 7.27 7.00 7.81 7.21 6.91 8.54 7.35 7.50 6.30 6.04 5.47 Note: Weighted means, with the weights given by the expansion factors in the surveys. Source: Author's calculations based on LSNIS surveys for 1985/86, 1991, and 1997. 140 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. Z corresponding roughly to secondary school age. Attendance levels for school- age children in urban areas are very high throughout, at 97 percent or higher for children ages 6-11 and 90 percent or higher for children ages 12-17 (see tables 1 and 2). By and large, most children go to school in urban Peru, and there is no indication that attendance levels were lower in the 1991 crisis year. But the extent to which children combine school and work changed drama- tically across years. The fraction of children who were working is consistently lower in 1991 than in 1985/86 or 1997. Differences are particularly large for older children. In the full urban sample, 31 percent and 20 percent of children ages 12-17 were both attending school and employed in 1985/86 and 1997, respectively. In 1991, by contrast, only 12 percent of children in this age group combined school and work. This pattern holds remarkably consistently across strata, for girls as well as boys, and for children in the richest and poorest income quintiles. Charting the mean number of grades passed by age for children exposed and unexposed to the crisis, without conditioning on attendance, shows that mean years of schooling are below the ideal on-track progression line. This is espe- cially clear at higher ages, because of dropouts and the cumulative effects of late school entry and repetition (figure 2). However, children exposed to the crisis have an average of 0.1-0.2 more years of schooling. There is no clear pattern whereby children unexposed to the crisis fall further behind or catch up with age. III. ECONOMETRIC SPECIFICATION AND RESULTS To control for changes in household characteristics across survey years, data from the three surveys are pooled to estimate probit regressions for contem- poraneous events-the probability of attending school or being employed. These regressions include dummy variables for the two noncrisis survey years (1985/86 and 1997) and other controls. The omitted category is therefore the crisis year of 1991. The changes in the probability associated with the 1985/86 and 1997 dummy variables are interpreted as year effects conditional on changes over time in household characteristics. Pervasive, predictable differ- ences across years could be consistent with a causal effect of the 1988-92 crisis on attendance and employment. The second part of the analysis reports the results of ordinary least squares (OLS) regressions in which the dependent variable is cumulative-the number of grades passed for age. These regressions include dummy variables for two survey years, as well as the measure of crisis exposure. The coefficients on the measure of exposure are interpreted as the marginal effect of a year of exposure to the crisis on the number of grades passed. For both parts of the analysis, specification (1) includes dummy variables for survey year, geographic stratum, gender, and a set of age dummy variables. The OLS regressions also include the measure of crisis exposure. Information on years Schady 141 F i (J U R E 2. Number of Grades Completed, by Age 1 2 -8 E 0 cow 8 / Ideal progression , 6 Exposed to crisis - Unexposed to crisis ci4 E z 0 7 8 9 10 1t 12 13 14 15 16 17 Age Note: Children are expected to start first grade at age 6 or 7. Therefore they should have passed first grade by age 8, second grade by age 9, and so on, unless they start school late or repeat a grade. Children are considered to be on track if they have completed at least this number of grades for their age, a generous definition because children who start school at age 6 should have completed first grade by age 7. Source: Author's calculations based on the LSMIS surveys for 1985/86, 1991, and 1997. of education of both parents (separately) is available for a subsample of chil- dren-those who live in households in which both parents are present and who are the sons or daughters of the household head. Specification (2) therefore adds controls for the education of both parents (separately), dummy variables for the age structure within the household, and a main effect in household size. The LSMIS survey data include identifiers for mapping households in the surveys to the districts in which they live.1' The fixed effects specifications in (3) limit the analysis to households living in districts that were included in the sample in 1985/86, 1991, and 1997 and include the variables in specification (2), as well as dummy variables for each district. These regressions therefore identify differences in attendance levels, employment levels, and educational attainment from within-district variation across survey years or birth cohorts. 11. Peru has more than 1,800 districts, which are the smallest political and administrative division in the country (roughly comparable to counties in the United States). According to the 1993 population census, average district population was 12,600 inhabitants. Some predominantly rural districts have fewer than 200 inhabitants, whereas a handful of urban districts in Lima have more than 100,000 inhabitants (Schady 2002, p. 419). 142 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 (Note that district fixed effects are not a satisfactory solution for the problem of differences in the sample in the rural sierra between the 1985/86 and 1997 surveys and the 1991 survey. The likely problem with the 1991 sample for the rural sierra is not that enumerators did not visit outlying districts but that they did not visit outlying areas within a given district.) One disadvantage of the estimations that include measures of parental education and those with district fixed effects is that both are limited to subsamples of the data. In the three regions considered in the analysis, data on the schooling of both parents are available only for 73.9 percent of 6-1 1-year-olds, and only 77.8 percent of 6-11-year-olds live in districts that were included in the samples in 1985/86, 1991, and 1997. The sample sizes for the within-district regressions are further reduced when all children in a district have the same outcome-for example, if all children ages 6-11 in a district attend school or if none of them is employed. Better controls therefore come at the cost of smaller sample sizes and less precision. Attendance and Employment For children ages 6-11 and for those ages 12-17 the probability of attending school is significantly lower in 1997 than in 1991, but not in 1985/86 (table 3). TABLE 3. Probability of Attending School Specification (1)a Specification (2)b Specification (3)' Children ages 6-11 1985/86 0.000 -0.000 0.000 1991 - - 1 997 -0.0 1 6*18't -0.0 12 * * -0.0 19-1' Pseudo R2 0.12 0.29 0.44 Number of observations 3,257 2,573 845 Children ages 12-17 1985/86 0.002 0.001 0.002** 1991 - - 1997 -0.017** -0.003` -0.002* Pseudo R2 0.17 0.26 0.30 Number of observations 4,393 3,092 2,130 Note: The dependent variable takes a value of one if a child is attending school and zero otherwise. The table reports changes in the probability of attending school at the means of other variables. Standard errors are corrected for heteroskedasticity and clustering. >#Significant at the 5 percent level. ***Significant at the 1 percent level. 'Includes geographic stratum and gender controls and a vector of age dummy variables. bSupplements specification 1 with variables for the education of both parents (separately), household size, the number of household members under age 3, ages 3-5, ages 6-8, ages 9-11, ages 12-14, and ages 15-17. cThis fixed effects specification limits the sample to households living in districts that were included in the sample in 1985/86, 1991, and 1997, and includes the variables in specification 2, as well as a dummy variable for each district. Source: Author's calculations based on LSMIS surveys for 1985/86, 1991, and 1997. Schady 143 TABLE 4. Probability of Being Employed Specification (1)a Specification (2)" Specification (3)' Children ages 6-11 1985/86 0.089 * 0.082* 0.054 * 1991 - - - 1997 0.056**t 0.078 ** 0.042** Pseudo R2 0.10 0.13 0.15 Number of observations 4,322 3,160 2,234 Children ages 12-17 1985/86 0.223 >* 0.161' * 0.168""" 1991 - - - 1997 0.138 * 0.157** 0.144*" Pseudo R2 0.07 0.11 0.14 Number of observations 4,405 3,100 2,253 Note: The dependent variable takes on a value of one if a child is employed and zero otherwise. The table reports changes in the probability of being employed at the means of other variables. Children are considered to be employed if they responded affirmatively to any of the following questions in the surveys: (1) "In the last 7 days, did you work as an employee for a business, corporation, government, a boss, or another individual?"; (2) "and in the last 12 months?"; (3) "In the last 7 days, did you work for yourself, [a cuenta propia, or self-employed], or as an unpaid family member, or on the farm?"; (4) "and in the last 12 months?" Standard errors are corrected for heteroscedasticity and clustering. -Significant at the 5 percent level. `*Significant at the 1 percent level. 'Includes geographic stratum and gender controls and a vector of age dummy variables. 'Supplements specification 1 with variables for the education of both parents (separately), household size, the number of household members under age 3, ages 3-5, ages 6-8, ages 9-1 1, ages 12-14, and ages 15-17. 'This fixed effects specification limits the sample to households living in districts that were included in the sample in 1985/86, 1991, and 1997, and includes the variables in specification 2, as well as a dummy variable for each district. Source: Author's calculations based on LSMS surveys for 1985/86, 1991, and 1997. But the differences are small-the probability of attending school is, at most, 2 percentage points lower in 1997 than in 1991.12 Differences in the probability of being employed are much larger than dif- ferences in the probability of attending school. On average a child age 6-11 was 4-9 percentage points more likely to be employed in either 1985/96 or 1997 than in 1991, and a child age 12-17 was 14-22 percentage points more likely (table 4). All of these effects are significant at the 1 percent level or higher. Looking at these results together suggests that the margin for adjustment in urban Peru is primarily in the extent to which children combine school and employment rather than in the extent to which they attend school. Results of 12. Note that children ages 12-17 in 1997 were actually exposed to the crisis. If the crisis compelled some of these children to leave school, they might not have returned to school thereafter. The comparison of the dummy variables for 1991 and 1997 is therefore less persuasive for this older group than for the children ages 6-11. An anonymous referee provided this insight. 1 44 THE WORLI) BANK ECONOMI(. REVIEW, VOL. 1 8, NO. Z regressions on differences across years in the probability that a child attends school and works relative to the probability that a child only attends school are very similar to those for differences in the probability of being employed because of the very high attendance rates (table 5). Children were much more likely to combine school and work in 1985/86 and 1997 than in the crisis year of 1991. (The results from multinomial regressions, unreported but available on request, show no significant differences across years in either the odds that children are employed only or in the odds that they are neither attending school nor employed. The only exception is in the odds of working only for 12-17- year-olds relative to attending school and working, which is significantly higher in 1997 than in 1991.) The regression results reported in tables 4 and 5 show a clear pattern of lower probabilities of child employment in 1991. The analysis then considers the kind of work children are doing-wage or nonwage employment. For children ages 6-11 the reduction in employment was almost exclusively a reduction in non- wage employment (table 6). For children ages 12-17 the picture is mixed. The proportion of children in both wage and nonwage employment was lower in 1991 than in either 1985/86 or 1997, although the bulk of the adjustment also TABLE 5. Probability of Both Attending School and Being Employed Specification (1)" Specification (2)" Specification (3)c Children ages 6-11 1985/86 0.084>: 0.078* * * 0.054V* 1991 - - 1997 0.054* 0.076W ' - 0.040** Pseudo R2 0.09 0.12 0.16 Number of observations 4,220 3,086 2,193 Children ages 12-17 1985/86 0.223 - 0.162 * 0.168'-' 1991 - - - 1997 0. 133'-- 0. 150*-- 0. 137*- Pseudo R2 ().07 0.11 0.14 Number of observations 4,113 2,939 2,135 Note: The dependent variable takes on a value of one if a child is both attending school and employed and zero if she is in school only. The table reports changes in the probabilitv of both attending school and being employed at the means of other variables. Standard errors are corrected for heteroscedasticity and clustering. **Significant at the 5 percent level. :""'Significant at the 1 percent level. aIncludes geographic stratum and gender controls and a vector of age dummy variables. ESupplements specification 1 with variables for the education of both parents (separately), household size, the number of household members under age 3, ages 3-5, ages 6-8, ages 9-11, ages 12-14, and ages 15-17. 'This fixed effects specification limits the sample to households living in districts that were included in the sample in 1985/86, 1991, and 1997, and includes the variables in specification 2, as well as a dummy variable for each district. Souirce: Author's calculations based on L.SMNS surveys for 1985/86, 1991, and 1997. Scbady 145 TABLE 6. Sample Means for Schooling and Employment Status, by Type of Employment 1985/86 1991 1997 Cbildren ages 6-1 1 Attending school and no employment 89.8 97.3 92.2 Attending school and wage employment only 0.7 0.5 1.0 Attending school and nonwage employment only 8.9 1.8 4.7 Attending school and both wage and 0.2 0.0 0.1 nonwage employment Not attending school and no employment 0.2 0.3 2.0 Not attending school and wage employment only 0.0 0.0 0.0 Not attending school and nonwage employment only 0.1 0.0 0.1 Not attending school and hoth wage and 0.0 0.0 0.0 noniwage employment Total 100.0 100.0 100.0 Cbild rent ages 12--) 7 Attending school and no employment 63.0 81.7 70.1 Attending school and wage employment only 7.5 2.8 5.9 Attending school and nonwage employment only 21.4 8.8 13.6 Attending school and both wage and 2.6 0.1 1.0 nonwage employment Not attending school and no employment 1.9 3.5 3.4 Not attending school and wage employment only 1.3 1.8 3.8 Not attending school and nonwage employment only 1.7 1.1 1.8 Not attending school and both wage and 0.6 0.2 0.3 nonwage employment Total 100.0 100.0 100.1) Note: Weighted means, with the weights given by the expansion factors in the surveys. .Souirce: Author's calculations based on isNos surveys for 1985/86, 1991, and 1997. took place in nonwage employment. The poor aggregate conditions in 1991 may have lowered demand for child employment in the nonwage sector either because of lower demand for goods and services produced by home enterprises, which frequently employ children wvithout remuneration, or because older workers lost their wage-paying jobs and displaced children from their nonwage jobs-or both. Poorer households are more likely to be credit-constrained than are their better-off counterparts. Although the means in table 2 show no clear differ- ences in outcomes between the first and fifth income quintiles, the regression coefficients reported in tables 3-5 might not hold for all income levels. To investigate this possibility, nonparametric (lowess) regressions are run of the probability of being both employed and in school on predicted log per capita annual household income (based on variables for age, gender, education of the household head, and household size and composition). (Predicted income is used because of the potential endogeneity of income in a child employ- ment regression, but the results are very similar using actual income.) The 146 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 FIGURE 3. Probability of Both Attending School and Being Employed for Children Ages 6-11, by Year 0 N-Dn 0 - 1985 - - X--- 1991 … ----- 1997 (m LOl .0 - e ~~~~~~\ "\ Q \ ~~~~~~~~~~~~~\\ 6 6.5 7 7.5 8 8.5 Log of per capita income Source: Author's calculations based on the LsNs surveys for 1985/86, 1991, and 1997. probability of being both in school and employed is lower in 1991 than in 1985/86 and 1997 across the entire income distribution (figures 3 and 4) 13 The regression line is flattest in 1991, suggesting that the employment con- straint may have been most binding for the poor. Still, the evidence for different responses to the crisis is limited. Additional (unreported) specifica- tions test for heterogeneity in the year effects by interacting the year dummy variables with a number of household characteristics, such as the education of parents, household size, the dummy variables for geographic stratum, and the gender of the child. The coefficients on these interaction terms are insigni- ficant, so we cannot rule out the null hypothesis of identical household responses to the crisis. Finally, the difference in slopes across years is most pronounced for children ages 6-11. Plausibly, jobs previously held by 6-11- year-olds were taken by 12-17-year-olds during the crisis, when jobs were scarce. 14 13. Comparable graphs for the fraction of children who are attending school only, essentially the complement of the graphs in figures 3 and 4, show that the fraction of children who are attending school only is highest in 1991 at all income levels. 14. An anonymous referee provided this insight. - - - - - Schady 147 FIGURE 4. Probability of Both Attending School and Being Employed for Children Ages 12-17, by Year Co0 E 5: N) 1985 0 U) --- 1991 CD ------- 1997 0 -o 2 6 7 8 9 Log of per capita income Souirce: Author's calculations based on the LSMIS surveys for 1985/86, 1991, and 1997. Educational Attainment Two sets of OLS regressions for the mean number of grades completed are presented in table 7. In the upper panel, every year of crisis exposure is constrained to have the same effect, whereas the effect is allowed to vary by year in the lower panel. The results for the continuous exposure variable suggest that every year of exposure to the crisis is associated with a 0.04 to 0.05 increase in the number of grades completed. That is, children who were of school age for the entire 1988-92 period would have completed about one-quarter more grades than those who were not of school age during the crisis. Because grades completed are measured in whole integers only, one way of interpreting this coefficient is that one out of every four or five children exposed to the entire crisis period has completed one more grade compared with children who were not exposed to the crisis. The dummy variable specifications in the lower panel of table 7 provide some evidence that school attainment does not increase proportionately with crisis exposure. A "low" amount of crisis exposure (1 to 2 years) appears to have no effect on the number of grades completed, whereas a "high" amount of crisis exposure (3 to 5 years) increases grade completion by about 0.2 years. This could 148 THE WORLD BANK ECONONIIC REVIEW, VOL. i 8, NO. Z TABLE 7. Average Number of Grades Completed, Ages 6-17 Specification (1)a Specification (2)b Specification (3)' Years of crisis exposure 0.050)" 0.044** 0.043' R2 0.81 0.86 0.87 Number of observations 8,646 6,203 4,917 Crisis exposure = 0 - - - Crisis exposure = 1 0.008 0.057 0.035 Crisis exposure = 2 -0.011 0.052 0.089 Crisis exposure = 3 0.204** 0.165* 0.173* Crisis exposure = 4 0.184** 0.236** 0.234*: Crisis exposure = 5 0.224>* 0.204* 0.199* R 2 0.82 0.86 0.87 Number of observations 8,646 6,203 4,917 Note: All specifications include dummy variables for the 1985/86 and 1997 survey years. Standard errors are corrected for heteroscedasticity and clustering. "Significant at the 10 percent level. *Significant at the 5 percent level. "**Significant at the 1 percent level. aIncludes geographic stratum and gender controls and a vector of age dummy variables. bSupplements specification 1 with variables for the education of both parents (separately), household size, the number of household members under age 3, ages 3-5, ages 6-8, ages 9-11, ages 12-14, and ages 15-17. cThis fixed effects specification limits the sample to households living in districts that were included in the sample in 1985/86, 1991, and 1997, and includes the variables in specification (2), as well as a dummy variable for each district. Source: Author's calculations based on LSMS surveys for 1985/86, 1991, and 1997. mean that it takes some time for children or their parents to adjust their schooling and employment decisions to a crisis. Too much should not be made of these more nuanced results, however. F-tests on the additional coeffi- cients fail to reject the null that these dummy variable specifications do not represent a significant improvement in fit over the more parsimonious specifica- tions in which crisis exposure enters the regression linearly. Interpretation of Results Macroeconomic crises can affect the total amount of schooling chosen, the timing of schooling, and the extent to which schooling is combined with work. The 1988-92 crisis in Peru appears to have reduced access to credit, which should have led to a reduction in human capital investment. The effect of the crisis on the marginal benefits from education is unclear in theory. Only if the crisis had been perceived as persistent and as disproportionately reducing the earnings and employment prospects of low-skilled workers could changes in the marginal benefit of schooling explain the observed patterns in enrollment and grade attainment. Yet the crisis does not seem to have had a disproportionate effect on less educated workers. Comparisons based on the 1985/86, 1991, and 1997 surveys suggest that incomes fell substantially for all categories of Scbady 149 households, whatever the amount of schooling of the household head."5 More- over, using data from annual labor force surveys conducted in Lima, Saavedra (1998) and Saavedra and Maruyama (1998) show that the rate of return to education estimated from a standard Mincerian regression fell noticeably during the crisis (from 0.11 in 1987 to 0.08 in 1991) and rose steadily thereafter (to about 0.13 in 1995). It therefore seems unlikely that changes in access to credit or in the marginal benefit to education can explain the patterns in school attendance and grade completion in Peru. The crisis appears to have affected education outcomes through a substantial reduction in forgone income-the opportunity cost of attending school. Real wages in urban areas dropped precipitously during the crisis. Saavedra (1998) estimates that the mean real monthly wage for informal sector workers in Lima fell from about 600 soles in 1987 to 200 soles in 1991, before recovering partially to about 400 soles in 1997. Formal sector workers suffered even more staggering declines, with real monthly wages plunging from 1,200 soles to 200 soles between 1987 and 1991 and recovering to 800 soles by 1997. These results suggest that there was a very steep decline in the opportunity cost of schooling in Lima during the crisis. Children who hold jobs in Peru in noncrisis years may have been less willing and able to combine school and work during the 1988-92 crisis. The extent to which a reduction in child employment frees up time for schooling or for child leisure is an unresolved issue in the literature on child labor. The three Peru LSNMS surveys have no time allocation data and so cannot be used to answer this question conclusively. But the results presented here are consistent with some substitution between employment and schooling. During the 1988-92 crisis children in Peru were less likely to combine school with work, and they were more likely to make adequate grade progress. Arguably, children who were not working could expend more effort in school. Two potential complications for interpreting the results are possible changes in public expenditures on education and changes in migration patterns, specifi- cally migration from rural to urban areas. An increase in public expenditures on education during a crisis could reduce the marginal cost of education (for example, if the increased expenditure takes the form of scholarships) or increase the marginal benefits (for example, if the increased expenditure improves the quality of education). 15. Glewwe and Hall (1998) report results based on a panel of households in Lima only included in the 1985/86 LsMs survey and in a follow-up survey in 1990. These results suggest that households in which the head had more schooling suffered smaller income losses. The analysis here finds no clear pattern. A simple tabulation of income by the education of the household head shows a greater drop in mean income in households with heads with some secondary education (-41 percent) than in households with heads with a primary education or less (-27 percent), or some tertiary education (-25 percent). The recovery in incomes between 1991 and 1997, by contrast, clearly favored households with more school- ing. Incomes between 1991 and 1997 increased by 2 percent for households with heads with a primary education or less, 19 percent for households with some secondary education, and 76 percent for house- holds with some tertiary education. 150 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 FIGURE 5. Recurrent and Capital Public Expenditures on Education, 1980-97 (constant 1997 soles) g LO~ a) U / 0 M co 4-- 0 0 o N ~ ~ ~ ~ ~ ~ ~ o- ___ / ~~~ __ __ 0 1985 1991 1997 Recurrent expenditures --- Capital expenditures Source: World Bank staff calculations based on data from the Peruvian Ministry of Economics and Finance. The Garcia government had no program for keeping children in school during the crisis. Indeed, public expenditures on education followed roughly the same pattern as per capita GDP -rising sharply between 1985/86 and 1987, dropping equally sharply between 1987 and 1991, and increasing steadily thereafter (figure 5). If anything, the constant attendance patterns and the improvements in age-specific grade completion during the crisis took place despite the changes in public expendi- tures on education. Moreover, it is unlikely that households compensated for changes in public expenditures by increasing private expenditures on education between 1985/86 and 1991 and reducing them thereafter. At best, households may have been able to protect these expenditures somewhat in the context of dramatic cutbacks in overall household income.16 Also, a decree passed between the 1995 and 1996 school years called for automatically promoting to second grade all children who completed first grade. This could bias downward the estimates for the impact of the crisis on grade attainment reported in this article. By construction, all 16. Data from countries affected by the 1998 East Asian crisis, for example, show a varied picture. In Indonesia private expenditures on education fell both in absolute terms and as a share of total household expenditures, whereas in the Republic of Korea private expenditures on education fell by less than overall household spending (World Bank 2000, p. 121). Scbady 151 first-graders 8 years old or younger in 1995, who stood to benefit from the automatic promotion policy, are defined as unexposed to the crisis because they would be 10 years old or younger in 1997. Migration is a concern because the sample is limited to Lima and the urban areas of the coast and sierra. In Peru migration has traditionally been from rural to urban areas, although there appears to have been an important migration from urban to rural areas in the mid-1990s. This reverse migration was a response to the improved security in the countryside as a result of the weakening of the Shining Path and Tupac Amaru Revolutionary Movement after 1992 and of government policies to encourage resettlement of abandoned rural areas. Lower mean incomes and lower adult education levels in rural areas suggest lower underlying propen- sities to attend school and complete grades among migrants than among the sedentary population. Rural to urban migration could thus diminish the propensity to attend school and make satisfactory grade progress of the 1991 sample relative to the 1985/86 sample, whereas urban to rural migration could diminish this propensity in the 1991 sample relative to the 1997 sample. To the extent that this is the case, the estimates presented here of the effect of the crisis would be downwardly biased. In the absence of migration the estimated coefficients on the measures of crisis exposure would have been even larger. VI. CONCLUSION This article examines the impact of the profound 1988-92 macroeconomic crisis in Peru on patterns of accumulation of human capital. The main finding is that households, including poor households, were very reluctant to make cutbacks in key human capital investments. There is no evidence of a drop in school atten- dance. Children exposed to the crisis were less likely to combine work with school and had completed more grades than children unexposed to the crisis. The results suggest that macroeconomic crises do not always slow human capital accumulation in developing economies. Of course, crises have serious consequences for household welfare. Income falls, and consumption may fall if households cannot smooth out the income shock. The health status of children may deteriorate. In Peru, Paxson and Schady (2004) find that the 1988-92 crisis led to a sharp increase in infant mortality and a deterioration in the nutritional status of children. The quality of education may suffer because of the cutbacks in private and public expenditure. The main message of these results is not therefore that there are no social costs of macroeconomic crises. Rather, the policy implications are first, given the variation in impacts reported in the literature, further research is needed to understand why macroeconomic crises lead to a slowdown in human capital accumulation in some countries but not in others. Second, in some middle- income countries, where aggregate reductions in incomes do not appear to lead to worse schooling outcomes, it may be more important to put in place policies to protect consumption and health rather than policies to reduce school dropouts during a crisis. 1 52 If IE L \Y )RI.) BANk I NOXI t ( I( Y II [\Y, (L. i 8, NO. 2 APl'PEN DI TAB[.E A- 1. Exposure to Economic Crisis by Age and Survey Year Age (years) Survev vear Exposure (nuniber of years) 6 1985/86 0 7 1985/86 0 8 1985/86 0 9 1985/86 0 10 1985/86 0 11 1985/86 0 12 198.5/86 0 13 1985/86 tI 14 1985/86 0 15 1985/86 0 16 1985/86 0 17 1985/86 () 6 1991 1 7 1991 2 8 1991 3 9 1991 4 1() 1991 5 11 1991 5 12 1991 5 13 1991 5 14 1991 5 15 1991 5 16 1991 5 17 1991 5 6 1997 0 7 1997 () 8 1997 0 9 1997 0 1() 1997 0 11 1997 12 1997 2 13 1997 3 14 1997 4 1 5 1997 5 16 1997 5 17 1997 5 Source: Author's calculations based on [I %ts surveys for 1985/86, 1991, and 1997. REFERBNCES Becker, Gary. 1964. IHumn)a Capital: A Theoreticala .nd Empirical Analysis witb Spccial Rel'rciicc to Education. Li , University of C I Press. Behrmani, Jere R., Suzanne Duryea, and Nliguel Szekely. 2000. "Schoolinig Ilnvesteints and Aggregate Conditions: A Household-Survey-Based Approach for Latin America aiid the Caribbean." Working Paper, Inter-Amnerican Development Banik, Washinigton, D.C. Scbadv 153 Deatoni, Angus. 2003. "Measuring Poverty in a (Growing World (or Measuring Growth in a Poor World)." N\K:tE Working Paper 9822. Cambrige, .Mass.: National Bureaul of Economic Research. De Ferranti, D)avid, Guillermo E. Perry, Indermit S. Gill, and Luis Serven. 2000. Seclurinzg Ooir Future in a Global Feoniorz01. Washington, D.C.: World Bank. Espafna, Sergio, Suhas Parandekar, and Maria Paula Savanti. 2002. "El impacto de la crisis en el proceso educativo en Argentina." Vorking Paper, World Bank, Washington, D.C. Filmer, Deoti. 1999. Educational Attainment and Fnirollmiet Pro/iles: A Resouirce Book Based on a,n Analysis of Demiographic and Hael/tb Sirveys Data. Washingtoii, D.C.: World Batik. Flug, Karniit, Antonio Spilinibergo, and Erik Wachteniheim. 1998. "Investments in Educarion: Do Economic Volatility and Credit Constraints Matter?" lournal of Developmient Leonomics 55(2):465-8 1. Glew'we, Paul, and Gillette Hall. 1994. "Poverty, InequLality and living Standards duiring Llnorthodox Adjustment: The Case of leru, 1985-1990." Leonooic Dei'elopmient and Cultural Chanige 42(4):689-717. . 1998. 'Are Some Groups Mlore Vulncral-le to Macroeconomic Shocks than Oth,rs? Hypothesis Tests Based on Panel Data from Perm." lournal of Developmient Economics 56( 1 :18 1-206. Goldin, Claudia. 1999. "Egalitarianism and the Returnls to Education during the Great Transformation of American Education." lourl7al of Political EcononlY 107(6):S65-S94. Jacoby, lianana G. 1994. 'Borrowing Constraints and Progress through School: Evidence from PerU." Rev'iew ol Econom7zics and Statistics 76(1:1 5 1-60. jacobv, Hanan G., and Emmanuel SkoLifias. 1997. "Risk, Financial Markets, anid Humian Capital in a Developing Country." Rev'ieiw' of Econiooiiic Stuidies 64(3): 311-35. Jensen, Robert. 2000. "Agricultural Volatility and Investments in Children." A,incrican Econiomic Review' Papers and Peoceedizgs 90(2):399-404. Paxson, Christina, anid Norhert R. Schady. 2002. "The Allocation and Imipact of Social Fuiids: Spending on School Infrastructure in lerul." W'orld Bank Ecozonozic Reivieiw 16(2):29-. 1 9. - . 2004. "Child Health and the 1988-92 Economic Crisis in Peru." PolicY Research Working Paper 3260. World Bank, Washington, D.C. Ravallion, Martin. 2003. ' L., .,.-, Aggregate Welfare in Developing Countries: Hoxv Well Do National Accounts and Surveys Agree?" Review of Ecoruniomics and Statistics 85(3):645-51. Rosen, Sherwini. 1977. "Hiuman Capital: A Survey of Empirical Research." In Ronald Ii.. 'I.i ,. ed., Research in Labor Economics. Greenwich, Corn.: JAI Press. Saavedra, jaine. 1998. "Crisis real o de expectativas? El empleo en el Perei antes y desprius de las reformas estruCtirales." GRADE Working Paper 25. Grupo de Andlisis para el D)esarrollo, 1Lima. Saavedra, Jaime, and Eduardo IMaruvansa. 1998. 'Los retornos a la educacihn y ai Ia experiencia en el Peru: 1985-1997." In Richard Webb anid Moises Venitocilla, eds., PobreZa v econonua social: Anals:s de U17n1 eneCUCest7 FNNIV- 1997. Limia: Instituito Cuanto. Schady, Norhert R. 2001. "Convexity and Sheepskin Effects in the Hluman Capital Earnings ELnction: Recent Evidence for Eilipino M\en." Policy Research Working Paper 2566. World Bank, Washington, D.C. 0 2(1002. "Picking the Ploor: Indicators for Geographic Targetinig in Pleru." Rcu'iew' of luComue z(and Wealth 48(3):417-33. Strauss, John, Kathleen Beegle, Agus Dwivinto, Yulia Herawati, Daan Pattinasarany, Elan Satriawan, Bondan Sikoki, Sukamdi, and Firman Witoelar. 2004. Indonesian Living Stanidairds befiore and after the Financial Crisis: Fiidence fr7om the uIndionesia lsamiiy lift' Surv'v. Santa. M\lonica, Calif.: RAND. Thomas, Dumncan, Kathleen Beegle, Elizabeth FrankenblLerg, Bondan Sikoki, John StraUss, and Graciela Teruel. 20(14. "Education in a (Crisis." journal of Developmenut Econuomnics -!74 1):53-85. Willis, Robert. 1986. "Wage Determinants: A Survey and Reinterpretation of Humran C'apital Earnings Functionis." In Orlev Ashenfeltter anid Richard LayaFad, eds., Hf-andbou,ok of l.ab'or Econiomuics. New 'tork: North Holland. 154 THE WORI D) BANK ECONOMIC REVIEW, VOL. I8, NO. I World Bank. 1995. "Peru: Countrv Assistance Strategy." Washington, D.C. 1 999a. Pern Edocation at a Crossroads: (i :i ,* and Opportunities for the 21st Centtury. Washington, D.C. -_____ 1999b. Poverty anzd Social Developments in Peru, 1994-1997. Washington, D.C. - 2000. East Asia: Recovery and Beyond. Washington, D.C. - . 200(1. World Dev'elopment Report 2000/2001: Attackinig Poverty. New York: Oxford University Press. The Distribution of Income Shocks during Crises: An Application of Quantile Analysis to Mexico, 1992-95 William F. Maloney, Wendy V. Cunningham, and Mariano Bosch Moving beyond the simple comparisons of averages typical of most analyses of house- hold income shocks, this article employs quantile analysis to generate a complete distribution of such shocks by type of household during the 1995 crisis in Mexico. It compares the distributions across normal and crisis periods to see whether observed differences were due to the crisis or are intrinsic to the household types. Alternatively, it asks whether the distribution of shocks during normal periods was a reasonable predictor of vulnerability to income shocks during crises. It finds large differences in the distribution of shocks by household types both before and during the crisis but little change in their relative positions during the crisis. The impact appears to have been spread fairly evenly. Households headed by people with less education (poor), single mothers, or people working in the informal sector do not appear to experience disproportionate income drops either in normal times or during crises. Mexico's economic collapse in 1994-95 led to massive declines in household incomes averaging roughly 30 percent. Though the "Tequila crisis" enjoyed a singular celebrity due to Mexico's proximity to the United States and the com- prehensiveness of its prior reforms, neither the large size nor the adverse social consequences of the impacts were unique. Within Latin America, Argentina and Colombia are presently in equally brutal downturns, and in the 1990s William F. Maloney is lead economist, Office of the Chief Economist for Latin America and the Caribbean at the World Bank; his e-mail address is wmaloney@worldbank.org. Wendy V. Cunningham is senior economist, Human Development Sector, Latin American and Caribbean Region at the World Bank; her e-mail address is wcunningham@worldbank.org. Mariano Bosch Mossi is a junior professional in the Office of the Chief Economist for Latin America and the Caribbean at the World Bank; his e-mail address is mboschmossi@worldbank.org. This article was prepared as a background paper for the regional publication Securing Our Future in a Globalized Economy and was partially funded by the Regional Studies Program of the Office of the Chief Economist for Latin America and the Caribbean at the World Bank. It was also an input into the World Bank report "Income Risk, Household Coping Strategies and Income Security Policy in Mexico." The authors thank Ishan Ajwad, Omar Arias, Francisco Ferreira, and Norman Hicks for insightful comments. They are also grateful to the Mexican National Institute of Statistics, Geography, and Information (INEGI for use of their data. INEGI is in no way responsible for any incorrect manipulation of the data or for any erroneous conclusions. THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO 2. © The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi: 10.1093/wber/lhhO38 18:155-174 155 156 THE WORID BANK ECONONIIC REVIEW, VOL. I8, NO. I Russia and several countries in Asia experienced notable collapses. Thus, identifying who is most affected in these types of crises is likely to be a central issue in the design of safety nets throughout the world. The causes of the Mexican crisis have been described at length elsewhere (see Gil-Diaz and Carstens 1996; Edwards 1998; Calvo and Mendoza 1996; Kamin and Rogers 1996; among others). Broadly speaking, liberalization of trade and capital markets in the mid-1980s was followed by an expansion of private borrowing and aggregate demand more generally and by an appreciating exchange rate, all in the context of an underregulated financial sector.' Spec- ulative pressure on the peso in late December 1994 led the central bank to allow it to float, and it stabilized at nearly double its previous value against the dollar. Prices rose 35 percent, and output fell 6.2 percent across 1995, with dramatic social consequences. Wages remained fixed in nominal terms, leading to a real decline of 25-35 percent, and unemployment, though low by global standards, almost doubled from 3.9 percent to 7.4 percent (figure 1). Using a rich panel data set, this article investigates which types of house- holds suffered the largest income shocks during the crisis. It is thus similar in spirit to Glewwe and Hall's (1998) study of consumption falls during Peru's crisis of 1985-90 and is related more generally to the emerging literature on vulnerability. However, a central theme of this article is that policy analysts need to move beyond estimates of average falls in income or consumption and focus on the entire distribution of positive and negative income shocks. As an extreme example of why this may be important, consider the hypothetical distributions of the income shocks of two groups whose median income change has the same value, mn (figure 2). In group A, real income falls more or less evenly across most members, as illustrated by the relatively tight and symmetric distribution of shocks affecting this group. But in group B some heads of households lose their jobs and experience a very large negative shock to household income whereas others experience more moderate negative or even positive shocks, as illustrated by a larger mass of the distribution to the right of the median and a long and thick tail to the left. Looking only at the central tendency of the data therefore obscures important information about who suffered most from the crisis and which groups deserve more attention from a policy perspective. 1. Though there remains ample room for discussion about which particular fundamentals were central, or even whether the crisis was driven by self-fulfilling expectations only loosely linked to fundamentals (Sachs and others 1996), what is important is that commentators do not consider it unique, and so its impacts are of interest for understanding crises more generally. Edwards (1998), for example, sees a strong parallel in the deficient credibility of the reforms in both the Mexican crisis of 1982 and the Chilean crisis of 1998. An alternative view faults excessive borrowing in an environment of unfounded optimism about the future course of the economy. See, for example, Conley and Maloney 1995. Maloney, Cunningham, and Bosch 157 FIGURE 1. Unemployment and Real Wages in Mexico, 1987-2002 110 8 Crisis 105 , 7 Real wages 100 Unempioyment 6 95- 5 90- 80 - , , : g g y / r, ,. ., , \ < 4 80 . .. . ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~3 75- 2 70- 65 60 ,,.....,,,,,...... ,,,, ..... ,,,,,. . ......,.., 0 N N r2 m m m 0 G2 (N2 °J c. 0 0 9S 9s to o ; N c,2, c; 2 8 8 0 0 . c us g2 us o us s m2 us us m us us f2 SC SC m f2 us us SC SC us f2 SC m 0 0 r 0 0 Note: The real wage is the nominal manufacturing wage deflated by the consumer price index. Unemployment figures are the official numbers covering the entire economy. Source: Mexican Institute of Statistics, Geography, and Information. FIGURE 2. Use of Quantile Regression as a Tool for Exploring the Distributions of Changes in Household Income Med-n 5(th (A& B) Group A Group B / / 0\ 2Oth (B6 h8(A) 80t* (B) m 0 Change mu household income M Source: Authors' analysis; see text. 158 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. 2 Quantile analysis offers a useful tool for uncovering this hidden information because it permits studying the distributions at several points, not just at the median.2 The results show significant differences in the distribution of income shocks by household type in Mexico, suggesting that such exploration is indeed valuable. Of perhaps secondary importance, the results also show that median or quantile regression is more robust to the extreme values that commonly emerge in this kind of exercise than traditional ordinary least squares (OLS) estimators and should probably be used routinely in estimating income changes. The second central theme of the article is that distributions of income shock across households during crises need to be compared against distributions during "normal" times, to see whether the observed differences are related to the crisis or are perhaps intrinsic to particular household types. This question can be turned around to ask whether looking at the distribution of income shocks during normal periods can help identify how vulnerable certain households might be to severe income shocks during a crisis. In the Mexican case, the relative distribu- tions do not change much across periods, with a few important exceptions. Finally, the analysis shows that several common stylized facts about who experiences the largest negative shocks during crises-particularly about how households headed by disadvantaged single mothers and informal sector work- ers are affected-are not supported by the data. The article offers some sugges- tive evidence on why these findings might be reasonable. I. QUANTILE ANALYSIS: STUDYING THE ENTIRE DISTRIBUTION OF INCOMIE SHOCKS Conditional mean regression estimators such as OLS are traditionally used to estimate linear relations among variables. Minimizing the squared sum of errors allows estimating the values of the parameters that predict the mean of the dependent variable, conditional on the chosen set of explanatory variables. However, asymmetries or heteroscedasticity in the distribution of errors may lead to substantially different estimates of the impact of the variables under study at different parts of the conditional distribution. Looking only at the central tendency (the mean, for instance) of the data may thus hide important elements of the story. Furthermore, if there are outliers or if the distribution of the disturbances is nonnormal, mean estimators may be inefficient and biased. These concerns can be partially addressed by estimating the conditional median regression, in which half the errors lie below and half above the fitted curve. Quantile analysis, introduced in Koenker and Bassett (1978), extends this analysis to estimating curves where approximately t percent of the residuals lie 2. See Buchinsky (1994, 1995) for a detailed discussion of quantile regression methods. Maloney, Cuinningham, and Bosch 159 below the regression line and (100- t) percent above. Thus, the tth quantile of Y conditional on X is given by (1) Qyi(r I Xi) =(T)Xi, where j(T ) is the slope of the quantile line and thus gives the effect of changes in X on the tth conditional quantile of Y. Estimation for different values of t (from 0 to 1) yields regression lines for various percentiles of the conditional distribu- tion of Y. Median regression (T = 0.5) gives the same results as OLS when the distribution is symmetric. The analysis presented next suggests that distributions of income shocks to households show strong evidence of heteroscedasticity and asymmetry and thus that OLS is inappropriate. The following straightforward specification is designed purely to capture the differential effects of a crisis on the income of different household types: (2) QlnvYi(C I Di,) = a((T) + z_j,(T)D,, where AInYi is the change in the log of per capita labor income of household i, Di is an indicator variable that takes a value of 1 for family i if it is a member of the i different household types at the beginning of the period and a value of 0 otherwise. These groups capture the education level and age of the head of the household, household composition, and work status of the head. As in standard OLS, conditioning on all these categories at once allows disentangling their separate effects-whether, hypothetically, informally employed household heads might be particularly hard hit at any quantile relative to the base group (see section II for details on the base group) captured by the constant, oc, or whether it is just that poorly educated people, who are disproportionately employed in the informal sector, are hit harder. In the simple case illustrated in figure 2, group A might be the base group, and the way group B differs from group A would be captured by including a single dummy variable D,B to distinguish each household in that group. Quantile analysis differs from traditional analysis by allowing the entire dis- tribution of shocks for each category to be parameterized. For group A both the 80th quantile, capturing the less negative and positive tail of the clistribution, and the 20th quantile, capturing the extreme negative shocks, are relatively close to the median. For group B not only is the variance greater-both the 20th and 80th quantiles are further from the median than for group A-but the 20th quantile is displaced far to the left, reflecting that some households experienced extreme losses such as would arise from loss of employment of the household head. The values of the coefficients 6j(T) in equation 2 at each quantile permit these different distributions to be sketched out. As in standard OLS regressions, at the median (-T = 0.5) a negative and significant coefficient on DiB, the dummy variable representing, for instance, those with incomplete primary education, would imply that the distribution of shocks to this group is centered to the left of (lower than) that of base group A in figure 2: "On average" the least educated were hit harder than the 1 60 I'HE WORL D BANK fCC0NO NOIIC RE VILI W, VOlI . i 8, N(). 2 base group. The other two quantiles help flesh out the rest of the distribution. At the 20th quantile a negative and significant coefficient would imply that the value at which the most negative 20 percent of shocks occurred is also shifted down relative to that for the omitted category: The less educated experience lower lows, and hence the tail of distribution B is extended to the left. The coefficient on the 80th quantile tells whether the group experienced higher highs than the base group-whether B has a longer upper tail than A. Together, these three coefficients describe the shape of the distribution of shocks for each group relative to the base group. Two sets of tests are used to see whether these coefficients are in fact signi- ficantly different from each other. First, the coefficients are plotted at each decile along with their standard errors (Koenker and Hallock 2001). Graphically, it is then possible to detect whether the coefficient of one quantile lies outside the confidence interval of another. More formally, an F-test is employed to determine whether the coefficients on the 20th, 50th, and 80th quantiles are statistically different (Gould 1997). The log difference specification in equation 2 is used because measurement of the magnitudes of change is neutral to the direction of change. The potential downside is that it is only a close approximation to the percentage change for small changes in Y. Because the primary interest is not the total change but the difference from the base category, the log approximation may turn out to be reasonable. However, Kennedy (1981) also shows that for the case of a semiloga- rithmic estimation like this one, calculating the true percentage change correspond- ing to the dummy regressors requires transforming the estimated coefficients by3 (3) 6'/ = exp(6j - [1/2]V[6jj) - 1, where 6j is the estimate from equation 2, V[6,] its estimated variance, and 6*, the estimated "corrected" value.4 This turns out to have a significant impact (about 20 percent) on the estimates of the base group changes, which show magnitudes of around 50 percent. As suggested, however, the adjustments to the estimates of the differential effects are generally small. The corrected values are reported here. Quantile analysis offers two other advantages over traditional techniques. First, by construction, the dependent variable and so the residuals are unlikely to be normally distributed. In fact, were percentage changes to be used, as is 3. Halvorsen and Palmquist (1980) show that in the case of semilogarithmic regressions, when a variable in logs is regressed on a set of dummv variable, the inability to differentiate across the noncontinuous dummy variahles means that their coefficients are biased estimators of their impact on the percentage change of the dependent variable. In the case of one dummy variable, In( Yo,/Y) = x + 6D is equivalent to Y,/Y, = exp()(1 + 6 )D, where 6* is the relative effect on the dependent variable of the presence of the factor represented b) the dummv variable. Thus, the corrected coefficient that we are interested in is 6- = exp(6) -- 1. Kennedv (1981 ) argues that in estimatinig 6' one must take into account that Eexp)6) = exp[(6 + 1)/2V(6)], providing the rationale for equation 3. 4. The authors are grateful to Omar Arias for bringing this point to their attention. A previous application of this technique to wage data in Latin America is in Arias (2001). MIahlonzey, (2Unninghamn, and Boscb 161 often done, the distribution would be bounded below at -1 and might have very long right tails. Second, median or quantile regression deals better with extreme values or outliers than do traditional regression techniques. Some very large percentage increases in income are found within the right tails. They might reflect measurement error, but they might also reflect a very low income in the denoiminator in the first period and a higher income in the second. This might be the case, for example, with self-employed workers, who show very high volatility in their incomes in general. Simple averages will be strongly affected by such values and standard olS even more so because it minimizes the squared residuals. But median or quantile regressions do not because they are not based on the distance of the residual from the regression line but simply on the number of observations on either side. More generally, this makes median or quantile tchniques desirable when outliers or extreme values are present. As a final methodological point, the very heteroscedasticity that quantile ana- Ivsis is supposed to reveal also implies that standard errors of Koenker and Basset's (1978) original formulation are underestimated and thus that the t-statistics are overstated (Rogers 1993). As in Gould (1992, 1997) bootstrapping techniques are used to generate the correct standard errors, and Davidson and Mackinnon's (2000) algorithm is used for determining the correct number of bootstraps. What Is the Linik to Vulnerability? As numerous analysts have noted, there is substantial variance in use of the term vildnerability.is In most cases what distinguishes the concept from poverty or depriva- tion, whiclh are based on the first moments of income or consumption, is its focus on the second moment, the variance, of these measures. It is not just where a household is now but the likelihood that it mav find itself in a worse position. This immediately moves the description of the ex ante distribution of shocks to center stage. Quantile analvsis can be an important tool for studying vulnerability to income shocks or to welfare losses where inference based on the assumption of a commlion distribution of shocks (as opposed to situations such as those described in figure 2) can be misleading. Such concerns are not obviated by the use of limited dependent variable techniques that look only at the probability of movements into poverty. If there is substantial heteroscedasticity in the errors, traditional logit- and probit-based approaches to identifyiig vulnerability to becoming poor will generate inconsistent estimates (Yatchew and Griliches 1984). This offers a rationale for analysis in a continuous context (Glewwe and Hall 1998; Cunningham and Maloney 2000). Repeated panels can be used to estimate the distribution of shocks during normal times for each household type. However, it is not obvious that this 5. See ligon and Sehecter (2002) for a review of the literature and also Gainanou and Miorduch (2002). See also Chaudhuri (20)02) and Clhaudlhuri and others (2002). For related litcrature, see Jalan and Ravallion (2000). 1 62 THE WORLD BANK ECONONMIC REVIEW, VOL. I8, NO. 2 information is useful for identifying who is likely to experience large income shocks during crises, when almost by the definition of crisis the underlying data generation process may have changed. This is explicitly tested in the Mexican crisis in the next section. In the event that the distributions change and prior information is therefore not useful, Glewwe and Hall (1998) and this article can then be seen as offering individual draws from the crisis data generation process. They can tell us who experienced a given shock in a particular event but not who is most likely to experience such a shock in repeated events, which is of greater interest for policy design. Identifying who is most vulnerable to income or consumption shocks during crises would require repeated observations of crises. This, in turn, is likely to be complicated by the fact that different precipi- tating causes and contexts (for instance, how labor markets adjust) may imply very different patterns of shocks to household income. II. DATA The analysis studies changes in per capita income of families across the period 1994-96 using the Mexican National Urban Employment Survey (ENEU). The ENEU has conducted extensive quarterly household interviews in the 16 major metropolitan areas from 1992 to the present, which includes the period of the Tequila crisis. The sample is selected to be geographically and socioecono- mically representative. The questionnaire is extensive in its coverage of issues traditionally found in such employment surveys, including participation in the labor market, wages, and hours worked. Additionally, a household identifica- tion variable permits construction of household incomes. The ENEU is structured as a rotating panel. Each quarterly sample includes five cohorts, each in a different stage of the interview cycle: one-fifth of the sample in its first interview, one-fifth in its last (fifth) interview, and three-fifths in intermediate stages. To construct the panels and ensure proper identification, individuals were linked by position in an identified household, level of educa- tion, age, and gender.6 Household incomes were then constructed by aggregat- ing across the reported household members. A panel was constructed with 9,877 households beginning in 1994:3 to 1995:3, covering the onset of the crisis and period of largest wage losses around the Tequila crisis. Dummy variables were included for the education level of the head (primary incomplete, primary complete, secondary incomplete, secondary complete); age (under 25 or over 45); more than the mean number of children (1.3) in the house- hold; and household structure (single mothers with children, single women without children, and single men without children). Three dummy variables were also 6. Using just the first variables to concatenate and following changes in sex across the panel led to mismatching (or misreporting) of less than 0.5 percent. Initial and final values have under 10 percent zeros, approximated by adding a 1 prior to logging. Malonev, Cunningham, and Bosch 163 included for the sector in which the household head works (informal self-employed, informal salaried, and a residual category for being out of the labor force, unem- ployed, or otherwise not earning an income).' The constant, a, captures the base group of households headed by married, middle-age, college-educated males work- ing in the formal sector, with less than the mean number of children.8 The quantile analysis provides a very complete description of the distribution of the movement of household incomes. Worth highlighting is the relatively fluid movement between income quintiles. For example, only about 30 percent of households were in the same income quintile after five quarters, and a similar share jumped from the bottom to the top two quintiles. There are several possible reasons. First, roughly 40 percent of household heads are self-employed, and as the next sections show, entrepreneurs experience much more volatility in income flows than do workers in the salaried formal sector, reflecting varying business conditions or personal preferences. This higher volatility is also a characteristic of cross-sectional data and thus is not related to the process of linking the panels. Second, there is a very high degree of mobility in the labor force, and frequent movements between formal and informal sector jobs imply large changes in reported income.9 Third, there may be measurement error arising from the survey process, imperfect recall by the interviewed household member, or noise in the tabulation process. It seems reasonable to assume that this error is not correlated with the household categories and so would not affect the description of each category's distribution relative to that of the base. III. RESULTS The OLS estimates of equation 2 indicate that the income of the base group, captured by the constant, fell about 52 percent over the one-year period (table 1, column a). To find the impact on other groups, and to see whether it differs statistically from the impact on the base group, the value of the dummy variable is added to the constant. Those with a incomplete primary education experienced income falls 36 percentage points less than those of the base group, or under a third of the base group. The income of informal self-employed household heads fell 5 percentage points more than that 7. The term informal is used here to refer to workers unprotected by labor laws. It includes owners of firms with fewer than 16 emplovees who do not receive social security or medical benefits (fewer than I percent have more than five employees) and employees in these small firms, identified as informal salaried workers. 8. Single mnen with children made up less than 0.7 percent of the sample; they are included in the base group . 9. Mlaloniey (1999) finds that voluntary movements into self-employment from formal salaried work lead to a 30 percent average jump in reported earnings. Some 70 percent of those who moved report moving voluntarily. In deciding what sector to work in, individuals compare welfare, not wages, so a move from a sector with benefits and job security to a microenterprise where firm mortality rates are very high dictates a substantial compensating wage and risk premium. This could be what is picked up in the high apparent mobility among income levels. 164 THE WORLD BANK ECONO,MIC REVIEW, VOL. IS, NO. 2 TABLE 1. Correlates of Log Changes in Household Income during the Mexican Crisis: Quantile Analysis, 1994:3-1995:3 OLS Regression Quantile Regressions Standard Adjusted (±3 SD)' 20% 50% 80% F-testb a b c d e f Primary 0.363*- 0.031 0.519*** 0.045 * 0.003 21.01 '* ' incomplete Primary 0.306`* 0.040 0.449* 0.065 - 0.007 14.18" * Secondary 0.185;- -0.05 0.252;'* 0.014 0.005 5.82'-> incomplete Secondary 0.192 0.016 0.259*** 0.031 0.012 3.20;- Young -0.014 -0.068 -0.065 -0.001 -0.009 0.67 Old -0.148 -0.068 -0.228**;- -0.045;-* -0.026 12.62k * More than 1.3 0.037 0.005 0.047 0.012 0.016 0.42 children Single mothers 0.024 0.034 -0.046 0.035 0.039 0.71 Single women 0.046 0.009 0.022 0.005 0.028 0.28 Single men -0.014 -0.072 -0.072 -0.004 0.013 0.27 Informal -0.053* -0.046 -0.227*** -0.071-- 0.125*;': 46.00`* self-employed Informal salaried 0.030 -0.006 -0.027 0.012 0.077;-* 3.11 X' No remuneration -0.197* -0.194*- -0.532~` -0.070'- 0.111 * ** 16.17** Base group -0.517`c -0.360*** -0.664* ` -0.292 - 0.058** 229.56*`- Number of 9,877 9,636 9,877 9,877 9,877 9,877 observations "Significant at the 10% percent level. -Significant at the 5% percent level. ***Significant at the 1% percent level. Note: The dependent variable is the change in log per capita income of the household between 1994:3 and 1995:3 regressed on characteristics of Mexican households and their heads. Standard errors were calculated using bootstrapping techniques from Gould (1992, 1997). Number of bootstraps were obtained by running the algorithm proposed by Davidson and Mackinnon (2000). Coefficients were corrected following Kennedy (1981). 'Observations with income growth outside ±3 standard deviations from the mean have been dropped. bThe null hypothesis is equality of the coefficients in the three quantile regressions estimated. Source: Authors' analysis based on data from one quarter panel of the Mexican National Urban Employment Survey. of the base group and the income of nonremunerated household heads fell 20 percentage points more. These results are illustrated in figure 3, which charts the OLS coefficient (solid lines) and confidence interval (broken horizontal lines), as well as the estimated coefficients at each quantile from the 10th to 90th (solid curves) along with its confidence interval (broken curves). The median regression (the 50th quantile) shows a different picture, however (see table 1, column d). The base group experienced a 29 percent drop in income, roughly half the estimate from the OLS regression. Examination of the data reveals that a few outliers are largely responsible for this difference. Maloney, Cunningham. and Bosch 165 When observations beyond plus or minus three standard deviations are dropped, the OLS estimate approaches the median estimates, at 36 percent (column b). Median regression is preferred to such trimming, however, because the trimming cutoff is arbitrary. For example, 74 of 241 observations that are classified as outliers are households headed by self-employed workers whose incomes show high variability even under normal circumstances and hence may be legitimate outliers. An additional 34 observations are in the nonremunerated category, for which a change to even small earnings could increase household income by a very large percentage. Because the median estimator is less sensitive to such extreme values, it avoids the arbitrary discard- ing of outliers. Among household categories with significant coefficients in the median regres- sion, households headed by workers with incomplete primary education saw their incomes fall by 4.5 percentage points less than the base group and those with completed primary school by 6.5 percentage points less. Households headed by older workers experienced drops 4.5 percentage points greater than did the base group, and the self-employed 7.1 percentage points greater. That these coefficients are significant is reflected in figure 3 by the fact that the confidence interval at the 50th quantile does not span the zero line. What also becomes apparent is the value of looking beyond the median to other quantiles. If in figure 3 the parameter value at one quantile lies outside the confidence interval of another, there are statistically significant differences in their values. Thus the impact of the explanatory variable differs across quan- tiles. As an example, for households with heads with incomplete or complete primary education, the 20th quantile estimates lie above and outside the con- fidence interval of the 50th, showing that they are significantly greater in value. That significant differences occur is confirmed by the F-test (see table 1), which shows that for all education categories, the old, the self-employed, and those with no remuneration, the hypothesis can be rejected that the coefficients at the 20th, 50th, and 80th quantiles are statistically the same. More careful examination at the 20th quantile-the value below which the most extreme shocks are found-shows that households with heads who are less educated were more cushioned than the base group of households whose heads were college educated, who show an income drop of 66 percent. For households whose heads had a primary or incomplete primary education, the 20th quantile is only around -15 percent. For complete and incomplete secondary education, it is roughly -40 percent. The negative coefficient suggests that shocks to income were harsher for the old and self-employed by 23 percentage points and by even more for those whose heads are without remuneration.' ° 10. Note that the overall effect for this group is a more than 100 percent drop in income, which seems implausible. This is due to the bad approximation properties of the log transformation before great percentage changes (even after using Kennedy's correction), such as when going from some income to no income at all. 1 66 THE WORLD BANK ECONOMIC REVIEW, VOL. i8, NO. 2 The 80th quantile, which completes the distribution, shows few differences from the pattern at the median. None of the education or household composition dummy variables is significant. For all three quantiles the distribution of the poorly educated relative to the base group of college educated does not corre- spond to what is sketched in figure 2. The distribution overall is shifted to the right at the median and has shorter, lower tails, but it appears to be the same as the base case above the median-it is simply compressed below it. Poorly educated workers, again conditional on all other characteristics, experienced equal or more moderate income shocks than the base group at every point of the distribution. The most strikingly distinct overall distribution is now that of families with self- employed heads whose coefficient at the 80th quantile is 12.5 percentage points FIGURE 3. Quantile and OLS Estimations of the Central Tendency and Distribution of Income Shocks by Household Characteristics of Mexican Panel ,~~~~~-" Q\ .. . . , \o lo 20 30 40 50 60 70 go 90 10 20 30 40 50 80 70 80 90 N~~~~0N Sewnda,y momplete S.mdxy complete N 10 20 30 40 00 60 70 0 90 10 20 30 40 50 00 70 B0 90 OconO, QOld nti Primnary Incomplete Primnary comoplete __-___--___-_ ,/ -- _' ff'~-a --_- 10 20 30 40 90 07 70 B0 90 nO 20 30 40 50 00 70 00 90 Ocanhii Qianirie Se ondruncomltSecar cOmlet Maloney, Cunningham, and Bosch 167 FIGURE 3. Conztinued Sg. ,,he sigl ,ma /~~~~~~ /. ___ ~~~~~~~~~~~~~~~-- - --------// 0 -.-.~ ------- - - ..."..lc 7 .. . / / 0o 20 70 40 00 00 70 e0 50 17 20 30 00 00 a0 70 80 900 Quanotie Ouarmin Single mother Single man 0 o - == , ~ ~-_ /7 I0 20 ,0 4O s0 1'n 3I0 O IIIo 9I7 s t S~~~~~~~~~~~~~infomlslfemployed >13ma secd -1 o_0 /~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 1o 20 30 40 00 50 70 80 so 0o 20 30 40 50 60 70 80 90 Quatilne Quansie Noe:hsfigue woemprned perce3t kidser B~~~~~ -. _ - o C=9f--___- - = = 0 n 20 30 N.60 7 0 9 0 20 3 0 s 0 7 0 s Quantile ~~~~~~~~~~~~~~~~~~~~~~Qiuaohiie nurnber of quantile regressions (10th-90th) for each of the variables in the estimation. The OLS coefficient and confidence intervals are included for comparison. Obviously, the OLS regression returns a constant coefficient over the different quantile regressions. A zero line in the y-axis is also reported to evaluate statistical signiificance. Sou), ce: Authors' analysis based on data from one quarter panel of the Mexican Urban Emplovment Survey. 168 THE WORLD BANK 1( ONOMII(. RFVIEW V' . I 8, No 2 higher than that of the base group, suggesting that they not only experienced a worse median shock, but that their variance is much greater, because they demon- strate higher highs (at the 80th quantile) and lower lows (at the 20th quantile). Somewhat surprisingly, the distribution of households headed by the salaried informal sector workers shows no difference from the base group of formal salaried workers except that their highs are 7.7 percentage points higher. Similarly, families with households headed by workers with no remuneration show a coefficient at the 80th quantile that is 1 I percentage points higher than that of the base group. The household composition variables are not significant at any quantile. The distribu- tions of shocks for families headed by single women, single men, or single mothers are indistinguishable from those for households headed by married men. IV. ARE "NORMAL' PERIODs GOOD PREDICTORS OF ClRSI S? The logical question to ask is whether these differential patterns of shocks are due to the crisis or whether they exist in normal times as well. Or, in the context of asking about the vulnerability of different groups to income shocks, can the distribution estimated during normal periods be taken as a predictor of what distributions will look like during crises? The same regressions are run again, this time using a year of complete panels to form a sample ending right before the crisis (table 2). This is constructed by combining four additional panels: beginning in 1992:2, 1993:1, 1993:2, and 1993:3, the last of which ends in 1994:4, immediately before the economic collapse that began after the devaluation in late December 1994. The results show that the distinct distributions exist in normal periods as well (first three columns of table 2). Household heads with primary, incomplete primary, and incomplete secondary education show values at the 20th quantile shifted to the right relative to values for the base group while the 80th quantile is shifted to the left-they have lower highs and higher lows although the median change is statistically the same. Put differently, households with less educated heads generally have lower variance in their household incomes. The reasons are not clear. On the upside, perhaps the possibilities for income growth are lower for the less well-educated. On the downside it is possible that fewer opportunities to smooth income through credit markets or savings dictate that households take measures to reduce downside volatility-for example, by putting additional workers in the labor market if the household head's job is lost. The reverse occurs for self-employed and salaried informal sector workers and for households whose head is not remunerated. In all three cases the 80th quantile is shifted to the right and the 20th to the left, suggesting a greater variance in earnings. The self-employed also have a median income change below that of the base group. Taken together, this seems consistent with the standard dynamics of a small firm sector: In any period some firms do exceptionally well, and some less well than their formal salaried counterparts, and as is the case in many countries, Mcaloney, (Cunningham, and Boscb 169 TABLE 2. Comparison of Shock Distributions hefore and during Crisis, Quantile Analysis Presaiiple Period Crisis Period 20% 50.%) 8X0% 20%/. 50% 80% Primary 0. 232' 0.009 -0.065 -- 0.232**' 0.036 0.073- incomplete Primary () IX3-'' 0.005 -(0O55 0.224'* 0.060*** 0.066** Secondary 0. 10- - 0.002 -0.033; 0.128* 0.012 0.039 incomplete Secondarv 0.079** 0.0)8 -0.016 0. 166 * 0.023 0.028 Young 0.041 0.008 -0.029 -0.102 -0.009 0.019 Old --0. 192'- --0.047*** 0.014 -0.046 0.001 -0.039 More than 1.3 0.028 0.0 1 - 0.004 0.019 -0.005 0.012 children Single mothers -0.047 (.((5 0.008 (0.001 0.030 0.031 Single women -- 0.163 3' -0.055* _-0.079;'-* 0.219'* 0.061 0.1 16"- Single men -0. 10,3 -0.030 (.010 1 0.031 0.026 (1.1)14 Informal -0.263' --0.069 0. 105*** 0.049 -0.0(5 0.018 self-employed Informal sIlaried -. 093'. 0.(20 0.077 " 0.071 -0.008 0.000 No remuLneratioi O0. 157 0. 059 9 0.242- -0.447*' '- -0.124*- -0.106*** Base group -0.436*>'- --0.0O I 0.529* i0.4(06 '- -0_280'- -0.308*** Numher of 41,676 41,676 41,676 41.676 41,676 41,676 observations 'Significant at the 10%, percent level. "' -Significant at the 5"%/ percent level. """Significant at the I1° percent level. Note: The table shows the coefficients of a regression in which five periods (1992:4-1993:4, 1993:1994:1, 1993:2-1994:2, 1993:3-1994:3, and 1994:3-1995:3) of household per capita income growth have been regressed on characteristics of Mexican households and their heads. The regression includes dummy variable interacted wvith one of each variable for the crisis period that goes from 1994:3 to 1995:3. The second columin for eachi period retrieves the coefficients of the interactioin of each variable with the dummv variable for the crisis period. Standard errors, were calculated usinig bootstrapping techniques from C,ould ( 1992, 1997). Number of bootstraps were obtainied bv runninig the algorithm proposed by Davidson and NMackinnon (2000). C. . ni r. were correctcd following KIennedY (I 98 1). Sooirce: Authiors' analysis based on five quarter panels of the Mexican National Urban Emplovnmen t Survcy. firm mortality rates are very high."l This could be argued as reflecting greater precariousness, but because almost 70 percent of workers entering the self- employed sector from formal salaried work report doing so voluntarily, this greater volatility is consistent with welfare improvement (Maloney 1999). 11. See Les\enson and Mialoney (1998). Fainzylher and others (2003) find that the rate of micro- enterprise o"wners returiniig to formal employnieit, one mCasure of enterprise mortality rates, is roughly equivalent in Mlexico to that in the lUnited States. 1 70 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 The same lower than base group growth in median incomes occurs for households headed by older workers. This may reflect a tendency toward declining incomes with retirement. The reverse logic may hold for families whose head shows no earnings at the beginning of the period. On average, many of those unemployed will find jobs, so the tendency of this group to increase its incomes above the median for the base case is as expected. Households headed by single women have their entire distribution shifted to the left, doing worse at every quantile. The pattern is similar for single men, although the coefficients are insignificant for the upper two quantiles. This may suggest that jobs held by young people generally have fewer possibilities for large gains, greater possibilities for layoffs, and mediocre performance at the median. Households headed by single mothers, on the other hand, appear to have the same distribution of shocks as does the base group. Though perhaps surprising, this finding is consistent with anthropological studies of Mexican families. Selby and others (1990) find that matrifocal families have higher per capita incomes, proportionally more family members in the workforce, and lower dependency ratios and that these families generally do as well as nonmatrifocal house- holds.'2 In fact, Chant (1985, p. 650) finds that "despite major structural constraints of the economic and social potential of matrifocal families, single parent units frequently fare better than male headed households." As a final observation, larger households tend to have higher income gains than smaller families. This may reflect greater effort by the household head or labor force entry by additional family members who contribute to the family pot. These findings require revisiting the conclusions of section III. The values and significance levels of interactive terms corresponding to the crisis period for the combined precrisis and crisis periods are shown in the last three columns of table 2. These coefficients can be interpreted as showing how group X did relative to the base group in normal periods compared with how group X did relative to the base group during the crisis. The simplest example is that the 12. "Matrifocal households are not worse off because of the absence of a man, despite discrimination against women in the work force and the difficulties that women have getting well paid employment. Although median household income is 140/ lower than non-matrifocal households, since they average one less member in the household, their per capita incomes are 8.2% higher. They put almost as many members into the paid work force (1.38 vs. 1.4) and the ratio of dependent to members in the work force is lower than the non-matrifocal.... The genealogical complexity of matrifocal households is quite outstanding, with more grandchildren, more siblings, and more outsider as members. Though they are smaller, they are not abandoned and alone and preyed upon by society. Most of them are viable, operating units, doing as well as non-matrifocal households in the very tough world of the 'popular classes' of urban Mexico.... We had thought that thev would be highly vulnerable and living on the margin of existence. We had been persuaded of this notion by interviews with people from nuclear or extended families, and case studies, carried out mostly in Oaxaca, of families in tragic circumstances, recently abandoned by a drunken and abusive male, cast upon their own resources, or eking out a pitiable existence taking in washing and making tortillas to sell to other households. To our great relief, it seems that such circumstances are relatively transitory, particularly if people are long-enough established to have kin-folk upon whom they can rely" (Selbv and others 1990, p. 95). Maloney, Cunningbain, and Bosch 171 coefficient on the incomplete primary education group in the crisis period of table 1 is the sum of the coefficient in the normal period in table 2 and the dummy interactive term in the second panel.'3 Not surprisingly, the results for the crisis period in table 2 reveal a shifting downward of the distribution of the base group, with median and 80th quantile following roughly equivalently, whereas the lower tail captured by the 20th quantile elongates significantly, perhaps capturing a rise in job loss. The interactive terms on the non-base group dummy variables can be seen as adjusting their distribution relative to the adjustment in the distribution of the base group. If there are no significant interactive terms on group X, that means that the distribution of the group changed in the same way that the base group's distribution changed. A revealing example is that none of the interactive terms is significant at any quantile for the self-employed. This means that the findings in section III of higher highs, lower medians, and much lower lows for this group were not a function of the crisis because the same results are found in normal times. It cannot be rejected that the distribution of this group relative to the base group remained statistically equivalent. In contrast to the previous findings, they cannot be said to have suffered unusually during the crisis. The same can also be said for informal salaried workers. Similarly, the distributions for the old and young, those with above-average-size families, single mothers, and single men track the base group, suggesting that these groups also do not appear to have suffered any more or less in either median or variance than did the base group. The performance of single mothers is again some- what surprising, yet consistent with Glewwe and Hall's (1998) findings for Peru. The same cannot be said of families with heads who earned no income. Their distribution shifted sharply left at every quantile relative to the movement of the base group. Although both median and 80th quantile were previously higher than for the base group, perhaps reflecting that the head of the household often got a job across the sample period, this effect is reversed at the median and sharply attenuated at the 80th quantile during the crisis, perhaps reflecting the increased difficulty of getting a job. Most striking is that the less educated the household head, the less likely the distribution is to follow the college educated base group to the left during the crisis. For heads with a primary or an incomplete primary education, the value at the 20th quantile falls only 17 percent, compared with 40 percent for the base group, and at the 80th quantile income falls only 23 percent compared with 30 percent for the base group. The less well-educated appear to do better during the crisis than in normal periods relative to the base group, experiencing lower lows, higher highs, and, for those who completed primary school, a higher median. The premium at 13. Because of the corrections discussed earlier, the summation is not always exact and often diverges significantly at the extreme quantiles. 1 72 THE WORLD BANK ECONOMIC REVIEW, VOl, 1 8, NO. 2 the median found for this group in table 1 is, in fact, a feature of the crisis. This may be due to a greater propensity for poorer families to put extra workers in the labor force during crises. The same effect appears, although diminished in magni- tude and statistical significance, for workers with a secondary education. V. CONCLUSIONS Quantile analysis was used to identify which households suffered the largest income falls during the 1995 crisis in Mexico. The analysis provides estimates of "average" shocks, which are more robust to outliers and the nonnormality of the distribution and more fully describes the heteroscedasticity and asymmetries in the distribution of shocks, which appear to be important in understanding who experiences the most extreme falls in income. Because the underlying distribution of shocks is central to the concept of vulnerability more generally, these tools are potentially very useful. The results suggest several stylized facts. First, even during normal times different groups have very different distributions of shocks. As an example, households headed by informal self-employed workers have great variance in incomes, consistent with a sector made up of small businesses, which experience great volatility. Yet most workers voluntarily enter the self-employed small business sector, suggesting that taking on this high variance is consistent with welfare improvement. Thus, in attempting to measure job quality, the second moment of incomes needs to be included as only one of numerous job char- acteristics considered in workers' decisions on sectoral choice. Second, these normal-period distributions of income need to be taken into account when evaluating the impact of a crisis. Once they are controlled for, households that enter the crisis with a head who is unemployed do substantially worse in the distribution of outcomes than during normal periods and compared with the change in the distribution of the base group. However, virtually every other group fared as well or better than they did in noncrisis times relative to how the base group did. The least educated appear to have some slight gains at the median and very substantial gains at the lower tails of the distribution- their negative shocks are smaller than those of the base group. Neither informal workers nor households headed by single mothers experienced especially severe shocks at any point of the distribution, even though for self-employed workers that appears to be the case when only the crisis period is examined. These differences in the relative shapes of the noncrisis distribution from the crisis distribution suggest that it is problematic to use normal-period distribu- tions to forecast who would be most vulnerable to large income shocks during crises. Furthermore, different precipitating causes and contexts may imply very different patterns of shocks during crises and thus even greater differences between the two. For instance, during the Mexican crisis labor markets adjusted primarily through wages, not quantities. Most wages were held fixed in nominal terms while inflation eroded their real value, and unemployment did not become Maloney, Cunningham, and Bosch 173 especially high by regional standards. This is consistent with the finding here that the costs of the crisis were spread relatively evenly across household types- thus the broadly similar movements in the distributions across groups from precrisis to crisis periods. This may not be the case during crises in which labor markets adjust through quantities and unemployment is not distributed evenly across household types. Finally, although it is worthwhile knowing which types of household experi- ence the largest shocks to income during a crisis, it is not easy to move from there to statements about welfare. If the poor have fewer savings or if they cannot borrow, a given income loss may lead to larger consumption falls than for the rich. The Mexican income shocks from the crisis were permanent in the sense that incomes did not begin to recover for at least three years, so such income-smoothing strategies would be ineffective in the long run: Consumption would follow income reasonably closely. Even if income shocks led to identical consumption losses, poor families might be less able to tolerate these than the better off, particularly if they are forced below the poverty line. Ideally, there would be some mapping income shocks to welfare, either in a continuous context or as is implicit in poverty line analysis. REFERENCES Arias, Omar. 2001. "Are Men Benefiting from the New Economy? Male Economic Marginalization in Argentina, Brazil and Costa Rica." Policy Research Working Paper 2740. World Bank, Washington, D.C. Buchinsky, M. 1994. "Changes in the US Wage Structure 1963-87: An Application of Quantile Regres- sion." Econocretrica 62(3):405-58. -_____. 1995. "Estimating the Asymptotic Covariance Matrix for Quantile Regression Models: A Monte Carlo Study." Journial of Ecw zonoetrics 68(2):303-38. Calvo, Guillermo A., and Enrique G. Mendoza. 1996. "Mexico's Balance of Payments Crisis: A Chroni- cle of a Death Foretold." Journal of International Economics 41(3-4):234-64. Chant, Sylvia. 1985. "Single Parent Families: Choice or Constraint?" Development and Change t6(4):636-56. Chaudhuri, Shubham. 2002. "Empirical Methods for Assessing Household Vulnerability to Poverty." Working Paper, Columbia University, Department of Economics, New York. Chaudhuri, Shubham, Jyotsna Jalan, and Asep Sryahadi. 2002. "Assessing Household Vulnerability to Poverty from Cross-Sectional Data: A Methodology and Estimates from Indonesia." Working Paper, Columbia Universitv, Department of Economics, New York. Conley, John P., and William F. Maloney. 1995. "Optimal Sequencing of Credible Reforms with Uncertain Outcomes." Journal of Development Econiomics 48(1):151-67. Cunningham, Wendy, and William F. Maloney. 2000. "Measuring Vulnerability: Who Suffered in the 1995 Mexican Crisis." Working Paper, International Bank for Reconstruction and Development, Washington, D.C. Davidson, Russell, and lames G. MacKinnon. 2000. "Bootstrap Tests: How Many Bootstraps?" Econo- metric Review's 19(1):55-68. Edwards, Sebastian. 1998. "Two Crises: Inflationary Inertia and Credibility." Economic Journal 1()S(448):680-702. Fajnzylber, Pablo, William F. Maloney, and Gabriel Montes. 2003. "iAc Micro Firm Dynamics: How Similar Are They to Those in the Industrialized World? Evidence from Mexico." Working Paper, 1 74 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, No. 2o World Bank, Latin America and Caribbean Region. Online document available at wwv ..rIJI ,.- / laceconomist. Gainanou, Gisele, and Jonathan Morduch. 2002. "Measuring Vulnerability to Poverty." Discussion Paper 2002/58. United Nations University, World Institute for Development Economics Research, Helsinki. Gil-Diaz, Francisco, and Agusting Carstens. 1996. "One Year of Solitude: Some Pilgrim Tales about Mexico's 1994-95 Crisis." Anierican Economic Reiiew 86(2):164-69. Glewwe, Paul, and Gilette Hall. 1998. "Are Some Groups More Vulnerable to Macroeconomic Shocks than Others? Hypothesis Tests Based on Panel Data from Peru." Journal of Deielopment Econontics 56(1): 181-206. Gould, W. W. 1992. "Quantile Regression with Bootstrapped Standard Errors." Stata Technical Bulletin 9:19-21. (Reprinted in Stata Technic-al Bulletin Reprints 2:167-76.) - 1997. "Interquantile and Simultaneous-Quantile Regression." Stata Technical Bulletin 38:14-22. (Reprinted in Stata Technical Bulletin Reprints 7:167-76.) Halvorsen, R., and Palmsquist, R. 1980. "The Interpretation of Dummy Variables in Semilogarithmic Equations." American Economnic Revieu 70(3):474-75. Jalan, J., and Ravallion, M. 2000. "Is Transient Poverty Different? Evidence for Rural China." Journal of Deielopment Studies 36(6):82-99. Kamin, Steven B., and John FI. Rogers. 1996. "Monetary Policy in the End-Game to Exchange Rate Based Stabilizations: The Case of Mexico." Journal of International Econo,niics 41(3-4):285-307. Kennedy, Peter E. 1981. "Estimation with Correctly Interpreted Dummy Variables in Semilogarithmic Equations." American Economic Re'ieu' 71(4):80)1. Koenker, R., and G. Bassett Jr. 1978. "Regression Quantiles." Econometrica 46(1 ):33-50. Koenker, Roger, and Kevin Hallock. 2001. "Quantile Regression." journal of Economic Pers pectives 15(4):143-56. Levenson, Alec, and William F. Maloney. 1998. "The Informal Sector, Firm Dynamics and Institutional Participation." Working Paper, World Bank, Washington, D.C. Ligon, Ethan, and Laura Schechter. 2002. "Measuring Vulnerability." Discussion Paper 2002/86. United Nations University, World Institute for Development Economics Research, Helsinki. Maloney, William F. 1999. "Does Informality Imply Segmentation in Urhan Labor Markets? Evidence from Sectoral Transitions in Mexico." World Bank Economic Revieuw 13(2):275-302. Rogers, W. H. 1993. "Calculation of Quantile Regression Standard Errors." Stata Technical Bulletin 13:18-19. (Reprinted in Stata Technical Bulletin Reprints 2:133-7.) Sachs, Jeffrey D., Aaron Tornell, and Andres Velasco. 1996. "The Mexican Peso Crisis: Sudden Death or Death Foretold." Joumral of International Economics 41(3-4):265-83. Selhy, H., A. D. Murphy, and S. A. Lorenzen. 1990. The Mexican Urblan Household, Organizing for Sclf- Defeose. Austin: University of Texas Press. Yatchew A., and Z. Griliches. 1984. "Specification Error in Probit Models." Reviei of Economics and Statistics 67(1): 134-9. Agricultural Tariffs or Subsidies: Which Are More Important for Developing Economies? Bernard Hoeknman, Francis Ng, and Marcelo Olarreaga This article assesses the impact of the world price-depressing effect of agricultural sub- sidies and border protection in OECD countries on developing economies' exports, imports, and welfare. Developing economy exporters are likely to benefit from reductions in such subsidies and trade barriers, whereas net importers may lose as world prices rise. A simple partial equilibrium model of global trade in commodities that benefit from domestic support or export subsidies is developed to estimate the relevant elasticities. Simulation results suggest that a 50 percent reduction in border protection will have a much larger positive impact on developing economies' exports and welfare than a 50 percent reduction in agricultural subsidies. Although there is significant heterogeneity across developing economies, the results suggest that efforts in the Doha Round of WTO negotiations should be directed at substantially reducing border protection. High tariffs and domestic support and export subsidies granted to farmers in high-income economies limit developing economy agricultural production and exports. Such policies boost production in high-income countries, depress world prices, exacerbate the volatility of world prices, and reduce the scope for import competition. High tariffs and domestic support policies may, however, benefit net importers of agricultural products by providing access to the subsidized commodities at lower prices.1 Thus national interests regarding global reform Bernard Hoekman is senior advisor in the Development Research Groop at the World Bank and a research fellow at the Centre for Economic Policy Research, London; his e-mail address is bhoekman@ worldbank.org. Francis Ng is an economist in the Development Research Group at the World Bank; his e-mail address is fng@worldbank.org. Marcelo Olarreaga is a senior economist in the Development Research Group at the World Bank and a research affiliate at the Centre for Economic Policy Research, London; his e-mail address is molarreaga@worldbank.org. rhe authors are grateful to Ataman Aksoy, Morvarid Bagherzadeh, Bijit Bora, Gopi Gopinath, Harrv de Gorter, Ashok Gulati, Tim Josling, Will Martin, Wojciech Stawowy, the editor, three anonymous referees, and participants at the conference "The Developing Countries, Agricultural Trade, and the World Trade Organization," Whistler, June 16-17, 2002, for helpful comments and suggestions. They also thank Lili Tabada for assistance in constructing the database. 1. This potential national welfare benefit is offset hy the higher price volatility created by support policies, as country-specific shocks mav he transferr-ed to world markets. Also, as noted by Falvey and Tyers (1989), even net importers of agricultural products can gain from world price increases in the presence of distortions that led them to become net importers. T'his article ignores both the impact of policies on world price volatility and the potential for trade-pattern reversals associated with distortions. IHE \X'ORI D BANK ECONO(IIC REVIEW, VOL. 18, NO 2, © The International Bank for Reconstruction and Development / THE WORI D BANK 2004; all rights reserved. doi: 1.1093/wber/lhh037 18:175-204 175 176 THIE WORLD BANK K.( ()N5NOi(1 RElVIEW, vIoE. r 8, NO. 2 of agricultural trade and support policies will differ. However, most analyses conclude that the overall gain to developing economies from reforming agricul- tural policies greatly outweighs the potential cost to countries that are signifi- cant net importers of subsidized agricultural products. Negotiations were launched in 2000 in the World Trade Organization (\X'TO) to reduce trade-distorting interventions in agricultural markets, both subsidies (domestic support and export subsidies) and border protection (tariffs and tariff rate quotas). Developing economies need to determine which instruments of agricultural protection are most detrimental to their interests and therefore where to direct their negotiating efforts on agriculture in the Doha Round.2 This article attempts to shed some light on the issue by assessing the relative impact of tariffs, domestic support policies, and export subsidies on exports, imports, and welfare in 144 countries, 120 of them developing. The article assesses the impact of a 50 percent global reduction in agricultural tariffs and compares this to a 50 percent cut in domestic support and export subsidies.3 In welfare terms, tariffs matter significantly more than subsidies in that tariff reductions generate greater welfare gains. In large part this is because of high tariff peaks in both Organisation for Economic Co-operation and Development (OECD) and developing economies in products subject to domestic support or export subsidies. More generally, domestic support tends to have a much smaller impact on world prices than tariffs.4 The econometric analvsis generates very small estimates of the elasticity of net imports with respect to domestic support. Although the estimated elasticity with respect to export subsidies is considerably larger, export subsidies are much smaller than domestic support, so that their overall impact on the trade and welfare of developing economies is relatively small. These findings support the analysis of Snape (1987), who emphasizes the importance of reductions in border barriers in cases where governments also use subsidies to support production. As border barriers are reduced, the cost to national treasuries of maintaining a constant level of agricultural support increases substantially. Thus, reducing border protection will discipline the ability of governments to employ subsidy programs. A partial equilibrium framework is used to estimate the impact of policy changes for a sample of 144 countries on world prices of agricultural commodities that benefit from domestic support or export subsidies in at least one WTO member. 2. Thle policy siniulationi reflects a conservative interpretation of the Doha Nlinisterial Declaration: "We commit ourselves to comprehensive negotiations aimiied at: substanitial improvements in market access, reduction of, with a view to phasing out, all forms of export subsidies; and substantial reductions in trade-distorting domestic support' (si[o Doha Ministerial Declaration. para 13, November 2001. 3. Tariffs include the global ad valoreni equivalent of specific tariffs and tariff rate quotas. Linear cuts in tariffs are assuLmned rather than a nonlinear tariff formula to facilitate comparison with cuts in domiiestic support and export subsidies. Francois and Nlartin (2003) discuiss differenit nonlinear tariff reductioii formulas that could be used in the Doha Rounld negotiatiois. 4. See, for example, Snape (1 987), who argues that for a giseen amount of domiiestic productioni, tariffs lead to lower domestic conisumption thani do dormiestic subsidies and therefore to lower world prices. Hoeknizalt, Ng, and Olarreaga 177 The analysis focuses on 267 commodities, defined at the six-cligit level of the Harmonized System (Hs),S that benefit from domestic support or export subsidies. The analysis is restricted to this set of commodities so as not to bias the findings. Because most countries apply tariffs to all agricultural products, not just those that are subsidized, any comparison of the effect of reducing tariffs on all agricultural goods with a reduction in support policies is likely to conclude that tariffs are more important for developing economies. The partial equilibrium approach allows assessment of the effects of policy changes on individual countries, including the low-income and least developed areas that are of particular concern to the development comrnunitv. Most of these countries are generally subsumed in regional aggregates in applied general equilibrium models. The partial equilibrium approach also allows the use of disaggregated six-digit uis trade and protection data, and it allows estimation of trade elasticities directly from the data instead of using elasticities that have been estimated outside the data and the simulation model.6 I. A(RICUILT1URAL TARIFFS, DoMESTIC StUP'PORT. AND EXIPORT SUBSIDIES There are 158 commodities at the tis six-digit level that benefit fromn direct domestic support in at least one w-io member. Total support reported to the WTO was some US$227 billion a year on average during 1995-98 (appendix table A-I). Around $108 billion of domestic support is in the so-called Green Box-support exempted from future reduction commitmenlts because it is considered not to distort trade. Domestic support is used primarily by OEC.D countries, accounting for more than 88 percent of total domestic support payments notified to the wTo. The Quad group (Canada, European Union, japan, and the United States) accounts for 84 percent. Their share of world support is greater than 99 percent in several products (milling products; animal and vegetable fats and oils; prepared fruit, vegetable, and nut products; beverages and spirits; silk; and certain vegetable textile fibers and varn). Developing economies also have become greater users of subsidies-thev account for 12 percent of total domestic support reported to the WTO during 1995-98 (see appendix table A-I). Major subsidizers include Brazil, Thailand, and Venezuela. Not surprisingly, the least developed countries report no direct donmestic support. Developing economies support many of the same commod- ities as high-income countries, but they also support activities that oDcD) mem- bers do not (live trees and flowers; coffee, tea, and spices; gums, resins, and other vegetable saps; cocoa; and miscellaneous edible preparations. 5. The lHarmonized Commodity Description and Coding Sxstem (usually referred to as the llarmo- nized System or Hs) is used for tariff classification, it is miade up of 1,241 headings groLiped into 96 chapters. It has 5,00() subheadings identified by a six-digit code. 6. For recenit computable general equilibrium stodies focusing on the same question, see Beghin and others (20(02), Dimaranan and others (2002), Rae and Strutt (2002), and Tokarick (2003). These studies obtain qualitatively similar results (border barriers matter more than domestic support). 178 THE WORLD BANK ECONOMI(C RF VIEWN', V')[ 1 8, No. 2 Meat, dairy products, cereals, and sugar account for the lion's share of domestic support, representing almost 82 percent of reported nonexempt domestic support. They represent 78 percent of the $88 billion in domestic support provided by the Quad, 97 percent of the total for other OECD countries, and 93 percent for developing economies. This product concentration across regions is reflected in the high correlation across products of the domestic support provided by the Quad and by other groups. The correlation coefficient is 0.87 between domestic support in the Quad and in developing economies and 0.67 in the case of the Quad and other OECD countries. Export subsidies are also concentrated in a limited number of commodities. There are 208 tariff lines at the HS six-digit level that are subject to export subsidies in at least one WTO member. Export subsidy commitments across WTO member totaled $18 billion in 1995-98-some 10 percent of total direct domestic support.7 The OECl) accounts for 83 percent of all export subsidy commitments made in the WTO. Developing economies account for the remainder-with least developed economies not reporting any export subsidies (see appendix table A-I). Export subsidies are concentrated in the same four products that benefit most from domestic support. Meat, dairy products, cereals, and sugar represent 80 percent of export subsidies granted by WTO members. The product correlation of export subsidies across the three country groups is lower than for domestic support, but still high and positive (and statistically different from zero). The correlation is 0.56 between export subsidies in the Quad and in other OECD countries and 0.51 between the Quad and developing economies. The average most favored nation tariff applied to agricultural products varies substantially, but in the majority of OECD countries it is more than double the average for manufactures. Products that receive domestic support or export subsidies tend to have higher average tariffs. In the Quad the average tariff on imports of products that are subsidized by at least one WTO member is 26 percent, compared with the average tariff for all agriculture of some 17 percent (appendix table A-2). High tariffs and tariff peaks for subsidized products are also observed in other OECD countries and in developing areas. The global pattern of protection of agriculture will have differential impacts on countries, depending on whether they are net producers or consumers of the affected commodities. A first cut at identifying the implications of global protec- tion on individual countries is to calculate the relative importance to them of exports and imports of the products that are subsidized by at least one WTO member. This reveals that the least developed economies are potentially much more affected than other countries: Goods that are subject to domestic support in at least one wVTo member constitute 18 percent of their exports on average, compared with 3-4 percent for other countries (appendix table A-3). Similarly, 17 percent of their exports are in categories that receive export subsidies by at least 7. Export subsidy commitments ar-e used rather than actuial suibsidy disbursements because data are more reliable. Hoekrnan. Ng, and Olarreaga 179 one WTO member, compared with 4 percent for developed areas and 5 percent for other developing economies. A similar pattern holds for imports-some 9-13 percent of imports of least developed economies involve products that are sub- sidized, compared with 3-4 percent for other countries. For many least developed economies, the potential incidence of subsidies is there- fore very high. For economies such as Benin, Burkina Faso, Burundi, Chad, Malawi, Mali, Rwanda, St. Kitts and Nevis, St. Lucia, Sudan, Tanzania, Uganda, and Zimbabwe, goods that are subsidized by one or more WTO members constitute 60-85 percent of total exports. But even among least developed economies there is important heterogeneity. For example, products that receive domestic support or export subsidies in other wro members constitute less than 1 percent of Mauritania's exports. I. ANALYTICAIL FRAMEWORK A simple partial equilibrium model is used to estimate the impact on exports, imports, and welfare of a reduction in tariffs, domestic support, or export subsidies. World markets are assumed to be perfectly competitive and integrated, with no further scope for arbitrage across countries. Products traded in world markets under the same HS six-digit classification are considered to be perfectly homogeneous. Each six-digit HS product category represents only a small share of the economy, so that changes in a particular category have only a negligible effect on other product markets.8 Import demand for each HS six-digit product of country c is given by (1) M= a, ^! [P., (I + tM) I + TC)] P" SC pl, where ac is a demand parameter in country c that captures size and all other factors influencing import demand, p,, is the price in the world market, tc is the tariff in country C,9 r, is the average transport cost from country c to the world market,'0 £d is the import demand elasticity, se is producer support in country c,11 I?J8 is 8. The setup is similar to those in Zietz and Valdes (1986) and Hoekman and others (2002). FHoekman and others discuss some of the caveats associated with the use of this type of model. Note that no account is taken of such issues as the potential impact of exclhange rate overvaluation, indirect taxes, anid other factors that may result in an overall antiagriculture bias and thus offset the effect of tariff protection or subsidy policies. Schiff and Valdes (2002) suggest that in many developing economies, antiagriculture hias due to such policies has declined, implying that direct instruments, such as tariffs and subsidies, are the major determinants of the magnitude of protection. 9. Tariffs (or ad v-alorem equivalents of different border protection measures) are likely to vary by exporting country. For example, the ad valorem equivalent of a specific tariff is likely to vary by exporting country. Because information on bilateral protection levels is lacking, thev are assumed to apply on a most-favored-nation basis. 10. This explains differences in import prices across countries as observed in the data. Note that transport costs are likely to vary across exporting countries. 11. Countries with no domestic support are assigned a $1 value to ensure that the import demand function is not undetermined. 1 80 THE WORLD BANK ECONOMI(: REVIEW, VOL. i 8, NO. 2 the elasticity of import demand to producer support, and oc, is defined as a residual for notational simplicity, incorporating all terms other than price on the left side of the second equality. Export supply for each HS six-digit product of country c is given by (2) x [-b, I ISCc p where b, is a supply parameter that captures size and other determinants of export supply, es is the export supply elasticity, XS is the elasticity of export supply with respect to domestic support, 1 e, is the export subsidy in country c, Ax is the elasticity of export supply with respect to export subsidies, and fc is a residual used for notational purposes. Transport costs to world markets are the same for exporters and importers of the same HS six-digit good in the same country. The simultaneous presence of tariffs, domestic support measures, and export subsidies may lead to both imports and exports of a homogeneous product for a given country. Because of data constraints, import demand and export supply elasticities for products and import and export elasticities with respect to domestic support and export subsidies are assumed to be identical for all countries. This has implications for the underlying domestic supply and demand elasticities of domestic support across the countries in the data set. For example, if the elasticities are relatively similar across countries and if consumption is only marginally affected by changes in domestic support, then import demand elasti- cities should vary across countries depending on the ratio of domestic produc- tion to exports. The underlying domestic demand and supply elasticities cannot be estimated, however, because production and consumption data are not avail- able at the HS six-digit level. The empirical analysis tests whether elasticities for the two subsamples of developing and developed economies are statistically different (see note 18). The equilibrium world price is obtained by solving for the world price in the world market clearing condition ( 3 ) P', = argsol [Zm xc = j = _ The change in the world equilibrium price following a reduction in tariffs, domestic support, or export subsidies is obtained by taking the total differential of equation 3 with respect to tI, sC, or ec. The percentage changes in the world price after a common percentage change in each policy are given by 12. Again, countries with no domestic support (or export subsidies) are assigned a $1 value to ensure that the export supply function is not undetermined. Hoekman, Ng, and Olarreaga 181 =z' =-£d + j t C for changes in tariffs c e - Fd Z x- E's 131 (4) P1 =-d- Kg + Xs for changes in domestic support c c ~~e e kx c and for changes in export subsidies, c where' denotes the percentage change in the variable, t is the common percent- age change in tariffs, s is the common percentage change in domestic support, and e is the common percentage change in export subsidies. The change in export revenue and import revenue associated with a change in tariffs, domestic support, or export subsidies is given by Xcr = (1 + p + k.sS c-I + Axec (5)SCe Inr = _(Cd- _ 1)p t Ct _ ,dc Sc where xc' is the percentage change in export revenue in country c, and mcr is the percentage change in import revenue in country c (evaluated at world prices). Note that if there is no producer support, export subsidies, or tariffs in country c, there will be no changes in export revenue or import revenue in this country apart from that induced by the change in world price after other countries change their policies. The change in welfare can be calculated by taking the integral of the import demand and export supply functions with respect to world prices and tariffs. Domestic support and export subsidies are assumed to be simple transfers from government revenue to producers. 1 The change in welfare for exporters and importers relative to initial export and import revenue is then given by iLL = 1 _ ((1 E s SC+ (1 ec C- 13. This implicitly assumes that the government revenue necessary for agriculture subsidies was collected on a lump sum basis (taxation is nondistortionary). Note that these welfare measures also assume that there are no other distortions in the relevant part of these economies. 182 THE WORLD BANK ECONOMIC REVIEW, VOL. i8, NO. I (6) wc ( rd ( Ed_i ) d Ad (1+ Pu, + t ltc. + t,) ( SC- + t,,hr + tc1+ + t?l + ti where w,i is the change in welfare in an exporting country relative to the initial export revenue, and wcm is the change in welfare in an importing country relative to the initial import revenue. The first term on the right side of the w, in equation 6 is the change in import consumer surplus, and the last two terms provide the change in tariff revenue. Note that the changes in welfare in equation 6 take into account shifts of domestic import demand and export supply functions following changes in relevant policies. The overall change in welfare can be obtained by adding up the two expressions in equation 6 after normalizing the two terms to the same base (exports, imports, total trade, or per capita income). III. EMPIRICAL METHODOLOGY The empirical methodology consists of three steps. First, import demand and export supply elasticities are estimated with respect to prices and subsidies (Ed, s d, 5, and XX). Next, the demand and supply parameters (a, and bc) are calibrated for each country and product at the HS six-digit level. Finally, the elasticities and calibrated parameters are used to measure the changes in world prices, export revenue, import revenue, and welfare following a 50 percent reduction in agricultural tariffs, domestic support, or export subsidies in all countries. The elasticities cannot be obtained simply by estimating the import demand and export supply functions in equations 1 and 2 because they are determined simultaneously in any country. Moreover, world prices are not observed, but only export and import unit values in each country, including transport costs. If traded quantities are measured with error (which is likely because customs revenue authorities are generally mostly concerned with the value of shipments), unit values will also be measured with error, which may bias the results.14 To avoid these problems, units are chosen so that the average world price of each product during 1995-98 is equal to 1. The net import demand function is then estimated across countries and products as the log difference of import demand and export supply for each country (measured in value terms due to the choice 14. There is no obvious instrument for unit values at the HS six-digit level. Hoekmnan. Ng, and Olarreaga 183 of units). Note that world price terms will then drop from the specification, as log(p",) = log( 1) = 0. Equations 1 and 2 are used to obtain the following estimating equation: (7) log(m1t ) -log(x,) = log(a,) - log(bc) clog(1 + tc) - (Ed -- Es)log(1 + tc) _ (pd + ks)log(s) - klog(ec). GDP and population in each country are used as controls for ac and bc. Product and country dummy variables are also included to control for other demand and supply factors that are common within countries (such as endowments) or across HS six-digit categories (such as international market structure).'i There are two problems with the estimation of equation 7: Transport costs are not directly observable, and only the sum of the elasticity of import demand and export supply with respect to domestic support can be retrieved. Assuming that transport costs to the world market are equal for exporters and importers, transport costs can be proxied by the ratio of export and import unit values. As long as the measurement error in unit prices is identical for exports and imports, the problems are addressed. On the second issue elasticities of import demand and export supply with respect to domestic support are assumed to be equal. l Because of the unbalanced nature of the data set, a between estimator is used, based on four-year averages over the period for which domestic support and export subsidy data are available, rather than annual data (the sample is thus a cross-section of countries and products). 17 IV. RESULTS Table 1 reports the results of the estimation of equation 7, introducing different types of subsidies in a stepwise fashion. It reports results using notifications by WTO members of nonexempt types of domestic support only, denoted s,-9 It also reports results using notifications on exempt (Green Box) types of domestic sup- port, which are non-product specific. For purposes of estimation these categories are allocated across products using the distribution of domestic support commit- ments by product. This type of domestic support is denoted as s,'1-3 . The results of estimating equation 7 when the two types of domestic support are combined are also reported. Testing for the adding up of the two types of domestic support was done by running a nonlinear regression (but without country and product dummy variables because of programming constraints) that included the following 15. As world prices are normalized to I by choice of units, the export supplv elasticitv can be retrieved bv using the information on transport cost and the estimate of the import demand elasticity in fronit of (I + t). 16. The simulations test for the robustness of results by varying the elasticities with respect to domestic support on the demand and supply side. 17. This is also due to the fact that ad valorem equivalents of specific tariffs havst oilv been estimated for 1999 in the OE(.I) (2000). 1 84 I HF WORLD BANK L(.ONOi(M RFVI \\., V01 . 1 8, No. 2. TABLE 1. Ordinary Least Squares Estimates of Price, Domestic Support, and Export Subsidy Elasticities Notification of Notificatioo of nonexempt exempt Notification domestic domestic of export support' support ' I + 2 subsidies 3 + 4 Variable (1) (2) (3) (4) (5) log(GDP) 0.34 0.33 0.33 0.33 0.33 (0.09)- (0.09)'" (0.09)*' (0.09)"* (0.09). log(Pop) -0.28 --0.28 -0.28 --0._8 -0.28 (0 O I1 ) ': ' 1 ) ( 0 I I ( 0. I 1 ) (0. I 1 ) log(l + t) -(£ ) -1.72 -1.76 -1. 74 -1.67 -1.64 (0. 17)- 0. 1 8) (0. 18)' ( 0. 18) (0 . 1 8) log( l + r) (E 1:) -0.88 --0.88 -0.88 -0.87 -0.87 ( 0 05S)(() () S) ( 0.0 5 )` r A 05 ) ( 0.0 5 )`- log(s')4 - (i; + i;S) -(.(9 (0.02)"- Iog(s° D ) (13 + 'iS) --0.07 (0.02)*" _* -i. + ( DS4-) --0.07 -0.04 (0.02)"" (0.02)" log(e) - -0.24 -0.23 (0 ()02) '- 0.03);I Product dummy Yes Yes Yes Yes Yes variables Country dummy Yees Yes Yes Yes Yes variables Adjusted R2 0.23 0.23 0.23 0.23 0.23 Number of 14,661 14,661 14,661 14,526 14,526 observations Number of HS 267 267 267 267 267 six-digit lines "Significant at the 5 percent level. ."Significant at the I percent level. Note: These results arc from estimationi of equation 7 in the text. The left side variable is the difference of the logs of import revenue and export revenue, £' is the price elasticity of import demand, Es is the price elasticity of export supply, ',1 + y' is the sumIl of elasticities of import demand and export supply with respect to domestic support, ),' is the elasticity of export supply with respect to export subsidies, GDP is the value of GDP, Pop is population, t is tariff expressed in percentage pohints, r is transport cost, s 'DS is nonexempt domestic support, s5)s"1 is WTO-exempted doniestic support (Green Box), and e is the export subsidy. Standard errors in parenthesis are White robust. 'Corresponds to WTO categories DS4-9; see data appenidix. "Corresponds to WTO categories DS 1-3; see data appentdix. Souirce: Authors' calculations based on data from sources described in the appendix. variable: log (s,-/'5' + S 4 . The parameter ¢ is statistically not different from 1. Results are also reported using export subsidies only and both export subsidies and the sum of the two types of domestic support. Hoek,lIn7o, Ng, and Olarreaga 185 Results across the five specifications generally yield an elasticity of import demand in the 1.64-1.72 range and an elasticity of export supply in the 0.77-0.88 range. The elasticities of domestic support are in the 0.04-0.09 range. The elasticity of net import demand with respect to export subsidies is estimated at 0.23 in the preferred specification (column 5 of table l) These pooled regressions assume common elasticities across different products, which is not necessarily the case. Table 2 reports the results of the estimation of equation 7, but with elasticities allowed to vary across five product groups: animal products (HS 0 1-04); vegetables, fruits, and nuts (ius 06-09); cereals and grains (Hs 10-14); processed food products (HS 1 5-24); and cotton and other textile fibers (HS 50-53). The product group-specific elasticities are used as the base estimates for the simulation exercises that follow. The overall estimates in column 5 of table I are used to test for the robustness of the results.'9 Note that the elasticities of net import demand with respect to domestic support are relatively small (around 0.05), suggesting that a reduction in domestic support across A-ro members will have only a small impact on world prices.20 The elasticities with respect to export subsidies tend to be much larger. Simuilation Resuilts The baseline simulations use the estimated coefficients in table 2 to calibrate import demand and export supply in each country. Changes in export and import revenue and welfare following a 50 percent cut in tariffs, domestic support, and export subsidies are then calculated for each country using equa- tions 5 and 6. Recall that the simulations are done for the 267 tariff lines at the HS six-digit level for which at least one countrv provides domestic support or export subsidies. (The overall agricultural universe includes more than 900 tariff lines at the HS six-digit level.) The increase in trade and welfare across Ol('D countries, developing economies (excluding least developed economies), and least developed economies is much larger for the tariff cut than for the reduction in domestic support or export 18. Testing for the homogeneity of the parameters across groups of countries in the estimation of the regression reported in coluinii 5 was done by splitting the sample into developed and developing econiomiiies. Results suggest that domestic suipport anld export subsidc elasticities are inot statistically differenit across these two subsamples, whereas price elasticities tend to be different. 19. For cereals and other grains, the coefficient captuiriig the import demand elasticitv is inisignificant and smaller than the coefficient on the difference of import demand and export supply price elasticities (which is significant). Thus the assumptioni that the export suppix elasticity is zero canniot he rejected. The export suIpply elasticity is therefore set to zero in the simulations for these products, and the import demanid elasticities are calibrated accordingly. 20. The implicit assuimption here in moving along the export supply and impor-t demand functionis that domestic support affects only the variable cost of farmers receiving the subsidy. This approach does not permit measuring the ! .. 1,i impact that domestic support mas have on ficed costs or on decisions to produce. Note that it is sometimes argued that domestic sLIpport that is decoupled from production maya still have an impact on production levels because it may affect farmers' decisions to enter a market. 1 86 I'HF WORLD BANK ECONO\I1(. REVILW, VOl I. 8, NO. ' TABLE 2. Seemingly Unrelated Regression Estimates of Price and Domestic Support Elasticities by Group of Products HS06-09 HSIO-14 HS1S-24 HS50-53 HSOI-04 Vegetables, Cereals and Food Silk, cotton, Animal fruits, and other processed and other products nuts grains products textile fibers Variable (1) (2) (3) (4) (5) log(GDP) 0.35 0.95 0.18 -0.42 1.76 (0).21) (0.09)*- (().21) (0.23) (0.20)- log(Pop) -0.48 - 0.98 -0.11 0.38 -1.11 (0.29) (0.09): ' (0.27) (0.34) (f.l9)' ' log(l +t)-(Ri) -1.14 -1.19 -0.39 -1.44 -1.29 (0.42)** (0.39) * (0.40) (0.30)' (2.00) log( + i) - (' --) El) -0.36 -0.76 -0.74 -1.36 -0.50 (0. I1) * (0.08) (0. 10)** (0 .10) (0.28) log(s DS 3+5S4-9 -0(.00 -0.13 -0.03 -0.05 -0.03 ( p'+X ) (0.03) (0.03):r (0.03) (0.03) (0.09) log(e) -X -0.16 -0.30 --0.25 -0.16 -0.14 (0.04) - (0.07) (0.06)* (0.04)** (0.13) Product dummy Yes Yes Yes Yes Yes variables Country dummy Yes Yes Yes Yes Yes variables Adjusted R24 0.37 0.33 0.28 0.23 0.30 Number of 1,764 5,110 2,377 4,694 485 observations Number of HS 37 87 49 77 14 six-digit lines 'Significant at the I percent level. Note: These results are from estimation of equation 7 in the text. The left side variable is difference of the logs of import revenue and export revenue, Fd iS the price elasticity of import demand, F' is the price elasticity of export supply, Xd + 2i is the sum of elasticities of import demand and export supplv with respect to domestic support, k' is the elasticity of export supply with respect to export subsidies, GDP is the value of GDP, Pop is population, t is tariff expressed in percentage points, T is transport cost, s OS4 4 is nonexempt domestic support, 5DSI3 is WTO- exempted domiiestic support, and e is the export subsidy. Group-specific elasticities were estimated using the information in the whole sample, letting the elasticities vary by group of products. Standard errors in parenthesis are White robust. Souirce: Authors' calculationis based on data from sources described in the appendix. subsidies (table 3).2 1 Reductions in domestic support and export subsidies have little impact on the exports, imports, or welfare of developing economies. A 50 percent tariff cut by WTO members, however, boosts developing economy welfare by $7.8 billion, exports by 9 percent, and imports by 6.8 percent. The effects for least developed economies are much lower, with a negative welfare effect and a 3.8 percent rise in exports and 3.4 percent rise in imports. 21. The increase in export revenue is not necessaril) equal to the increase in import revenue because these are measured at different prices (due to the presence of transport costs). Hoekman, Ng, and Olarreaga 187 TABLE 3. Impact of a 50 Percent Cut in Tariffs, Domestic Support, and Export Subsidies across All WTO Members (267 Products) Change in exports Change in imports Change in welfare Value Value TIotal Per Country group ($ millions) %' ($ millions) % ($ millions) capita ($) Cut in tariffs OECD countries 5,938 5.4 8,846 8.1 13,419 17.04 Developing countries 7,782 9.0 6,522 6.8 7,695 1.90 Least developed countries 130 3.8 125 3.4 -11 -0.03 Cit in doimnestic support OECD countries -20 -0.0 232 0.2 53 0.07 Developing countries 96 0.1 75 0.1 167 0.04 Least developed countries 24 0.7 -2 -0.0 12 0.03 Ctut in export subsidies OECD countries -689 -0.6 -392 -0.4 206 0.26 Developing countries 41 0.0 -520 -0.5 -238 -0.06 Least developed countries 20 0.6 -18 -0.5 -14 -0.03 Soufrce: Authors' calculations based on data from sources described in the appendix. These average changes mask substantial heterogeneity across countries. For example, in Benin, Burundi, Guinea-Bissau, Paraguay, and Uganda the increase in export revenue associated with a 50 percent reduction in domestic support and export subsidies is two to three times larger than the increase associated with a 50 percent reduction in tariffs. The increase in welfare after a 50 percent cut in domestic support is also larger in these countries (figures 1 and 2). The highest percentage increases in exports following a 50 percent tariff reduction are found in Caribbean and Central American countries, reflecting their specialization in such commodities as edible fruits and vegetables, processed foods, and sugar-categories that see the largest expansion in demand in percentage terms (see appendix table A-1). With a few exceptions (such as the Republic of Congo, Malawi, and Mauritius), African countries tend to register only limited increases in exports and welfare. Most developing economies will benefit (some significantly) from a 50 percent tariff reduction for the 267 products in the sample (see figure 1). The largest gainers include Caribbean and Central American countries and some Cairns group members, such as Argentina, Chile, Colombia, and Malaysia. Among the winners, welfare gains are generally much larger under a 50 percent tariff cut (most countries are above the 450 diagonal). Some developing economies may lose from a 50 percent tariff cut-including net food importing countries, such as Algeria, Kuwait, Oman, Paraguay, Qatar, Saudi Arabia, and Singapore. These countries also tend to lose from a 50 percent reduction in subsidies, but to a much smaller extent because subsidy cuts lead to much smaller increases in world prices. 1 88 THE WORLD BANK ECONOMIC REVIEW, VOL. 1 8, NO. 2 FIGURE 1. Changes in Welfare in Developing Economies with Cuts in Tariffs or Subsidies ($ per capita) 44.3666 - bIz Gain from subsidy cuts! Lose from subsidy cuts / Gain from tariff cuts mus cri Ci) Gain from tariff cuts ca Cu guy klma vct CF brb ecu B~ ~ ~ ~ ~~~~~~y mar su phn arg______--------~ ---- ------- pry C dit kw qat XC- mac C brn ) CYP Gain from subsidy cuts/ sgP Lose from tariff cuts Lose from subsidy cuts Lose from tariff cuts -33.2544 hkg I I I II -4.63391 7.36886 Change in welfare with cuts in subsidies Note: The dotted 450 diagonal line shows which of the two reforms provides the greater welfare gains. The country names corresponding to the country codes are in the first column of table A-3 in the appendix. Source: Authors' calculations based on data from sources described in the appendix. For least developed economies, the welfare impacts tend to be relatively small, and a large number of least developed economies would experience a (modest) decline in welfare after a global 50 percent reduction in tariffs. Malawi would be the largest beneficiary in gains per capita from both types of reform.22 With respect to specific products a global 50 percent cut in tariffs generates a relatively large increase in developing economy exports of edible vegetables, fruits, and nuts (HS 07-08), sugar (HS 17), prepared vegetables and fruits (Hs 20-21), and tobacco (Hs 24). For least developed economies, the largest increases occur in meat (HS 02), sugar (HS 17), and miscellaneous edible preparations (Hs 21).23 22. This is driven by the composition of Malawi's export bundle. The largest increase occurs for tobacco and sugar exports. All these numbers should be interpreted with caution because import demand and export supply elasticities are assumed to be homogenous across countries. Splitting the sample into developing and developed economies yields smaller export supply elasticities for developing economies, suggesting an overestimate for exporters in developing areas. 23. The analysis does not include the trade restrictiveness associated with sanitary or phytosanitary barriers, as well as other nontariff barriers that may hinder trade and limit export expansion. For recent analysis of these issues see Wilson and Abiola (2003). Hoekman, ITg, and Olarreaga 189 FIGURE 2. Changes in Welfare in Least Developed Economies with Cuts in Tariffs or Subsidies ($ per capita) 2.46834- mwi Lose from subsidy cuts / Gain from subsidy cuts / Gain from tariff cuts Gain from tariff cuts com hti O) gmb mrt Gain from subsidy cuts! Lose from tariff cuts 0) C _C Lose from subsidy cuts / C) Lose from tariff cuts dj -6.4912 -mdv -.89609 .412529 Change in welfare with cut in subsidies Note: The dotted 450 diagonal line shows which of the two reforms provides the greater welfare gains. The country names corresponding to the countrv codes are in the first column of table A-3 in the appendix. Souirce: Authors' calculations hased on data from sources described in the appendix. These results assume that policy changes are undertaken on a global basis, including by developing economies. The impacts of a 50 percent cut in tariffs, domestic support, and export subsidies by OECD members only are qualitatively similar, in that changes in exports, imports, and welfare are much larger for the reduction in tariffs, supporting the conclusion that border barriers are particu- larly important for developing countries (table 4). Global reform spanning all WTO members tends to bring larger gains for those that benefit than do reforms by OECD countries only (figures 3 and 4). This emphasizes the importance for developing economies of participating in agri- culture reform efforts. Again, there is some heterogeneity among developing and least developed economies, but benefits from reform by all wrO members are always greater than benefits from OECD reform only (countries on the right side quadrants in figures 3 and 4 are all above the 450 line, implying that welfare gains would be larger under reform by all WTO members). These simulation results ignore trade preferences. This omission is important, because it implies that benefits may be overstated for developing economies that enjoy meaningful preferential access to protected markets. It also means that the losses to some least developed economies from global agricultural policy reform 1 90 1 HE WORLD BANK EC(ONO\MIC REVIEWX, VOI. I 8, NO. Z TABLE 4. Impact of a 50 Percent Cut in Tariffs, Domestic Support, and Export Subsidies in the OECD (267 Products) Change in exports Change in imports Change in welfare Value Value Total Per Country group ($ millions) % ($ millions) % ($ millions) capita ($) Cut in tariffs OECD countries 2,388 2.1 9,478 8.6 12,784 16.24 Developing countries 4,379 5.0 -424 -0.4 705 0.17 Least developed countries 44 1.3 -15 -0.4 0 (.00 Cut in doniestic support OECD countries -130 -0.1 301 0.3 237 0.30 Developing countries 172 0.2 -55 -0.1 -88 -0.00 Least developed countries 7 0.2 -1 -0.0 2 0.01 Cut in export su/bsidies OECD countries -957 -0.9 --252 -0.2 491 0.62 Developing countries 435 0.5 -451 --0.5 -488 -0.12 Least developed countries 3 0.1 -15 -0.4 -22 --0.05 Source: Authors' calculations hased on data from sources described in the appendix. may be understated. Although the overall impact of ignoring preferences will be negligible because the countries that get deep preferences are small and have only a marginal impact on world trade, additional research and analysis are needed. The potential problem of preference erosion is limited to a small number of countries. Few developing economies benefit significantly from trade prefer- ences, especially in the product categories that are the subject of this analysis. As noted by Hoekinan and others (2002), preference margins tend to be lowest for products for which tariffs are highest-because these are mostly "sensitive" categories. Recent research has also documented that utilization rates of pre- ferences are often much below 100 percent. Even where there are benefits (rents), a significant share is captured by importers and retailers, not the intended beneficiary countries.24 Nonetheless, some countries, especially least developed economies that have recently been granted duty- and quota-free access to large OECD markets, stand to lose from preference erosion in certain markets. An obvious example is sugar. Although the predicted increases in the world price of sugar following global reform will benefit net exporters to some extent and offset the loss in prefer- ential access to some extent, the existence of preferences implies that the losses to the least developed economies that currently benefit from preferences in such highly distorted markets may be greater (and the gains smaller) than suggested by the analysis here. This is not likely to be the case for other developing 24. See, for example, Brenton (2003), Inama (2003), Mattoo and others (2002), Tangermann (2002), and Ozden and Reinhardt (2003). Hoeknzan. Ng, and Qlarreaga 191 FIGURE 3. Changes in Welfare in Developing Economies Resulting from Reforms by All w-ro Members or by OECD Countries Only ($ per capita) 50.8774 biz o Gain froni "'TO cuts Gain from W TO cuts / E Lose from OECD cuts Gain from OECD cuts or EIc o guy _ v~ ~ ~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ct*-- 9-- C ecu , E agtg gPa St "aP cyLose from X1TO cuts! ._ sgp .Gain from OECD cuts Lose from N\:TO cuts! tc Lose from OECD cuts ° -37.7485 hkg II I II -16.785 28.577 Change in welfare with reform in the OECD only Note: The dotted 450 diagonal line shows whiich of the two reforimis provides the greater xvelfare gains. The country names corresponding to the country codes are in the first column of table A-3 in the appendix. So(/rce: Authors' calculations based on data from sources described in the appendix. economies that are predicted to benefit from global reform, however, because these countries are not granted the type of deep preferences that have been offered to least developed countries. Sensitiv'itv Analysis Given the various assumptions made about elasticities, several sensitivity analyses were performed. First the results in table 3 were reestimated using the elasticity estimates provided for the whole sample in column 5 of table I (instead of the elasticity estimates by product reported in table 2). The results in table 3 were also reestimated using extreme values for the elasticities of domestic support on the import and export sides (either zero or the total value of the estimated coefficient instead of half the estimated coefficient in table 2). When the elasticities estimated for the whole sample are used, a 50 percent tariff cut leads to a 20 percent smaller increase in exports for developing economies and a 15 percent smaller increase for least developed economies. A 50 percent cut in domestic support leads to a 25 percent greater increase in exports for developing economies and a 20 percent smaller increase for least 1 92 TIl l FXWORI I BXNK E(.ON().\I. R.F0 IL\ , V1-. I 8, NO. 2 FIGURE 4. Changes in Welfare in Least Developed Economies Resulting from Reforms by All wrro Members or by 0ECD Countries Only ($ per capita) 2.88087 fn Gain from WVT() cuts Gain from WTO cuts E Lose from OECD cuts E Gain from OLCD cuts a) E 0) E o 02 gmb Lose from WTO cuts/ GIain fromo OCCD cuts (3) ~~~~~~~Lose from WNTO cuts .Lose from OECD cuts C. 0) Cm dj, -7.38729 md, -3.26791 1.77725 Change in welfare with reform in the OECD only Note: The dotted 450 diagonal line shows which of the two reforms provides the greater welfare gains. The count-y names corresponding to the country codes are in the first column of table A-3 in the appendix. Sozurce: Authors' calculations based on data from sources described in the appendix. developed economies. A 50 percent cut in export subsidies leads to a 22 percent greater increase in exports for developing economies and a 15 percent smaller increase for least developed economies. The qualitative results remain: The increase in exports and welfare for developing economies is much larger under the 50 percent tariff cut than under the 50 percent reduction in domestic support or export subsidies. Simi- larly, for least developed economies the increase in exports is two times larger under the 50 percent tariff cut than under domestic support or export subsidy cuts. Developing economies experience welfare gains when tariffs are cut but welfare losses when domestic support or export subsidies are cut. For least developed economies the changes in welfare are almost negligible. As noted earlier, the elasticity of domestic support on import demand and export supply cannot be empirically identified separately, only their sum. To test the sensitivity of the assumption that the two are equal, each in turn is assumed to be zero and the coefficient estimate is then assumed to identify the other. The estimated changes in export revenue, imports, and welfare are not affected by these modifications (with import demand and export supply recali- brated accordingly). For a 50 percent cut in domestic support the increase in Hoeknman, Ng, and Olarreaga 193 exports by developing economies is 20 percent greater, and the welfare loss is 22 percent smaller when the domestic support elasticity of export supply is assumed to be zero and the export increase is 60 percent smaller and the welfare loss is 25 percent greater. The qualitative results remain the same. For least developed economies the estimated change in exports is affected only marginally under both scenarios. VI. CONCIUSION Tariffs matter much more than subsidies for an impact on world prices. Although there is some heterogeneity across countries, the positive welfare effect of reducing tariffs on products that are also affected by agricultural support is a multiple of what can be obtained from an equivalent cut in domestic support or export subsidies. This reflects not only the high tariff peaks in OECD countries but also developing economies' own use of tariffs to protect domestic production. Developing economies generally have low levels of domestic support and export subsidies, reflecting both budget constraints and a more neutral policy stance toward agriculture. The analysis suggests that attention in the Doha Round should focus on reducing border protection in both OECD countries and developing economies. This does not mean that subsidies are unimportant. Decoupling domestic sup- port from production and eliminating export subsidies are also both very important. Tariffs are often the only instrument of intervention that developing economies have to respond to the effects of OECi) subsidy policies. An important dimension of agricultural support policies that has been ignored here-the impact on price volatility-plays a major role (Valdes and Foster 2002). Tariff protection can shelter farmers from import surges in periods when world prices drop significantly. Whatever the source of the exo- genous shock that drives prices down, the adjustment may fall disproportion- ately on residual (non-OECD) markets because support policies shelter OECD farmers from the shock. Unilateral liberalization of agricultural trade in coun- tries such as India has been argued to have been politically unsustainable as farmers were subjected to large world price swings and import surges of com- modities subsidized by OECD countries (Gulati and Narayanan 2002). Many developing economies oppose further agricultural trade liberalization as long as OECD countries continue to provide large-scale support for their farmers. Substantial reductions in OECD agricultural support policies are therefore import- ant not only because reductions generate direct benefits for the many developing economies that are net exporters but also because OECD reductions are critical for creating the political support to allow developing country governments to continue to pursue welfare-improving agricultural trade policy reforms. Reductions in production subsidies and elimination of export subsidies in OECtD countries are necessary, although nor sufficient, for developing economies to reap significant 194 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. 2 gains from the current wTo negotiations on agriculture. At the same time, Anderson (2002) notes, if OECD members were to move seriously on subsidies, developing economies would also need to reduce protection. Without their own liberalization measures, the negative welfare effects for countries that experience terms of trade losses from subsidy removal would likely be greater. Because the simulations suggest that a number of countries will lose from global agricultural policy reforms, compensation mechanisms should accom- pany liberalization and removal of domestic support. The same is true for countries for which preference erosion may be important. Compensation mechanisms could include additional aid for trade and support for adjustment. Such measures should include actions that improve the functioning of input, downstream, and factor markets to support efforts by farmers to expand output in response to the rise in prices. Also important are measures to reduce costs for farmers, such as actions to improve the efficiency of services-finance, insur- ance, transport, storage, packaging. The cost-increasing effect of inefficient services can be substantial-as illustrated in recent research on transport ser- vices (Fink and others 2002; Francois and Wooton 2001). If such accompanying measures are taken, the resulting supply response may transform some countries from net importers to net exporters, attenuating the negative effects estimated herein (Anderson 2002). Finally, it is important to bear in mind that the analysis has been limited to subsidized commodities. The Doha Round negotiations span all trade, includ- ing nonsubsidized agricultural products and manufactures. The welfare findings generated by this analysis are therefore not particularly relevant, except to indicate that the countries that lose from reforms that affect subsidized agricul- tural products will need to identify other areas in which they can generate offsetting gains. In principle, this should be straightforward given the large negotiation set established in Doha. APPENDIX: DATA SOURCES All trade data are from UN Comtrade database (both value figures and unit prices). Data for countries that did not report trade data to Comtrade are mirrored using notifications by their trading partners. Tariffs are drawn from the United Nations Conference on Trade and Devel- opment (UNCTAD) and w-ro as provided in the UNCTAD/World Bank World Integ- rated Trade Solution (WITS) system. This database does not include the ad valorem equivalent of specific tariffs. For OECD countries this information comes from OECD (2000). The OECD ad valorem equivalents of specific tariffs use exclusively out-of-quota tariffs. Quota information is not available. This may bias some estimates because some import prices may be higher if exporters benefit from in-quota lower tariffs. GDP (in U.S. dollars) and population data are drawn from the World Bank's World Development Indicators database. Hoeknzan, Ng, and Olarreaga 195 The source of domestic support data is the WTO (document GIAG/NGIS/1). National currency data were converted to U.S. dollars using the period average exchange rate reported in the International Monetary Fund International Finiani- cial Statistics. The export subsidy data come from the W\TO (document G/AG! NG/S/5/Rev.1). The product classification in each country's WTO notification is arbitrary, and therefore these classifications are mapped to the HS six-digit classification system. In most cases this can be done through a one-to-one mapping. When the domestic support and export subsidies reported cover several six-digit tariff lines, the subsidy was distributed across the relevant tariff lines using the share of the reporting country's exports as weights. The con- cordance file is available on request. Only 30 W\TO members have made domestic support reduction commitments under the Agreement on Agriculture, but all members are required to notify WTO of domestic support. Compliance is weak-in 1995 only 75 percent of WTO members that were required to notify did so. In 1996 and 1997 the coverage dropped to around 50 percent. For 1998 only 28 percent of WTO members had notified by March 2000. However, most countries that did not notify in 1997-98 had very little or no support in 1995-96, so coverage of the data spans the major users. The incomplete reporting problem was addressed by using the average aggregate measure of support (AMS) reported for whatever years are available. The empirical analysis therefore involves an unbalanced panel. Domestic support notified to the WTO includes exempt and nonexempt meas- ures. There are nine categories of support, designated DS1 through DS9. DS1 covers measures that WTO members have placed in the Green Box and are there- fore exempt from reductions (the Green Box categories are defined in annex 2 of the Agreement on Agriculture). DS2 covers measures for developing economies that are exempt from reduction commitments under Article 6.2 of the Agreement on Agriculture relating to development programs. DS3 covers direct payments under production-limiting programs under Article 6.5. Categories DS4-DS9 cover measures that are not necessarily exempt from reduction commitments. DS4 refers to nonexempt support that is below the de minimis level (as set out in Article 6.4). The remaining categories included in the total ANIS of WTO members include market price support (DS5), nonexempt direct payments (DS6), other product-specific support (DS7), and any support measured using the equivalent measurement of support methodology (DS8). Finally, where relevant, a total figure for non-product-specific support is also given (DS9). Because exempt subsidies are not product specific, these were also mapped into product-specific subsidies using as weights the product-specific commit- ments that each country made in the Uruguay Round. Non-product-specific support and export subsidies are divided evenly into all products exported by the country concerned. All products shown in notifications to the WX'TO are included, whether or not the support is below the de minimis level for the member concerned. Thus, total ANIS may exceed total WTO commitments for a country. TABLE A- 1. Domestic Support and Export Subsidies by Product in Selected Country Groups, 1995-98 Domestic supporta Export subsidiesh All WTO Quad Other Developing All WITO Quad Other Developing members (4) OECD country members (3) OECD country (90) share (5) (81) (24) share (5) (16) HS-2 product ($ million) (%) share (%) share (%) ($ million) (%) share (%) share (%) ON 01 Live animals 365 67 0 33 101 16 22 62 02 Meat and edible meat offal 19,505 88 10 2 2,972 87 2 11 04 Dairy products, eggs, honey 15,494 82 11 6 4,689 87 10 4 05 Products of animal origin NES 1 0 0 100 06 Live tree and other plant; flowers 15 0 0 100 27 0 0 100 07 Edible vegetables and roots 5,307 95 2 3 235 10 1 67 08 Edible fruit and nuts, melons 4,731 98 1 2 303 28 6 65 09 Coffee, tea, mate, and spices 101 0 0 100 15 0 0 100 10 Cereals 41,230 79 2 20 5,606 93 0 7 11 Milling products, malt, starches 191 99 0 1 599 0 0 100 12 Oil seed, oleagic fruit; miscellaneous grain 663 83 0 17 123 66 0 34 13 Lac, gums, resins, and other vegetable saps 0 0 0 100 0 0 0 100 14 Vegetable plaiting materials, vegetable products NES .. 0 0 0 100 15 AnimalUvegetable fats and oils and products 2,159 100 0 0 191 69 0 31 16 Meat, fish, and seafood preparations NES 125 0 5 95 17 Sugars and sugar confectionery 8,530 79 1 21 1,013 79 0 26 18 Cocoa and cocoa preparations 0 0 0 100 36 0 0 100 19 Cereal, flour, starch, and milk preparations 34 0 0 100 20 Prep of vegetable, fruit, nut products 1,069 100 0 0 230 6 0 94 21 Miscellaneous edible preparations 3 0 0 100 129 0 91 9 22 Beverages, spirits, and vinegar 2,388 100 0 0 284 77 0 23 23 Residues and waste from food industry 486 80 0 21 5 84 0 16 24 Tobacco and manufactured tobacco products 968 92 1 7 139 70 0 30 33 Essential oils and resinoids, perfumes .. 0 0 0 100 35 Albuminoids, modified starches 11 0 0 100 41 Raw hides and skins 6 0 0 100 43 Furskins and artificial fur .. 0 0 0 100 50 Silk 20 99 0 1 0 0 0 100 51 Wool, animal hair, horsehair yarn 27 86 13 1 15 0 0 100 52 Cotton 991 96 0 5 74 0 0 100 53 Vegetable textile fibers NES 141 99 0 1 .. .. 98 Other incorporated products 14,788 12 0 88 933 84 13 3 All above agricultural products 119,172 84 4 12 17,897 79 4 17 Memo item Domestic support in Green Box' 108,052 81 5 15 Note: Numbers in parentheses are the number of countries in the group; the European Union is counted as one. NES is not elsewhere specified. aDomestic support is defined as the sum of direct amount in WTO DS4-9 categories. See data appendix. bExport subsidies are based on budgetary outlay commitments from 24 WTO member countries' notifications. cThe sum in WTO DS1-3 categories. Source: Based on WTO documents G/AGINGIS/1 and.G/AG/NG/S/5/Rev.1. TABLE A -2. Average (%) Most Favored Nation Tariff Equivalent and Maximum Rate on Subsidized Agricultural Products, 1995-98 Quad countries (4) Other OECD countries' (5) Developing countries (76) HS-2 product Average Maximumh Average Maximumh Average Maximum 01 Live animals 11 128 79 588 lO 80 02 Meat and edible meat offal 27 157 59 480 25 362 04 Dairy products, eggs, honey 92 343 79 506 22 159 05 Products of animal origin NES 1 1 13 65 06 Live tree and other plant; flowers 4 17 19 186 19 71 07 Edible vegetables and roots 25 865 29 409 21 284 08 Edible fruit and nuts, melons 9 219 10 210 24 157 09 Coffee, tea, mate, and spices 4 17 2 11 21 213 10 Cereals 66 719 37 446 12 158 11 Milling products, malt, starches 95 1,403 72 553 18 171 12 Oil seed, oleagic fruit; miscellaneous grain 12 654 8 230 10 255 13 Lac, gums, resins, and other vegetable saps 4 30 1 5 14 75 14 Vegetable plaiting materials, vegetable products NES 0 1 12 66 15 Animal/vegetable fats and oils and products 10 81 16 179 18 188 16 Meat, fish, and seafood preparations NES 18 99 17 83 26 100 17 Sugars and sugar confectionery 44 209 19 55 25 135 18 Cocoa and cocoa preparations 9 65 29 530 19 102 19 Cereal, flour, starch, and milk preparations 7 24 16 100 23 100 20 Prep of vegetable, fruit, nut products 20 163 30 380 25 100 21 Miscellaneous edible preparations 41 302 44 609 22 185 22 Beverages, spirits, and vinegar 10 42 6 20 38 1,050 23 Residues and waste from food industry 5 87 35 364 12 55 24 Tobacco and manufactured tobacco products 29 111 10 42 34 907 33 Essential oils and resinoids, perfumes 0 1 15 75 35 Albuminoids, modified starches 4 8 4 4 14 68 41 Raw hides and skins 0 0 I1 100 43 Furskins and artificial fur 3 6 5 5 15 43 50 Silk 59 236 0 1 10 80 51 Wool, animal hair, horsehair yarn 9 55 0 0 6 30 52 Cotton 5 21 0 0 6 37 53 Vegetable textile fibers NES 1 11 0 0 8 53 All above agricultural products 26 1,403 28 609 21 1,050 Memo item Domestic support in Green Box' 17 1,403 27 609 17 1,750 Note: Numbers in parentheses are the number of countries in the group. The European Union is counted as one. NES is not elsewhere specified. al igh-income OECD countries excluding Quad countries. bHighest applied tariff. Source: Based on UNCTAD TRAINS data through WITS; OECD 2000. 200 THE WORLD BANK ECONOMIC REVIEW, VOLE. I 8, NO. 2 TABLE A-3. Average (%) Trade Shares of Products Affected by WTO Members' Agricultural Subsidies, 1995-98 Domestic support Export subsidies Economy (code) Exports Imports Exports Imports Albania (alb) 8.8 8.2 10.6 22.0 Algeria (dza) 0.0 20.0 0.1 23.8 Angola (ago) 0.3 7.2 0.1 16.2 Antigua (atg) 6.0 2.8 6.0 5.2 Argentina (arg) 25.6 2.3 29.3 3.0 Australia (aus) 17.0 1.4 16.3 1.7 Bahrain (bhr) 0.1 3.7 0.2 7.3 Bangladesh (bgd) 2.2 12.7 0.4 13.3 Barbados (brb) 21.9 5.9 20.9 8.1 Belize (blz) 46.6 5.9 47.5 7.3 Benin (ben) 84.7 5.9 78.2 12.8 Bolivia (bol) 11.3 5.2 19.9 7.4 Brazil (bra) 13.1 6.7 20.1 6.7 Brunei (brn) 0.0 1.8 0.0 3.7 Bulgaria (bgr) 6.9 5.4 9.5 6.3 Burkina Faso (bfa) 75.5 7.9 73.8 9.2 Burundi (bdi) 72.8 10.9 72.2 10.8 Cameroon (cmr) 24.7 8.6 27.4 8.5 Canada (can) 3.4 2.1 3.9 2.8 Central African Republic (caf) 24.8 4.4 24.8 12.2 Chad (tcd) 82.5 3.5 82.7 13.5 Chile (chl) 14.2 3.2 12.6 4.2 China (chn) 1.9 4.0 2.1 4.6 Colombia (col) 32.0 7.2 34.2 7.7 Congo, Dem. Rep (zar) 10.6 8.6 7.3 13.8 Congo, Rep. (cog) 1.1 3.9 1.0 7.2 Costa Rica (cri) 37.5 5.9 44.5 4.7 C6te d'lvoire (civ) 48.7 11.0 57.2 13.0 Croatia (hrv) 2.2 5.0 2.8 6.2 Cuba (cub) 50.7 13.4 53.8 17.t Cyprus (cyp) 24.5 4.2 20.0 8.8 Czech Republic (cze) 1.9 3.6 3.1 3.4 Djihouti (dji) 9.5 9.9 10.1 20.9 Dominica (dma) 57.7 7.1 32.4 8.6 Dominican Repub (dom) 10.2 7.4 12.0 9.3 EU-15 (eec) 31.2 5.7 38.3 6.6 Ecuador (ecu) 2.2 4.9 4.3 5.0 Egypt, Arab Rep (egy) 11.0 17.1 9.1 12.8 El Salvador (slv) 42.0 8.3 24.7 10.7 Estonia (est) 6.3 6.3 3.5 7.8 Fiji (fji) 37.9 6.5 36.1 7.7 Gabon (gab) 0.1 4.9 0.1 6.7 Gambia, The (gmb) 11.0 17.5 3.3 20.6 Ghana (gha) 32.4 4.7 39.0 7.1 Grenada (grd) 14.5 7.2 23.3 12.4 Guatemala (gtm) 48.6 7.2 42.0 9.0 Guinea (gin) 7.8 19.1 7.8 17.2 (Co,ltiniued) Hoekman, Ng, and Olarreaga 201 TABLE A-3. Continuied Domestic support Export subsidies Economv (code) Exports Imports Exports Imports Guinea-Bissau (gnb) 39.8 4.2 39.6 12.7 Guyana (guy) 33.1 6.7 33.2 10.4 Haiti (hti) 12.2 14.9 17.9 22.7 Honiduras (hnd) 43.9 9.0 26.7 6.2 Hong Kong, China (hkg) 0.1 1.5 0.6 2.3 Hungary (hun) 5.6 2.3 6.7 2.1 Iceland (isl) 7.9 2.6 0.7 4.3 India (ind) 8.4 2.4 9.9 2.5 Indonesia (idn) 2.8 8.9 4.0 7.9 Iran, Islamic Rep. (irn) 1.0 10.3 0.7 14.5 Israel (isr) 4.1 3.3 5.0 3.6 Jamaica (jam) 12.1 5.5 13.8 7.9 Japan (jpn) 0.1 4.9 0.2 6.2 Jordan (jor) 6.2 12.3 5.3 15.4 Kazakhstan (kaz) 0.0 0.0 12.3 5.9 Kenya (ken) 48.7 8.4 36.2 10.5 Korea, Rep. (kor) 0.3 3.6 0.6 4.1 Kuwait (kwt) 0.0 4.0 0.1 6.5 Kyrgyz Republic (kgz) 24.1 4.7 25.1 8.6 Latvia (Iva) 2.0 5.7 3.8 7.2 Lithuania (Itu) 6.7 5.0 7.1 7.5 Macao (mac) 0.3 2.5 0.5 6.7 Madagascar (mdg) 26.6 7.8 25.8 6.2 Malawi (mwi) 75.7 4.4 77.0 9.6 Malaysia (mys) 0.5 3.4 1.4 3.2 Maldives (mdv) 2.1 8.3 0.2 10.8 Mali (mli) 84.5 5.8 84.3 10.6 Malta (mlt) 1.2 3.0 0.6 3.4 Mauritania (mrt) 0.7 13.8 0.1 28.9 Mauritius {mus) 24.6 8.5 25.4 7.5 NMexico (mex) 3.0 4.3 4.5 4.3 Mongolia (mng) 12.0 2.2 12.5 10.6 Morocco (mar) 9.0 13.8 9.7 11.0 19ozambique (moz) 0.0 0.0 36.5 18.2 Myanmar (mmr) 23.6 1.0 17.1 6.7 New Zealand (nzl) 24.1 3.0 25.3 3.3 Nicaragua (nic) 40.0 8.7 35.2 15.8 Niger (ner) 17.2 12.9 2.3 13.4 Nigeria (nga) 1.8 7.5 2.0 10.9 Norway (nor) 0.3 2.8 0.5 2.6 Oman (ornn) 0.7 5.8 0.2 5.7 Pakistan (pak) 7.0 6.7 9.1 7.3 Panama (pan) 38.8 4.1 19.9 1.4 Papua New Guine (png) 15.1 2.6 12.0 3.0 Paraguay (pry) 55.1 3.5 27.0 11.9 Peru (per) 19.3 9.6 9.4 11.8 Philippines (phi) 5.6 4.5 3.6 6.6 Poland (pol) 2.7 4.9 3.9 4.4 (Continued) 202 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. Z TABLE A-3. Continued Domestic support Export subsidies Economy (code) Exports Imports Exports Imports Qatar (qat) 0.0 2.3 0.0 4.3 Romania (rom) 4.9 3.8 4.4 5.5 Russian Federation (rus) 1.4 6.9 1.5 12.4 Rwanda (rwa) 59.0 18.2 59.2 18.5 Saudi Arabia (sau) 0.1 6.3 0.2 9.0 Senegal (sen) 7.6 16.1 7.8 19.1 Sierra Leone (sle) 6.6 8.7 6.8 15.8 Singapore (sgp) 0.6 1.2 0.7 2.3 Slovak Republic (svk) 2.1 3.0 2.7 4.2 Slovenia (svn) 0.9 3.6 1.6 5.1 Solomon Islands (slb) 9.1 1.7 3.3 5.1 South Africa (zaf) 6.4 3.2 7.6 4.0 Sri Lanka (Ika) 2.1 9.1 3.6 10.5 St. Kitts and N (kna) 77.5 5.6 31.2 6.8 St. Lucia (Ica) 63.8 5.3 61.2 10.2 St. Vincent and (vct) 57.3 10.5 34.4 8.8 Sudan (sdn) 60.1 8.6 48.6 12.5 Suriname (sur) 11.6 5.5 9.6 6.8 Switzerland (che) 0.5 3.2 1.4 3.1 Taiwan, China (oan) 0.2 3.6 0.6 3.7 Tanzania (tza) 67.8 5.0 60.0 7.8 Thailand (tha) 7.0 2.8 8.0 2.5 Togo (tgo) 42.5 6.3 41.8 13.7 Trinidad and To (tto) 2.1 6.4 3.6 7.6 Tunisia (tun) 4.0 6.9 4.5 7.3 Turkey (tur) 10.5 5.0 8.1 4.8 Uganda (uga) 63.3 7.5 83.6 7.4 United Arab Emi (are) 1.0 3.1 1.8 4.9 United States (usa) 5.2 1.8 5.2 2.5 Uruguay (ury) 23.0 6.2 25.2 6.4 Venezuela (ven) 0.8 8.0 1.2 8.7 Zambia (zmb) 8.1 4.0 7.1 5.6 Zimbabwe (zwe) 59.3 3.0 48.8 3.7 All countries (143) 3.6 3.7 4.4 4.4 OECD countries (23) 3.1 3.3 4.0 3.9 Developing countries (90) 4.2 4.2 5.0 5.0 Least developed countries (30) 17.8 8.9 16.7 13.1 Note: Intra-EU trade is excluded. Source: Based on partners' data from UN COMTRADE statistics. A number of limitations are associated with the AMS data (see de Gorter 2002 for a careful discussion of problems associated with measurement of the AMS). Two problems are that the time period for which data are available is short and reporting is incomplete, especially for more recent years. Another problem is that the economic relevance of the AMS time series is limited given the use of the Hoekman, Ng, and Olarreaga 203 fixed 1986-88 benchmark for calculating price supports. However, assuming that changes in world prices over time have a proportional effect on the ams figures of all WTO members, the double log specification used to estimate the import demand and export supply and the focus on percentage changes in variables should not significantly affect our results. REFERENCES Anderson, Kym. 2002. "Trade Liberalization, Agriculture and Poverty in Low Income Countries." Paper presented at the Trade and Industrial Policy Strategies (tiPs) Annual Forum, September 9-11, Johannesburg. Available online at www.tips.org.za. Beghin, John, David Roland-Holst, and Dominique van der Mensbrugghe. 2002. "How Will Agricultural Trade Reforms in High-Income Countries Affect the Trading Relationships of Developing Coun- tries?" Working Paper, World Bank, Washington, D.C. Brenton, Paul. 2003. "Integrating the Least Developed Countries into the World Trading System: The Current Impact of EU Preferences under Everything hut Arms." Policy Research Working Paper 3018. World Bank, Washington, D.C. de Gorter, Harry. 2002. "The ANIS and Domestic Support in W'TO Trade Negotiations on Agriculture: Issues and Suggestions for New Rules." Working Paper, World Bank, Washington, D.C. Dimaranan, Betina, Thomas Hertel, and Roman Keeney. 2002. "OECD Domestic Support and the Devel- oping Countries." GTAP Working Paper 1161. Purdue University, Department of Agricultural Eco- nomics, Center for Global Trade Analysis, West Lafavette, Ind. Falvey, Rod, and Rod Tyers. 1989. "Border Price Changes and Domestic Welfare in the Presence of Subsidized Exports." Oxford Economic Papers 41(2):434-51. Fink, Carsten, Aaditya Mattoo, and Cristina Neagu. 2002. "Trade in International Maritime Services: How Does Policy Matter." World Bank Economic Ret'iew 16(1):120-45. Francois, Joseph, and Will Martin. 2003. "Formula Approaches for Market Access Negotiations." World Economy 26(1):1-28. Francois, Joseph, and Ian Wooton. 2001. "Trade in International Transport Services: The Role of Competition." Review of International Economics 9(2):249-61. Gulati, Ashok, and Sudha Narayanan. 2002. X I - Import Competition When Developing Coun- tries Liberalize Trade: The Indian Experience." Working Paper, International Food Policy Research Institute, Washington, D.C. Hoekman, Bernard, Francis Ng, and Marcelo Olarreaga. 2002. "Eliminating Excessive Tariffs on Exports of Least Developed Countries." World Bank Econonzic Rceietv 16(1):1-21. Inama, Stefano. 2003. "Trade Preferences and the wrio Negotiations on Market Access." Working Paper. United Nations Conference on Trade and Development, Geneva. Mattoo, Arvind, Devesh Roy, and Arvind Subramanian. 2002. "The Africa Growth and Opportunity Act and Its Rules of Origin: Generosity Undermined?" Policy Research Working Paper 2908. World Bank, Washington, D.C. OECD (Organisation for Economic Co-operation and Development). 2000. "Post Uruguay Rounds Tariff Regimes: Achievements and Outlook." OECD, Paris. Ozden, Caglar, and E. Reinhardt. 2003. *'The Perversity of Preferences." Policy Research Working Paper 2955. World Bank, Washington, D.C. Rae, Allan, and Anna Strutt. 2002. "The Current Round of Agricultural Trade Negotiations: Why Bother about Domestic Support?" Paper presented at the 5th Annual Conference on Global Economic Analysis, Taipei, June 5-7. Schiff, Maurice, and Alberto Valdes. 2002. "Agriculture and the Macroeconomy." In Bruce Gardner and Gordon Rausser, eds., Handbook of Agricuzltural Econonmics, vol. 2B. Amsterdam: North Holland. 204 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. Z Snape, Richard H. 1987. "The Importance of Frontier Barriers." In H. Kierzkowski, ed., Protection and Competition in International Trade. Oxford: Blackwell. Tangermann, S. 2002. "The Future of Preferential Trade Arrangements for Developing Countries and the Current Round of wTo Negotiations on Agriculture." Working Paper. Food and Agriculture Organ- ization, Rome. Tokarick, Stephen. 2003. "Measuring the Impact of Distortions in Agricultural Trade in Partial and General Equilibrium." Working Paper. International Monetary Fund, Washington, D.C. Valdes, Alberto, and William Foster. 2002. "On the Management of Price Risk in the Context of Trade Reform in lODCS." Working Paper. World Bank, Washington, D.C. Wilson, John, and Victor Abiola. 2003. "Standards and Global Trade: A Voice for Africa." Working Paper. World Bank, Washington D.C. Zietz, Joachimlz, and Alberto Valdes. 1986. "The Potential Benefits to LDCS of Trade Liberalization in Beef and Sugar by Industrialized Countries." Weltwirtschaftliches Archiv 122(1):94-112. The Earnings Effects of Multilateral Trade Liberalization: Implications for Poverty Thomas W. Hertel, Maros Ivanic, Paul V. Preckel, and John A. L. Cranfield Most researchers examining poverty and multilateral trade liberalization have had to examine average, or per capita effects, suggesting that if per capita real income rises, poverty will fall. This inference can be misleading. Combining results from a new international cross-section consumption analysis with earnings data from household surveys, this article analyzes the implications of multilateral trade liberalization for poverty in Indonesia. It finds that the aggregate reduction in Indonesia's national poverty headcount following global trade liberalization masks a more complex set of impacts across groups. In the short run the poverty headcount rises slightly for self- employed agricultural households, as agricultural profits fail to keep up with increases in consumer prices. In the long run the poverty headcount falls for all earnings strata, as increased demand for unskilled workers lifts incomes for the formerly self-employed, some of whom move into the wage labor market. A decomposition of the poverty changes in Indonesia associated with different countries' trade policies finds that reform in other countries leads to a reduction in poverty in Indonesia but that liberal- ization of Indonesia's trade policies leads to an increase. The method used here can be readily extended to any of the other 13 countries in the sample. Poverty reduction is an increasingly important consideration in the deliberations on multilateral trade liberalization, and it has been accorded an important position in the Doha Development Round of the World Trade Organiza- tion (WTO) negotiations.' Globkom and the World Bank sponsored a confer- ence in Stockholm in October 2000 aimed at assessing the state of the art in Thomas W. Hertel and Paul V. Preckel are professors and Maros Ivanic is a graduate research assistant in the Department of Agricultural Economics at Purdue University; their e-mail addresses are hertel@purdue.edu, preckel@purdue.edu, and ivanicm@purdue.edu. John A. L. Cranfield is assistant professor in the Department of Agricultural Economics and Business at the University of Guelph; his e-mail address is jcranfie@uoguelph.ca. The authors acknowledge support from the Development Research Group at the World Bank. Specifically, they would like to thank Will Martin for championing this work and making the household surveys available. They would also like to thank seminar participants at the World Bank, Cornell University, the 2001 GeWisola conference in Braunshweig, the Globkom confer- ence in Stockholm, and the Inter-American Development Bank conference on trade policy in the Americas, as well as Francois Bourguignon, L. Alan Winters, and three anonymous reviewers for valuable comments. 1. See also the survey by Winters (2000) and the handbook on trade liberalization and poverty by McCullogh and others (2001). THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO 2, Q The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi:10.1 093/wber/lhhO39 18:205-236 205 2 06 THE WORLD BANK ECONOMIC REVIEW, VOI. i 8, NO. 2 quantitative policy research on trade policy and poverty. It drew together eco- nomists working with household surveys (Levinsohn and others 1999; Case 1998; Friedman 2001; lanchovichina and others 2000, 2002) and researchers using computable general equilibrium (CGE) models with a poverty focus (Devarajan and van der Mensbrugghe 2000; Harrison and others 2002b).2 Conference participants came to the realization that factor markets, so critical for determin- ing the trade-poverty linkage, have been neglected in much of the poverty research (Harrison and others 2002b; Hertel and others 2000). The same point was emphasized in the path-breaking CGE-based work of Adelman and Robinson (1978) and in a recent overview by Decaluwe and others (1999). More recently, Bourguignon and others (2002) have more fully developed the role of labor markets in determining the impact of macroeconomic shocks on poverty. One reason for the historic neglect of factor markets in much of the research on poverty stems from the preference for focusing on the expenditure side of household surveys, because of the greater reliability of spending data for measuring poverty (Lipton and Ravallion 1995). Although valuable for poverty measurement, this focus is inadequate for counterfactual analysis of policies and poverty, which require proper treatment of the earnings effects. The factor price effects of trade policy changes are often substantial, particularly for households that do not have diversified sources of income but rely on sector-specific factors. For this reason, it is important to stratify households, identifying those with specialized earnings patterns (see also Decaluwe and others 1999). In the long run, an increase in returns to labor and capital employed in one sector will attract more resources to that sector, spreading the gains more widely. The Stolper-Samuelson theorem states that if the benefiting sector is relatively intensive in unskilled labor, a rise in the relative price of this sector's output will boost economywide unskilled wages over the long run, benefiting wage earners not directly employed in that sector. This distinction between the short- and long- run earnings impacts of trade liberalization is a central theme of this article. The work presented at the Globkom-World Bank conference on trade policy and poverty and other recent work reveal another deficiency.3 Linking multilateral trade liberalization and poverty requires a multiregion approach. But the country specificity of household surveys and their inconsistency with the multiregion models used for trade policy analysis make such analysis difficult. The simplest approach to assessing the poverty impacts of multilateral trade liberalization sidesteps this problem by focusing solely on the average or per capita effects of trade liberalization (see, for example, Cline 2004). Thus the entire income distribution is assumed to shift in parallel, based on the predicted change in per capita income. To the extent that this equiproportional shift in 2. Other good examples are offered by Lofgren (1999) and Evans (2001). 3. See also the recent survev of studies on1 trade and poverty by Reimer (2002). Hertel and others 207 income following trade liberalization is positive, it will likely lift some house- holds above the poverty line, so the poverty rate is predicted to fall. Decaluwe and others (1999) extend this approach by identifying different household types (smallholder farmers, landless laborers) and evaluating the per capita income change for each stratum independently. However, they continue to assume a parallel shift in the income distribution for each stratum. The national per capita approach is unlikely to be satisfactory, particularly in the short run when returns to specific factors are differentially affected by trade liberalization. The Decaluwe and others (1999) approach works fairly well in the short run, but not in the long run, when differences in the composition of self-employed earnings across the income spectrum come into play. Recently, studies have sought to map the price changes from a CGE model directly to the survey data, circumventing the need to aggregate households. Chen and Ravallion (2003) combine disaggregated household survey data with trade liberalization results from the Global Trade Analysis Project (GTAP) model of global trade. Thus, they are able to draw conclusions about the impacts of accession on individual household types and locations, which is very attractive from a policy point of view. However, the well-known inconsistencies between survey data and national accounts data (which the trade models use) frequently give rise to contradictory predictions for national per capita outcomes. One contribution of this article is to show how to reconcile per capita earnings and spending patterns between the two frameworks, enabling consistent predictions of national per capita impacts of trade liberalization by both the trade model and the survey-based microsimulation analysis. The methodology developed here was designed explicitly for multiregion analysis. By capitalizing on a newly available methodology for estimating household spending patterns across both countries and income strata within countries, spending behavior can be summarized parsimoniously. This allows more attention to be focused on the earnings side of the problem, which is critical to the results. Although the treatment of factor markets is rudimentary compared with the recent work of Bourguignon and others (2002), it has the virtue of being operational across a wide range of countries and household surveys. This approach has been implemented for 14 countries and could readily be applied to many more countries where income surveys are available. This article begins by examining the pattern of earnings specialization in the sample of developing countries. The systematic patterns that emerge motivate the subsequent stratification of households. The analytical framework for the study consists of a microsimulation model, built on the household survey data and used to assess impacts on individual households, and a global trade model, used to generate price changes. A key part of the research exercise involves modifying the trade model and adjusting the two databases to be mutually consistent and produce the same national per capita outcomes. The model is then used to analyze the impact of global trade liberalization on poverty in one of the sample economies, Indonesia. The article concludes with a discussion of 208 THE WORLD BANK ECONONIH REVIEW, VOL. 18, NO. Z the strengths and limitations of this approach for linking global trade liberal- ization and poverty in developing economies. I. SPECIALIZATION OF EARNINGS IN A SAMPLE OF DEVELOPING ECONOMIES A key premise of this article is that in the short run household incomes will be differentially affected by global trade liberalization, depending on their reliance on sector-specific factors of production. For example, a household that earns all of its income from a family-run farm will be heavily dependent on the prices of agricultural products. If agricultural prices fall, household members may even- tually find other employment, but this is likely to be difficult in the short run, particularly if members are not currently employed off-farm. This close link between farm household welfare and agricultural prices has also been observed in studies using annual household survey data (Chen and Wang 2001). Because of the potential importance of specialized earnings sources in the analysis of short-run impacts of trade liberalization, their prevalence was exam- ined using national household survey data for 14 countries in Africa, Latin America, and Southeast Asia. The surveys were selected on the basis of their availability through the World Bank, recent coverage (summers of 2001 and 2002), thorough treatment of household earnings, and match with the country coverage in the trade modeling database from GTAP version 5. This was the largest group that could be assembled at the time based on these criteria. The unit of analysis is the household, with equal sharing of income within the household assumed to obtain income on a per capita basis.4 The share of households with specialized earnings from agriculture (house- holds that earn 95 percent or more of their income from agricultural profits) was plotted against GDP per capita, measured in purchasing power parity (PPP) terms (figure 1). Because workers in these households both work full-time in agriculture and are self-employed, they are unlikely to switch quickly to other activities if returns to farming fall. ILikewise, because they are fully employed in agriculture, they are unable to quickly increase the amount of effort devoted to farming if returns rise, short of reducing their leisure time. There is a negative correlation between GDP per capita and the share of house- holds specialized in agriculture. In Malawi, the poorest country in the sample, nearly 40 percent of households have specialized earnings from agriculture. In Chile, the richest country in the sample, only about 6 percent of households have specialized earnings from agriculture. Although there are a few exceptions, for 4. The equal sharing assumption is clearly problematic, as it is only a special case of what would be found in a more general bargaining model of intrahousehold behavior (Bourguignon and Chiappori 1994). This assumption tends to understate income inequality, although the impact on poverty measures is less clear (Haddad and Kanbur 1990). Hertel and otibers 209 FIGURE 1. Share of Households with Agriculture-Specialized Earnings and GDP per Capita 40 0 Madawi 8" 30 130 S S 20 Zambia Indonesia § 0 05 Uganda Philippines Colombia | 10 1 0 Bangladesh Peru Chile Vietnam Venezuea T nalan M * ~~~~Mexcoo Brazil 0 0 2.000 4,000 6,000 8,000 10,000 GDP per capita (PPP US$) Souirce: Authors' calculations based on data from 1993-2002 household surveys and World Bank data on countries' (DP. many developing economies the agriculture-specialized segment of the population is substantial, and the share is generally inversely related to per capita GDP. But how distinct is this agriculture-specialized group? Does it warrant indi- vidualized treatment in the analysis? Consider the case of Indonesia, which falls in the middle of the sample. Although it is neither among the poorest countries nor among those with the highest share of agriculture-specialized households, it does have a significant proportion of agriculture-specialized households. A plotting of the distribution of households in the Indonesian survey with the data arranged by share of household income derived from agricultural profits (x axis) and log of income level (y axis) shows a bimodal distribution for agricultural specialization (figure 2). Although a majority of households receives little or no profit income from agriculture, a substantial minority (21 percent of the population) receives virtually all its income from self-employment 210 THE WORLD BANK ECONOMIC: REVIEW, VOL. i8, NO. 2 FIGURE 2. Agriculture-Specialized Earnings in Indonesian Households and Income Level 0' 0 0 n Highest -:: . . F t ~~~~~~~~~LoWeSL Share of agricultural profits in household's income (percent) Source: Authors' calculations based on data from Indonesia 1993 household survey. TABLE 1. Structure of Poverty Headcount in Indonesia by Earnings-Based Stratum (percent) Indicator Agriculture Nonagriculture Wages Transfers Diversified Total Share in total population 21 15 18 1 44 100 Share in total poverty 34 11 7 3 45 100 Poverty rate in stratum 25 11 5 30 16 15 Source: Authors' calculations based on Indonesia 1993 household survey data. in farming (table 1), so isolating this group in a specific stratum looks like a good idea.5 It looks like an even better idea based on the share of the impoverished population in this category. Although agriculture-specialized households 5. Note that although figures 1-5 use households as the unit of observation, table l uses population, assuming the equal sharing of income among household members. Thus this table reports that 21 percent of the population resides in the agriculture-specialized stratum. Hertel and otbers 211 account for only about one-fifth of the population in Indonesia, they account for more than one-third (34 percent) of individuals with per capita incomes of less than $1 a day (see table 1). Clearly, the fate of these households under trade liberalization could have an important effect on the national poverty rate. The other type of household specialization that appears to be correlated with national per capita income in the sample is wage and salary specialization (figure 3). People in this group work for others. Because these wage-specialized households are wholly reliant on labor income, their earnings will be closely tied to changes in market wages. There is a strong positive correlation with per capita GDP at PPP prices. The poorest countries tend to have relatively few such households (less than 5 percent of total households in Uganda and Vietnam), whereas the richer countries tend to have more than 25 percent of their house- holds in this category. This positive correlation is hardly surprising. Increased specialization is expected, along with more efficient formal labor markets, as countries become more developed. FIGURE 3. Share of Households with Wage-Specialized Earnings and GDP per Capita 40 Mexico 30~~~~~~~~~~ 30 0 A bVenezuela Brazil Cdombia j20 Cl tQ Zzmbia o5 Indonesia P 10 Maw Philippines Theland Bangladesh 0 Uganda Vetnam 0 2,000 4.000 6,000 8.000 10,00 GDP per capita (PPP US$) Source: Authors' calculations based on data from 1993-2002 household surveys and World Bank data on countries' GDP. 212 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 FIGURE 4. Wage-Specialized Earnings in Indonesian Households and Income Level C' a Highest i/_ ~~~~~~~~~~~~~~~ ~~Log of ,, ,. _ ~ ~ ~ = LowesL Share of wages in household's total income (percent) Source: Authors' calculations based on Indonesia 1993 household survey. A plotting of a three-dimensional distribution of households in the Indonesia survey highlighting the share of household income obtained from wages and salaries again shows sharp peaks at the two extremes (figure 4). Although most households are not specialized in this dimension, there is a significant cluster above the 95 percent earnings share from wages and salaries. These households account for about 18 percent of the population and about 7 percent of the population under the poverty line (see table 1). Thus, in contrast to agriculture- specialized households, the wage-specialized households are disproportionately nonpoor. This is also evidenced in the modest stratum poverty headcount of 5 percent for wage-specialized households compared with 15 percent for the nation as a whole. In addition to agriculture- and wage-specialized households, some house- holds also specialize in nonagricultural profits (self-employed in nonagricultural sectors) and transfers and some households are not specialized (diversified households). Although the relative size of the transfer-specialized group appears to be positively correlated with per capita income, the household shares for the Hertel and others 213 other two categories do not. For Indonesia, households wholly reliant on non- agricultural profits (for 95 percent or more of total income) account for 15 percent of the population and 11 percent of the poor, with a poverty rate somewhat below the national average (see table 1). Households specialized in transfers are a much smaller group (1.3 percent of the population) but dispro- portionately poor (2.6 percent of the national poor). Finally, the diversified group represents about 45 percent of both the total and the impoverished populations in Indonesia. II. IMPUTING RETURNS FROM PROFIT-TYPE INCOME FOR LONG-RUN ANALYSIS Because returns (net of tax) to comparable factors of production are expected to be equalized across sectors in the long run, the boost to self-employed agricul- tural labor that occurs in the short run when world food prices rise will be shared with nonagricultural labor as more workers are drawn into farming (or fewer leave agriculture). Alternatively, when nonfarm wages rise, this improve- ment is expected to be shared eventually with self-employed farm laborers as they seek off-farm jobs. Thus, to identify the long-run impacts on individual households and groups, it is necessary to assess the underlying factor endowments of the self- employed population. How much of observed agricultural and nonagricultural "profits" can be attributed to unskilled labor, and how much represents returns to land or capital? This type of earnings imputation is notoriously difficult, but it is essential to an analysis of the impacts of trade liberalization on poverty in the long run. The first step in splitting reported agricultural and nonagricultural income into returns to capital, labor, and land is to impute income for each household member as the average wage of all workers in the economy with the same personal characteristics (age, education, skill, and industry) who earned wage income only. This imputed labor income for all household members involved in the family business is then subtracted from the reported profits of the household business. For greater accuracy, agricultural and nonagricultural business income are carefully separated, so that only imputed agricultural wages are subtracted from agricultural profits and only imputed nonagricultural wages are subtracted from nonagricultural profits. If no information on the nature of imputed wage is available, the industrial classification for the head of the household is used. Next, the remaining profit-type agricultural income is split into returns to capital and returns to agricultural land. If the sum of imputed wages exceeds total reported business income, all operating surplus is classified as returns to labor and the capital return for this business is set to zero to avoid negative flows. Then property rental income directly reported by households is added to this composite to estimate total returns to capital and agricultural land. Because nonfarm land is treated as part of the nonfarm capital composite, this completes the task of imputation for nonagricultural profits. Ideally, information on 2 14 THE WORLD BANK ECONOMIC REVIEW, VOL. T 8, NO. Z FIGURE 5. Imputation of Labor Earnings for Self-Employed Agriculture Households in Indonesia 100 Land 90 80 70 - Capital Labor 0 0) E 5 O 40 X 30 20 10 0- 1.00 2.00 3.00 4.00 5.00 S.0C Income (millions of rupiahs) Source: Authors' calculations based on Indonesia 1993 household survey. household farm land holdings would be used to split this remaining category of income for the agricultural enterprises, but attempts to do so failed. Therefore, a simple share-based split of nonwage, profit-type income in agriculture is used. To make it consistent with the trade liberalization simulation analysis, the national average share of farmland in total nonlabor agricultural earnings was taken from the GTAP 5 database. For Indonesia, the results of this imputation procedure show that for most households employed in agriculture more than 80 percent of their income represents returns to family labor (figure 5).6 The residual share of income attributed to capital and land is greatest at the lowest and highest income levels. The very poorest households have a very low endowment of human capital, because they have not only low incomes but also a low share of imputed labor within that low income level.7 The same U-shaped relationship between capital's 6. This imputation was also undertaken for the other countries in the sample, and the results are discussed in Ivanic (2003). 7. Since children under 12 years of age are typically excluded from the employment questions on household surveys, it is also possible that these apparent returns to non-labor inputs are really returns to child labor. This pattern emerges in many of the other countries as well. Hertel and otbers 215 FIGURE 6. Imputation of Labor Earnings for Self-Employed Nonagriculture Households in Indonesia 100 90 80 Capital 70 X) 602 E 50- 40 30 20 10- C ~~~~Labor o 4 20 0 5 10 15 20 25 Income (millions of rupiahs) Souirce: Authors' calculations based on Indonesia 1993 household survey. share of imputed income and per capita household income is found for the nonagriculture households, but the shares for capital income are much larger, especially at the higher income levels (figure 6). One common problem in household surveys is underreporting of income. In a study of income distribution in Organisation for Economic Co-operation and Development (OECD) countries Atkinson and others (1995) find evidence of systematic income underreporting of 10-20 percent. Underreporting is likely to be even greater in developing economies, as is evident when the share of estimated gross factor income in agriculture and nonagriculture are computed for Indonesia.8 The survey shows the share of gross factor income earned in agriculture to be nearly half, whereas it is only about 20 percent in the national accounts.9 To reconcile these two databases in this key dimension by adjusting for underreporting of capital income, nonagricultural profit-type income reported by the wealthiest households in the survey is adjusted to reflect the same agriculture-nonagriculture mix of earnings as in the GTAP 5 database. This 8. To match up with the survey data definitions, agriculture is defined as including fisheries and forestry. 9. This figure comes from the 1997 (GTAP 5 database for Indonesia. It is based on an updated version of the 1993 Indonesia input-output table (Biro Pusat Statistik). 2 1 6 THE WOR I D BANK FC ONONI IC RFVIEW, VOL. i 8, NO 2 TABLE 2. Estimated Shares of Gross Factor Earnings in Indonesia (percent) Factor Agriculture Nonagriculture Total tlnskilled wage 5.1 12.7 17.8 Skilled wage 0.0 7.8 7.8 Profits 14.5 59.9 74.5 Land 1.5 0.0 1.5 Capital 5.3 37.9 43.2 Natural resources 0.9 2.0 2.8 Imputed unskilled labor 6.9 9.6 16.5 Imputed skilled labor 0.0 10.4 10.5 Tlotal 19.6 80.4 100.0 Souirce: Authors' calculations based on Indonesia 1993 house- hold survey and cl1AP 5 data. approach is supported by the work of Mistiaen and Ravallion (2003), who find that underreporting of income is greatest for the rich."' The factor composition of GTAP'S gross national earnings is next adjusted in line with that suggested by the survey data (Ivanic 2003). This has the important consequence of increasing the share of skilled labor in the economy. Also, as a result of the substantial imputed returns to self-employed labor, the capital intensity of the GTAP database is reduced for Indonesia."l The resulting matrix of gross factor earnings shares for Indonesia shows that imputed labor income accounts for more than a third of profits (table 2). Total labor earnings are about equally divided between wages and salaries and imputed labor income. Skilled labor accounts for about half the imputed income in the nonagricultural sector, but almost none in agriculture, which relies almost entirely on unskilled labor. III. ANALYTICAL FRANMEWORK Analysis of the impacts of trade liberalization on the poor begins with the specification of a utility function and an associated consumer demand system for determining household consumption and the maximum utility attainable by the household at a given set of prices and income. A modified version of the GrAP global trade model is then used to generate the price changes for the microsimulation analysis. 10. An alternative would be to increase all nonfarm profit-type income by the same proportion. However, when that is done, the inicome adjustimienit is sufficienit to lift most of the affected houscholds out of poverty. This does not seem realistic, and given the focus on poverty, the adjustment was made for the richest households instead. 11. The apparently excessivelv capital intensity of the (l-A, database for- developing economies, particularly in Southeast Asia, points to a pervasive problem of underestimating returns to self-employed labor. Using the survey data in this way promises to improve the (.OAF database for these coiitries. Hertel and otbers 217 Specifying the Microsimiulationz Model The utility of the household at the poverty line (here referred to as the marginal household) is defined as the poverty level of utility. If a household's utilitv falls below this level following trade liberalization, it is considered to have fallen into poverty. If a household rises above this level of utility, it is considered no longer in poverty. The poverty level of utility can also be used to compute the poverty gap-the amount of transfer required to lift extremely poor households up to the poverty line, permitting them to achieve the poverty level of utility. This approach to determining the poverty line appears preferable to that proposed by Decaluwe and others (1999), who identify a basic needs bundle of goods and implement it in a linear expenditure system (LES) model of con- sumption in which households below the poverty line cannot substitute among consumption items when prices change. In contrast, the approach used here permits such substitution and does not rely on the somewhat artificial definition of a fixed basket of basic needs. This study uses Rimmer and Powell's (1992a, 1992b, 1996) implicit directly additive demand system (AIDADS) to represent consumer preferences because it captures expenditure patterns across the global income spectrum (see also Cranfield and others 2000, 2002). AIDADS has been widely estimated on inter- national cross-section data, and it performs well out of sample compared with other demand systems (Cranfield and others 2003). This functional form may be viewed as a generalization of the popular, but restrictive LES. But unlike the LES, AIDADS allows for nonlinear Engel responses while maintaining a parsimonious parameterization of consumer preferences. The budget share form of AIDADS iS given as (1) A, = (p,,7y, /y) + ( [Lc, + PI1 exp(u)]/[ I + exp(u) )( - p'y/y) V n. where X.,, is the budget share of good n; p,, is the price of good n; y is income; y,, x,,, and 3,, are unknown parameters; it represents utility; and the final term represents the share of discretionary spending in total income. The following parametric restrictions are used to ensure well-behaved demands: 0 < x,,, PI, < l for all n, and Y 1l=. If ox,, =,, for all goods, then AIl)ADS simplifies to the LES. By replacing the values of the marginal budget shares in the LES with more general terms that are functions of a value that varies with utility, AIDAI)S allows for marginal budget shares that vary across per capita expenditure levels in a fairly general manner. Moreover, the average budget shares from AIDADS also vary nonlinearly across expenditure while remaining within the unit interval. This study also draws on recent work by Cranfield (1999) and Cranfield and others (2004), who estimate the parameters of a complete demand system while simultaneouslv using data on the distribution of expenditure by quintile to permit recovery of the unobservable distribution of expenditures for each quintile. This approach requires data typically used in demand system estimation (prices 218 TIHE W'ORLD BANK ECONOMIC REN'IEW, VOL. I8, NO. Z TABLE 3. Estimated AIDADS Parameter Values, Noncountry Specific Parameter Grains Livestock Other food Nondurables Durables Services '1 0.002 0.000 0.000 0.000 0.000 0.000 x~ 0.233 0.203 0.333 0.151 0.035 0.044 1 0.000 0.051 0.047 0.262 0.113 0.528 Sozurce: Authors' estimates based on data from the International Comparison Project and the available data on expenditure inequality. and per capita quantities and expenditures), in addition to survey-based infor- mation on the distribution of expenditure (or income). Rather than estimating a model that predicts a budget share for each good on a per capita basis in each observation, the framework approximates the distribution of expenditure, esti- mates demand system parameters consistent with the demand and expenditure data (including the distribution information), and predicts budget shares for each good across expenditure levels within each national observation. Consumption, price, and expenditure data from 113 countries in the 1996 International Comparisons Project (icp) data set are used for the demand sys- tem portion of the model (Kravis and others 1982). Survey data from the 14 countries for which this formation is available are supplemented with quintile data from the Deininger and Squire (1996) database and various issues of the World Bank's World Development Report (1992, 1993, 2000).12 It is important to note that the recovered expenditure distribution aggregates back to the per capita expenditure levels in the icP data, as well as reproducing the observed distribution of total expenditure. The icp consumption and price data are aggregated up to six goods: staple grains, livestock products, other food products, other nondurable goods, durable goods, and services. The emphasis on food products is appropriate for this poverty-focused study, because poor households spend a large share of their income on food. The non-country-specific estimated AIDADS parameters (table 3) may be used to predict spending patterns across the income spectrum, even in countries where household expenditure surveys are not available. This makes the AIDADS estimation of parameters particularly well suited to a multicountry analysis of trade and poverty. A few observations about these parameters are in order. First, the estimated subsistence budget shares (Yn) for all products except for staple grains are zero. Second, the parameters i, and K3n represent estimates of the bounds of the AIDADS marginal budget shares. So a, = 0.233 indicates that at low income levels grains account for as much as 23.3 cents of each additional 12. In the cases in which there are no original survey data and only quintile data are available, a finer distribution of expenditure is approximated across 15 expenditure levels for each observation in the icp data set. These 15 expenditure levels are equally allocated across the five quintiles, so that there are three expenditure levels within each quintile. Hertel and others 219 TABLE 4. Calibrated AIDADS Parameter Values for Indonesia Parameter Grains Livestock Other food Nondurables Durables Services 'Y 0.002 0.000 0.000 0.000 0.000 0.000 ce 0.314 0.151 0.308 0.135 0.045 0.046 1 0.000 0.038 0.043 0.231 0.142 0.547 Source: Authors' estimates based on data from the International Comparison Project and the available data on expenditure inequality. dollar of expenditure, whereas 3n, = 0 suggests that at high income levels changes in expenditure on grains are negligible with an increase in expenditure. Live- stock and other food also show values of an > 1, thereby dictating that marginal expenditures fall with income, whereas nondurables, durables, and services show values suggesting that marginal expenditures rise with income. As is generally done in microsimulation studies, this international estimation of the AIDADS parameters is followed by a calibration step in which the values of oCn and 3n, are altered to ensure that predicted demands equal observed demands for the country in question.13 While some of the parameters change considerably (most notably grains and livestock products, see table 4), they retain the qualitative relationships that were observed in the cross-section estimates (see table 3). The predicted expenditure patterns for households in Indonesia across the full spectrum of expenditure show a monotonically declining pattern for the grains budget share, as expected from the parameters in table 4, and for livestock and other food, albeit at a slower rate (figure 7). The budget shares for nonfood goods, including nondurables, durables, and services, follow an increasing pattern. A natural alternative to using the AIDADS function would be to predict expen- diture patterns from the survey data, either by econometrically estimating a demand system (although it would be difficult without obtaining separate price data) or by using budget shares to create a local measure of welfare changes, as in Chen and Ravallion (2003). This would not be possible in 6 of the 14 sample countries, because expenditure data are not available in the surveys for those countries. However, such data are available for Indonesia, and preliminary comparisons indicate that the predicted pattern of expenditure is consistent 13. After a common AIDADS demand system is estimated across all countries in the icp sample, the parameters are adjusted on a country-by-country basis to match observed per capita spending patterns. First, the subsistence budget share is taken as fixed, and the ratio of the actual discretionary budget share to the fitted discretionary budget share is used to rescale the remaining parameters of the AUDADs demand system. Second, the level of utility and a scaling parameter in the AIDADS utility function are calibrated to match the observed expenditure pattern in each country. The result is a country-specific AIDADS utility function and demand system that matches actual consumption. 220 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 FiGURE 7. Predicted Budget Shares for Indonesia Using Calibrated Parameters, 1997 100 l cZ 700 Services 60 ' Durable e -. 0 Nondurable 0*50 cs, 50 _, > , *~~~~~~~~~~~~~~ Other food n 40 - - 'U Livestock 30 U Grain : 20 E n 10 0 N O e 1 (D CS) O N n 0t 3 m N- XO N- L LL 'T LO LO U) LO LA) LO LO (D CO CO (D (O (D co (D (c N- r- Log of nominal expenditure Source: Authors' estimates based on IcP 1996 data updated to 1997. with the pattern from the survey."4 Given the convenience of working with an explicit demand system, as well as the consistency obtained by using the same demand system in the global trade model, this appears to be a sound choice. Next, the per capita utility function, which is common across all individuals within each country, can be used to specify the household microsimulation model, which involves maximizing per capita utility, subject to a per capita budget constraint, based on the households' overall endowments. Choose (xlk, .. ., xik, .. , xn), where i indexes the commodities and k house- holds. To maximize uk (2) subject to Yi lU,(xik,Uk) = 1, (3) Ui(Xik, Uk) = Pik(Uk) ln([xik - yi]/A exp[uk]) V i, (4) (Pik(uk) = [ci + exp(uk)]/[1 + exp(uk)], and (5) ,1P,xik = yk = EfWfE' - Ef6fPfE + T' Y In this formulation, equations 2 and 3 define the implicitly additive AIDADS utility function with parameters ci, Pi, y,, and A and the marginal budget share as 14. Because the base years and commodity definitions differ between the ICP data and the survey data, the per capita expenditure shares also differ. Hertel and others 221 defined by equation 4. Equation 5 is the per capita budget constraint, with income defined net of depreciation and inclusive of any transfers,15 where Wf is the wage paid to endowment Ekf, 6i is the geometric rate of depreciation for endowment EPf (zero for noncapital items), Pf is the cost of replacing depreciable endowment f (the capital goods price), and Tk is the transfer rate for household k, which is assumed to be a constant share of net national income, Y. The subsequent analysis uses survey-based observations on endowments and transfers. The depreciation rate for capital stock is obtained from the national accounts. Trade liberalization will alter the wages associated with each endow- ment, the price of capital goods, and transfers. The resulting level of income for household k can be computed using equation 5. Once the new income level is known, it can be combined with the new vector of commodity prices to compute expenditure on each good, and hence individual demands, using equation 1. Equations 2-4 are then used to compute per capita utility, and the postliberaliza- tion utility level is used in computing the change in poverty headcount. Equations 1-5 can also be used to compute the transfer necessary to bring a given impover- ished household up to the poverty line. Modeling Trade Liberalizationz In theory, the microsimulation model could be used in conjunction with any policy simulation framework capable of yielding the requisite price changes. In practice, however, there are substantial challenges in marrying the two analy- tical frameworks. Most important, the two models must be consistent in their characterization of earnings and spending-no small task as the preceding discussion of the earnings data in table 2 shows.16 This analysis uses a modified version of the GTAP global trade model (Hertel 1997) to generate the price changes to be fed into the microsimulation analysis. The modifications are aimed at obtaining consistency in national per capita con- sumption between the global trade model and the microsimulation framework. Building on the GTAP model has several advantages. First, it is a global model, so it is capable of producing results from global trade liberalization scenarios- an important objective of the analysis undertaken here. Second, it is a relatively standard CGE model, assuming perfect competition and differentiated products in international trade. Owing in part to this simplicity, GTAP is the most widely used global trade model, with more than 1,000 users around the 15. The only taxes that are explicitly modeled are indirect taxes, which are reflected in the difference between consumer prices and gross factor earnings. 16. Note that the postsimulation incidence analysis abstracts from the potential impact that the resulting changes in income distribution might have on relative prices. Given the relativelv modest shifts in income, coupled with modest differences in consumption shares, the resulting approximation error is unlikely to be very severe. This issue could be resolved if the disaggregated households were directly incorporated into the trade policy model (Cogneau and Robilliard 2000). This would be a major under- taking in a relatively detailed ccE model of the global economy. 222 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. 2 world. Demonstrating how to make it consistent with the microsimulation model opens the door to other analysts interested in addressing distributional issues. A final reason for using this framework is the regional disaggregation in the GTAP 5 database, which is large and continually expanding (13 regions in version 1, 66 regions in version 5, and a projected 87 regions in version 6). Once gross factor earnings in the microsimulation and GTAP models have been reconciled (see previous discussion), several further adjustments are required to fully reconcile these analytical frameworks. First, in the specification of con- sumer demand in the GTAP model the constant difference of elasticities demand system is replaced by the econometrically estimated AIDADS demand system discussed previously,'7 making specification of consumer demand in the two frameworks fully consistent. Because the icp-based consumer expenditure shares are evaluated at consumer prices, whereas the GTAP consumption vector is evaluated at producer prices, wholesale, retail, and transport margins applied to goods destined for private consumption must be explicitly modeled. These are modeled using a Cobb-Douglas production function, which combines the producer good and margin services to produce the consumer good. This is important, because these margins can perform an important insulating role when trade liberalization alters world prices and hence domestic producer prices (Winters and others 2003). Depreciation, a critical component of the macroeconomic accounts, is absent from the survey data. To reconcile the net income effects of trade liberalization between the two frameworks, national depreciation is shared out among the households in the microsimulation model in proportion to estimated gross earnings from capital."8 A final problem relates to transfer payments, which are unobserved in the GTAP 5 database but are assumed to be proportional to net national income. Accordingly, government spending, tax revenues, and foreign borrowing, which are explicitly modeled in GTAP, are also tied to net national income in the model closure adopted in the subsequent simulation analysis.19 Following Harrison and others (2002a, 2002b) forgone tariff revenue is replaced by a value-added tax to maintain tax's share in net national income.20 17. Because the AIDADS parameters are modified to replicate observed per capita consumption in each region, the parameters differ by country (see previous discussion). 18. National depreciation is obtained from the (;TAP S data base. This estimate comes originally from the World Bank. The share of depreciation in gross capital income is computed and applied to the microsimulation data base. 19. This fixed-share assumption for government spending is not strictly true in the standard closure for version 6.1 of the (lTAIP model because of the nonhomotheticity of private consumption. To make this hold exactly, a preference shift for regional household utility function is introduced such that the shares of private and public consumption and savings in net national income are fixed. 20. CTAP users will recognize that the textile and apparel quota rents are treated as export taxes in the model. However, these rents rarely accrue in full to the government, so they have been omitted from the tax replacement equations. Hertel and others 223 IV. PROTECTION ESTIMATES AND THE PRICE EFFECTS OF MULTILATERAL TRADE LIBERALIZATION The GTAP 5 database, documented in Dimaranan and McDougall (2002), incor- porates relatively recent information on merchandise trade and agricultural protection. Agricultural tariffs are derived from the Agricultural Market Access Database and are for 1998 (Gibson and others 2002). Nonagricultural tariff data are for 1997, or the most recent year, and come from the World Integrated Trading System maintained by the United Nations Conference on Trade and Development and the World Bank. The only nontariff trade barriers in the database relate to export measures. Agricultural export subsidies for 1998, as reported to the wro, are incorporated, as are the quota rents associated with restrictions on textile and apparel exports to North America and Europe.21 The trade liberalization experiment removes the tariffs and quotas. It does not attempt to capture the impact of prospective liberalization of direct trade in services or barriers to international investment or the movement of people in the services sectors. The liberalization experiment also leaves domestic agricultural subsidies in place. Modeling these subsidies requires considerable care given the decoupled nature of many of these programs. These will be tackled in future work. Indonesia's 1998 tariffs in primary agriculture are lower than those of other developing economies and far lower than those for developed economies (table 5). Indonesia's tariffs are higher for processed food imports, particularly beverages and tobacco products, putting it on a par with other developing economies for this combined category of imports. Indonesia's average tariffs on textiles and apparel products are relatively high, as are its tariffs on motor vehicles. This study explores the impact of trade liberalization using both short-run and long-run closures. The short run assumes that wage and salaried laborers are mobile across sectors, but that capital, land, and self-employed labor are immobile. The returns to this group of factors are combined into sectoral "profits" corresponding to the agriculture and nonagriculture profits reported in the household surveys. The long-run closure assumes that self-employed labor is perfectly mobile and perfectly substitutable with wage labor of the same skill category. It also assumes that capital is perfectly mobile across sectors, whereas farm land is partially mobile across uses within the agricultural sector.22 Aggregated price changes for global trade liberalization are reported relative to the model's numeraire, the average price of primary factors worldwide (table 6).23 Consider first the total effects. A rise in primary factor prices in Indonesia (both 21. For ease of comparison in table 5, these have been placed on a cost, insurance, and freight basis and combined with the average import tariffs on textiles. 22. In this long-run closure, the elasticity of transformation of agricultural land across uses is set at -1.0, the default value in the GTAP parameter file. 23. Note that the model generates these price changes for all regions in the model. Although this study explores the implications for Indonesia only, the analysis could easily be extended to the other 13 countries for which survey data have been incorporated into the model. 224 THE WORDI) BANK ECON0OMIC RF\VIFW, VOl.. i 8, N(. ' TABLE 5. Average Tariff Rates for Indonesia, Developing Economies, and Developed Economies Product Indonesia Developing Economies Developed Economies Primary agricultural goods 7 27 13 Rice 6 14 91 Wheat 2 24 64 Feedgrains 4 6() 14 Other agriculture 8 23 11 Oilseeds 7 37 25 Raw sugar n.a. n.a. n.a. Meat and livestock n.a. n.a. n.a. Raw milk 2 25 3 Forestry 1 3 1 Fishing 8 9 3 Fats and oils 10 28 9 Processed food 15 30 20 Meat 17 35 23 Dairy 7 30 25 Rice 0 19 72 Sugar 1 8 38 Other processed food 16 25 18 Beverages and tobacco 86 47 11 Textiles, apparel 16 16 11 Textiles 16 16 7 Apparel 25 15 14 Other manufactures and mining 10 9 2 Automobiles 42 20 2 Electronics 9 5 2 Other manufactures 5 9 2 Wood and paper 6 9 1 M9ining 3 4 0 Petrochemicals 8 10 2 Metals 9 9 2 Note: n.a. is not applicable (nontraded goods). Souirce: Authors' calculations based on CTAP 5 data. short- and long-run) means that Indonesia experiences a real appreciation as a result of this trade liberalization experiment. That is, increased demand for Indonesia's exports bids up its prices relative to the world average. On the com- modity side, Indonesian food prices rise as developed countries reduce agricultural protection and the European Union and the United States curb their exports of subsidized products. This price hike is not offset by the reduction of Indonesia's relatively modest agricultural tariffs. In contrast, the producer prices of durables and nondurables fall substantially in the short run, resulting in a change in relative prices between food and nonfood merchandise. Not surprisingly, the price of services moves closely with wage rates, which rise strongly in Indonesia relative to the rest of the world as a whole. The rise in the price of services means that the (consumer) price changes for margin-inclusive nonfood goods are moderated. TABi,E 6. Short-Run Aggregated Market Price Changes in Indonesia as a Result of Trade Liberalization Short run Indoinesia's Own Liberalization Liberalization by Developed Economies Liberalization bv Developing Economics Long run Agricultural Nonagricultural Agricultural Nonagricultural Agricultural Nonagricultural Item Commodities Comimiodities C.ommodities Commodities Commodities Comniodities Total Item Total Factors Factors Agricultuiral --0.8 -].7 3.9 2.0 0.1 0.0 3.5 Land 5.3 profit Nonagricultural -(.3 -0.1 1.5 3.2 0.1 -0.3 4.1 Capital 5.0 profit UJnskilled labor -0.4 0.( 2.1 3.4 0.1 0.1 5.3 Unskilled wage 6.1 Skilled labor -0.4 0.1 1.6 3.3 0.1 0.1 4.8 Skilled wagc 5.3 Public transfers -0.5 -(0.2 2.0 3.1 0.1 -0.2 4.4 Public transfers 5.7 Private transfers -0.5 -0.2 2.0 3.1 (.1 0.2 4.4 Private transfers 5.7 Commodities Commodities Prodctcer prices Producer prices Staple grains -0.5 0.1 3.4 2.6 0.1 -0.3 5.4 Staple grains 7.1 Livestock -1.4 0.2 2.5 2.3 --0.3 -0.3 3.0 Livestock 5.3 Other food -1.3 0.8 5.0 1.7 0.5 -0.7 6.1 Other food 6.3 Nondurables -0.1 -3.7 0.7 2.8 -0.1 0.7 -1.2 Nondurables 0.7 Durables 0.0 -9.8 0.1 1.3 -0.1 -0.4 -8.8 Durables --9.7 Services -0.2 0.8 1.3 2.4 0.0 (-01 4.2 Services 5.8 Margin services --0.2 0.8 1.3 2.4 0.0 -0.1 4.2 Margin services 5.8 Consuimer prices ConstU??ter prices Staplc grains -0.4 0.2 3.1 2.5 0.1 -0.3 5.2 Staple grains 6.8 Livestock - 1.2 0.3 2.3 2.3 0.1 -0.6 3.2 Livestock 5.4 Other food -1.1 0.8 4.5 1.8 0.2 -0.3 5.9 Other food 6.2 Nondurables -0.1 -3.0 0.8 2.8 0.0 -0.7 -0.3 Nondurables 1.6 Durables -0.1 -2.2 1.0 2.1 0.1 0.2 0.5 Durables t.4 Services -0.2 0.8 1.3 2.4 0.1 -0.1 4.2 Services 5.8 Source: Authors' calculations based on GIAI' S data. 226 THE WORLD BANK ECONOMIC REVIEW, VOL. 1 8, NO. 2 Because the AIDADS demand system employed in the postsimulation analysis is estimated at consumer prices, the vector of consumer price changes (bottom panel of table 6) is pertinent to this evaluation of household welfare. The total short-run price impacts of global trade liberalization are decom- posed into the components attributable to Indonesia's own liberalization, as well as those attributable to liberalization in other developing and developed economies using the method of Harrison and others (1999). For each country group the effects of farm and food liberalization (agriculture) are distinguished from the effects of other merchandise trade liberalization (nonagriculture). The most striking finding of this decomposition is the dominance of developed country trade liberalization in Indonesian earnings impacts (see table 6). Agri- cultural profits are largely driven by developed country agricultural liberaliza- tion, whereas nonagricultural profits are dominated by the elimination of developing economy nonagricultural trade barriers. The effects on earnings are much stronger than those of Indonesia's own liberalization. Furthermore, although trade liberalization in other developing countries has an ambiguous effect on factor returns in Indonesia, all of these Indonesian earnings respond positively to liberalization in developed areas. The decomposition of aggregated commodity price effects of trade liberal- ization shows that developed country trade liberalization accounts for the bulk of the food price increase (see table 6). For durables and nondurables, developed country liberalization results in modest producer price increases, but these are more than offset by the price-depressing effect of Indonesia's own liberalization. For services, both Indonesia's own liberalization and developed country liberal- ization contribute to the price rise. In the long run, with self-employed labor and capital mobile across sectors, their returns increase somewhat more than in the short run, and the increase in agricultural profits is concentrated in land rents (see table 6). Food prices rise more evenly in the long run, with the movement of land, labor and capital encouraging increased product supplies, particularly for staple grains (rice). Nondurable producer prices now rise slightly rather than falling. The producer price of durables falls somewhat more in the long run and the rise in the price of services is somewhat larger when factors are fully mobile. V. POVERTY IMPACTS This section turns to the microsimulation results to examine the impact of trade liberalization on different types of households in Indonesia. It begins with an analysis of the poverty impacts and then explores the impacts across the entire income distribution. Summary Measures of Poverty Sbort-Run Effects. In the short run the number of poor is projected to fall for the nonagriculture-, wage-, transfer-specialized, and diversified strata, with the Hertel and others 227 TABLE 7. Short- and Long-Run Impacts of Global Trade Liberalization on the Poverty Headcount in Indonesia (percentage change) Agriculture Nonagriculture Labor Transfer Specialized Specialized Specialized Specialized Diversified Total Indonesia's own liberalization Agricultural 0.4 -0.6 0.4 -0.3 0.1 0.0 commodities Nonagricultural 3.1 -0.2 -0.4 0.1 1.3 1.6 commodities Developed country liberalization Agricultural 2.4 2.6 1.5 1.2 -0.1 -0.5 commsodities Nonagricultural 0.4 -2.0 -2.4 -1.4 -1.0 -0.7 commodities Developing country liberalization Agricultural -0.1 (I.1 0 0 0 0.0 commodities Nonagricultural -0.8 -0.3 -1.0 -0.4 -0.7 -0.7 commnodities Short-run total 0.7 -0.5 -2.8 -0.9 -0.6 -0.3 change in poverty headcount Long-run total change --1.1 -1.2 -1.6 -0.8 -1.2 -1.2 in poverty headcount Souri-ce: Authors' calculations based on Indonesia 1993 household survey and G;TAP 5 data. sharpest percent change (-2.8 percent) in the labor-specialized stratum (table 7). The short-run poverty headcount among agriculture-specialized households increases as a result of trade liberalization. Agricultural profits rise by less than the increase in prices for staple grains and other food products due to the sharp increase in wages. Services prices also rise at a faster rate. However, the national impact of this slight increase in poverty among agriculture-specialized households is offset by declines in poverty in other strata. Consequently, national poverty falls 0.3 percent in the short run as a result of global trade liberalization. Disaggregation by type of trade reform of the short-run poverty headcount impacts by stratum shows that different types of trade liberalization have different effects on poverty in the various strata (see table 7). Liberalization of Indonesia's own trade policies increases poverty in Indonesia. The commodity price declines following the tariff cuts are insufficient to offset declining incomes as Indonesia experiences the real depreciation required to restore external balance in the wake of increased imports. The largest reductions in national poverty result from developed economy agricultural trade liberalization (-0.5 percent) and nonagricultural trade 228 THE WORLD BANK ECONOMIC REVIEW, VOL. I 8, NO. 2 liberalization (-0.7 percent). Liberalization of developed country trade lowers poverty among the agriculture-specialized and diversified households, which account for nearly 80 percent of the poor in Indonesia. Thus, these reductions in poverty determine the overall outcome, and national poverty falls despite the rise in poverty among other household groups. Liberalization of nonagricultural trade by developed economies lowers poverty in every group except the agri- culture-specialized households. Thus, liberalization of nonagricultural trade by developed countries complements agricultural liberalization. Economywide trade liberalization, which results in broad-based reductions in poverty, is preferable to narrow sector-specific measures, which tend to benefit one group at the expense of others. Liberalization of the agricultural trade policies of other developing economies does not affect the short-run national poverty rate in Indonesia because these trade policies have a negligible impact on prices in Indonesia (table 6). How- ever, liberalization of developing economy nonagriculture trade reduces poverty across all strata in Indonesia because developing economy tariffs on manufac- tures are so high. Long-Run Effects. The most striking thing about the long-run poverty head- count impacts is the greater degree of uniformity across strata (see table 7). Whereas the short-run impacts include both increases and reductions in pov- erty, with great variations in the extent of reductions (from -0.5 percent to -2.8 percent), the long-run impacts are relatively similar across all strata, with poverty rates falling by -1.1 percent and -1.6 percent in non-transfer- dependent households. This happens because once the imputed returns to self- employed labor are accounted for, poor households are essentially reliant on unskilled wages (recall figures 5 and 6). If real unskilled wages rise, then poverty falls. This point is also emphasized in the analysis of Harrison and others (2002a) for Brazil. Impacts across the Income Distribution A more comprehensive analysis of the impacts of alternative trade liberalization measures on household welfare across the income spectrum involves computing the equivalent variation (EV) of the ensuing price and income changes by solving equations 2-5 for the transfer required to give each household the postreform level of utility at prereform prices. This EV is subsequently normalized by initial income to show the percentage gain across the income spectrum. If this curve is rising, it indicates a regressive effect-that trade liberalization benefits the wealthy more than the poor-and if it is falling, it indicates a progressive effect, with greater benefits for the poor than for the rich. The distribution of EV impacts across the income spectrum in Indonesia, under the short-run closure, is plotted for households in each stratum, arranged from poorest to richest, with a line connecting the households in each stratum (figure 8). For example, because the richest agriculture-specialized household is in the 87th percentile, the EV line for this stratum terminates there. The richest Hertel and others 229 FIGURE 8. Equivalent Variation (Relative to Initial Income) due to Global Trade Liberalization, by Stratum 19 J uJ 4 E 09 CL -o 1 10 2,0 3,0 40 So 60 70 80 90 10 -06 Household (percentile of total population) Source: Authors' calculations based on microsimulation model for Indonesian households. households in the survey are in the nonagriculture-specialized stratum. The poverty line, based on the poverty headcount information reported in table 1, is shown as the 15th percentile. There is a clear upward slope for all strata, indicating that the rich benefit more from trade liberalization than do the poor. As anticipated from the poverty results, the poor households in the agriculture-specialized stratum are hurt in the short run by global trade liberalization. Only the wealthiest house- holds in this stratum gain. In contrast, the wage-specialized households benefit across the entire income spectrum as wages rise relative to commodity prices. Members of the other three strata also gain, though more modestly. The EV curves all have the same shape (see figure 8). This shape is inherited from the consumption side of the story. Recall that the consumption parameters in the microsimulation model are independent of the stratum-all households in Indonesia have the same parameters. What distinguishes their consumption bundle (given common prices) is their per capita income and hence their utility level. Furthermore, based on the estimated parameters, household spending at the lowest income levels is dominated by staple grains and other food products, whereas at the highest income levels durable goods and services are more important. At the lowest income levels the consumption component of the relative EV measure ("total" line in figure 9) amounts to about -4 percent of income. The bulk of this is due to higher prices for staple grains and other food products, 230 THE WORLD BANK ECONOMIC REVIEW, VOL. i S, NO. Z FIGURE 9. Consumption Impacts (Relative to Initial Income) of Global Trade Liberalization, by Stratum 05 Nondurables 10 20 30 40 50 60 70 9-0 0- , ' ,"0 .: Staple grains o _1 5 Other food o -2 2 5- 0 -3 E u15 Total 0 -4 - -4 s Household (percentile of total population) Souirce: Authors' calculations based on microsimulation model for Indonesian households. which experience the highest consumer price increases (5.4 percent for grains and 6.1 percent for other food products) and claim a large share of the poorest consumers' budgets. The relative contribution of these food price increases to the relative EV diminishes at higher income levels due to the smaller budget share devoted to these items (recall figure 7). The change in livestock price is more modest, as is its budget share in Indonesia, so its contribution to the consump- tion price EV is smaller. The prices of durables and nondurables change little under trade liberalization. The main nonfood price impact is in the services aggregate, which rises strongly due to the higher wages. The services impact on the consumption component of the relative EV is small at the lowest income levels because of its small share in consumers' budgets. However, the services impact is quite important at the highest income levels, where it dominates the total impact on wealthy consumers. Overall, the shape of the total consumption component of EV takes its cue from food prices, although their effect is offset somewhat by the increase in service prices. For factor earnings' contributions to households' EV the short-run earnings story is nearly invariant across the income spectrum for specialized households. Thus the self-employed nonagriculture-specialized households' EV curve in figure 8 is essentially a parallel upward shift of the agriculture-specialized households' EV line, with the difference owing to the differential change in nonagricultural and agricultural profits. Similarly, the transfer-specialized Hertel and others 231 households represent another incremental shift upward. The case is different for the diversified household EV curve, since these households fare less well than nonagricultural-specialized households at the lowest income levels (see figure 8), but gain much more at the highest income levels-even more than the wealthy transfer-specialized households. This suggests a change in the mix of diversified household income as per capita household earnings rise. This issue is explored further by disaggregating the short-run income side of the relative EV results in figure 8 for the diversified household stratum. At low- income levels agricultural profits contribute the most to the relative change in income (figure 10). However, this contribution declines steadily as income levels rise until it is superceded in importance just after the 50th percentile by the contribution of unskilled wages and nonagricultural profits. At the highest income level this component of the relative change in diversified household income has become nearly the least important one. Overall, the relative income change increases as the income percentile rises. This follows from the interaction between the relative importance of the different sources of earnings with their respective factor prices. As incomes in this stratum rise, earnings shift toward factors that are more favorably affected by trade liberalization. Finally, the change in mix of contributing factors to the wage-specialized households' short-run, relative earnings changes is explored (figure 11). Because these households have a blend of skilled and unskilled labor, it is not surprising FIGURE 1 0. Composition of Earnings Price Effects within the Diversified Stratum 5 Total 4 E 0 3 .E o 2- E Transfers 0 - 10 20 30 40 50 60 70 D0 90 100 Depreciation Household (percentile of stratum population) Souirce: Authors' calculations based on microsimulation model for Indonesian households. 232 THE WORLID BANK FCONONMIC REVIEW, VOL. I 8, NO. 2 FIGURE 11. Composition of the Earnings Price Effects within the Wage-Specialized Stratum 6 Total 5 0 S 0 1 10 20 30 40 50 60 70 80 90 100 -1 Household (percentile of stratum population) Source: Authors' calculations based on microsimulation model for Indonesian households. to see that the relative importance of skilled labor rises with incomes. Because unskilled wages rise slightly more than skilled wages, the total earnings curve slopes slightly downward. This, in turn, explains why the total EV curve for the wage-specialized households in figure 8 loses some of its dominance over the transfer- and nonagricultural-specialized household strata as incomes rise. IV. CONCLUSIONS Assessing the impact of multilateral trade liberalization on poverty is challen- ging. As Winters (2000, p. 43) notes: "Tracing the links between trade and poverty is going to be a detailed and frustrating task, for much of what one wishes to know is just unknown. It will also become obvious that most of the links are very case specific." Winters lays out a general framework for thinking about the impact of trade policy on poverty. This study is similar in spirit to Winters's effort. It recognizes that the definitive assessment of the impact of trade liberalization on poverty must be done on a case-by-case basis. It also recognizes the need for a set of internationally comparable estimates of the global impact on poverty in a range of different countries and develops a tractable methodology for providing this. The approach relies on detailed earn- ings data from household surveys; an econometrically estimated demand system Hertel and others 233 that reflects the changes in consumption patterns across the income spectrum and that provides a natural vehicle for analysis of household welfare and poverty, particularly in the context of multicountry analyses; and a micro- macro consistent framework for projecting the price impacts of global trade liberalization. The approach is applied to an assessment of the consequences of global liberalization of merchandise tariffs, agricultural export subsidies, and quotas on textiles and clothing. To fully develop the analysis, the study focuses on the consequences for Indonesia. The approach could readily be applied to any of the other 13 countries for which microconsistent databases have been assembled. For Indonesia the national headcount measure of poverty is reduced following global trade liberalization in both the short and the long run. However, the aggregate reduction in Indonesia's national poverty headcount masks a more complex set of impacts across different strata. In the short run the poverty headcount rises slightly for self-employed, agriculture-specialized households, as the rise in farm profits is outpaced by consumer prices. There- fore, this study echoes Kanbur's (2000) call for disaggregated analyses of poverty impacts. In the long run, however, the poverty headcount in Indonesia falls for all strata, because the increased demand for unskilled workers lifts incomes for the formerly self-employed, some of whom move into the wage labor market. The study also decomposes the change in headcount poverty in Indonesia associated with different countries' trade policies. Liberalization of other coun- tries' trade policies leads to a reduction in the national poverty headcount in Indonesia. Liberalization of Indonesia's own trade policies-particularly those protecting the nonagricultural sectors-leads to an increase in national poverty. In summary, the framework developed here fills an important gap in researchers' toolkits for analysis of the poverty impacts of multilateral trade liberalization. By stratifying households according to earnings specialization, it captures a great deal of the diversity relevant to trade policy impacts while preserving analytical tractability and comparability across countries. Another important contribution of this work is to show how consumer spending across the income spectrum can be characterized using a single, econometrically estimated demand system. Once calibrated to match observed national spending patterns, this demand system yields a unique poverty level of utility that provides an ideal benchmark for evaluating changes in poverty rates using anv member of the class of poverty measures proposed by Foster and others (1984). Although this approach is illustrated by an analysis of the poverty impacts of trade liberalization in Indonesia, it can be applied to other countries. This approach should enrich traditional analyses of multiregion trade liberal- ization, making them more relevant for policymakers who are increasingly concerned about the consequences of such actions for poverty in developing economies. 234 THE WORLD BANK ECONONIIC REVIEW, VOL. i 8, NO. 2. REFERENCES Adelman, Irma, and Sherman Robinson. 1978. Income Distribution Policy: A Computable General Equilibrium Model of South Korea. Stanford: Stanford University Press. Atkinson, A. B., L. Rainwater, and T. M. Smeeding. 1995. Income Distribution in OFCD Countries. Paris: OECD. Biro Pusat Statistik. 1993. "SUSENAS: Indonesia's Socio-Economic Survey." Jakarta, Indonesia. Bourguignon, Francois, and Pierre Andre Chiappori. 1994. "Income and Outcomes: A Structural Model of Intra-Household Allocation." In The Measurement of Household Welfare, ed. R. Blundell, 1. Preston, and I. Walker. Cambridge: Cambridge University Press. Bourguignon, Fran,ois, Anne-Sophie Robilliard, and Sherman Robinson. 2002. "Representative vs. Real Households in the Macro-economic Modeling of Inequality." Paper prepared for the Conference on Frontiers in Applied General Equilibrium Modeling, Yale University, April 5-6, New Haven, Conn. Case, Anne. 1998. "Income Distribution and Expenditure Patterns in South Africa." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, 2000, Stockholm. Chen, Shaohua, and Martin Ravallion. 2003. "Welfare Impacts of China's Accession to the World Trade Organization." World Bank Economic Review 18(1):29-57. Chen, S., and Y. Wang. 2001. "China's Growth and Poverty Reduction: Trends between 1990 and 1999." Policy Research Working Paper 2651. World Bank, Washington D.C. Cline, W. 2004. Trade Policy and Global Poverty. Washington, D.C.: Institute for International Economics. Cogneau, Denis, and Anne-Sophie Robillard. 2000. "Growth, Distribution and Poverty in Madagascar: Learning from a Microsimulation Model in a General Equilibrium Framework." Discussion Paper 61. International Food Policy Research Institute, Trade and Macroeconomics Division, Washington, D.C. Cranfield, J. A. L. 1999. "Aggregating Non-Linear Consumer Demands-A Maximum Entropy Approach." Ph.D. diss., Purdue University, Department of Agricultural Economics, West Lafayette, Ind. Cranfield, John A. L., Thomas W. Hertel, and Paul V. Preckel. 2000. "Multi-lateral Trade Liberalization and Poverty." Paper prepared for the Conference on Poverty and the International Economy organ- ized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. Cranfield, J. A. L., J. S. Eales, T. W. Hertel, and P. V. Preckel. 2003. "Model Selection When Estimating and Predicting Consumer Demands Using International, Cross Section Data." Empirical Economics 28(2):353-64. Cranfield, J. A. L., P. V. Preckel, J. S. Eales, and T. W. Hertel. 2002. "Estimating Consumer Demand across the Development Spectrum: Maximum Likelihood Estimates of an Implicit Direct Additivity Model." Journal of Development Economics 68(2):289-307. - 2004. "Simultaneous Estimation of an Implicit Directly Additive Demand System and the Distribution of Expenditure-An Application of Maximum Entropy." Economic Modelling 21: 361-85. Decaluwe, B., A. Patry, L. Savard, and E. Thorbecke. 1999. "Poverty Analysis within a General Equilibrium Framework." Working Paper 99-06. Universite Laval, Department d'economique, Centre de Recherche en Economie et Finance Appliquees, Quebec, Canada. Deininger, K., and L. Squire. 1996. "A New Data Set Measuring Income Inequality." World Bank Economic Review 10(3):565-91. Devarajan, Shantayanan, and Dominque van der Mensbrugghe. 2000. "Trade Reform in South Africa: Impacts on Households." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. Hertel and others 235 Dimaranan, Betina V., and Robert A. McDougall. 2002. Global Trade, Assistance, and Protectiont: The GTAP 5 Data Base. West Lafayette, Ind.: Purdue University, Department of Agricultural Economics, Center for Global Trade Analysis. Evans, David. 2001. "Identifying Winners and Losers in Southern Africa from Global Trade Policy Reform: Integrating Findings from (TAP and Poverty Case Studies for a Zambian Example." Paper prepared for the Economics and Social Research Council Development Economics/International Economics Conference, Nottingham University, April 5-7, Nottingham, UK. Foster, J. E., J. Greer, and E. Thorbecke. 1984. "A Class of Decomposable Poverty NMeasures." Econometrica 52(3):761-66. Friedman, Jed. 2001. "Differential Impacts of Trade Liberalization on Indonesia's Poor and Non-Poor." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. Gibson, P., J. Wainio, and D. Whitley. 2002. "Agricultural Tariff Data." In Globlal Trade, Assistance, and Protection: The GTAP 5 Data Base, ed. Betina V. Dimaranan and Robert A. McDougall. West Lafayette, Ind.: Purdue University, Department of Agricultural Economics, Center for Global Trade Analysis. Haddad, L., and R. Kanbur. 1990. "How Serious is the Neglect of Intra-household Inequality?" Econonzic Journal 100(402):866-81. Harrison, Glen W., Thomas F. Rutherford, and David G. 'ITarr. 2002a. il,..' .1 Multilateral and Unilateral Trade Policies OF IRFlOSUR for Growth and Poverty Reduction in Brazil." Working Paper 3051. World Bank, Washington, D.C. - 2002b. "Trade Liberalization, Poverty and Efficient Equity." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliameiltary Commission on Global Development, October 20-21, 2000, Stockholm. Also published in 2003 Journal of Development Economics 71:97-128. Harrison, W. Jill, J. Mark Horridge, and K. R. Pearson. 1999. "Decomposing Simulation Results with Respect to Exogenous Shocks." Paper presented at the Second Annual Conference on Global Economic Analysis, June 20-22, Copenhagen. Hertel, Thomas W., ed. 1997. Global Trade Analysis: Modeling and Applications. New York: Cambridge University Press. Hertel, Thomas W., Paul V. Preckel, and John A. Cranfield, 2000. "Multilateral Trade Liberalization and Poverty Reduction." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. lanchovichina, Elena, Alessandro Nicita, and Isidro Soloaga. 2000. "Implications of Trade Reform on the Distribution of Income in Mexico." Paper prepared for the Conference on Poverty and the Inter- national Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. -_____ 2002. "Trade Reform and Poverty: The Case of Mexico." World Economy 25(7):945-73. Ivanic, Maros. 2003. "Reconciliation of the GTAP and Hiousehold Survey Data on Factor Earnings." GTAI' Research Memorandum #5. Purdue University, Department of Agricultural Economics, Center for Global Trade Analysis, West Lafayette, Ind. Available online at www.gtap.agecon.purdue.edu/ resources/res-display.asp?recordidD=1408. Kanbur, R. 2000. "Income Distribution and Development." In Handbook of lncoine Distribution, ed. A. B. Atkinson and F. Bourguignon. Amsterdam: Elsevier. Kravis, I. B., A. W. Heston, and R. Summers. 1982. World Product and Incomne: Inteniational Coinpar- isons of Real Gross Prodiuct. Baltimore: Johns Hopkins University Press. Levinsohn, James, Steven Berry, and Jed Friedmani. 1999. "Impacts of the Indonesian Economic Crisis: Price Changes and the Poor." Paper prepared for the Conference on Poverty and the International Economy organized by World Bank and Swedish Parliamentary Commission on Global Development, October 20-21, Stockholm. 236 THE WORLD BANK ECONONIIC REVIEW, VOL. I 8, NO. z Lipton, M., and M. Ravallion. 1995. "Poverty and Policy." In Handbook of Development Economics, ed. J. Behrman and T. N. Srinivasan. Amsterdam: Elsevier. Lofgren, Hans. 1999. "Trade Reform and the Poor in Morocco: A Rural-Urban General Equilibrium Analysis of Reduced Protection." Discussion Paper 38. International Food Policy Research Institute, Trade and Macroeconomics Division, Washington, D.C. McCulloch, N., L. A. Winters, and X. Cirera. 2001. Trade Liberalization and Poverty: A Handbook. London: Center for Economic Policy Research. Mistiaen, J., and M. Ravallion. 2003. "Survey Compliance and Distribution of Income." Policy Research Working Paper 2956. World Bank, Washington, D.C. Reimer, Jeffrey J. "Estimating the Poverty Impacts of Trade Liberalization." Policy Research Working Paper 2790. World Bank, Washington, D.C. Rimmer, M. T., and A. A. Powell. 1992a. "Demand Patterns across the Development Spectrum: Estimates of AIDADS." Working Paper OP-75. Monash University, Centre of Policy Studies, Clayton, Victoria, Australia. -_____. 1992b. "An Implicitly Directly Additive Demand System: Estimates for Australia.' Working Paper OP-73. Monash University, Centre of Policy Studies, Clayton, Victoria, Australia. -____. 1996. "An Implicitly Additive Demand System." Applied Economics 28(12):1613-22. Winters, L. Alan. 2000. "Trade, Trade Policy, and Poverty: What Are the Links?" Discussion Paper 2382. Centre for Economic Policy Research, London. Winters, L. A., N. McCulloch, and A. McKay. 2003. "Trade Liberalisation and Poverty: The Empirical Evidence." Discussion Paper 88. University of Sussex, Institute of Development Studies, Falmer, Brighton, UK. World Bank. 1992. World Development Report 1992: Development and the Environment. New York: Oxford University Press. -____ 1993. World Development Report 1993: Investing in Health. New York: Oxford University Press. - 2000. World Development Report 2000/2001: Attacking Poverty. Oxford: Oxford University Press. Asymmetries in the Union Wage Premium in Ghana Niels-Hugo Blunch and Dorte Verner There is little evidence on the size of the union wage premium in developing economies. The article uses a matched employer-employee data set for Ghana and adopts a quantile regression approach that allows the effects of unionization to vary across the conditional wage distribution. It is shown that if there are intrafirm differences in unionization, there does appear to be a premium among poorer paid workers in the formal sector. Although this cannot be given a causal interpretation, it suggests import- ant issues about how unions may affect one part of the labor market. I did not begin my research on Ghanaian trade unions with the idea that organized labor could rescue this small, developing West African state from its economic difficulties. But neither did I have the pre-conceived notion, today popular in some quarters, that unions are irrelevant to the struggles of Third World peoples to maximize their own freedom. -Paul S. Gray (1981) Experience in developed economies suggests that unions may be a mechanism for providing a positive work environment by reducing labor turnover and wage nego- tiation costs and promoting worker training, increased benefits, and higher produc- tivity (Standing 1992). Unions have also been found to reduce wage inequality and wage discrimination (Chaykowski and Slotsve 2002; Freeman 1980; Panagides and Patrinos 1994; Standing 1992). Much less is known about unions in developing economies, particularly in Sub-Saharan African countries.1 Such analysis seems especially warranted for these countries as a basis for policy proposals. Formal sector jobs are scarce and wages are generally low, leading both to poverty and-because of low tax revenues-low levels of goods and services from the public sector. To shed some light on these issues in developing economies, this article analyzes wage determinants in Ghanaian manufacturing industries, focusing on Niels-Hugo Blunch is a consultant in the Social Protection Unit of the Human Development Network at the World Bank and a Ph.D. candidate at The George Washington University; his e-mail address is nblunch@worldbank.org. Dorte Verner is a senior country economist in the Social Development Family of the Latin America and Caribbean Region at the World Bank; her e-mail address is dverner@world- bank.org. The authors thank Francois Bourguignon, Sudharshan Canagarajah, Donald Parsons, and participants at a seminar at the Centre for Labour Market and Social Research, Aarhus, Denmark, for helpful comments and suggestions. Comments and suggestions from three anonymous referees and a journal review board member helped greatly improve this article. 1. See, however, Kristensen and Verner (1999), Rama (2000), and Schultz and Mwabu (1998). THE WORLD BANK ECONOMIC RFVIEW, VOL. 18. NO 2, O The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi:10. 1093/wber/lhhO40 18:237-252 237 238 THF WORLD BANK ECONOMIC REVIEW, VOI . i 8, NO. Z the association of union membership and wages and the possible asymmetries in this association across the conditional wage distribution. Ghana is an ideal candidate for such analysis because of its long history of active labor unions.2 1. HYPOTHESES A union premium could arise through three possible channels. First is the direct effect through individual union membership. This is the standard union premium, well known from the empirical literature on the union relative wage effect begin- ning with Lewis (1963). Second, because this does not take into account possible spillovers to nonunion workers, the analysis here allows unionism to potentially affect all workers through an industry-level union density variable. Third, there might be an additional union effect arising from training (Booth and Chatterji 1998; Booth and others 1999), because unions may promote training more than management does. For example, unions can have a longer time horizon than management, which may focus on maximizing profits and stock values in the short term, or unions may counterbalance the firm's ex post monopsonistic power in wage determination. Thus where unions are active there may be increased recognition of the value of training, so that trained workers receive a wage premium.4 (Booth and Chatterji 1998 and Booth and others 1999 develop theoretical models in which a key prediction is that union workers receive more training and higher returns to training than do nonunion workers.) The hypothesis here is that union effects are more likely at the lower part of the wage distribution, because unions are generally seen as (and generally perceive themselves as) proponents of workers' rights and wages for the poorer or less skilled segment of the workforce. This view is also in line with previous research, which generally finds that unions reduce wage inequality and wage discrimination. Chamberlain (1994) finds distinct asymmetries in the union premia in U.S. manu- facturing, which is 28 percent at the bottom decile but declines monotonically to 0.3 percent at the top decile. This contrasts with an ordinary least squares (OLS) estimate of the mean union premium of 15.8 percent, driven primarily by the bottom decile of the conditional distribution (for Ghana, the result is still more 2. Unionism is low or even absent in many countries of Sub-Saharan Africa because of the economic dominance of smallholder agricultuLre and the historically prohibition against independent trade unions in many countries. Ghana, however, has a long tradition of labor unions, originating with the many guilds and artisan associations in the early nineteenth century (Gray 1981). The potential synergies between labor organizations and government were anticipated early, even before Ghana's independence from British rule in 1957: "As early as 1930, Lord Passfield (Sidney Webb) had noted in a dispatch that regu- lation of wage laborer organizations was of importance, and that colonial governments should act to facil- itate the passage of unions into constitutional channels" (Gray 1981, p. 14). 3. To avoid a causal interpretation, we use the term onionn Twage premiuim rather than onion relative wage effect, which has previously been widely used in the literature but unfortunately might suggest a causal relationship. 4. Alternatively, unions may enable trained workers to extract the rents created by workers. Bltinch aned Verner 239 striking, with the mean union premium disappearing altogether). More recently, Card (1996) finds similar results for the United States and Chaykowski and Slotsve (2002) for Canada. Thus, although there is some evidence of asymmetry in the union wage premium for developed countries, such asymmetry may be particularly relevant for developing economies, where the high costs of monitoring may prevent effective monitoring of adherence to minimum wage legislation. Unions in developed areas may bargain more on behalf of workers from the middle part of the wage distribution because minimum wage legislation is generally adhered to. Thus for developing and developed economies alike, the upper part of the wage distribution would not be expected to exhibit a positive and statistically significant association between wages and individual union membership. Analyzing the impact of unions across the entire wage distribution is thus important for at least two reasons (Chaykowski and Slotsve 2002). First, it may shed light on whether unions have an impact on particular socioeconomic groups in the labor market, enriching the social and economic view of unions, including whether unionism should be reinforced or retrenched. Second, such analyses provide insight into which workers benefit most from unions and, therefore, help identify which groups unions may most successfully appeal. Before getting on with the analysis, some words of caution about interpreting the results are in order. First, because matched employer-employee panel data sets for Sub-Saharan Africa are rare, the empirical analysis is limited to cross- section data-as is the case for most wage equation studies for Sub-Saharan Africa. That means that individual effects, which might otherwise overcome the possible selection of workers into unions (on unobservables such as type or ability; see Card 1996 for a study for the United States using panel data), cannot be isolated. Second, to account for selection into unions using cross-section data requires an exclusion restriction (or instrument). This is also not feasible for the present analysis because there are no readily available variables that affect selection into unions but do not affect wages directly. Controlling for firm size should somewhat mitigate the biases arising from the nonrandom place- ment of unions (at the aggregate level) into sectors or industries,5 but the possibility of selection of workers into unions at the individual level based on unobservables (ability, motivation) remains. The results should therefore be seen as suggestive rather than as explicitly causal.6 5. Previous studies of the association of union membership and wages have not always been able to control for firm size. Schultz and Mwabu (1998), for example, analyzing the association of individual union membership and wages and unemployment in South Africa using a quantile r-egression framework, note the importance of this variable with respect to the nonrandom placement of unions but are unable to include the variable in their analyses because it is not part of their data set. 6. It may be that unions are simply more active in higher-wage sectors. Again, because high-wage sectors are also typically found to be those with larger firms, controlling for firn size should at least mitigate some of this bias. 240 THE WORLD BANK ECONOMIC IRVIFE, VOL i S, No. 2 II. DATA AND DESCRIPTIVE ANALYSIS The data are from the Regional Program on Enterprise Development survey for Ghana, organized by the World Bank and funded by the British Overseas Development Administration and conducted by the Centre for the Study of African Economies at the University of Oxford and the University of Ghana at Legon in 1994.' Although the survey yields information on firms and workers in 180 manufacturing firms, missing observations for one or more variables resulted in an effective estimation sample of 683 workers in 108 firms. The main variables applied in this study include a "core" of (log) monthly wages (including allowances); the standard human capital variables of age and age squared (to capture potential general experience), tenure in the firm and tenure squared (to capture potential specific experience), highest level of education completed, training variables (whether a worker has received training within the firm), and occupational control variables and firm-level control variables, most notably firm size by number of employees (the appendix lists the variables and their definitions).8 The possible existence of a union wage premium may be determined in various ways. The first thing to consider is the channel through which unions may affect wages: Is there an individual, direct effect through union membership, or is there an indirect effect through the degree of unionization at an establishment or in an industry as a reflection of the bargaining power of a union within a firm or industry, thus allowing for spillovers to nonunion workers? This is examined as an empirical question, through a dummy variable for individual union membership and an industry-level union density variable. The aggregate union wage premium (spillover mechanism) is modeled using the degree of unionization at the industry rather than at the firm level largely because collective bargaining in Ghana occurs mostly at the more aggregated industry level (Gray 1981). This approach further ensures that collinearity is not likely to be a serious problem, because these two union variables are only weakly correlated (a simple correlation of 0.08), whereas the individual union membership and firm union density variables have a simple correlation of 0.76, so including both of these in the same regression could result in substantial collinearity problems.9 7. Surveys were conducted in 1992, 1993, and 1994. Although the survevs have a panel structure for collecting firm-level data, the data on workers were collected as independent cross-sections. The analysis therefore includes only the most recent of the three surveys, treating the data as cross-sectional for both firms and workers. 8. Larger establishments tend to pay higher wages than smaller establishments,!. .,;r...1,, for other factors (Schaffner 1998; Velenchik 1997). Not controlling for- this, therefore, could lead to substantial omitted-variable bias. 9. Maloney and Ribeiro (1999) include the union density variable at the firm level, arguing that this is a proxy for the bargaining power of the union over the firm's rents. However, the high correlation between this variable and the dummy variable for individual union membership in the Ghana sample means that inclusion of both would likely yield problems with multicollinearity, and so it does not appear valid to simultaneously allow the two different channels of a union relative effect (individual and aggre- gate spillover) proposed here. Bliin,cb and Verner 241 There may still be an endogeneity issue related to the use of the variable of individual union membership, however. What is needed is a good instrument that may help explain individual union membership without also explaining wages. Because the data set lacks such a variable, the analysis follows Schultz and Mwabu (1998) in arguing that it is beyond the scope of the data to endogenize union membership, including explaining who gets a union job and who does not, and to explain the extent to which unions enhance the productivity of workers with the same observable characteristics. As a result, any estimated union wage premium may overstate or understate the true union wage premium. Because the estimated coefficient picks up the effect from the omitted variable as well, the omission of any variable influencing wage determination that is positively correlated with the union variable will cause the estimated coefficient on the union variable to be upwardly biased, whereas the omission of a nega- tively correlated variable will cause the estimated coefficient to be downwardly biased. For example, it may be conjectured that nonunion jobs are typically found in smaller establishments or industries because it takes a certain size for an establishment or industry to become interesting for a union in terms of potential members. Hence, failure to control for firm size may cause bias in the union premium. However, the present data set permits incorporating firm size as an explanatory variable, which likely decreases such bias considerably. Furthermore, it may be conjectured that firms that have become unionized respond to unionization by carefullv vetting prospective new workers so as to employ even higher quality workers than before because of increased wage demands. Because "quality" or "ability" is unmeasurable, omitted variable bias is again possible in the union wage premium estimates. The same applies to the labor turnover argument: If unions reduce labor turnover, and this effect cannot be directly observed and included as an explanatory variable, the result is once again omitted variable bias in the union premium estimate. With a panel data set one solution is to extract the individual fixed effect of the data, thus mitigating the potential bias (see, for example, Card 1996). But with a cross-section data set, as in this case, unionism has to be modeled based on individual union membership (and possibly a variable of union densitv at the firm or industry level, constructed from this). Even though the data do not allow for fully endogenizing union membership, it is possible to shed some light on whether the endogeneity of individual union membership will pose severe problems for the subsequent analyses by doing some descriptive work. Data on individual and firm characteristics across union membership reveal that union workers earn higher wages than nonunion work- ers, not controlling for other characteristics (table 1). However, union workers (and their firms) are different from nonunion workers (and their firms) in several ways (although not always statistically significantly so). They are better educated, older, and more experienced (within the firm). They are also more likely to have a permanent contract. At 127 employees, the average firm size of unionized workers is much larger than that of nonunionized workers, at 242 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. Z TABLE 1. Worker and Firm Characteristics across Individual Union Membership Nonunion worker Union member Variable Mean SD Mean SD Difference t-Statistic Log(wages) 10.522 0.077 10.752 0.054 0.230-- 2.65 Age 34.1 0.870 36.1 0.949 2.0 1.58 Gender 0.204 0.032 0.153 0.031 -0.050 1.21 None 0.094 0.021 0.094 0.030 0.001 0.01 Primary 0.037 0.012 0.015 0.008 -0.023 1.53 Middle 0.509 0.032 0.446 0.045 -0.064 1.15 Secondary 0.123 0.024 0.124 0.030 0.001 0.03 Vocational 0.102 0.017 0.183 0.032 0.081- 2.29 Polytechnic 0.073 0.016 0.124 0.033 0.051 1.44 Professional 0.042 0.012 0.015 0.008 -0.027* 1.98 University 0.021 0.007 0.000 0.000 -0.021*r 2.99 Production 0.443 0.040 0.376 0.039 -0.067 1.22 Administration 0.104 0.024 0.193 0.030 0.089* 2.44 Commercial 0.062 0.014 0.079 0.020 0.017 0.70 Professional 0.046 0.012 0.025 0.015 -0.021 1.06 Support staff 0.098 0.030 0.178 0.036 0.081 1.86 Mlanager 0.127 0.017 0.099 0.026 -0.028 0.93 Experience (years) 5.9 0.562 8.3 0.926 2.4;` 2.27 Permanent contract 0.969 0.012 1.000 0.000 0.031- 2.50 Received training in firm 0.320 0.045 0.282 0.049 -0.038 0.59 Accra 0.611 0.056 0.629 0.088 0.017 0.19 Number of employees in firm 64.8 12.072 126.7 27.601 61.8*' 2.21 *Significant at the 10 percent level. -Significant at the 5 percent level. "'"Significant at the I percent level. Note: Number of observations in the full sample = 683. The clustering of workers within firms is taken into account in the calculation of SDs and t-statistics for Hn: Difference = 0. Source: Regional Program on Enterprise Development for Ghana (Wave III 1994). 65 employees. Firms of unionized workers are also more likely to be located in Accra, the capital. That firm size is an important correlate of individual union membership is reassuring, because including this variable in subsequent analyses is likely to take care of many of the concerns with nonrandom placement of unions, at least at the aggregate level. However, selection is still possible and even probable at the individual level, where workers may select into unions based on preferences or latent personal characteristics or be chosen by employees based on latent personal characteristics, all of which are unobserved by the researcher. Again, although the data prevent a detailed analysis of individual determinants of union membership, the finding from the previous descriptive analysis that union workers (and their firms) are quite different from nonunion workers (and their firms) on observable characteristics suggests the likelihood of differences BluncVh and Verner 243 on unobservable characteristics such as ability and motivation as well. Yet additional calculations reveal that union membership is fairly evenly distributed across occupations and industries. For example, a production worker is only slightly more likely to be member of a union than a manager (26.3 and 24.7 percent). 10 Because 98 percent of the sample has a permanent contract (table A-1), including that as an explanatory variable may appear fruitless because of the low variation. However, there is the possibility of an indirect effect from training. Employers would be more inclined to invest in training for employ- ees who are likely to stay with the firm for some time, so the quantity or quality of training may be different for workers with a permanent contract. The variable for contract status is thus interacted with training. Because only about 30 percent of the sample both has a permanent contract and received training, this is likely to have some explanatory power because of its higher variation. In sum, because the data are unsuitable for making explicit causal inferences about the effect of unions on wages, the analysis explores only the association between union membership and wages, keeping in mind the differing character- istics of union workers (and firms) and nonunion workers (and firms) and the possible nonrandom placement of unions into high-wage sectors. III. MFTHODOLOGY The theoretical framework for the analysis is standard human capital theory. An individual builds up knowledge and skills through education and experience (Becker 1964; Mincer 1974). Formally, the economic model is derived from the theory of individual demand for schooling, which views education as an invest- ment in human capital (Becker 1964). In the traditional human capital litera- ture, wages are determined by education and other individual characteristics. Because the Ghanaian data set allows inclusion of union- and firm-level vari- ables, the standard Mincerian wage function is augmented with union- and firm-level characteristics (I) Wi = W(I,, F, Uj), where W, the wages of individual i, is the dependent variable; I is a vector of individual characteristics (such as age and age squared proxying general experi- ence; tenure in the firm, capturing firm-specific experience; the level of educa- tion; and gender); F is a vector of characteristics for the firm of individual i, including the size of the firm (measured by the number of employees) and geographical location; and U is a vector of variables capturing the possible 10. The industry-level union densities are 19.8 for wood, 25.7 for food, 32.1 for textiles, and 33.1 for metal. 244 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 union wage premium for individual i, as measured by individual union member- ship or the union density of the industry or firm (as discussed in the previous section). Lewis (1963) defines the union-nonunion wage differential (referred to here as the union wage premium) asI (2) r, = (Wx - W I W. The possible existence and magnitude of the wage differential depend on the extent to which the union affects the wages of union workers relative to the wages of nonunion workers. At one extreme the union may, through its bar- gaining power, merely extract and subsequently share the existing rents of the firm (in the form of profits) with its members. At the other extreme, the union may generate rents through its potential adverse effects on labor turnover and the costs of wage negotiations between management and workers. In reality, a combination of these two effects would be expected. The theory is silent on the precise empirical implementation of this notion of a union wage premium, which is left to individual researchers. This study uses quantile regression analysis. It enables simultaneously estimating the marginal effects for different quantiles (where the quantile of interest may be chosen arbitrarily) of the dependent variable, thus exploring the entire conditional distribution. The main advantage here is the semi-parametric nature of the approach, which relaxes the restrictions on the parameters being constant across the entire conditional distribution of the dependent variable. A priori, the union wage premium might be expected to differ across the conditional wage distribution. For example, unions may be bargaining mainly on behalf of workers at the lower end of the wage distribution. As noted, this is especially likely in devel- oping economies, where minimum wage legislation may not always be strictly adhered to because of the high costs of monitoring. Furthermore, it seems likely that the returns to education and tenure, for example, would differ across the wage distribution. For example, education might be thought to be a more important determinant at higher quantiles than at lower quantiles.12 The method has other virtues, as well. Allowing the parameter estimates for the marginal effects of the explanatory variables to differ across quantiles of the dependent variable achieves greater robustness to potential heteroscedasticity. This contrasts with OLS regression analysis, which requires homoscedasticity 11. The union-nonunion wage differential or union wage premium-which is the focus here-is only one aspect of the economics of unions. For an excellent recent survey of the economics of unions in a broader context, see Booth (1995). See also Lewis (1963, 1986), Freeman and Medoff (1984), and Pencavel (1995). 12. This turns out to be the case. An extended version of this article with the full set of results is available online at www.niels-hugo.dk. Bluncb and Verner 245 (indeed, in most of the empirical literature, the presence of homoscedasticity is merely a maintained hypothesis). Additionally, when the error terms are non- normal, for instance, quantile regression estimators may be more efficient than least squares estimators. Furthermore, because the quantile regression objective function is a weighted sum of absolute deviations, it yields a robust measure of location. As a consequence, the estimated coefficient vector is not as sensitive to extreme observations of the dependent variable. The method, developed by Koenker and Basset (1978), can be formulated as (3) Yi = XiB00 + uoi = Quante(Y,IXi) = X, 0O where Quante(YilXi) denotes the Oth conditional quantile of Y given X for individual i. In general, the Oth sample quantile (O < 0 < 1) of Y solves (4) min!=n{ E 0|Yi-X il + E (1-0)|Y,-X|} 1: Y,>X' i:y,Significant at the 5 percent level. *'-Significant at the 1 percent level. Note: Number of observations = 683. R2 is pseudo-R2 for quantile regressions and standard R2 for OL.s regressions. The SEs are bootstrapped using 200 replications, running a clustered bootstrap (where the clusters are the firms) with the quantile regression run within each draw. Source: Authors' calculations based on data from Regional Program on Enterprise Development for Ghana (Wave III 1994). worker is trained, the degree of unionization in the industry affects individual wages positively, hinting at unions' bargaining power in extracting some of the rents from training from the firm and sharing it with members.17 This effect 17. For unions to bring about higher wages, there must first be some surplus arising from the employment relationship, and the union must have sufficient bargaining power to induce the firm to share this surplus in the form of higher wages. It should be noted that unions can also sometimes increase or decrease this surplus. 248 IT-IF WORI1) BANK lCONOMI( RFVIFW', VOLI . i 8, NO 2 persists at the higher end of the wage distribution as well. The union wage premium associated with individual union membership decreases slightly when the industry union density variables are included, hinting at their impor- tance in modeling the union wage premium in Ghana. The Olis results are strikingly different from the quantile regression results. The OLiS estimates indicate the presence of only an indirect aggregate union premiurn and no direct premium from individual union membership for all three main models (see table 2). These findings substantiate the importance of the quantile regression approach as an alternative or a complement to the more traditional ol s-based analysis. Our findings correspond to those of Chamberlain (1994) but are even more pronounced. Chamberlain reports a mean union premium (ois estimate) of 15.8 percent for U.S. manufacturing firms, with a 28.0 percent premium at the bottom decile of the wage distribution and a 0.3 percent premium at the top decile, implying that the mean union wage premium is driven primarily by the bottom decile of the conditional distribution. In the OLS regression here, the union premium vanishes altogether. The quantile regression reveals that this happ;ens because only the bottom decile of the conditional distribution receives a union wage premium. Schultz and Mwabu (1 998) also find distinct asymmetries across the conditional distribution of South African wages. Again, however, the results for Ghana are even more pronounced. In addition to the union wage premium for the bottom decile, Schultz and Mwabu find a statistically significant union wage premium at the top decile of the conditional distribution (1 1 percent for Africans and -24 percent for whites), as well as a statistically significant and positive union premium of 60 percent for the oi,s regressions (for Africans only), which this study does not. The data in the Shultz and Mwabu study, however, do not allow controlling for firm size, which may be an important determinant of union placement. Excluding firm size from the core model of this study (model 1 a in table 2) yields a statistically significant union wage premium at the median and increases the individual union premium to almost 56 percent (explO.4431 - l). Doing so also yields a statistically significant union premium for the OLS regression. Although this may help explain the discrepancies between the Schultz and Mwabu results and the present results, that is not to say that the two sets of results would necessarily have been similar had Schultz and Mwabu had firm-size data in their South Africa dataset. Rather, the main point is that the union premium may be substantial even after controlling for firm size, as our results indicate. Even if firm size is included as an explanatory variable, concerns remain about the possible selection of workers into unions at the individual level based on unobservables (ability, motivation). Because union and nonunion workers have different observable characteristics (see table 1), they might well have different unobservable characteristics (ability, motivation) as well. Again, that implies that the results here are more suggestive than explicitly causal. Bl/}znc?, alndS IVernter 249 V. CONC LLJSION This study establishes important associations between wages and individual union membership and uniOln density at the industry level of manufacturing workers in Ghana. Findings of distinct asymmetries in the association of union membership and wages across the conditional wage distribution of workers in Ghanaian manufacturing are consistent with the expectation that unions bar- gain mainly on behalf of workers at the lower end of the wage distribution. The analysis finds evidence of a union premium related to individual union member- ship and an additional individual union premium coming througlh training, interpretable as unions promoting training or bargaining and sharing with members some of the rents obtained by firms. An additional spillover effect to trained nonunion and union workers comes through the degree of unionization of the industry interacted with training. OLS, the workhorse in this line of inquiry, yields strikingly different results: The union wage premiunm disappears. This highlights the importance of quantile regression techniques, especially where there is a presumption that the impacts may differ across the conditional distribution of the dependent variable. Although this study captures parts of the nonrandom placement of unions at the aggregate level by including firm size, concerns related to the possible selec- tion of workers into unions based on unobservable characteristics of individuals (ability, motivation) remain. Further research is required to shed additional light on several issues. Is there a causal relationship, or is the observed positive association between wages and union membership and union density at the industry level due simply to a nonrandom placement of unions? Do unions generate the rents that are subsequently awarded to union members through increased wages by reducing turnover and wage negotiation cost and increasing productivity, or do they merely act as the voice of the union members, extracting already existing rents from firms through their bargaining power, rents that the workers could not obtain on their own? Although suggestive, the data limitations of this study do not allow for explicit causal answers to these questions. Some evidence suggests that union workers are more productive than their nonunion colleagues in manufacturing in general (Sapsford and Tzannatos 1993). This appears to be a potentially fruitful avenue for further research in Ghana. However, all such efforts are severely handicapped by the lack of data. Data needs should be considered in the design of future surveys aimed at collecting matched employer-employee data sets. Getting at the dual caus- ality with respect to union placement into certain sectors or industries is data intensive as well. Drawing causal conclusions on the impact of unions on wages (and productivity, if data were also available) requires panel data on workers and possibly on firms to allow extracting worker-, firm-, sector-, or industry-specific effects. 250 THE WORLD BANK ECONONIC REVIEW, VOL. 18, NO. 2 APPENDIX: RECiRESSION VARIABLES AND SAMPLE Mean CHARACTERISTICS TABLE A-1. Regression Variables and Their Definitions Variable Definition Log(wages) Log of monthly wages of individual Individual variables Age Age of individual Age squared Age of individual squared Gender dummy variable I if a woman, 0 otherwise Education Dummy variables for highest level of education completed (1 if as stated, 0 otherwise; "no completed education" is the reference group) Primary Primary education Middle Middle education Secondary Secondary school Vocational Vocational education Polytechnic Polytechnic or technical education Professional Professional education University University Occupation Dummy variables for occupation (1 if as stated, 0 otherwise; "production worker" is the reference group) Administration Administrative staff Commercial Commercial and sales staff Professional Professional staff Support Support staff Manager Managerial staff Tenure (years) Years of tenure in the firm Tenure squared Years of tenure in the firm squared Training I if trained in the firm, 0 otherwise Tenure x training Interaction of tenure and training Permanent I if permanent contract, 0 otherwise Permanent x training Interaction of permanent and training Union membership 1 if member of a union, 0 otherwise Union membership x training Interaction of union membership and training Firnl variables Union density (industry) Share of the firms in an industry with at least 1 unionized worker Union density x training Interaction of union density and training Log(firm size) Log of the number of firm employees Accra 1 if firm is in Accra, 0 otherwise Note: The reduced core model excludes "training" and "union density" and their interaction terms, the intermediate model excludes "union density" and their interactions terms, and the full model includes all the variables. Source: Authors' model. Bllunch and Verner 251 TABLE A-2. Sample Mean Characteristics Variable Mean SD Log(wages) 10.590 0.062 Age 34.7 0.700 Age squared 1331.0 53.176 Gender 0.189 0.025 Individual characteristics Education None 0.094 0.018 Primary 0.031 0.009 Middle 0.490 0.026 Secondary 0.123 0.019 Vocational 0.126 0.016 Polytechnic 0.088 0.016 Professional 0.034 0.009 University 0.015 0.005 Occupation Production worker 0.423 0.031 Administration 0.130 0.020 Commercial 0.067 0.012 Professional 0.040 0.009 Support 0.122 0.025 Nlanager 0.119 0.015 Tenure (years) 6.6 0.504 Tenure squared 87.8 12.926 Training 0.309 0.036 Tenure training 2.012 0.279 Permanent 0.978 0.009 Permanent x training 0.300 0.035 Union membership 0.296 0.039 Union membership x training 0.083 (.019 Firn characteristics Log(number of employees) 3.735 0.119 Union density x industry 0.289 0.005 Union density x training 0.198 0.012 Accra 0.616 0.052 Notes: The sample consists of 683 workers. The clustering of workers within firms has been taken into account in the calculation of SDs. Sozurce: Authors' calculations based on data from Regional Program on Enterprise Development for Ghana (Wave III 1994). REFERENCES Becker, Gary S. 1964. Human Capital. Chicago: Chicago University Press. Booth, Alison L. 1995. The Economics of the Trade Unionz. Cambridge: Cambridge University Press. Booth, Alison L., and Monoiit Chatterji. 1998. "Unions and Efficient Training." Econonzic journal 108(447):328-43. Booth, Alison L., Marco Francesconi, and Gylfi Zoega. 1999. "Training, Rent-sharing and Unions." Discussion Paper 2200. Centre for Economic Policy Research, London. 252 THE WORLD BANK ECONONMIC REVIEW, VOL. I 8, NO. 2. Card, David. 1996. "The Effect of Unions on the Structure of Wages: A Longitudinal Analysis." Econometrica 64(4):957-79. Chamberlain, Gary. 1994. "Quantile Regression, Censoring and the Structure of Wages." In Advances in Econometrics, ed. Christopher Sims. New York: Elsevier. Chaykowski, Richard P., and George A. Slotsve. 2002. "Earnings Inequality and Unions in Canada." British Journal of Industrial Relations 40(3):493-519. Freeman, R. 1980. "Unionism and the Dispersion of Wages." Industrial and Labor Relations Review 34(1):3-23. Freeman, Richard B., and James L. Medoff. 1984. What Do Unions Do? New York: Basic Books. Gould, W. W. 1992. "sgll.1: Quantile Regression with Bootstrapped Standard Errors." Stata Technical Bulletin 9:19-21. Reprinted in Stata Technical Bulletin Reprints 2:137-39. - . 1997. "sg70: Interquantile and Simultaneous-Quantile Regression." Stata Technical Bulletin 38:14-22. Reprinted in Stata Technical Bulletin Reprints 7:167-76. Gray, Paul S. 1981. Unions and Leaders in Ghana: A Model of Labor and Development. New York: Conch Magazine Limited. Koenker, R., and G. Basset Jr. 1978. "Regression Quantiles." Econometrica 46(t):33-50. - . 1982. "Robust Tests for Heteroscedasticity Based on Regression Quantiles." Econometrica 50(1):43-61. Kristensen, N., and D. Verner. 1999. "Labor Market Distortions in Cote d'lvoire: A Quantile Regression Analysis of Employer-Employee Data." Africa Regional Labor Market Study, World Bank, Washington, D.C. Lewis, H. Gregg. 1963. Unionism and Relative Wages in the United States. Chicago: Chicago University Press. - . 1986. Union Relative Wage Effects: A Survey. Chicago: Chicago University Press. Maloney, W. F., and E. P. Ribeiro. 1999. "Efficiency Wage and Union Effects in Labor Demand and Wage Structure in Mexico: An Application of Quantile Analysis." Policy Research Working Paper 2131. World Bank, Washington, D.C. Mincer, Jacob. 1974. Schooling, Experience, and Earnings. New York: Columbia University Press. Panagides, Alexis, and Harry Anthony Patrinos. 1994. "Union-Nonunion Wage Differentials in the Developing World: A Case Study of Mexico." Policy Research Working Paper 1269. World Bank, Washington, D.C. Pencavel, J. 1995. "The Role of Labor Unions in Fostering Economic Development." Policy Research Working Paper 1469. World Bank, Washington, D.C. Psacharopoulos, George. 1985. "Returns to Education: A Further International Update and Implica- tions." Journal of Human Resources 20(4):583-97. - 1994. "Returns to Education: A Global Update." World Development 22(9):1325-343. Rama, Martin. 2000. "Wage Misalignment in CFA Countries: Were Labour Market Policies to Blame?" Journal of African Economies 9(4):475-511. Rogers, W. H. 1992. "sgll: Quantile Regression Standard Errors." Stata Technical Bulletin 9:16-19. Reprinted in Stata Technical Bulletin Reprints 2:133-37. Sapsford, David, and Zafiris Tzannatos. 1993. The Economics of the Labor Market. London: Macmillan. Schaffner, J. A. 1998. "Premiums to Employment in Larger Establishments: Evidence from Peru." Journal of Development Economics 55:81-113. Schultz, T. Paul, and Germano Mwabu. 1998. "Labor Unions and the Distribution of Wages and Employment in South Africa." Industrial and Labor Relations Review 5l(4):680-703. Standing, G. 1992. "Do Unions Impede or Accelerate Structural Adjustment? lndustrial versus Company Unions in an Industrialising Labour Market." Cambridge Journal of Economics 16:327-54. Velenchik, A. D. 1997. "Government Intervention, Efficiency Wages, and the Employer Size Wage Effect in Zimbabwe." Journal of Development Economics 53:305-38. Governance Matters III: Governance Indicators for 1996, 1998, 2000, and 2002 Daniel Kaufmann, Aart Kraay, and Massimo Mastruzzi Six dimensions of governance are estimated covering 199 countries ancd territories for four periods: 1996, 1998, 2000, and 2002. The indicators are based on several hundred individual variables measuring perceptions of governance drawn from 25 data sources constructed by 18 organizations. These individual measures are assigned to categories capturing key dimensions of governance. An unobserved-components model is used to construct six aggregate governance indicators in each of the four periods. Point esti- mates of the dimensions of governance are provided as well as the margins of errors for each country for the four periods. Methodological issues are also addressed, including tests for potential biases, and the interpretation and use of the data, given the estimated margins of errors for the indicators. The data and a Web-based graphical interface are available online at www.worldbank.org/wbi/governance/govdata2002/index.html. This article presents estimates of six dimensions of governance for 199 countries and territories for 1996, 1998, 2000, and 2002 developed in the context of an ongoing project to measure governance across countries. Section I describes the data used in developing this round of the governance indicators, which include several new sources. Data sources used in the earlier studies were updated forward to 2002 and backward to 1996, and previously estimated indicators for 1998 and 2000 were revised to reflect the new data. The aggregation procedure, described in section II, provides not only estimates of governance for each country but also measures of the precision or reliability of these estimates. Although the new data have improved the precision of the governance indicators, the margins of error remain large relative to the units in which governance is measured, so that comparisons across countries and especially over time should be made with caution. Measurement error is not unique to these indicators but is pervasive among all measures of governance and Daniel Kaufmann is director at the World Bank Institute; his e-mail address is dkaufmann@worldbank. org. Aart Kraay is senior economist in the Development Research Group at the World Bank; his e-mail address is akraay@worldbank.org. Massimo Mastruzzi is research analyst at the World Bank Institute; his e-mail address is mmastruzzi@worldbank.org. The authors thank Merli Baroudi, David C(ieslikowsky, Peter Cornelius, Rui Coutinho, Guy Dunn, Rich Fullenbaum, Joel Hellman, Phil Keefer, Marta Lagos, Marc Levy, Satish Mannan, Fiona Paua, Adrian Shute, Craig Webster, and Jennifer Windsor for providing data and answering numerous questions; Steve Knack, Clay Lowery, and Steve Radelet for their comments; and Erin Hoffmann for assistance. The support and collaboration of the World Economic Forum, the U.S. State Department, the Netherlands government, and Italian Trust Funds is appreciated. THE WORLD BANK ECONOMIC REVIEW, VOL. 18. NO 2, ©D The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi:10. 1093/wber/lhhO4I 18:253-287 253 254 THE WORLD BANK ECONOMIC REVIEW, VOL . 1 8, NO. 2 institutional quality. An advantage of the measures used here is that explicit margins of error reflecting this measurement error can be computed. Section III examines issues related to the construction and use of the governance indicators, such as the usefulness of subjective measures of governance relative to alternatives. It also empirically investigates the importance of ideological biases in expert assessments of corruption, finding little evidence that they exist. To illustrate the consequences of the substantial margins of error associated with the governance indicators, the aid allocation rules proposed for the U.S. government's Millennium Challenge Account, which rely on these measures, are examined in section IV. Section V explores the limited evidence available on global trends in governance, and section VI compares the Control of Corruption indicator estimated here with the widely used Corruption Perceptions Index produced by Transparency International. I. MEASURING GOVERNANCE In this study governance is defined broadly as the traditions and institutions by which authority in a country is exercised. This includes the process by which governments are selected and replaced, the capacity of the government to formu- late and implement sound policies, and the respect of citizens and the state for the institutions that govern economic and social interactions among them. This defini- tion guides the construction of the governance indicators for this study. Governance Clusters Data on perceptions of governance from a large number of sources are organized into six clusters corresponding to six dimensions of governance. The first two clusters are intended to capture the process by which those in authority are selected and replaced. One cluster, referred to as "voice and accountability," includes indicators of the political process, civil liberties, and political rights. These indica- tors measure the extent to which citizens are able to participate in the selection of governments. This category also includes indicators measuring the independence of the media, which serves the important monitoring role of holding those in authority accountable for their actions. The second cluster, "political stability and absence of violence," combines several indicators measuring perceptions of the likelihood that the government in power will be destabilized or overthrown by unconstitutional or violent means. This cluster captures the idea that the quality of governance is compromised by the likelihood of a wrenching change in govern- ment that directly affects the continuity of policies and undermines the ability of citizens to peacefully select and replace those in power. 1. There is some ambiguitv about the normative direction of a few of the suhcomponents this indicator. For example, a few sources ranik countries such as C.uba and the Democratic Republic of Korea highly in terms of their political stability, which simply reflects the longevity of the governments in power. Kaufmann, Kraay, aznd Mastruzzi 255 The next two clusters summarize indicators of the ability of the government to formulate and implement sound policies. The "government effectiveness" cluster combines the quality of public service provision, the quality of the bureaucracy, the competence and independence of the civil service, and the credibility of the government's commitment to policies. The focus is on inputs the government needs to produce and implement good policies and deliver public goods. The second cluster, "regulatory quality," focuses on the policies themselves. It includes measures of the incidence of market-unfriendly policies, such as price controls or inadequate bank supervision, and perceptions of the burdens imposed by excessive regulation in areas such as foreign trade and business development. The last two clusters summarize the respect of citizens and the state for the institutions that govern their interactions. "Rule of law" includes indicators that measure how well agents abide by the rules of society. These include perceptions of the incidence of crime, the effectiveness and predictability of the judiciary, and the enforceability of contracts. Together, these indicators measure a society's success in developing an environment in which fair and predictable rules form the basis for economic and social interactions and pro- perty rights are protected. The final cluster, "control of corruption," measures perceptions of corruption, conventionally defined as the exercise of public power for private gain. The focus of the various sources differs somewhat, ranging from the frequency of additional payments needed to get things done to the effects of corruption on the business environment and 'grand corrup- tion" in the political arena. Corruption is often a manifestation of a lack of respect of both the corrupter (typically a private citizen or firm) and the corrupted (typically a public official or politician) for the rules that govern their interactions. Souirces of Governance Data Some 250 individual measures from 25 sources produced by 18 different organ- izations are used in constructing the 2002 indicators (table 1; further details on each source and on how questions from each source were assigned to the six governance clusters are available online at www.worldbank.org/wbi/governance/ govdata2002/index.html). These organizations include international organiza- tions, political and business risk-rating agencies, think tanks, and nongovernmen- tal organizations. Six new data sources are included in 2002: Afrobarometer, a survey of individuals in 12 African countries; Reporters without Borders, an assessment of press freedoms compiled by an international journalist organization; Human Rights, a numerical coding of assessments of certain dimensions of human rights as reported by the U.S. State Department and Amnesty International (first reported in Cingranelli and Richards 2001 and subsequently updated and expanded by Craig Webster); World Markets Online, a commercial risk-rating agency; Voice of the People, a citizen survey sponsored by Gallup International; and the World Bank's Country Policy and Institutional Assessment (CPIA), TABLL I Sources of Governance Data Availability Country Source Publication Code Type coverage' Represenitative 1996 1998 2000 2002 Afrobarometer Afroharometcr Survey AFR Survey 12 x Business Environment Business Risk Service BRI Poll 50 x x x x Risk Intelligence Business Environment Qualitative Risk Mlcasure QLMN Poll 115 x x x N x Risk Intelligence Columbia Univcrsitv State Capacity Project CDU Poll 98 x x x Economist Intelligencc UnIit Country Risk Service Elt! Poll 11 5 x x x x European Bank for Reconstruction Transition Report EBR Poll 26 x x x x and Redevelopment Freedom House Nations in Transition FEIT Poll 27 x x x x Freedom House Freedom in the World FRH Poll 192 x x x x N Gallup International Gallip Millenniuiim Survey GMS S 60 x Gallup International 50th Anniversary Survey GALLUP Survey 44 N Gallup International Voice of the People Survey GAL Survey 46 x Heritage Foundation/ Economic Freedom Index HER Poll 161 x x x x x Wallstrect Journal Institute for Mlanagement World Competitiveness WCY Survey 49 x x x x and Development Yearbook Latinobarometro Latinobarometro Surveys LBO Sul-vey 17 x x x Political Risk Services International Country Risk Guide PRS Poll 140 x x x x x PriceWatcrhouseCoopers Opacity Index PWC Survey 35 x Reporters Without Borders Reporters Without Borders RSF Poll 138 x x Global Insight's DRI Country Risk Review DRI Poll IlI x x x x x NMcGiraw-Hill State Dcpartmiient/ HIuman Rights Report HUM Poll 159 x x x x x Amensty International World Bank Business Environmenit and BPS Survey 18 x x Enterprise Performance Survey World Bank World Business Environment Survey WBS Survev 81 x x x World Bank Country Policy and Institutional CPIA Poll 136 x x x x Assessments World Economic Forum Global Competitiveness Report GCS Survey 75 x x x x World Economic Forum Africa C(ompetitiveness Report GCSA Survey 23 x World Mlarkets Research Center World Markets Oninie WlIO Poll 186 x x 'Countries included in the most recently available version of source. Kauifnaznn, Kraay, and Nlastruzzi 257 an assessment of country performance constructed by World Bank country economists. Two of the new sources are also available for earlier years (Human Rights and CPIA). To improve the comparability of the governance indicators over time, indicators for 1998 and 2000 were revised to incorporate these sources. Two minor sources used in the past and no longer available were discarded.3 For convenience the revised indicators are referred to as indicators for 1998 and 2000, even though the measures are based on data for a two-year period (1997-98 and 2000-01). A subset of indicators is also available for 1996. These were used to construct new aggregate governance indicators for 1996. Two categories of sources are used: polls of experts and surveys of business- people or citizens. The choice between polls and surveys involves tradeoffs between cross-country comparability and firsthand knowledge of local con- ditions.' Polls of experts are designed to provide comparable results across countries through elaborate benchmarking. However, their reliability depends on the ability of a small group of experts to provide accurate assessments of the governance dimensions being measured.5 Surveys typically draw on the responses of large numbers of respondents with direct knowledge of local conditions. However, to the extent that ostensibly identical survey questions are interpreted differently by respondents with different cultural or socioeco- nomic backgrounds, it can be difficult to make cross-country comparisons.6 How representative the sources are of the world as a whole is also important. A number of sources cover a large sample of developed and developing eco- nomies, whereas others cover very narrowly focused samples. Many of the poorest and smallest countries tend not to be covered by commercially oriented 2. Transparency International's Corruption Perceptions Index (cr1) is not used as a component of our aggregate corruption indicator because the (-PI is itself an aggregate of sources that are already included in the corruption indicator constructed here. 3. These are the Central Euiropean Econoromic Retviewu, which rated a sample of transition economies and ceased publication after our only use of this source in the 1997/98 indicators, and the Political and Economic Risk Consultancy, which has also discontinued its rating of a small number of Asian eco- nomies. Dropping these sources does not affect country coverage, and it makes the aggregate indicators more comparable over time. 4. For a more detailed discussion of the advantages and disadvantages of polls of experts relative to surveys of market participants, see Kaufmann and others 1999b, 2002, and Kaufmann and Kraav 2002. 5. Most of the polls of experts cover large groups of raters. For example, the Economist Intelligence Unit, based in London, draws on the views of a worldwide network of correspondents for its assessments, as does Freedom House, which is based in New York, and Reporters without Borders, based in France. Other polls of experts have a narrower institutional affiliation for their respondents. For example the Europeani Bank for Reconstruction and Development Transition Report ratings are based primarily on the assessments of its staff in London, and the State Department component of our Human Rights measure reflects the views of Ul.S. State Department employees. 6. The three main sources of firm-level survev data (the Geneva-based World Economic Forum's Global Comipetitiveness Report, the Lausanne-based Institute for Mvanagement Development's World Competitiveness Yearbook, and the Washington-based World Bank's Business Environment Surveys) interview primarily domestic rather than foreign-owned firms in the countries thev cover. 258 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 polls because they are relatively unattractive to foreign investors. Because there is a strong positive association across countries between governance and per capita incomes, this difference between sources makes it difficult a priori to compare indicators from sources that cover sets of countries with very different income levels. Similarly, there may be regional differences in governance that hamper simple comparisons across sources. For example, it is difficult to com- pare a governance rating based only on transition economies with one based on a broad set of countries. As discussed in Kaufmann and others (1999a), the methodology used here to construct aggregate governance indicators takes these differences in country coverage into account, transforming the data from indi- vidual sources into common units for aggregation across sources. Table 1 identifies whether sources are considered representative or nonrepresentative. Aggregation Methodology Implicit in how the data have been organized is the view that within each cluster each indicator measures a similar underlying basic concept of governance. There are considerable benefits from combining related indicators into an aggregate governance indicator for each cluster. First, the aggregate indicators span a much larger set of countries than any individual source, permitting comparisons of governance across a broader set of countries. Second, aggregate indicators can provide more precise measures of governance. Third, it is possible to construct quantitative measures of the precision (and thus margins of error) of both the aggregate governance estimates for each country and their components. An extension of the standard unobserved-components model is used to combine the component indicators of each governance cluster into an aggregate governance indicator. The model expresses the observed data in each cluster as a linear function of the unobserved common component of governance, plus a disturbance term capturing perception errors or sampling variation in each indicator.7 Thus the observed score of country j on indicator k, y(j,k), is assumed to be a linear function of unobserved governance, g(j), and a disturbance term, E(j,k): (1) y(j,k) = (k) + 0(k) [g(j) + (j, k)], where a(k) and P(k) are unknown parameters that map unobserved governance g(j) into the observed data y(j,k). As a choice of units, g(j) is assumed to be a random variable with mean zero and variance one. The error term is assumed to follow a normal distribution with zero mean and the same variance across countries but a different variance across indicators: E[F(i,k)2] = (r2E(k). Finally, the errors are assumed to be independent across sources: E[e(j,k) e(j,l)] 0 for 1 different 7. The same methodology was used to construct previous versions of indicators; for detail, see Kaufmann and others (1999a). Unobserved components models were pioneered in economics by Goldberger (1972), and the closely related hierarchical and empirical Bayes models in statistics hy Efron and Morris (1971, 1972). Katifinann, Kraay, and Mastruzzi 259 from k. This imposes the identifying assumption that the only reason two sources might be correlated is that both are measuring the same underlying unobserved governance dimension.8 The disturbance term F(j,k) captures two sources of uncertainty in the rela- tionship between true governance and the observed indicators. First, the parti- cular aspect of governance covered by indicator k is imperfectly measured in each country, reflecting either perception errors on the part of experts (in the case of polls), or sampling variation (in the case of survevs). Second, the relationship between the concept measured by indicator k and the correspond- ing broader aspect of governance may be imperfect. For example, even if the aspect of corruption covered by indicator k (such as the prevalence of "improper practices") is perfectly measured, it may be a noisy indicator of corruption if there are differences across countries in what are considered to be "improper practices." Both sources of uncertainty are reflected in the indicator-specific variance of the error term, u (k). The estimate of governance for a country produced by the unobserved- components model is the mean of the distribution of unobserved governance conditional on the K(j) observed data points for that country. This conditional mean is the following weighted average of appropriately rescaled scores of each component indicator: (2) E[g(j) Iy(jjl), . .., y(j, K(j))] = Ski) w(k) - [(y(j,k) - ot(k))/O(k)], where the weights applied to each source k, w(k) = cr,(k)-2/[l + yK=', (r,(k)-2] are inversely proportional to the variance of the error term of that source. Precision weighting results in efficiency gains relative to the alternative of simply averaging rescaled scores from each source for each country. The stand- ard deviation of this conditional distribution is also reported as an indicator of the confidence in this estimate: (3) SD[g(j)Iy(,1). y(j,K(i))] (I +ZK>)(u(k) ) 8. For some pairs of sources, this assumption may not be literally true. For example, it will be violated if different risk-rating agencies base their assessments on the assessments of other agencies included in the sample. To the best of our knowledge, we have excluded any source of governance data found to be explicitly based on another one of our sources. Nevertheless, the possibility of correlated errors remains. The main consequence would be that our standard errors will be biased downward, see Kaufmann and others (1999a) for an example. This underscores the importance of caution in comparing governance estimates across countries and over time. 260 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 This standard deviation is declining in the number of individual indicators in which a particular country appears and increasing in the variance of the dis- turbance term on each of these indicators. The assumptions of the unobserved-components model ensure that the distribu- tion of governance in each country is normal, conditional on the data for that country. Therefore, these conditional means and standard deviations for each coun- try have a natural interpretation: there is a 90 percent probability that the true level of governance in a country is in an interval of plus or minus 1.64 times the reported standard deviation centered on the point estimate itself. This range is referred to as a 90 percent confidence interval around the estimate of governance for a country.9 Implementing his approach requires estimates of all the unknown survey- specific parameters, ot(k), 3(k), and a 2(k). These are computed in a two-stage procedure. In the first stage, maximum likelihood methods are applied, using only the representative sources to retrieve the parameters for each governance cluster. This is a standard application of the unobserved-components model. The many nonrepresentative sources cannot be included in the first stage of the estimation procedure because the distribution of unobserved governance in the subset of countries covered by these surveys is not the same as that in the world as a whole. As a result, for these sources governance in the countries covered by these surveys cannot be assumed to follow a standard normal distribution, as is required by the maximum likelihood procedure. The parameters of the nonrepresentative sources are obtained in the second stage of the procedure. In this stage the preliminary estimates of governance based only on representative sources are treated as an observable proxy for governance, and the parameters of interest for the nonrepresentative sources are obtained by regressing these indicators on observable governance (directly estimating equation 1).1° All the estimated parameters of the unobserved-components model are then used to construct a final set of estimates of governance. The resulting estimates have an expected value across countries of zero and a standard deviation across countries of one. Due to sampling variability, this will not be exactly true for any individual governance indicator in any period. To avoid confusion about the units of the governance indicators, the estimates of governance are rescaled by subtracting the mean across countries and dividing by the standard deviation across countries for each indicator, so that each indicator has a mean of zero and a standard deviation of one in each period. 9. This is a slight abuse of terminology, because these are not confidence intervals in the usual sense of a stochastically varying interval centered around a fixed unknown parameter. Rather, governance is treated as a random variable, and the 90 percent confidence interval is simply the 5th and 95th percentiles of the conditional distribution of governance, given the observed data. 10. Getting consistent estimates of the parameters of the nonrepresentative sources requires adjust- ment for attenuation bias caused by the fact that the observable proxy for governance is a noisy indicator of true governance. Fortunately, the information on the standard errors associated with the governance estimates obtained in the first stage can be used to do this. Kaufmann, Kraay, and Mastruzzi 261 It is also important to note the assumption that the distribution of unob- served governance is the same in every period. In particular, this imposes the restriction that the mean or world average of governance is the same in each period. As a result, the indicators are not informative about global trends in governance (see section V), although they are informative about changes in countries' relative positions over time. II. GOVERNANCE INDICATORS FOR 1996, 1998, 2000, AND 2002 This section presents aggregate governance indicators for the six indicators for all four periods and also examines changes in the indicators over time. The data are available online at www.worldbank.org/wbi/governance/govdata2002/ index.html. Levels of Governance Worldwide The governance estimates are normally distributed with a mean of zero and a standard deviation of one in each period, implying that virtually all scores lie between -2.5 and 2.5 (higher scores indicate better outcomes). As discussed, this also implies that the aggregate estimates convey no information about global trends in governance. They are, however, informative about changes in countries' relative positions over time. The voice and accountability indicator covers 199 countries for 2002, the largest set among the six indicators (table 2).11 Four indicators cover 195 countries, and one, political stability, covers 186 countries. Over time, there has been a steady increase in the number of countries covered by each indicator. The number of data sources has increased as well, raising the median number of sources available per country, which ranges from six to eight in 2002 compared with four to six in 1996. The proportion of countries in the sample with govern- ance estimates based on only one source has declined considerably, from an average of 15 percent in 1996 to 10 percent in 2002. Because the 2002 indicators now cover virtually all countries in the world, no major improvements in country coverage are expected in the future. An important consequence of this expanded data availability is that the standard errors for the governance indicators have declined. In 1996 the average of the standard errors ranged from 0.26 to 0.39, whereas in 2002 they ranged from 0.19 to 0.27 (see table 2). Moreover, the average standard errors for the revised 1998 and 2000 indicators are also lower than the previous estimates, again reflecting the incorporation of more data for more countries. These 11. A few of the entities covered by the indicators are not independent states (French Guyana; Hong Kong, China; Martinique; Puerto Rico; and West Bank and Gaza). A handful of very small independent principalities (Andorra, Monaco, and San Marino) are also included. For convenience, all 199 entities are referred to as countries. 262 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. 2 TABLE 2. Summary Statistics on Governance Indicators Voice and Political Government Regulatory Rule of Control of Year accountability stability effectiveness quality law corruption Overalla Number of countries 1996 192 165 180 182 167 151 173 1998 192 166 184 185 186 184 183 2000 192 166 185 186 186 185 183 2002 199 186 195 195 195 195 194 Median number of sources per country 1996 4 4 4 4 6 4 4 1998 4 4 4 4 7 5 5 2000 5 6 5 4 8 6 6 2002 7 6 6 6 8 7 7 Proportion of countries tvith only one data source 1996 0.15 0.13 0.21 0.15 0.07 0.18 0.15 1998 0.14 0.10 0.19 0.13 0.11 0.18 0.14 2000 0.14 0.06 0.18 0.13 0.10 0.17 0.13 2002 0.10 0.11 0.10 0.10 0.10 0.10 0.10 Average standard error 1996 0.26 0.39 0.28 0.34 0.26 0.29 0.30 1998 0.25 0.32 0.30 0.34 0.25 0.25 0.28 2000 0.25 0.33 0.27 0.35 0.22 0.26 0.28 2002 0.21 0.27 0.22 0.22 0.19 0.21 0.22 'A simple average of the six indicators. Source: Authors' calculations based on sources listed in table 1. declines in margins of error illustrate the benefits to precision from constructing composite indicators based on as much information as possible. Despite these improvements the margins of error associated with estimates of governance remain substantial relative to the units in which governance is mea- sured. This is illustrated by plotting countries in ascending order according to their point estimates of two selected dimensions of governance in 2002 on the x axis and according to estimates of governance and the associated 90 percent confidence interval described above on the y axis for each indicator (figure 1). The confidence intervals vary across countries because countries appear in differ- ent numbers of sources with different levels of precision and are large relative to the units in which governance is measured. This point is emphasized by the horizontal lines in figure 1, which delineate the quartiles of the distribution of governance estimates. Even though the differences between countries in the bottom and top quartiles are substantial, the number of countries that have 90 percent confidence intervals that lie entirely within a given quartile is not large. Thus many of the small differences in estimates of governance across coun- tries are not likely to be statistically significant. For many applications it is therefore more useful to focus on the range of possible governance values for each country (as summarized in the 90 percent confidence intervals shown in figure 1) than on the point estimates. For two countries at opposite ends of the Kaufmann, Kraay, and Mastruzzi 263 FIGURE 1. Selected Aggregate Indicators of Governance, 2002 Voice and accountability 2.5 2 1.5 - 0.5 2 0= 5. ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~~. C 0 > ,, x -1.5 > > <,0 5 0 -2.5 o m~~~~~ 0 0.2 0.4 0.6 0.8 1 Percentile rank with the midpControl of Corruption 2 - 1.5-> 0)~~~~~~~~~~~~~~~~~~ s -of g wt '-0.5 > 0 -1.5-f - M> i 0 0.2 0.4 0.6 0.8 Percentile rank Note: Vertical bars show the statistically likely range of values of governance for each country, with the midpoint of each bar cotresponding to the best single estimate. Selected countries are labeled. Source: Authors' calculations based on sources listed in table 1. scale of governance whose 90 percent confidence intervals do not overlap at all, it is clear that differences in governance between them are significant. For pairs of countries that are closer together and whose 90 percent confidence intervals overlap, the significance of estimated differences in governance is less clear. 264 THE WORLD BANK C(ONOMIC REVIEW, VOL. i 8, NO. 2 Differences across countries in the margins of error associated with govern- ance estimates are due to cross-country differences in the number of sources per country and differences in the precision of the sources in which each country appears. As equation 3 shows, the precision of governance estimates increases with the number and precision of sources. Across countries the standard error of the governance estimate for a country declines at the rate of the inverse of the square root of the number of sources, consistent with the assumption that errors are uncorrelated across sources. In practice, this means that a 90 percent confidence interval for a country with only one source will be roughly twice as large as the 90 percent confidence interval for a country appearing in the median number of seven sources. On the precision of sources, recall that for each source the variance of the error with which it measures the unobserved true level of governance is esti- mated and the inverse of these estimated variances is used to weight sources when constructing the aggregate score for each country. This means that more precise sources (in the sense of providing less noisy signals of governance) receive more weight in the aggregate indicators, minimizing the variance in the estimates of governance for each country. There is considerable variation in the weights assigned to different sources (reflecting substantial differences in the estimated precision of each source),"2 and these differences are reflected in the differences in the margins of error associated with governance scores for each country. Table 3 summarizes the weights applied to each source for a hypothetical country appearing in all sources in constructing the corresponding aggregate indicator. The weights for a country appearing in a subset of sources would be proportional to the ones reported for those sources. Changes over Time in Estimates of Governance The observed change in governance for a given country between two points in time can be attributed to four factors: changes in the perceptions of governance recorded in the underlying sources available in both periods, changes in the weights applied to different sources in each period, changes to the set of sources for a country, and the addition of new countries to the aggregate indicator that systematically rate better or worse than the country in question (recall that indicators measure only countries' relative positions). Changes in scores assigned to countries by underlying sources is the clearest reason for changes in governance for a country over time. Interpreting changes due to the other three reasons is more difficult and involves tradeoffs. Consider the reweighting of sources that occurs from year to year as the observed 12. In estimating the unobserved components model, the estimated precision of each source reflects the extent to which that source is correlated with other sources. In the empirical framework, errors are assumed to be uncorrelated across sources. As a result, sources that tend to be highly correlated with other sources are more informative, and hence have lower error variances, than sources that tend to be only xveaklv correlated with other sources. Kanifozann, Kraay, and Mastruzzi 265 correlations among sources change. The reweighting ensures the most precise estimates of the level of governance for each year, but some of the changes over time for a given country will reflect this reweighting rather than changes in the underlying indicators. However, these changes in weights account for only a small fraction of the variance of changes in governance estimates on average (Kaufmann and others 2002). There are also tradeoffs in interpreting the changes in governance estimates from the addition of new sources for a country. Adding new data sources improves the precision of estimates of governance in a country at a single point in time. However, if the new sources rate a country significantly differently from existing sources, this can result in changes in estimates of governance that reflect the inclusion of new information on the previous period rather than actual changes in governance. To reduce the effects of this source of variation in governance estimates and to improve the precision of estimates for past years, previous indicators have been recalculated incorporating all the data used for this analysis. Nevertheless, the 2002 indicators also reflect the information embodied in a few new sources relative to 2000 and previous periods, and this provides a further reason why changes over time should be interpreted with caution. i Bias resulting from the addition of new countries to the aggregate indicator can be removed from comparisons of governance estimates over time by limiting such comparisons to changes in countries' percentile ranks for the same set of countries for both periods. This procedure is a useful robustness check when considering changes over time in a specific country or set of countries. In practice, however, this source of bias is relatively small, especially when com- paring 2002 with 2000, because there are now only small changes in the number of countries covered between these two periods. A final issue concerns the statistical significance of observed changes in the aggregate indicators. The basic observation is that changes in the estimates of governance tend to be small relative to the cross-country differences in levels of governance. It is difficult to be more precise about the statistical significance of changes in governance because of the aggregation procedure. For each period the aggregation procedure summarizes knowledge about governance in a given country in terms of the distribution of unobserved governance conditional on the data for that country. The mean of this conditional distribution is used as the best estimate of the level of governance in a country, and the standard deviation of this distribution is used to summarize the precision of the know- ledge about governance for that country. However, when the aggregation procedure is repeated in successive periods, no information is produced about the joint distribution of governance in successive periods. Without this joint 13. On the Web site displaying the data (www.worldhank.org/wbi/governance/govdata2002/ index.html), users may identify the sources of governance data used for each country, indicator, and period. 266 THE WORLD BANK ECONOMIC REVIEW, VOL. i 8, NO. 2 TABLE 3. Weights Used to Aggregate Governance Indicators Voice and accountability Political stability Government effectiveness 1996 1998 2000 2002 1996 1998 2000 2002 1996 1998 2000 2002 Representative sources CDU - - 0.03 0.04 - - 0.07 0.09 - - 0.07 0.05 DRI - - - - 0.25 0.17 0.19 0.14 0.08 0.07 0.12 0.07 EIU 0.31 0.12 0.09 0.18 0.13 0.14 0.23 0.21 0.19 0.18 0.20 0.21 FRH 0.14 0.22 0.14 0.12 - - - - - - - - HER - - - - - - - - - - - - HUM 0.05 0.10 0.06 0.04 0.10 0.10 0.08 0.04 - - - - PRS 0.09 0.14 0.14 0.07 0.09 0.25 0.07 0.07 0.05 0.02 0.07 0.05 RSF - - - 0.02 - - - - - - - - WBS - 0.01 0.00 - - 0.07 0.01 - - 0.06 0.03 - WMO - - - 0.06 - - - 0.17 - - - 0.13 Nonrepresentative sources AFR - - - 0.01 - - - - - - - 0.01 BPS - - - - - - - - - - - 0.01 BRI - - - - 0.22 0.10 0.12 0.11 0.09 0.06 0.08 0.06 EBR - - - - - - - - - - - - FHT 0.30 0.37 0.49 0.39 - - - - 0.31 0.15 0.12 0.24 GAL - - - 0.01 - - - - - - - - GCS - - - 0.02 0.13 0.07 0.04 0.05 0.06 0.09 0.09 0.06 GCSA - - - - - 0.04 - - - 0.15 - - GMS - - 0.02 - - - 0.01 - - - 0.00 - LOB 0.08 - - 0.01 - - - 0.03 0.00 - - 0.01 OPF - - - - - - - - - - - - PIA - - - - - - - - 0.13 0.15 0.12 0.06 QLM - - - - - - - - - - - - WCY 0.01 0.00 0.01 0.01 - - 0.13 0.06 0.05 0.04 0.06 0.04 - not available. Note: See table 1 for source codes. The weights used in constructing the aggregate governance indicators correspond to those that would be applied for a hypothetical country appearing in all of the available sources for that indicator. The weights are proportional to the inverse of the variance of the estimate of measurement error for each source (see discussion in text). For a country appearing in fewer sources, the relative weights applied to each source will be the same as the relative weights implicit in this table. Source: Authors' calculations based on sources listed in table 1. distribution, precise probabilistic statements cannot be made about changes in governance over time.14 Instead, a useful rule of thumb is to focus on changes in governance for countries in which the 90 percent confidence intervals in the two periods do not overlap. This can be illustrated by plotting the 2002 score on the x axis and the 14. Extensions of the aggregation procedure along the lines of dynamic unobserved component models could in principle provide information about the joint distribution of governance over time. We have not yet attempted to implement this idea with our data. Kainfmann, Kraay, and Mastruzzi 267 Regulatory quality Rule of law Control of corruption 1996 1998 2000 2002 1996 1998 2000 2002 1996 1998 2000 2002 - - - - - - 0.03 0.03 - - 0.08 0.06 0.09 0.08 0.23 0.03 0.03 0.06 0.08 0.09 0.09 0.05 0.09 0.06 0.23 - - 0.21 0.34 0.08 0.06 0.20 0.24 0.07 0.11 0.12 0.10 0.05 0.09 0.06 0.05 0.08 0.07 0.08 - - - - - - - - 0.02 0.02 0.02 0.01 - - - - 0.03 0.03 0.06 0.09 0.03 0.04 0.03 0.02 0.05 0.01 0.03 0.03 - 0.00 0.01 - - 0.05 0.02 - - 0.05 0.07 - - - - 0.26 - - - 0.11 - - - 0.09 - - - - - - - - 0.02 - - - 0.00 - - - 0.00 - - 0.08 0.01 - - - - 0.08 0.06 0.09 0.06 0.03 0.01 0.01 0.01 0.07 0.23 0.06 0.11 - - - - - - - - - - - - 0.16 0.17 0.27 0.13 - 0.18 0.23 0.22 - - - - - - - 0.01 - - - - 0.14 0.19 0.09 0.04 0.07 0.10 0.14 0.07 0.09 0.12 0.10 0.06 - 0.15 - - - 0.00 - - - 0.17 - - - - - - - - 0.00 - - 0.04 0.02 - _ _ _ _ 0.00 - 0.00 - - - - 0.06 0.08 0.08 0.21 0.07 - 0.11 0.10 0.04 - 0.11 0.08 0.04 - - - - 0.11 0.12 - 0.07 0.10 0.10 - 0.11 0.21 0.16 0.20 0.10 0.10 0.08 0.08 0.06 0.36 0.07 0.09 0.08 2000 score on the y axis and by drawing the 45° line that distinguishes countries with declines in the quality of governance from countries with improvements in governance (figure 2). Countries with large changes in governance relative to their margins of error in each period are highlighted, and the 90 percent confidence intervals in each period is indicated by vertical and horizontal lines. The score for each country that appears in the 2002 indicators but not in the 2000 indicators is plotted along the 450 line, giving a visual summary of the distribution of governance among the countries added to the sample in 2002. The number of countries with large changes in governance over this brief period is quite small. This is not surprising given the short period under consideration and the gradual nature of most changes in governance. Many of 268 THE WORLD BANK ECONONMIC REVIEW, VOL. I8, NO. Z FIGURE 2. Changes Over Time in Selected Governance Estimates, 2000 to 2002 Rule of Law 8o 3- C4C N .C e 2 | -3 -2 G . .1 2 3 Governance rating 2002 -2 - -3 - Political stability and absence of violence CO3 2 UA .* AG K -3 1 ~1.2 3 o *NM ~~~~~~Governance rating 2002 -2- -3 Note: Highlighted countries are those in which the 90 percent confidence intervals in the two periods do not overlap. The corresponding confidence intervals in 2000 and 2002 are indicated as vertical and horizontal bars. The 45° line demarcates the difference between countries showing declines in governance (above the line) and those showing improvements (below the line). Source: Authors' calculations based on sources listed in table 1. Kaufmann, Kraay, and Mastruzzi 269 the changes are understandable in light of events over these two years. For example, Argentina's recent financial crisis is reflected in strong declines in perceptions of governance across the board. Similarly, the recent turmoil in Zimbabwe is associated with a sharp decline in perceptions of the rule of law. For the United States, declines in political stability and absence of violence reflect heightened concerns about terrorism in the aftermath of the terrorist attacks of September 11, 2001. In Sri Lanka reductions in sectarian violence drive an improved score in this category. The reasons for some of the other changes highlighted in figure 2 are less obvious. These are examined in more detail by tallying the number of sources available in both periods that move in the same or opposite direction as the aggregate indicator or that register no change (table 4). The overall rate of agreement between changes in the sources and the direction of change in the aggregate indicator is calculated as the agreement ratio (the ratio of number of agreements to the total number of changes in both directions). The agreement ratio is quite high for countries with large changes in govern- ance, with an average across all countries and indicators of 0.79. This provides some confidence that for countries with large changes in governance estimates, the changes are driven primarily by changes in underlying sources. For only four countries is the agreement ratio less than one-half-Belarus and Iraq for regulatory quality and Madagascar and West Bank and Gaza for control of corruption. Belarus's surprisingly high score in 2002 is driven primarily by the very strong responses from firms in the Business Environment and Enterprise Performance Survey. Iraq illustrates an unusual case in which the reweighting of sources has a substantial effect on changes over time. In both periods Iraq has one of the worst scores in the world, so the large change in its score reflects no real improvement during the period but rather the much lower weight assigned to the source that rated Iraq highest in 2002. For control of corruption the large improvement observed in Madagascar and the large decline in West Bank and Gaza are both driven entirely by changes in the set of sources in which these countries appear. Madagascar appears in one new source that rates it highly (World Markets Online), and it does not appear in the 2002 version of the State Capacity Project, which gave it a poor score in 2000. West Bank and Gaza fares well on the World Business Environment Survey in 2000, a source that is not available in 2002, and scores poorly on the only source available for 2002, World Markets Online. Unlike these large changes in governance, which reflect primarily changes in the underlying sources, the majority of smaller changes reflect a combination of all four reasons for variation discussed. For the remaining smaller changes in governance between 2000 and 2002 not reported in table 4, the average agree- ment ratio across all countries ranges from 0.57 to 0.64 for the six indicators, substantially lower than the agreement ratio for large changes. This suggests a need for greater caution in interpreting the small changes in governance estimates typical from one period to the next. Although changes over longer 270 THE WORLD BANK ECONOMIC REVIEW, VOL. 1 8, NO. 2 TABLE 4. Large Changes in Governance Estimates, 2000 to 2002 Governance Sources available Changes in sources score in both periodsa between 2000 and 2002 Indicators and countries 2002 2000 Agree No change Disagree ratio Added Dropped Voice and accountability Sierra Leone -0.57 -1.36 2 2 0 1.00 2 0 Political stability Sri Lanka -0.90 -1.80 3 0 3 0.50 1 0 Namibia 0.46 -0.72 3 0 1 0.75 3 1 Argentina -0.74 0.46 7 0 1 0.88 2 1 C6te d'lvoire -2.04 -0.88 3 0 1 0.75 2 1 Georgia -1.90 -0.85 3 0 0 1.00 1 2 Israel -1.35 -0.47 4 0 3 0.57 1 0 Kyrgyz Republic -1.21 -0.03 1 1 0 1.00 1 1 United States 0.34 1.26 5 1 1 0.83 1 2 Venezuela -1.20 -0.48 4 1 2 0.67 3 1 Government effectiveness Dominica 0.32 -0.86 1 0 0 1.00 1 0 Argentina -0.49 0.30 7 1 0 1.00 1 1 Egypt, Arab Rep. -0.32 0.35 4 2 1 0.80 1 1 Gambia, The -0.81 0.25 1 0 1 0.50 2 0 Tunisia 0.65 1.32 4 0 1 0.80 1 1 Regulatory quality Afghanistan -1.82 -3.57 1 0 0 1.00 1 0 Belarus -1.67 -2.65 1 1 2 0.33 3 1 Iraq -2.31 -3.36 0 0 3 0.00 2 0 Moldova 0.80 0.14 5 0 0 1.00 3 1 Russian Federation -0.30 -1.55 6 0 1 0.86 3 1 Congo, Dem. Rep. -1.77 -2.87 3 0 0 1.00 1 1 Argentina -0.84 0.44 5 0 1 0.83 2 1 Bangladesh -1.05 -0.02 2 2 0 1.00 3 1 Cameroon -0.88 0.12 3 1 0 1.00 2 1 El Salvador 0.04 1.12 2 1 1 0.67 2 1 Zambia -0.60 0.43 3 0 1 0.75 2 1 Rule of law Samoa 0.94 -0.14 1 0 0 1.00 1 0 Argentina -0.73 0.18 8 1 2 0.80 2 4 Cote d'lvoire -1.21 -0.53 2 3 1 0.67 2 1 Georgia -1.17 -0.56 4 3 0 1.00 2 2 Namibia 0.45 1.21 2 3 1 0.67 2 1 Zimbabwe -1.33 -0.73 6 2 0 1.00 1 0 Control of corruption Madagascar 0.14 -0.80 0 1 1 0.00 1 2 Belarus -0.78 -0.07 4 1 0 1.00 1 2 Malawi -0.91 -0.22 2 2 0 1.00 2 2 Namibia 0.21 1.16 3 0 1 0.75 4 1 West Bank and Gaza -0.99 0.76 0 0 0 - 1 1 Average 3.11 0.78 0.81 0.79 1.78 1.00 - not available. Note: Sources of changes in estimates of governance between 2000 and 2002 for each country for which the 90 percent confidence intervals for the level of governance in the two periods do not overlap. aThe number of individual sources that agree or disagree with the direction of change of the aggregate indicator. Source: Authors' calculations based on sources listed in table 1. Kaufmann, Kraay, and Mastruzzi 271 periods of time, such as from 1996 to 2002, would be expected to be more informative, this is partly offset by the larger changes in the composition of the sources between the periods. III. USE OF PERCEPTIONS-BASED SOURCES AND POTENTIAL IDEOLOGICAL BIAS The construction of the aggregate governance indicators described herein relies exclusively on subjective, perceptions-based measures of governance. For many of the key dimensions of governance, such as corruption or the confidence that property rights are protected, objective data are almost by definition impossible to obtain, so there are few alternatives to subjective data. Why Perceptions-Based Meastires Are Used Consider corruption. As an illegal activity, there are no direct measures of its prevalence. Various indirect measures are possible, but none are without diffi- culty. For example, relying on the frequency of references to corruption in the media will reflect not only the prevalence of corruption but also the freedom and objectivity of the press. Similarly, trials for corruption will reflect the competence and independence of the police and the judicial system and not exclusively the prevalence of corruption. Recently, a few studies have attempted to assess corruption by looking for patterns in objective data that can only be consistent with corruption, such as variations in the procurement prices paid for homogeneous medical inputs across hospitals in Buenos Aires (Di Tella and Shargrodsky 2003) and gaps between existing stocks of public infrastructure and past infrastructure spending across Italy (Golden and Picci 2003). Though interesting, such exercises have enormous data requirements, and cross-country measures of corruption based on this idea are unavailable.'5 Objective measures may be available for some other dimensions of govern- ance, but they are not without weaknesses. Objective data on elections can be used to measure democratic participation. But there is considerable variation across countries in the extent to which election outcomes reflect the will of voters. Measuring the extent to which elections are subverted, whether through intimidation or manipulation of results, returns quickly to the realm of percep- tions-based data. This is just one example of the important distinction between de jure and de facto situations regarding governance. Countries may have extensive formal protections of property rights but little or no enforcement. 15. Furthermore, these within-country measures based on prices, assets, and expenditure patterns are typically a proxy of the combined effect of the extent of mismanagement, inefficiency, and corruption. Disentangling the pure effect of corruption is far from simple. 272 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. I For example, most countries now have formal independent anticorruption commissions or agencies, but their effectiveness varies greatly. Subjective perceptions of governance often matter as much as the legal reality. For example, Hellman and Kaufmann (2004) develop a measure of "crony bias" or unequal influence across firms based on firms' perceptions of undue influence on political decisionmaking exerted by powerful firms. They find that perceived unequal influence has a strong negative impact on a firm's assessment of public institutions and on its behavior toward those institutions, resulting in less use of the courts to resolve business disputes, lower enforce- ability of court decisions, lower levels of tax compliance, and higher levels of bribery. Thus, inequality of influence not only damages the credibility of insti- tutions but also affects the likelihood that firms will use and provide tax resources to support such institutions, thereby perpetuating the weakness of such institutions and the likelihood of capture by the influential. Finally, recent studies have yielded a profusion of results linking objective measures of the structure of institutions to a range of governance outcomes.16 Although the studies have greatly expanded understanding of the institutional determinants of development, these objective measures of institutional quality do not lend themselves well to the construction of aggregate governance indi- cators. The measures typically do not have normative content on their own. They assume normative content only in the context of a particular empirical analysis linking them with a particular outcome. For example, although meas- ures of decentralization may be correlated with the incidence of corruption across countries, the explanatory power of this variable is generally not suffi- ciently strong to consider decentralization as a reasonable proxy for corruption. None of this is to suggest that the subjective data used here are problem-free. Perceptions-based questions about governance can be vague and open to inter- pretation. For example, a well-crafted question on corruption asks firms for the estimated share of bribes in the annual spending of firms like theirs. By contrast, generalized opinion questions, such as a citizen's perception of the overall tolerance of the population for corruption, are less informative for constructing aggregate indicators of governance. Today, studies like this can rely on more specific, better crafted, and to an extent experiential questions. For instance, the Global Competitiveness Report survey of firms contains much more specific questions about corruption and governance than even during the mid-1990s, and some are of a quantitative and experiential nature (such as percentage of senior management time spent dealing with public officials). 16. A nonexhaustive list includes the links between decentralization and corruption, the effect of the structure of the legal system on financial market development, the effect of checks and halances in the political system on regulatory and fiscal performance, and the effect of democratic institutions on a wide variety of development outcomes. Kau/ltmann, Kmraa, aznd Mastruzzi 273 Potential Ideological Biases in Perceptionzs Data A potential drawback of information collected from polls of experts is that it may reflect the ideological tendencies of the institutions compiling the performance ratings. This may not be a major concern for the sources used for this analysis. The high degree of correlation among virtually all of the sources is difficult to reconcile with a systematic ideological bias among one or more sources. Never- theless, it is useful to investigate the extent to which differences in assessments are related to observable measures of the ideology of the government in power in each country. This is done as follows. Surveys of firms or individuals are assumed not to be tainted by ideology, because they reflect the views of a large number of respon- dents in each country. However, a poll of a smaller number of experts affiliated with a particular institution may reflect that institution's ideology. The effects of ideology can therefore be identified by looking at correlation across countries of the ideology of the government in power and the difference in the percentile ranks assigned to countries by a poll of experts and a survey of individuals and firms. This approach was applied to several polls of experts, using the World Business Environment Survey for 2000 as a benchmark survey assumed to be unaffected by respondent ideology. Government ideology was measured using an indicator variable of the political orientation of the government in power (taken from Beck and others 2001) that takes on the value 1 for left of center, 2 for centrist, and 3 for right of center. The difference between the percentile rank of a country on a poll of experts and its rank on the World Business Environment Survey is regressed on the indicator variable measuring the ideology of the government in that country, for several polls of experts (table 5). All variables are measured in 2000, the most recent year for which the ideology variable is available.'7 The coefficient on the ideology variable will therefore capture the extent to which a given poll of experts rates countries with left- or right-wing governments systematically differently from the survey. A positive coefficient indicates that the poll tends to rate right-of-center governments more highly relative to the survey, whereas a negative coefficient indicates a bias toward left-of-center governments.'8 The Heritage Foundation is the only source that appears to have a consistent ideological bias, assigning relatively higher scores to countries with right-of-center governments than the corresponding surveys. However, this ideology bias is fairly moclest, resulting in about a 7-10 percentage point higher ranking for a right-of-center government 17. For voice and accountability, Gallup Millennium Survey is used instead of World Busilness Enviroinment Survey as the comparator survey, because the World Business Environment Survev ques- tions on voice and accountability capture the extent to which firms have a voice in policymaking, which is considerably narrower than most other polls. 18. Because most of the countries are coded as left or right of center, almnost identical results are obtained if dummy variables for left- and right-of-center governments are included instead. 274 THE WORLI) BANK ECONOMIC RFVIEW, VOL. i 8, No. 2 TABLE 5. Ideology Regressions for 2000 PRS PIA EIU DRI CDU BRI QLM HUM EBR HER FRH Voice and accountabilityv Ideology -2.78 -1.64 -1.72 3.67 -0.83 0.59 0.46 0.27 0.68 0.23 Observations 44 43 28 46 46 (no.) Adjusted R2 _0.01 -0.02 -0.04 -0.01 -0.02 Political stabllity Ideology 12.37 8.86 8.54 4.97 3.15 12.11 2.68" 1.80* 1.87, 0.93 0.61 2.52" Observations 52 51 46 42 25 56 (no.) Adjusted R2 0.1 0.04 0.05 -0.01 -0.02 0.09 Government effectiveness Ideology -1.84 -0.66 -2.38 1.86 -7.12 1.64 0.64 0.16 0.68 0.48 1.90* 0.25 Observations 52 47 51 46 42 25 (no.) Adjusted R2 -0.01 -0.02 -0.01 -0.02 0.05 -0.04 Regulatory quality Ideology 8.05 13.3 3.22 6.55 10.24 1.57 2.08** 0.45 0.88 1.77 Observations 52 47 46 15 56 (no.) Adjusted R2 0.02 0.07 -0.02 -0.01 0.04 Rule of law Ideology 1.52 3.39 5.61 5.67 4.68 7.32 6.47 5.32 7.42 0.41 0.73 1.65 1.46 1.21 1.65 1.63 1.19 1.91* Observations 52 47 51 46 42 25 49 56 56 (no.) Adjusted R2 -0.02 -0.01 0.03 0.02 0.01 0.05 0.03 0.01 0.05 Control of corruption Ideology 3.05 1.4 0.31 0.57 -2.21 2.83 1.84 0.63 0.34 0.1 0.18 0.68 0.46 0.58 Observations 52 47 51 46 42 25 49 (no.) Adjusted R2 -0.01 -0.02 -0.02 -0.02 -0.01 -0.03 -0.01 "Significant at the 10 percent level. '*Significant at the 5 percent level. Note: See table I for source codes. Results of cross-country regressions of difference in percentile rank between each poll of experts and the corresponding question from the World Business Environment Survey on an indicator variable taking the value I if the government of a country is left of center, 2 if it is center, and 3 if it is right of center. Percentile ranks are on a scale from 0 to 100, based on the sample of countries common to each pair of sources. The table reports the slope coefficient and t-statistic. 'Uses a question from the Gallup Millennium Survey instead of World Business Environment Survey. Source: Authors' calculations based on sources listed in table 1. Kazufmannt, Kraav, and Mastrnzzi 275 than for a center government. Moreover, in all cases the ideology variable explains only a trivial fraction of the difference in assessments between polls and surveys, suggesting that the importance of ideological biases in polls is quite small overall. IV. MARGINS OF ERROR AND CLASSIFYING COUNTRIES ACCORDING TO GOVERNANCE PERFORMANCE Margins of error are not unique to subjective indicators but are pervasive in all efforts to measure governance. The margins of error complicate the use of governance indicators for classifying countries according to governance perform- ance. Classifications based on individual indicators or even on a single aggregate indicator inevitably run the risk of misclassifying countries due to the margins of error inherent in all indicators. Margins of Error Are Not Uniquie to Subjective Data One of the strengths of the governance indicators reported here is the ability to construct explicit margins of error for the estimates of governance for each country. These margins of error are not unique to subjective or perceptions- based measures of governance, however, but apply to most other measures of institutional quality and to many other socioeconomic indicators as well. That measurement error is pervasive is obvious from the range of "preliminary" estimates of basic variables such as real GDP growth produced even in countries with high-quality statistical systems. Consider, for example, recent efforts to construct measures of governance that rely on objective and quantifiable data rather than exclusively on perceptions- based data sources. Knack and Kugler (2002) argue that variables such as waiting time to obtain a telephone line and number of telephone faults can serve as proxies for public administrative capacity, that degree of government reliance on trade taxes can serve as a proxy for the ability of the government to broaden its tax base, or that volatility in budgetary expenditure and revenue shares are indicative of a volatile policy environment. Clague and others (1999) argue that the proportion of currency in circulation held in the banking system is a good proxy for protection of property rights. Djankov and others (2002, 2003) use cross-country data on the number of administrative procedures required to start a business and the number of legal procedures required to collect an unpaid debt to capture the complexity of the regulatory and legal environment. Although such measures can in principle provide an accurate measure of the underlying concept they attempt to assess, their usefulness as a measure of broader notions of governance depends on how well the underlying concept corresponds to such broader ideas of governance. For example, the number of procedures required to start a business may not be a good indicator of the complexity or burden of regulation in other areas. Similarly, the willingness of individuals to hold currency in banks reflects confidence in a very particular set 276 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 of property rights but does not necessarily capture other dimensions of property rights protection, such as confidence in the police and judicial system. This is not a specific drawback of such objective measures. All measures, sub- jective and objective, are necessarily imperfect proxies for broader notions of gov- ernance. But it does reinforce the importance of margins of error for objective indicators as well, to the extent that they are used as proxies for broad concepts of governance such as those measured here using subjective data.'9 A simple calculation can give a sense of the order of magnitude of margins of error for objective indicators. Suppose that there are two noisy indicators y on a common unobserved concept of governance, g:y, = g + Fi, i = 1,2. If the variance of the unobserved measure of governance is normalized to be one, the correlation between the two observed indicators is p = [(1 + ±r 2)(1 + ur22)]-l/2. Suppose that indicator 1 is one of the subjective governance indicators presented here, for which the variance of the meas- urement error, cr12, is known, and that indicator 2 is one of the objective indicators already described. Then from the observed correlation between the two indicators, the variance of measurement error in the objective indicator, r22, can be inferred. This calculation is done for several objective governance indicators (table 6). The implied standard deviation of measurement error in the objective indicator is calculated under three assumptions: that the estimate of the standard devia- tion of measurement error in the subjective indicator is correct (assumption A), that the subjective and objective indicators have the same standard deviation of measurement error (assumption B), and that the standard deviation of measure- ment error in the subjective indicator is twice as large as that in the objective indicator (assumption C). The actual standard deviation of measurement error for the subjective indicator is also calculated, computed as the average across all countries of the country-specific standard errors in our governance indicators. For all indicators and for all three sets of assumptions, the implied standard deviation of measurement error in the objective indicators is much higher than the corresponding standard deviation of the subjective governance indicator. Under benchmark assumption A, the implied margin of error for the objective indicators is 7-15 times larger than that of the subjective indicators. This clearly exaggerates the difference in the precision of subjective and objective indicators because it compares a single objective indicator with an aggregate of several subjective measures, and as discussed, aggregation improves precision. But this is only part of the story. The government effectiveness and regulatory quality indicators have a median of six sources per country, and the rule of law indicator has a median of eight sources. This can explain why the standard deviation of measurement error of the objective sources might be 6 = 2.4 to 8 = 2.8 times higher than that of the corresponding subjective indicators, but it still cannot explain 19. These margins of error should, of course, also reflect measurement error in the raw data on which they are based-for example, the nontrivial measurement error in macroeconomic variables such as the money supply or the composition of public expenditures. TABLE 6. Imputed Margins of Error for Objective Governance Indicators Corresponding Absolute Implied margin of error for objective indicatora Actual margin of subjective indicator value of error for Objective indicator for 2002 correlation Assumption A Assumption B Assumption C subjective indicator Telephone wait time Governance effectiveness 0.56 1.43 0.88 0.58 0.21 Phone faults Governance effectiveness 0.32 2.92 1.47 1.00 0.21 Trade tax revenue Governance effectiveness 0.50 1.68 1.00 0.67 0.21 Budgetary volatility Governance effectiveness 0.50 1.68 1.00 0.67 0.21 Revenue source volatility Governance effectiveness 0.49 1.71 1.01 0.67 0.21 Contract intensive money Rule of law 0.57 1.39 0.86 0.57 0.19 Contract enforcement Rule of law 0.40 2.25 1.22 0.82 0.19 Regulation of entry Regulatory quality 0.50 1.67 1.00 0.66 0.22 Aggregate objective indicator Governance effectiveness 0.73 0.88 0.60 0.39 0.21 aAssumption A: the estimate of the standard deviation of measurement error in the subjective indicator is correct. Assumption B: the subjective and objective indicators have the same standard deviation of measurement error. Assumption C: the standard deviation of measurement error in the subjective indicator is twice as large as that in the objective indicator. Source: Authors' calculations based on sources listed in table 1. 278 THE WORLD BANK ECONOMIC REVIEW, VOL. I8, NO. Z all of the difference in the precision of the indicators. Even with an aggregated objective indicator for government effectiveness, the implied standard deviation of measurement error is still four times as large for the objective indicator as for the subjective one (last row of table 6), though the benefits of aggregation should be roughly comparable for the two indicators, with a median of five sources per country for the objective indicator and six for the subjective indicator. Assumptions B and C are designed to be more favorable to the precision of the objective indicators. Assumption B discards the information in the margins of error that were constructed for the subjective indicator and simply makes the neutral assumption that the subjective and the objective indicators have the same standard deviation of measurement error. This reduces the implied stand- ard deviation of measurement error for the objective indicator relative to the benchmark assumption A, but it remains large at 0.6 for the composite objective indicator and higher for the individual indicators. Assumption C weights things even further in favor of the objective indicators, by assuming that the objective indicator is twice as precise as the subjective indicator. Yet substantial estimates of the standard deviation of measurement error remain, on the order of 0.4 and higher for individual objective indicators. This simple calculation underscores and helps quantify the intuitive notion that all governance indicators, not just the subjective ones constructed here, are subject to nontrivial margins of error. Care should be taken in making govern- ance comparisons based on any such measures, and wherever possible it is desirable to construct explicit margins of error to aid in these comparisons. Margins of Error and the U.S. Millennium Challenge Account To illustrate the importance of taking margins of error into account when classifying countries by level of governance, this section examines the criteria for country eligibility for the new aid program of the U.S. government, the Millennium Challenge Account (MCA). The MCA allocates funds to countries that "govern justly," "invest in people," and "promote economic freedom," placing governance issues center stage (U.S. Department of State 2002).20 The alloca- tion criteria draw heavily on a number of cross-country measures of the quality of governance, including five of the six governance indicators presented here (all but the political stability indicator). The first round of countries eligible for MCA funds was selected using these criteria in 2004. The allocation criteria are designed to ensure that funds go to low-income countries with relatively sound policies and institutions. The process starts with the 74 countries that are eligible for concessional lending from the International 20. Details on the NlCA can be found online at www.mca.gov. See Radelet (2003) for a detailed discussion of the NMCA. The mCA is not the only example of explicit use of governance indicators. For example, the World Bank uses its internal assessments of countries' policy performance, the Country Policy and Institutional Assessment, to allocate concessional lending from its International Development Association. Kaufmcann, Kraay, anid Mastruzzi 279 Development Association that have per capita incomes of less than $1,435 in 2001 21 This set of countries is rated according to 16 performance criteria covering three dimensions of performance: governing justly (6 criteria), investing in people (4 criteria), and promoting economic freedom (6 criteria). Four of the governance indicators constructed here (voice and accountability, government effectiveness, rule of law, and control of corruption) are performance indicators under the "governing justly" performance dimension; a fifth governance indicator, regulatory quality, is included under promoting economic freedom. To qualify for assistance, countries must be in the top half of all potentially eligible countries on at least half of the performance criteria under each of the three dimensions of performance. Countries must also be in the top half of potentially eligible countries according to the corruption rating from the gov- ernance indicators described in this article. This rule is designed to ensure that resources are channeled to countries that are performing well in a variety of dimensions of governance and in which corruption especially is relatively low. Though an objective and monitorable set of criteria for determining eligibility is highly desirable, both for allocating aid and for creating clear incentives for potential aid recipients, the substantial margins of error associated with govern- ance estinmates mean that it is difficult to assign countries to a definitive perform- ance category based on their estimated level of governance. Recognizing this, the MCA criteria do not require countries to pass the median hurdle on all indicators. However, the allocation rule requires countries to score in the top half of all relevant countries on the corruption indicator, which would constitute a hard hurdle for eligibility. Corruption is surely an important factor in allocating aid, but a simple in-or-out rule runs the risk of misclassifying some countries because of the large margins of error. The MNICA fact sheet recognizes this possibility and provides some flexibility for softening this hard hurdle (U.S. Department of State 2002). To get a sense of the risk of misclassifying countries using a single measure such as corruption, the 74 potential MICA countries are ranked according to their scores on the 2002 control of corruption indicator developed here. This ranking is plotted on the x axis (marked by diamonds), and the 90 percent confidence intervals for each country are shown as a vertical line on the right hand y axis (figure 3). For a substantial fraction of countries, the median score (indicated as a heavy black horizontal line on the graph) falls within the 90 percent con- fidence interval. For only 11 of the 37 countries in the bottom half of the sample 21. A number of countries with per capita income grater than $1,435 are currently eligible for ID\ lending under the small island economies exception, but these will not be eligible for the M\c.A during the first vear. The group of 74 countries is based on data on ida eligibility, available online at www.worldbank.org/ida and per capita gross national income in U.S. dollars in 2001 using World Bank Atlas exchange rates, availahle in the World Bank's World Development Indicators. In the second year eligibility will expand to all countries with per capita incomes of less than $1,435, and in the third year to all countries with per capita incomes of less than $2,975. 280 THF WORLD BANK FCONOMIC REVIEW', VOL. 1 8, NO. 2 FIGURE 3. Using Governance Indicators to Allocate Aid for the Millennium Challenge Account * @ _ -0.5 0.75 A- - 0 o ~~ ~~~~~~~~A A -0 D 0.5 ' on g+++++++W++ogsOgliiiigtpK2; Median corruption score - .2 AM 0.25- A " 0 _ -R-^ ^ ' ' ' ' ' ' ' 5-2 0 ~~~~~~~~~~~~~~~~~~~~~-2 5 0 0.1 0.2 0.3 0.4 05 0.6 0.7 0.8 0.9 1 Country rank (0-1) Note: All 74 countries eligible for the first round of the MCA are shown by their corruption rank (on the horizontal axis) and their corruption scores (on the right-hand vertical axis). Diamonds indicate the corruption score, and the vertical lines for each country indicate the 90 percent confidence interval for corruption. The squares, triangles, and circles indicate the probability (on the left axis) that a country's corruption score is in the top half of the sample. Squares indicate countries where this probability is less than 25 percent, circles where it is between 25 percent and 75 percent, and triangles where it is greater than 75 percent. Source: Authors' calculations based on sources listed in table 1. is the 90 percent confidence interval fully below the median score. For the remaining 26 countries for which the confidence intervals include the median, there is at least a 5 percent probability that their scores are actually in the top half of the sample. Similarly, only 17 of the 37 countries in the top half of the sample have confidence intervals that are fully above the median score, whereas for the remaining 20 countries there is at least a 5 percent probability that their scores are actually in the bottom half of the sample. Thus, for the majority of the 74 countries there is a nontrivial probability that they could be mistakenly classified in the wrong half of the sample based on their point estimates of governance alone. The probability that a country's true unobserved level of governance falls in the top half of the sample is plotted on figure 3. Not surprisingly, for the worst rated countries the probability of falling into the top half of the sample is close to zero (marked by squares). Similarly, Kaufmann, Krmay, and Mastruzzi 281 the best rated countries almost certainlv belong in the top half (the circles). However, for a large intermediate range of 20 countries there is a nontrivial probability that they belong in either the top or bottom half of the sample. For these countries it seems prudent to rely on additional sources of information in making INMCA eligibility decisions. This also underscores the need for some flex- ibility in the POCA allocation rule and the need for this flexibility to be symmetric. Not only should countries that barely miss the list of better performers be given special consideration, as currently proposed in the NICA fact sheet, but countries that barely make the list of better performers also merit further scrutiny. V. GLOBAL TRENDS IN GOVERNANCE This section presents the limited evidence available on trends in global averages of governance. Because the means of the governance indicators were rescaled to equal zero in each period, the aggregate indicators are by construction informa- tive about countries' relative positions around the average but uninformative about trends in global averages of governance. To discuss trends in governance worldwide requires going back to the underlying sources of governance data.22 For the six dimensions of governance, data are reported from up to four major underlying sources available for each of the four periods 1996, 1998, 2000, and 2002: Economist Intelligence Unit Country Risk Service, DRI Country Risk Review, International Country Risk Guide, and Global Competitiveness Report (table 7). The first three are polls of experts that cover a large set of countries with a consistent methodology from year to year and can therefore be expected to be informative about overall trends. Global Competitiveness Report covers a smaller set of countries, but it is the only survey of individuals that is available in all four periods. To maximize comparability across sources and over time, the focus is for the first three sources on the set of countries common to these three sources for all periods. For the fourth source the focus is on the smaller set of countries available in each period and a small number of survey questions that have been consistently available over time. The underlying data have been rescaled to run from zero to one, and for each source and governance component the score is reported on the same question or average of questions used in the aggregate indicator. Table 7 reports the average across all countries of each source in each year, the standard deviation across countries for each source, and the t-statistic associated with a test of the null hypothesis that the world average score is the same in the first and last periods. Several sources report substantial declines in world averages of the six dimensions of governance. The DRI Country Risk 22. By conistruction, the standard deviation of the aggregate governance estimates is equal to one in each period, and so these aggregate indicators also cannot be used to assess whether there has been global convergence in governance. TABLE 7. Global Trends in Governance, Selected Sources t-Statistic for World Average SD across Countries Mean Difference Number of from Last Year ol Indicator and Source Countries 1996 1998 2000 2002 1996 1998 2000 2002 to First Yeara Voice and accountability DRI 112 - - - - - - _ EIU 112 0.41 0.42 0.42 0.46 0.30 0.32 0.31 0.28 1.3 PRS 112 0.67 0.66 0.66 0.66 0.23 0.25 0.25 0.25 -0.3 GCS - - - - - - - - - - Political stability DRI 102 0.82 0.81 0.74 0.70 0.18 0.18 0.24 0.27 -4.1 EIU 102 0.55 0.53 0.58 0.55 0.29 0.29 0.30 0.28 0.1 PRS 102 0.80 0.75 0.74 0.76 0.14 0.19 0.17 0.13 -2.0 GCS - - - - - - - - - - Government effectiveness DRI 102 0.59 0.58 0.50 0.48 0.27 0.26 0.30 0.30 -2.9 EIU 102 0.41 0.47 0.46 0.41 0.30 0.24 0.24 0.30 0.0 PRS 102 0.63 0.67 0.59 0.67 0.24 0.12 0.10 0.15 1.5 GCS 51 0.45 0.48 0.42 0.31 0.14 0.15 0.11 0.13 -5.2 Regulatory quality DRI 106 0.83 0.84 0.79 0.76 0.14 0.14 0.18 0.20 -3.0 EIU 106 - - - - - - - - - PRS 106 - 0.63 0.60 0.76 - 0.20 0.22 0.21 4.6 GCS 51 0.51 0.58 0.59 0.58 0.14 0.15 0.15 0.15 2.5 Rule of lauw DRI 102 0.73 0.73 0.67 0.65 0.20 0.20 0.23 0.24 -2.6 EIU 102 0.50 0.53 0.51 0.54 0.27 0.29 0.30 0.27 1.1 PRS 102 0.77 0.67 0.68 0.66 0.20 0.24 0.22 0.23 -3.8 GCS 51 0.67 0.64 0.59 0.21 0.23 0.24 -1.7 Control of corruption DRI 102 0.61 0.60 0.54 0.53 0.25 0.26 0.30 0.31 -1.8 EIU 102 0.37 0.37 0.36 0.38 0.31 0.33 0.32 0.32 0.2 PRS 102 0.62 0.52 0.48 0.42 0.20 0.22 0.22 0.20 -7.2 GCS 51 0.56 0.57 0.59 0.57 0.13 0.14 0.11 0.14 0.3 - Not available. Note: See table I for source codes. Reports trends in cross-country averages of selected components of governance indicators. For EIU, DRI, and PRS, the sample of countries is restricted to those that appear in all three sources in all four periods, to ensure comparability over time and across indicators. For GCS averages are reported across countries of selected individual questions available in all four rounds, with attention restricted to countries available in all four periods. 't-Statistic associated with a simple test for equality of global averages in the first and last available periods. Source: Authors' calculations based on sources listed in table 1. 284 THE WORLD BANK ECONOMIC REVIEW, VOL. i8, NO. Z Review in particular shows statistically significant declines in all five indicators in which it appears. International Country Risk Guide reports significant declines in world averages for political stability, rule of law, and control of corruption but improvements in regulatory quality and government effective- ness. Among polls, the Economist Intelligence Unit Country Risk Service alone consistently does not report any significant trend. Finally, the single survey, Global Competitiveness Report, reports significant deterioration in political stability and government effectiveness and smaller declines in rule of law. It is not clear how much importance to ascribe to these trends in world averages. On one hand, these statistics represent the only information available on trends over time and so should be taken seriously. On the other hand, the disagreement among sources on the direction of global trends is striking-overall 8 averages improve or remain the same and 11 decline. Looking only at statistically significant changes, however, shows that declines in governance averages outnumber increases 10 to 2 (and both of the statistically significant increases are in regulatory quality). All that can be cautiously concluded, therefore, is that there is certainly no evidence of any significant improvement in governance worldwide and, if any- thing, that the evidence is suggestive of a deterioration in key dimensions such as control of corruption, rule of law, political stability, and government effective- ness. It can therefore be safely concluded that the (relative) governance estimates for a country do not underestimate absolute trends because there is no evidence of a worldwide improvement. VI. COMPARISONS WITH TRANSPARENCY INTERNATIONAL'S CORRUPTION PERCEPTIONS INDEX Transparency International's pioneering Corruption Perceptions Index (cPI), like the indicators presented here, is an average of ratings reported by a number of perceptions-based sources. In content, the primary differences are that the CPI relies on a subset of the sources used here and it treats multiple years of data from the same source as separate sources in the aggregation procedure.23 In particular, the 2002 cPi is based on 10 distinct data sources but uses between two and three years of data from some of them and treats them as separate sources, to arrive at a total of 15 components. In contrast, the control of corruption indicator constructed here is based on 14 distinct sources, using only data from 2002 and without using multiple years from the same source. The CPI also differs in its approach to aggregation (see Lambsdorff 2002 for details). It uses a percentile-matching approach to put data in common units, a simple average of rescaled scores as the estimate of corruption for each country, and a nonparametric bootstrapping procedure (discussed later) to generate measures of precision for the aggregate indicator. The control of corruption indicator, by 23. See Kaufmann and others (2003) for more detailed discussion of the two indicators. Kaufmann, Kraay, and Mastruzzi 285 contrast, uses an unobserved-components model to transform individual sources into common units (this is the role of the s'S and 3's in equation 1), construct an appropriately weighted average of sources as the aggregate score, and produce margins of error to summarize the precision of the estimates of governance. The estimates of corruption based on the two indicators come to very similar results, with a correlation above 0.9. However, the 2002 cPi covers only 102 countries, because it discards countries with fewer than three data sources. In contrast the control of corruption indicator covers 195 countries, or nearly twice as many. The margins of error generated by the two approaches are similar as well: The average width of a 90 percent confidence interval is 0.94, or 9.4 percent of the range of units from 0 to 10, for the cPI, and 0.71, or 14 percent of the range from -2.5 to 2.5, for the control of corruption indi- cator. These figures are not fully comparable, however, because the control of corruption indicator covers many more countries, many with only one or two sources of data, and hence would be expected to have somewhat larger margins of error. If only the set of 102 countries appearing in the cPi are included, the average width of a 90 percent confidence interval for the control of corruption indicator is 0.52, or 10 percent of the range of this index-almost identical to the margins of error in the cPi relative to its scale of units. However, the apparent similarity in the precision of the two indicators is likely to be the result of two offsetting biases in the Transparency International meth- odology. First, the bootstrapping approach understates the margins of error by overstating the precision of estimates of corruption for countries with relatively few sources. The intuition for this is straightforward. When the number of obser- vations is small, bootstrapped standard errors will understate true standard errors because the observed data are less likely to span the full range of variation in the underlying data generating process. This suggests that the cPI margins of error should spuriously be smaller than those reported for the control of corruption indicator here. In particular, for a country with only three data sources, numerical simulations suggest that the cPI approach will understate the standard error (over- state the precision) of the corruption estimate by about 40 percent. Second, the estimates of corruption produced by the unobserved-components model used here are a precision-weighted average of individual sources, whereas the cPI approach is based on a simple average. Because precision weighting improves the accuracy of the estimates of corruption, cPI margins of error should correctly be expected to be larger than those reported here. However, according to the estimates in this article, the differences in estimated precision across sources in the cPI are sufficiently small that the benefits of precision weighting are relatively small. VII. CONCLUSION This article presents substantially expanded and updated indicators of six dimen- sions of governance for 1996, 1998, 2000, and 2002. Large numbers of individual 286 THE WORLD BANK ECONOMIC REVIEW, VOL. 1 8, NO. 2 sources were aggregated, both expanding country coverage and improving the precision of the aggregate indicators. Nevertheless, margins of error remain sub- stantial relative to the units in which governance is measured-especially import- ant to consider when classifying countries according to levels of governance. An important methodological observation is that there are few alternatives to subjective, experiential data for measuring certain dimensions of governance. Objective indicators of governance, although also very useful, have implicit mar- gins of error on the same order of magnitude as those associated with subjective aggregates. The importance of ideological biases in the perceptions data from polls of experts was empirically investigated and for the most part discounted. Finally, the limited evidence on trends over time in governance worldwide is difficult to interpret, but it can be said with some confidence that there is little evidence of improvements in global governance over the period considered. As this research project on measuring cross-country differences in governance continues, additional data should become available to enable further improve- ments in precision. The broader objective is to provide a set of monitorable indicators of governance for individual countries to benchmark themselves against other countries and over time. Limitations will remain, however, in what can be achieved with this kind of cross-country, highly aggregated data. Such data cannot substitute for in-depth, country-specific governance diagnos- tics as a basis for policy advice to improve governance in a particular country. REFERENCES Beck, Thorsten, George Clarke, Alberto Groff, Phillip Keefer, and Patrick Walsh. 2001. "New Tools in Comparative Political Economy: The Database of Political Institutions." World Bank Economic Revietr' 15(1):165-76. Cingranelli, David, and David Richards. 2001. Coding Government Respect for Human Rights. Manual version 1. Binghamton, N.Y.: State University of New York. Clague, Christopher, Philip Keefer, Stephen Knack, and Mancur Olson. 1999. "Contract-Intensive Money: Contract Enforcement, Property Rights, and Economic Performance." Journal of Economic Growth 4(2):185-211. Di Tella, Rafael, and Ernesto Shargrodsky. 2003. "The Role of Wages and Auditing during a Crackdown on Corruption in the City of Buenos Aires." Journal of Lawv and Economics 46(1):269-92. Djankov, Simeon, Rafael La Porta, Florencio Lopez-de-Silanes, and Andrei Shleifer. 2002. "The Regula- tion of Entry." Quarterly Journal of Economics 117(1):1-37. - . 2003. "Courts." Quarterly Journal of Economics 118(2):453-517. Efron, Bradley, and Carl Morris. 1971. "Limiting the Risk of Bayes and Empirical Bayes Estimators- Part 1: The Bayes Case." Journal of the American Statistical Association 66:807-15. 1972. "Limiting the Risk of Bayes and Empirical Bayes Estimators-Part 2: The Empirical Bayes Case." Journal of the American Statistical Association 67:130-39. Goldherger, Arthur. 1972. "Maximum Likelihood Estimation of Regressions Containing Unobservable Independent Variables." International Economlic Review 13(1):1-15. Golden, Miriam, and Lucio Picci. 2003. "Proposal for a New Measure of Corruption, and Tests Using Italian Data." Working paper, University of Bologna, Italy. Hellman, Joel, and Daniel Kaufmann. 2004. "The Inequality of Influence." In Trust in Transition, ed. J. Kornai and S. Rose-Ackerman. New York: Palgrave-Macmillan. Kaufnzann, Kraav, andl Mastruzzi 287 Kaufmann, Daniel, and Aart Kraay. 2002. "Growth without Governance." Econoonzia 3(1):169-230. Kaufmann, Daniel, Aart Kraav, and Pablo Zoido-Lobat6n. 1999a. "Aggregating Governance Indicators." Policy Research Working Paper 2195. World Bank, Washington, D.C. Available online at www.worldbank.org/wbi/governance/pubs/aggindicators.html. .1999b. "Governance Matters." Policy Research Working Paper 2196. World Bank, Washington, D.C. Available online at www.worldbank.org/wbi/governance/pubs/govmatters.html. 2002. 'Governance Matters 11-Updated Indicators for 2000/01." Policv Research Working Paper 2772. World Bank, Washington, D.C. Available online at wwvw.worldbank.org/wbi/governance/ pubs/govmatters3.html. - . 2003. "Governance Matters Ill: Governance Indicators for 1996-2002." Policy Research Work- ing Paper 3106. World Bank, Washington, D.C. Knack, Steve, and Mark Kugler. 2002. "Constructing an Index of Objective Indicators of Good Govern- ance." Working paper, World Bank, Washington, D.C. Lambsdorff, Johann. 2002. "Background Document to 2002 Corruption Perceptions Index." Available online at ww' . .1 J. .. -l.. '-''_'cpi_fd.pdf. Radelet, Steven. 2003. Challenging Foreign Aid: A Policvrnaker's Guide to the Millenniunm Challeinge A4ccount. Washington, D.C.: Center for Global Development. U.S. Department of State. 2002. "Fact Sheet: The Millennium Challenge Account." Washington, D.C. Wirtz, Ronald. 2002. "Gross Domestic Product: Understanding News from Noise." Region. Federal Reserve Bank of Minneapolis. Available online at http://minneapolisfed.org/pubs/region/02-06/ gdp.cfm. HUMAN DEVELOPMENT REPORT 2004 . - . S ..... (- I/li J 'sitlllt i dh li mitv (a Iil/h:c ii a ,aid siv diill~ ~ ~~hla w ntie ... Ins Devldopmaes,tl -Programm ievialccalc,ala I,a _- I~~~~~~~(, . l,- +,ptii it ( )pening iwh anl .i'll, 'i ,. III tVitall III] L, h tI ecu hulin llI.i 3|e- - 1l\l}~~~,whyw "ld .¢l ithuil.l lihli 1 - ), Nold l [auicaw Ainl:t,, l See, II xaluill"l i lItc Ill.at tIII- *1 )lllit- , " globl .dio.ti"lrl lui.\, I'lo ulght (o the ~~~~~~~'0 1.l l -t 1. 1 -11, * ag-'ld III I I l ai ld (uniltlri,sa.lik . Th sein,litStle1l( v r ~ ~~~~~~~ l)i wng (,n I1 . lat I-i nxll-lll I1f an 1-al go ernelil1ts xi(ll( ound .h ")I Id,II Ic w - /¢b1-(,//pl{>f)ti, y-- eNid,tluc-ha.scd I, "I'titilit,onal. L'i idi(ial, and sis(eeto-11, m1it _i ~ ~ ~ ~ ~ 1 .ll -ljl~ ~~~~~I ' ptIn I ,( .l>i lilitlo-.d ltlselt) .and( llllltlcllt1,11 d ... eI... 1.Icv ill ffi xi~~~~~~~~~~~~~~,I wN I IIIl'> I' lli- c I) I, d.I, t II I)l a, 11I ill I: g I I 1 WORLD DEVELOPMENT REPORT 200S World Bank ,., ~~~~~~~~(1- riM t;# IdI:i,p,,.l r:z., 21I0d Inio l(l l ll axt, .n depll lind e(t'laOldit; o th _AB '- 1 -All - i ( 1\ I. ..1. Ili( t |1l At sha (pe ninen 1 11 t I-I | li, ic :and behaNi o, 1. affecti ng tli, 11111.1tiliCilt~ ~ ~ ~ ~ ~ il lil,i 11 -a'1111('(1 t ltr , 1 and to ililtln't ill, b 3 9 >- illsebilile~~~~~~~~lin ti'ls.l- 4 lle'l moi'lit(ieY 'pull g szIll, .111,1 recluce poNctl-. 11(i .~~~~~~~~~~~R 1- t lw t( 11 te 1 opl)-Il ... iti,s .il I1( (lial(le gt' (.k, "I( b,R (S'k-Nciopting z i -= :~~~~~~~nt1 : old (lgge l, j)I:I,f(;icll'l.lgoi- II.1 ') -I, ltinigl .wgl- 1 buli)ld( ,n - S .^L ~~~~~~i li l il.l'i'ml im- il) p)asl\-t 1,V(1 . 1D,w 1pincm Rep t .l- .nd di).m ,ii n