80517 Volume 25 • Number 2 • 2011 ISSN 0258-6770 (PRINT) ISSN 1564-698X (ONLINE) THE WORLD BANK ECONOMIC REVIEW Volume 25 • 2011 • Number 2 THE WORLD BANK ECONOMIC REVIEW Has India’s Economic Growth Become More Pro-Poor in the Wake of Economic Reforms? Gaurav Datt and Martin Ravallion Are The Poverty Effects of Trade Policies Invisible? Monika Verma, Thomas W. Hertel, and Ernesto Valenzuela Corruption and Confidence in Public Institutions: Evidence from a Global Survey Bianca Clausen, Aart Kraay, and Zsolt Nyiri Agricultural Distortions in Sub-Saharan Africa: Trade and Welfare Indicators, 1961 to 2004 Johanna L. Croser and Kym Anderson Thresholds in the Finance-Growth Nexus: A Cross-Country Analysis Hakan Yilmazkuday The Value of Vocational Education: High School Type and Labor Market Outcomes in Indonesia David Newhouse and Daniel Suryadarma Pages 157–359 Disability and Poverty in Vietnam Daniel Mont and Nguyen Viet Cuong www.wber.oxfordjournals.org 2 THE WORLD BANK ECONOMIC REVIEW editors Alain de Janvry and Elisabeth Sadoulet, University of California at Berkeley assistant to the editor Marja Kuiper editorial board Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on August 19, 2013 Harold H. Alderman, World Bank (retired) William F. Maloney, World Bank Pranab K. Bardhan, University of California, David J. McKenzie, World Bank Berkeley Jaime de Melo, University of Geneva Scott Barrett, Columbia University, USA Juan-Pablo Nicolini, Universidad Torcuato di Asli Demirgüç-Kunt, World Bank Tella, Argentina Jean-Jacques Dethier, World Bank Nina Pavcnik, Dartmouth College, USA Quy-Toan Do, World Bank Vijayendra Rao, World Bank Frédéric Docquier, Catholic University of Martin Ravallion, World Bank Louvain, Belgium Jaime Saavedra-Chanduvi, World Bank Eliana La Ferrara, Università Bocconi, Italy Claudia Paz Sepúlveda, World Bank Francisco H. G. Ferreira, World Bank Joseph Stiglitz, Columbia University, USA Augustin Kwasi Fosu, United Nations Jonathan Temple, University of Bristol, UK University, WIDER, Finland Romain Wacziarg, University of California, Paul Glewwe, University of Minnesota, Los Angeles, USA USA Dominique Van De Walle, World Bank Ann E. Harrison, World Bank Christopher M. Woodruff, University of Philip E. Keefer, World Bank California, San Diego Justin Yifu Lin, World Bank Yaohui Zhao, CCER, Peking University, Norman V. Loayza, World Bank China The World Bank Economic Review is a professional journal used for the dissemination of research in development economics broadly relevant to the development profession and to the World Bank in pursuing its development mandate. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantita- tive development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. Consistency with World Bank policy plays no role in the selection of articles. The Review is managed by one or two independent editors selected for their academic excellence in the field of development economics and policy. The editors are assisted by an editorial board composed in equal parts of scholars internal and external to the World Bank. World Bank staff and outside researchers are equally invited to submit their research papers to the Review. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 25 † 2011 † Number 2 Has India’s Economic Growth Become More Pro-Poor in the Wake of Economic Reforms? 157 Gaurav Datt and Martin Ravallion Are The Poverty Effects of Trade Policies Invisible? 190 Monika Verma, Thomas W. Hertel, and Ernesto Valenzuela Corruption and Con�dence in Public Institutions: Evidence from a Global Survey 212 Bianca Clausen, Aart Kraay, and Zsolt Nyiri Agricultural Distortions in Sub-Saharan Africa: Trade and Welfare Indicators, 1961 to 2004 250 Johanna L. Croser and Kym Anderson Thresholds in the Finance-Growth Nexus: A Cross-Country Analysis 278 Hakan Yilmazkuday The Value of Vocational Education: High School Type and Labor Market Outcomes in Indonesia 296 David Newhouse and Daniel Suryadarma Disability and Poverty in Vietnam 323 Daniel Mont and Nguyen Viet Cuong 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 25, 3 Issues, 2011): Institutions—Print edition and site-wide online access: £168/ $252/E252, Print edition only: £154/$231/E231, Site-wide online access only: £140/$210/E210; Corporate—Print edition and site-wide online access: £251/$376/E376, Print edition only: £230/$345/E345, Site-wide online access only: £209/$314/E314; Personal—Print edition and individual online access: £43/$64/E64. US$ rate applies to US & Canada, EurosE applies to Europe, UK£ applies to UK and Rest of World. There may be other subscription rates available; for a complete listing, please visit www.wber.oxfordjournals.org/subscriptions. Readers with mailing addresses in non-OECD countries and in socialist economies in transition are eligible to receive complimentary sub- scriptions on request by writing to the UK address below. Full prepayment in the correct currency is required for all orders. Orders are regarded as �rm, and payments are not refundable. Subscriptions are accepted and entered on a complete volume basis. Claims cannot be con- sidered more than four months after publication or date of order, whichever is later. All subscriptions in Canada are subject to GST. Subscriptions in the EU may be subject to European VAT. If registered, please supply details to avoid unnecessary charges. For subscriptions that include online versions, a proportion of the subscription price may be subject to UK VAT. Personal rates are applicable only when a subscription is for individual use and are not available if delivery is made to a corporate address. Downloaded from http://wber.oxfordjournals.org/ at International Monetary Fund on August 19, 2013 The current year and two previous years’ issues are available from Oxford University Press. Previous BACK ISSUES: volumes can be obtained from the Periodicals Service Company, 11 Main Street, Germantown, NY 12526, USA. E-mail: psc@periodicals.com. Tel: (518) 537-4700. Fax: (518) 537-5899. CONTACT INFORMATION: Journals Customer Service Department, Oxford University Press, Great Clarendon Street, OxfordOX2 6DP, UK. E-mail: jnls.cust.serv@oup.com. Tel: þ44 (0)1865 353907. Fax: þ 44 (0)1865 353485. In the Americas, please contact: Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oup.com. Tel: (800) 852-7323 (toll-free in USA/Canada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals Customer Service Department, Oxford University Press, Tokyo, 4- 5-10-8F Shiba, Minato-ku, Tokyo, 108-8386, Japan. E-mail: custserv.jp@oup.com. Tel: þ 81 3 5444 5858. Fax: þ 81 3 3454 2929. POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published three times a year, in February, June, and October, by Oxford University Press for the International Bank for Reconstruction and Development/THE WORLD BANK. Send address changes to The World Bank Economic Review, Journals Customer Service Department, Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Periodicals postage paid at Cary, NC and at additional mailing of�ces. Communications regarding original articles and editorial management should be addressed to The Editor, The World Bank Economic Review, The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland. ENVIRONMENTAL AND ETHICAL POLICIES: Oxford Journals, a division of Oxford University Press, is committed to working with the global community to bring the highest quality research to the widest possible audience. Oxford Journals will protect the environment by implementing environmentally friendly policies and practices wherever possible. Please see http://www.oxfordjournals.org/ethicalpolicies.html for further information on environmental and ethical policies. DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org. PERMISSIONS: For information on how to request permissions to reproduce articles or information from this journal, please visit www.oxfordjournals.org/jnls/permissions. ADVERTISING: Advertising, inserts, and artwork enquiries should be addressed to Advertising and Special Sales, Oxford Journals, Oxford University Press, Great Clarendon Street, Oxford, OX2 6DP, UK. Tel: þ 44 (0)1865 354767; Fax: þ 44(0)1865 353774; E-mail: jnlsadvertising@oup.com. DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are those of the respective authors and contributors and not of the International Bank for Reconstruction and Development/THE WORLD BANK or Oxford University Press. Neither Oxford University Press nor the International Bank for Reconstruction and Development/THE WORLD BANK make any representation, express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal responsibility or liability for any errors or omissions that may be made. The reader should make her or his own evaluation as to the appropriateness or otherwise of any experimental technique described. The World Bank Economic Review is printed on acid-free paper that meets the minimum require- PAPER USED: ments of ANSI Standard Z39.48-1984 (Permanence of Paper). INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed and/or abstracted by CAB Abstracts, Current Contents/Social and Behavioral Sciences, Journal of Economic Literature/EconLit, PAIS International, RePEc (Research in Economic Papers), and Social Services Citation Index. COPYRIGHT # 2011 The International Bank for Reconstruction and Development/THE WORLD BANK All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without prior written permission of the publisher or a license permitting restricted copying issued in the UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Typeset by Techset Composition Limited, Chennai, India; Printed by Edwards Brothers Incorporated, USA. Has India’s Economic Growth Become More Pro-Poor in the Wake of Economic Reforms? Gaurav Datt and Martin Ravallion The extent to which India’s poor have bene�ted from the country’s economic growth has long been debated. A new series of consumption-based poverty measures spanning 50 years, including a 15-year period after economic reforms began in earnest in the early 1990s, is used to examine that issue. Growth has tended to reduce poverty, including in the postreform period. There is no robust evidence of more or less poverty responsiveness to growth since the reforms began, although there are signs of rising inequality. The impact of growth is higher when using poverty measures that reflect distribution below the poverty line and when using growth rates calculated Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 from household surveys rather than national accounts. The urban-rural pattern of growth matters for the pace of poverty reduction. However, in marked contrast to the period before the reforms, urban economic growth in the period after the reforms has brought signi�cant gains to the rural poor as well as the urban poor. India, poverty, inequality, economic growth. JEL codes: I32, O15, O40 There has been much hope that India’s economic reforms starting in the early 1990s would bring more rapid poverty reduction. Growth has cer- tainly accelerated, with GDP per capita rising at 4 –5 percent since 1991, up from barely 1 percent in the 1960s and 1970s and 3 percent in the 1980s. However, as research has shown, the sectoral pattern of growth matters to its impact on poverty in India. The green revolution stimulated pro-poor rural growth.1 In the past, both the urban and rural poor gained from growth in the rural sector, while urban growth had adverse Gaurav Datt (gdatt@worldbank.org) is a senior economist in the Economic Policy and Poverty Sector, South Asia Region, at the World Bank. Martin Ravallion (corresponding author; mravallion@worldbank. org) is director of the Development Research Group at the World Bank. The authors are grateful to Pranab Bardhan; Ann Harrison; Ashok Kotwal; Rinku Murgai; Abhijit Sen; Anand Swamy; participants at seminars at the University of Adelaide, University of California at Berkeley, and Monash University; and the journal editor and three referees for helpful comments. The authors are also grateful to Dandan Zhang for excellent research assistance. These are the views of the authors and should not be attributed to the World Bank. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. 1. Datt and Ravallion (1998) found that farm productivity growth reduced rural poverty. Earlier support for this view includes Ahluwalia (1978, 1985); van de Walle (1985); Bhattacharya, Coondoo, and Mukherjee (1991); and Bell and Rich (1994). Dissenting views include Saith (1981) and Gaiha (1995). THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 157 –189 doi:10.1093/wber/lhr002 Advance Access Publication February 15, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 157 158 THE WORLD BANK ECONOMIC REVIEW distributional effects in urban areas and no discernable impact on rural poverty (Ravallion and Datt 1996). The disappointing outcomes for the poor from nonfarm growth have also been traced to India’s socioeconomic inequalities in access to schooling.2 However, though past research points to the importance of rural economic growth for poverty reduction in India, postreform growth has not favored the rural sector. Several observers have pointed to both geographic and sec- toral divergence in India’s postreform growth (Bhattacharya and Sakthivel 2004; Jha 2000; Datt and Ravallion 2002; Pur�eld 2006). This has meant that much of the nonfarm economic growth bypassed the sectors and states where it would have had the most impact on poverty, based on a model calibrated to prereform data (Datt and Ravallion 2002). By this view, the composition of the higher growth would mean that it bypassed many of India’s poor. Against this view is the conjecture that India’s growth process has Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 changed—implying a new set of parameters in the relationship between growth and poverty reduction. Ravallion and Datt (1996) studied a period when policy emphasized rapid development of the capital goods sector in a largely closed economy, on the assumption that the capital stock and industrial struc- ture could be manipulated exogenously through central planning, even in a largely market-based economy.3 The strategy was also founded on “trade pessi- mism�—the beliefs, grounded in the experiences of colonialism, that India could not compete in global markets until its domestic capital stock was much larger and that foreign (Western) countries could not be trusted as a source of essential goods. These beliefs were questioned in both academic and policy circles at the time, and the poor economic performance as the years passed seemed to substantiate that skepticism.4 The success of China’s promarket reforms starting in 1978 further fueled doubts in the 1980s about India’s econ- omic strategy. The policy debate raged for many years, but it was a balance of payments crisis that triggered more extensive reforms in the early 1990s. Trade liberaliza- tion was combined with efforts to support higher productivity in the private sector.5 Supporters argued that these reforms would allow India to exploit its comparative advantage in labor-intensive goods and services, directly bene�ting 2. Ravallion and Datt (2002) found a strong interaction effect between the initial level of human development at the national level and the nonfarm growth rate in determining poverty reduction at a national level. 3. On the history of thought on development strategies and their implications for poverty, with speci�c reference to India, see Lipton and Ravallion (1995). 4. Some observers in India at the time questioned these assumptions, raising concerns about labor absorption (given high population growth) and (hence) poverty reduction; in particular see Vakil and Brahmanand (1956). Chakravarty (1987) provides an insightful account of the history of thought on India’s ( prereform) development strategy. 5. On India’s reform agenda since the early 1990s, see Ahluwalia (2002) and Panagariya (2008). Datt and Ravallion 159 the poor. The reforms would “favour the poor by beginning to remove the per- vasive bias that exists against the employment of unskilled labour� (Joshi and Little 1996, p. 221). The hope was that the postreform urban economy would be more effective in reducing both urban and rural poverty. However, there are also reasons to question whether the new policy environ- ment would put India on a new path of rapid poverty reduction. The greater openness to external trade came with suf�cient productivity growth to ensure higher growth of national output.6 But new inequality-increasing forces also appear to have emerged, and several observers have reported evidence of rising consumption inequality since the early 1990s.7 This may well reflect the ante- cedent inequalities in other “nonincome� dimensions, particularly in human capital, which can mean that the poorest are largely left behind; these inequal- ities were far greater in India around 1990 than in China around 1980.8 Intuitively, rising inequality will attenuate the impact of growth on poverty, though this effect is ambiguous in theory; for example, an increase in a stan- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 dard measure of inequality, such as the Gini index, need not mean an increase in the proportion of people living in poverty (ceteris paribus)—that depends on precisely how the Lorenz curve shifts with the change in inequality (Datt and Ravallion 1992). Some observers have also questioned whether the postreform growth process has ful�lled expectations that it would increase aggregate demand for unskilled labor and (hence) help reduce poverty. They point out that the fastest growing sectors of India’s economy have tended to be more intensive in capital and skilled labor, notably the booming business services sector. This pattern of growth is hardly what the “comparative advantage� arguments of reform advo- cates in the 1980s predicted as the outcome of India becoming a more open economy. Given that an argument for reform is that it should make growth more labor intensive, it is interesting to see what happened to employment in India. The 1999–2000 survey of employment by the National Sample Survey Organization (NSSO) suggested a slight deceleration in employment growth, although the latest available survey for 2004–05 suggests that employment growth was virtually the same from 1993–94 to 2004–05 as in the preceding 10 years (Panagariya 2008, p. 146). These comparisons 6. Eswaran and Kotwal (1994, chapter 7) argue that domestic productivity growth is key to the outcomes for poor people from trade openness in India. The sequencing of reforms was important, and India’s reformers wisely emphasized domestic reforms (such as industrial delicensing) before external reforms (Bhagwati 1993). 7. Evidence of rising inequality in India since 1991 is reported in Ravallion (2000), Deaton and Dre` ze (2002), and Sen and Hiamnshu (2004a, b). There was no trend increase, or decrease, in consumption inequality over the period up to about 1990 (Bruno, Ravallion, and Squire 1998). ` ze and Sen (1995) on the constraints stemming from India’s meager 8. See the discussion in Dre human development attainments at the outset of its current reforms and the contrast with China. Also see Chaudhuri and Ravallion (2006) on the distinction between “good� and “bad� inequalities in China and India and the discussion of inequality of opportunity in World Bank (2005). 160 THE WORLD BANK ECONOMIC REVIEW are clouded because of the large share of employment in the informal sector, for which reliable measurement is dif�cult, and because the reforms themselves may induce output and employment to shift to the informal sector.9 Even more relevant is the observation that the nonfarm sectors that are relatively intensive in unskilled labor—trade, construction, informal manufacturing—fared better in the post-1991 period than earlier (Kotwal, Ramaswami, and Wadhwa 2009). The nonfarm sector’s aggregate demand for unskilled labor appears to have increased after the reforms, even though the most dynamic sectors have been intensive in skilled labor. And these newly created relatively unskilled nonfarm jobs typically pay more than agri- cultural labor.10 The importance of rising rural nonfarm employment and incomes is also suggested by the �nding of Foster and Rosenzweig (2004a, b) that nonfarm wages and salaries associated with the rapid growth of the rural factory sector Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 was the fastest growing component of rural incomes during 1971–99 (especially during 1982–99). Moreover, the growth in nonfarm wages and sal- aries and rural in industrial activity was highest where growth in agricultural yields was lowest. This is consistent with the hypothesis that mobile capital sought relatively low-wage areas to produce tradables in response to demand fueled by urban growth.11 Another potential channel through which India’s postreform urban econ- omic growth could affect rural poverty is public �nance. Higher economic growth rates generate higher tax revenues, which can support propoor spend- ing. In recent years, rural antipoverty programs have expanded considerably, notably under the National Rural Employment Guarantee Act, which aims to provide 100 days of unskilled work to any rural family that wants to work at the statutory minimum wage rate in agriculture. This program is �nanced through general taxation. It is clear from these observations that arguments can be made for and against any claim that the economic reforms have helped reduce poverty in 9. Similarly, Sen (2009) shows that employment in the formal (“organized�) manufacturing sector did not rise after trade liberalization. However, this is a moot point as 80 percent of manufacturing employment is in the informal sector (Kotwal, Ramaswami, and Wadhwa 2009). 10. For evidence on this point, see Jacoby, Rabassa, and Skou�as (2010), who �nd a 25 percent differential in farm and nonfarm wages after controlling for age, experience, and education. 11. Kotwal, Ramaswamy, and Wadhwa (2009) point to the limits of nonfarm employment growth in reducing the labor to land ratio in agriculture suf�ciently to produce a rapid increase in agricultural wages. The faster growth in nonagricultural wages over agricultural wages suggests the need for a rural labor market model that can explain a premium on nonfarm jobs. That such a premium exists is suggested by some recent evidence; for instance, World Bank (forthcoming) reports a rising premium of casual nonfarm wages over agricultural wages from 25 –30 percent in 1983 to 45 percent in 2004– 05. Lanjouw and Murgai (2009) further document that education levels are higher among casual nonfarm rural workers than among agricultural workers, which suggests that education plays a role in helping one segment of the rural workforce to better access the growing nonfarm jobs. Datt and Ravallion 161 India. To help inform this debate, this article addresses the following ques- tions: Has India’s higher growth rate since the early 1990s delivered a higher pace of progress against absolute poverty? Has the responsiveness of poverty to growth changed in the postreform period? Has the poverty impact of the urban-rural composition of growth changed? In particular, is there any sign that urban economic growth has been more propoor since the reforms than before them? Section I outlines the concepts and methods used in this study. Section II describes the dataset, which updates the data set constructed for Ravallion and Datt (1996), along with some improvements in the estimation methods. Section III presents the results and their implications. Section IV draws some conclusions. I. CONCEPTS AND METHODS Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 The analysis uses three poverty measures. The head-count index is given by the percentage of the population who live in households with per capita con- sumption below the poverty line. The poverty gap index is the mean distance below the poverty line expressed as a proportion of that line, where the mean is formed over the entire population, counting the nonpoor as having zero poverty gap; this can be interpreted as a measure of the depth of poverty. The squared poverty gap index, introduced by Foster and others (1984), is the corresponding mean of the squared proportionate poverty gaps. Unlike the poverty gap index, the squared poverty gap index is sensi- tive to distribution among the poor, in that it satis�es the transfer axiom for poverty measurement (Sen 1976). The squared poverty gap index can be thought of as a measure of the severity of poverty. All three measures are among those proposed for measuring poverty by Foster, Greer, and Thorbecke (1984). As for virtually all poverty measures in practice, this class of measures can be written as functions of the survey mean relative to the poverty line and the relative distribution of income, as represented by the Lorenz curve (see, for example, Datt and Ravallion 1992 and Kakwani 1993). (The term “relative distribution� refers to all effects on poverty that are transmitted through changes in the Lorenz curve.) When the poverty line is �xed in real terms, the poverty measure (Pt) is strictly decreasing in the mean (mt) for any given rela- tive distribution (though the elasticity can vary greatly, depending on the initial mean and Lorenz curve). For example, the elasticity of the headcount index to growth in the mean, holding relative distribution constant, is given by one minus the elasticity of the cumulative distribution function evaluated at the 162 THE WORLD BANK ECONOMIC REVIEW poverty line. However, a higher growth rate may also entail a shift in distri- bution for or against the poor. Of interest here is the total effect of growth on poverty, allowing distribution to change, rather than the partial effect, holding relative distribution constant.12 Assuming that the poverty measure can be derived as a differentiable function of the mean, allowing relative distribution to change with the mean, the interest is in estimating the growth elasticity of poverty reduction, de�ned by: d ln Pt ð1Þ p; d ln mt where p is estimated by the regression coef�cient of ln Pt on ln mt across the available time series, allowing the error term to be autocorrelated and heteroskedastic.13 When both the dependent and the independent variables are estimated from the same survey data, the possibility of bias arises because measure- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 ment errors in the survey can be passed on to both variables. Overestimating the mean will tend to underestimate poverty. (The sign of the bias is ambig- uous in theory, given that there is also an attenuation bias in the estimate of p.) An instrumental variable (IV) estimator is also used, in which the instruments exclude any variables derived from the same survey as the dependent variable. This is also helpful for controlling the effect of changes in survey design. The urban-rural composition of growth and poverty reduction are also examined. In India, as in most developing countries, the rural sector has a higher incidence of extreme poverty and accounts for a substantially higher share of absolute poverty than the urban sector (Ravallion, Chen, and Sangraula 2007). Also in common with most (growing) developing economies, India’s trend rate of growth has been higher in the nonfarm sectors than in agriculture. The fortunes of poor people in urban and rural areas are linked. The scope for the urban economy to absorb wage labor from rural areas has long been seen as a key factor in poverty reduction. Labor mobility can yield an equili- brium relationship between the real wages of similar workers, entailing “hori- zontal integration� in earnings and income distributions, with the living standards of people at similar levels of living but in different sectors causally related. Such integration can also arise without labor mobility. Proximity to 12. Analytic formulae for the partial elasticities (holding relative distribution constant) are found in Kakwani (1993). On the conceptual distinction between partial and total elasticities in this context, see Ravallion (2007). Also see the discussion of alternative de�nitions of this elasticity in Heltberg (2004). 13. A dynamic model (with lags in Pt and ln mt) is not feasible given the uneven spacing of the time series. However, there is little choice but to assume even spacing when implementing the corrections to the standard errors for serial correlation. Datt and Ravallion 163 urban areas enhances demand for the outputs of the rural economy.14 The living standards of households in different sectors but sharing similar factor endowments will tend to move together to the extent that trade in goods attenuates differences in real factor prices. The fact that the rural sector pro- duces food some of which is consumed in the urban sector can mean that agri- cultural growth boosts urban welfare by lowering food prices (to the extent that domestic food markets are only weakly integrated with global markets). Transfers can also produce horizontal integration. The existence of such horizontal integration suggests that changes ema- nating from the urban sector can have powerful effects on levels of living in the rural sector and vice versa. This can also entail distributional effects, notably when the distributions of absolute levels of living in different sectors overlap imperfectly (share a positive density over certain, compact, intervals of the range of living standards but not others). The urban sector of a developing country will often include an elite that has no counterpart Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 in the rural sector. When combined with shared poverty in the overlapping interval of the distribution, this uneven overlap of urban-rural distributions can have strong implications for how an increase in incomes in one sector spill over to affect both average levels of living and relative distribution in the other sector. The urban-rural decomposition of poverty is also of interest. The relevant measures of poverty can be additively decomposed using population weights, such that the national level of poverty at date t is given by: ð2Þ Pt ¼ nut Put þ nrt Prt ðt ¼ 1; ::T Þ where nit is the population shares and Pit the poverty measures for sector i ¼ u, r (for urban and rural). This property of additivity is exploited in testing whether the sectoral composition of growth matters by estimating the following regression on the discrete data: D ln Pt ¼ pu sm m utÀ1 D ln mut þ pr srtÀ1 D ln mrt ð3Þ þ pn ðsm m rtÀ1 À sutÀ1 nrtÀ1 =nutÀ1 ÞD ln nrt þ 1t ðt ¼ 2; . . . ; T Þ where D is the discrete time difference operator, smit ¼ nitmit/mt is sector i’s share of mean consumption at date t, and mit is the mean for sector i. The pu, pr par- ameters can be interpreted as the impact of (share-weighted) growth in the urban and rural sectors, while pn gives the effect of the population shift from rural to urban areas—interpretable as a “Kuznets effect� following Kuznets (1955). To motivate this test regression, notice that, under the null hypothesis of 14. Lanjouw and Murgai (2009) and World Bank (forthcoming) argue that India’s urban economic growth has exerted a pull on the rural economy through diversi�cation into rural nonfarm activities. 164 THE WORLD BANK ECONOMIC REVIEW pu ¼ pr ¼ pn ¼ p, equation (3) collapses to: ð4Þ D ln Pt ¼ p D ln mt þ 1t Thus, under this null hypothesis, it is the overall growth rate that matters, not its composition. Rejecting this null tells us that the composition of growth is a signi�cant factor in poverty reduction. Whether economic growth in one sector affects distribution in the other sector is also tested, estimating the following system (dropping time subscripts for brevity): ð5:1Þ sP m m m m u D ln Pu ¼ pu1 su D ln mu þ pu2 sr D ln mr þ pu3 ðsr À su nr =nu ÞD ln nr þ 1u ð5:2Þ sP m m m m r D ln Pr ¼ pr1 su D ln mu þ pr2 sr D ln mr þ pr3 ðsr À su nr =nu ÞD ln nr þ 1r ðsp p m r À su nr =nu ÞD ln nr ¼ pn1 su D ln mu Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 ð5:3Þ þ pn2 sm m m r D ln mr þ pn3 ðsr À su nr =nu ÞD ln nr þ 1n where sPit ¼ nit Pit /Pt and pi ¼ pui þ pri þ pni, so that summing equations (5.1), (5.2), and (5.3) yields equation (3). Equation (5.1) shows how the composition of growth and population shifts affect urban poverty; equation (5.2) shows how they affect rural poverty; and equations (5.3) shows the effect on the population shift component of D logP. Only equations (5.1) and (5.2) are estimated.15 I I . D ATA To address the questions posed in this article, it is desirable to have a reasonably long time series of household surveys; a short series can be deceptive for infer- ring a trend.16 India provides rich time series evidence for testing and quantify- ing the relationship between the living standards of the poor and macroeconomic aggregates. Among developing countries, India has the longest series of national household surveys suitable for tracking living conditions of the poor. At the time of writing, distributional data on household consumption in India could be assembled from 47 surveys spanning 1951–2006. Though some of the earliest surveys had smaller sample sizes and covered shorter periods, the surveys are large enough to be considered representative at the urban and rural levels as well as nationally. And because the basic survey instruments and 15. Equation (5.3) need not be estimated separately since the parameters can be inferred from the estimates of equations (5.1), (5.2), and (3) using the adding-up restriction. These three equations are estimated as single equations, although there may be some ef�ciency gains from estimating them as a system. 16. For example, the �rst survey (1992) available in the postreform period indicated a substantial increase in poverty, fueling much debate about the wisdom of reforms. Datt and Ravallion (1997) questioned this inference at the time, arguing that the 1992 survey was deceptive about trends. Datt and Ravallion 165 methods have changed little (though there are some comparability problems, addressed below), the surveys should be comparable over time. The period of analysis in Ravallion and Datt (1996) ended two years after India’s economic reforms began. This article adds 14 more rounds of National Sample Surveys (NSS). Though the data are not ideal, there are now suf�cient postreform data to revisit the question of whether India’s higher growth rates have delivered the promise of a higher rate of progress against poverty. While attribution to reforms per se is clearly problematic, revisiting those earlier �nd- ings using these new data spanning 15 years of the postreform period offers some insight into whether India’s progress against poverty has accelerated or decelerated. Survey Data A new and consistent time series of poverty measures for rural and urban India Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 over 1951–2006 was derived for this study, based on consumption distri- butions from 47 household surveys (rounds 3 –62) conducted by the NSSO. This series improves greatly on the most widely used time series on poverty measures in India to date based on Ahluwalia (1978, 1985).17 The pre-1991 data also differ in some respects from the dataset constructed in Ravallion and Datt (1996), as noted below. Some of the early survey rounds (notably rounds 4 –12) covered periods con- siderably shorter than a year. These rounds were aggregated to broadly conform to a year-long survey period. Rounds 4 and 5, 6 and 7, 9 and 10, and 11 and 12 were pair-wise aggregated using the number of survey months covered as weights.18 Thus, with these combined rounds, the dataset has 43 observations over 1951–2006. As is well established practice for India and elsewhere, real consumption expen- diture per person is used to measure household standard of living. The underlying survey data do not include incomes, though it can be argued that current con- sumption is a better welfare indicator of living standards than is current income. While the surveys are highly comparable over time by international standards, there is a comparability problem in the rounds since the early 1990s. While most of the surveys used a uniform recall period of 30 days for consumption items, seven of the survey rounds (55–60 and 62) used a mixed-recall period, with one week recall for some items (such as food) and one year for others (mainly nonfood items). Preliminary investigation found that the mixed-recall period reduced the log of the headcount index at a given level of mean consumption by 17. Prior to Ravallion and Datt (1996), work on poverty and growth in India had relied on poverty measures in Ahluwalia (1978), which contained estimates of poverty measures for rural areas for only 12 survey rounds spanning 1956– 57 to 1973– 74. Ahluwalia (1985) extended this by another round (1977–78). 18. For instance, the headcount index for combined rounds 6 (for May –September 1953) and 7 (for October 1953– March 1954) is 5/11th of the headcount index for round 6 plus 6/11th of the headcount index for round 7. 166 THE WORLD BANK ECONOMIC REVIEW about 0.2 and that the effect is (highly) signi�cant.19 This is probably because the shorter recall periods for food in the mixed-recall period give higher reported food spending, which has a higher budget share for poorer households. All the regressions include a control for mixed-recall period survey rounds. Urban-rural classi�cation is from the NSSO.20 Over such a long period, some rural areas would have become urban. To the extent that rural (nonfarm) economic growth may contribute to the evolution of successful villages into towns, this process might produce a downward bias in estimates of the (abso- lute) elasticities of rural poverty to rural economic growth. The impact on the urban elasticities could go either way, depending on the circumstances of new urban areas relative to old ones. There is little choice but to use the NSSO’s classi�cation, however, since the unit record data are unavailable for the full period covered by this exercise (nor is it clear what the best corrective would be if there were access to that data). The population numbers are from the censuses and assume a constant Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 growth rate between censuses. They are also centered at the mid-points of the survey periods. The trend increase in the urban population share was 0.24 per- centage point a year in the period 1951–2006 (with a robust standard error of 0.04). In the 40 years after 1950, the urban sector’s population share rose from 17 percent to 26 percent, and it reached 29 percent by 2005. Poverty Lines and Price Indices The rural and urban poverty lines used here are those originally de�ned by the India Planning Commission (1979) and endorsed by the Expert Group on Estimation of Proportion and Number of Poor (India Planning Commission 1993). These lines were set at a per capita monthly expenditure of 49 rupees (Rs) for rural areas and Rs 57 for urban areas at 1973–74 prices, correspond- ing to per capita total expenditure needed to attain caloric norms of 2,400 cal- ories per person per day in rural areas and 2,100 in urban areas.21 19. Regressing the change in the log of H across 42 rounds on the change in the log of the survey mean and the change in a dummy variable for the mixed-recall period rounds (MRP) yielded a regression coef�cient of –0.20 with a t-ratio of 16.7. (Note that since the other variables in the regression are in differences not levels, the MRP dummy variable is also differenced.) Similarly, mixed-recall period rounds tended to yield signi�cantly lower inequality (as measured by the Gini index) in both rural and urban areas. 20. The NSS has followed the Census de�nition of urban areas, which is based on several criteria including a population greater than 5,000, a density of at least 400 people per square kilometer, and three-fourths of the male workers engaged in nonagricultural pursuits. 21. An expert group constituted by the India Planning Commission (2009) recently recommended a higher poverty line for rural areas for 2004/05 while retaining the of�cial line for urban areas. Thus, the implied urban–rural cost of living differential at the poverty line is lower than that in this study. The new rural line was not used in this study because it showed zero cost of living difference at the poverty line in the 1970s when the poverty lines were backcast using the study’s urban and rural deflators, which is not plausible. Datt and Ravallion 167 Rural and urban price indices are needed to update (and backcast) these poverty lines for different survey periods. Since the analysis is con�ned to the all-India level, so are the deflators.22 Following well-established practice, the deflators are based on the all-India Consumer Price Index for Industrial Workers (CPIIW) for urban areas and the all-India Consumer Price Index for Agricultural Laborers (CPIAL) for rural areas.23 Deaton (2008) argues that between the 1999–00 and 2004–05 rounds, the of�cial CPIAL underestimated the rate of rural price inflation because the food component of the index underestimated the rate of food price inflation and the index assigned too much weight to food during a period when food prices were falling relative to nonfood prices. (Potentially similar problems arise for the CPIIW, although Deaton found these to be of less concern for that period.) Deaton’s comparison of the CPIAL with his survey-based food price index using median unit values of food items from the two surveys offers support for his claim that the CPIAL underestimated the rate of food price inflation.24 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 However, Deaton’s method cannot be used here because the household-level data needed to construct unit values–based food price indices are not accessi- ble for the long period of the analysis. And feasibility aside, there are concerns about using unit values over time (and across space). The quality of consump- tion could change, which would change the unit value even if prices were unchanged; for example, if the quality of rice consumed rises over time, the unit values will suggest price inflation even when there is none. However, Deaton is right to stress the importance of properly weighting food when measuring poverty. This study weighted both the food and the nonfood components of the CPIAL and CPIIW using the survey-based (rural and urban) food shares that can be calculated from the published grouped data for NSS rounds. It used the food share at the poverty line (similar to one set of Deaton’s price indices25), which is conceptually more appropriate for measur- ing poverty. More precisely, the food and nonfood components of the CPIAL and CPIIW for any round were reweighted by the predicted food and nonfood shares for the rural and urban areas at the poverty line in the preceding round. 22. Thus, this study does not use any state-level price indices or poverty lines, which have been subject to criticism (Deaton 2003; Deaton and Tarozzi 2005). 23. While the analysis covers a long period back to 1951, the all-India CPIAL is available from September 1964 and the all-India CPIIW from August 1968. For the earlier years, we rely on our past work on constructing a consistent rural and urban price index series, using the state-level CPIALs and the Consumer Price Index for the Working Class, a precursor to the CPIIW (see Datt 1997 for details). This series also corrects for �rewood prices in the CPIAL, which had remained unchanged in the published CPIAL data since 1960–61. The �nal CPIIW and CPIAL are averages of monthly indices corresponding to the exact survey period of each NSS round. 24. The unit value is the ratio of expenditure on a type of goods to quantity. This is the price only if there is just one good of that type; in practice, the categories differ in quality. 25. Deaton (2008) presents price indices using both average food shares and estimated food shares at the poverty line. The estimated food shares are derived from a regression of food shares on the log of per capita consumption and its squared value using unit-record data. 168 THE WORLD BANK ECONOMIC REVIEW Predicted food shares are derived from grouped data on budget shares, using a regression for the previous round of food budget shares as a cubic function of the cumulative proportion of the population ranked by per capita monthly total expenditure. Poverty line food shares for the current round were then derived as predictions at the estimated headcount index for the previous round.26 Since the published grouped data on budget shares are available only from round 14 (July 1958–June 1959), the reweighting started with round 15 (July 1959–June 1960) using the predicted poverty line food shares for round 14. The reweighted indices for successive rounds were then combined to form the �nal chain price indices for rural and urban areas. These indices correspond to the evolving food and nonfood budget shares of people near the poverty line and thus help attenuate errors due to the use of outdated consumption patterns (in the of�cial price indices) to measure current inflation for the poor. These price indices can be compared with other recent work on this subject. First, the rates of rural and urban inflation implied by these indices can be Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 compared with those in Deaton (2008) and with of�cial price indices (CPIAL/ CPIIW) for 1999–2000 (55th round) to 2004–05 (61st round), the only period for which the Deaton indices are available. Deaton �nds a higher rate of rural inflation (14 percent) over this period than that implied by the of�cial price indices or the revised indices in this study (both at 11 percent). The urban rates of inflation are similar across all three sets of indices.27 The food share in the current study’s rural index (71 percent) is similar to that in the CPIAL (69 percent), and both are higher than Deaton’s (65 percent). Thus, the CPIAL’s food share in rural areas in 2004–05 is not inappropriate for the current study’s poverty line, despite this study’s use of a higher urban food share (see the statistical appendix, available at http://wber.oxfordjournals.org/, for details). But the bulk of the difference is due to Deaton’s use of a food price index based on unit values instead of the CPIAL food index based on actual prices.28 As mentioned, since survey-based food price indices over the longer period of the current analysis cannot be constructed, further comparisons cannot be made for the earlier prereform period. A second comparison is with the survey unit value–based urban to rural (Tornquist) price indices estimated by Deaton (2003) for the 43rd (1987–88), 50th (1993–94), and the 55th rounds (1999–2000), which are 111.4, 115.6, and 115.1 (with rural equal to 100 in each round), as against this study’s higher estimates of 133.0, 131.7, and 136.2. However, two observations are 26. Thus, for instance, for the 43rd round, the food share regression was estimated for the 42nd round, and the poverty line food share for reweighting the price index for the 43rd round was estimated as the prediction from this regression at the headcount index for 42nd round. In the case of mixed-recall period survey rounds, the regression for the most recent round with a uniform recall period was used. 27. The urban to rural price index of this study (with the 55th round as the base) lies between those for the of�cial price indices and Deaton’s (2008). 28. The numbers reported in Deaton (2008) imply that 75 percent of the difference between his deflators and the CPIAL is due to his use of unit values; the rest is due to the weights. Datt and Ravallion 169 pertinent. First, Deaton’s indices are food price indices while this study’s indices are general price indices; the relative price of food has certainly not been constant, as shown by Deaton’s own work. Second, this study’s starting point is the of�cial poverty lines for 1973–74, which imply a 16 percent urban to rural price differential. This differential increased to 33 percent by 1987–88 and remained roughly constant till 1999–2000, the relative constancy over this period being analogous to Deaton’s estimates. Thus, as far as the change in the urban to rural price ratio is concerned, comparison is possible only over essen- tially the postreform period for which this study’s estimates are similar to Deaton’s deflators. National Accounts Private �nal consumption expenditure and net domestic product data are from the national account system (NAS). Imperfect matching between the survey Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 periods and the annual accounting periods used in the NAS makes it harder to detect the true effect of aggregate growth on poverty. To mesh the NAS data with the NSSO poverty data, the annual NAS data were linearly interpolated to the mid-point of the survey period for different rounds. Following Ravallion and Datt (1996), both NAS and NSS data are used in the same regressions only for the period 1958 onward, because the shorter survey periods of the early rounds lead to poor mapping between NSS rounds and NAS annual data for that period. The NSS series of mean household consumption per capita does not fully reflect the gains in mean consumption indicated by the NAS from the early 1990s onwards. The overall elasticity of the NSS mean consumption to NAS consumption is 0.48 (t ¼ 4.03) in a regression of consumption growth from the NSS on consumption growth from the NAS, with controls for changes in whether the round used mixed-recall periods and changes in the log ratio of the rural price index to the NAS deflator. The elasticity is signi�cantly less than unity. It is also lower in the post-1991 period, declining from 0.57 (4.47) in the pre-1991 period to 0.45 (t ¼ 3.29). However, the null hypothesis that the elasticities are the same for the two subperiods cannot be rejected. To investigate further the source of divergence between NAS and NSS con- sumption per capita data in the two subperiods, the difference between the NAS and the NSS mean consumption growth rates were also regressed on dummy variables for pre- and post-1991 subperiods and on pre- and post-1991 per capita net domestic product growth rates. (All regressions include controls for change in the dummy variable for a mixed-recall period round as well as change in the log ratio of the rural price index to the NAS deflator.) These tests con�rmed that the divergence in mean consumption growth rates was greater in the post-1991 period, although the difference between the two subperiods is not statistically signi�cant. The divergence between NAS and NSS mean con- sumption growth rates tends to be higher the higher the per capita net domestic 170 THE WORLD BANK ECONOMIC REVIEW product growth rate, an association that is somewhat stronger in the post-1991 period. It is dif�cult to fully assess the role of NSSO methods in this divergence from NAS consumption. By international standards, those methods appear to have changed little over decades. That is probably good news for comparabil- ity, although it does raise questions about whether NSSO methods are in accord with international best practice. However, it is notable that the multiple-recall period rounds of the NSS have narrowed the gap between the NAS and NSS consumption aggregates. When the difference over time in the log of the NSS mean is regressed on the corresponding difference in NAS consumption and the change in the dummy variable for mixed-recall period rounds, the coef�cient is 0.055 (t ¼ 4.14). This suggests that NSS design may account for at least some of the discrepancy between the two data sources. Some of the gap between the consumption aggregates from these two sources is undoubtedly due to errors in NAS consumption, which is determined Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 residually in India after subtracting other components of domestic absorption from output at the commodity level. There are also differences in the de�nition of consumption, and NAS consumption includes components that should not be in a measure of household living standards.29 Some degree of underreport- ing of consumption by respondents, or selective compliance with the NSS’s randomized assignments, is inevitable. However, it is expected that this is more of a problem for estimating consumption by the rich (notably in urban areas) than the poor.30 If so, then it is not clear that there will be much bias in the poverty measures based on the surveys.31 For the same reason that the consumption aggregates from the NSS are diverging from the private consumption component of domestic absorption in the NAS, one cannot rule out the possibility that the NSS is underestimating the increase in inequality in India. I II. RE S U LTS This section presents an overview of trends in the variables of interest, both for the entire 50-year period and for the periods before and after 1991. It also pre- sents estimated growth elasticities of poverty reduction, separately for urban and rural areas and for their interaction. Trends There can be no doubt that growth has accelerated in the postreform period. The trend rate of growth in India’s net domestic product per capita was 1.63 29. For further discussion of the differences between the two data sources, see Sundaram and Tendulkar (2001), Ravallion (2000, 2003), Sen (2005), and Deaton (2005). 30. There is evidence from other sources consistent with that expectation; see Banerjee and Piketty (2005) on income underreporting by India’s rich. 31. For a more complete discussion of this issue, see Korinek, Mistiaen, and Ravallion (2006). Datt and Ravallion 171 percent during 1958–91 (with a robust standard error of 0.06 percent) and 4.28 percent (0.18 percent) during 1992–2006.32 Similarly, the annual rate of growth of private consumption per capita from the NAS rose from 1.21 percent before 1991 to 3.13 percent after. The acceleration in the survey-based per capita consumption growth—though less than that in mean income or con- sumption from the NAS—is also notable, from 0.68 percent a year before 1991 to 1.33 percent after . By sector, the highest growth rates in output in the period after 1991 were in the tertiary sector (primarily services and trade), fol- lowed closely by manufacturing, while agriculture continued to lag. The sector that gained the most between the two periods was services; agriculture showed little or no improvement in growth (Chaudhuri and Ravallion 2006).The main long-run structural shift in India’s economy has been out of agriculture into services, a trend that continued after 1991. What about poverty? The headcount index and the squared poverty gap for both urban and rural sectors exhibit neither a trend increase nor a trend Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 decrease in rural poverty until about 1970, when a trend decrease emerged (�gure 1). Sustained, though uneven, progress against poverty had clearly emerged in India before the economic reforms starting in the early 1990s. Comovement is strong between the urban and rural measures, and there is clear indication of a declining absolute difference between the poverty measures for urban and rural areas after about 1970.33 Indeed, the urban squared poverty gap overtakes the rural index by the end of the period. In common with other developing countries (Ravallion, Chen, and Sangraula 2007), in India poverty has been urbanizing over time, as the share of the poor living in urban areas has risen. Only about 15 percent of India’s poor lived in urban areas in the 1950s, but about 28 percent did in 2005–06. However, because more than 70 percent of the population still lives in rural areas, the rural sector accounted for the bulk of national poverty at the end of the period—72 percent of the total number of poor, 68 percent of the aggregate poverty gap, and 65 percent of the aggregate squared poverty gap. The number of poor people has declined since the early 1990s, primarily as the number of poor in rural areas has declined. Over the entire 50-year period, the exponential trend in poverty reduction—the regression coef�cient of the log poverty measure on time— was 1.3 percent a year for the headcount index, rising to 2.2 percent for the poverty gap and 3.0 percent for the squared poverty gap. For the period before 1991, the trends were 1.1 percent for the headcount index, 32. These are based on regressions of log net domestic product per capita on time. Here and elsewhere, following Boyce (1986), the two growth rates are estimated as parameters of a single regression constrained to ensure that the predicted values were equal in 1992 (to avoid an implausible discontinuity). The supplemental appendix (available at http://wber.oxfordjournals.org/) contains a fuller analysis of trends. 33. The regression coef�cient of rural H minus urban H on time after 1970 is – 0.231 percentage point a year (t ¼ – 4.617); for SPG it is – 0.062 (t ¼ – 9.545). 172 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Poverty Measures for India Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. 2.1 percent for the poverty gap, and 2.8 percent for the squared poverty gap; for the period after 1991 the corresponding trends were 2.4 percent, 3.4 percent, and 4.2 percent. So exponential trends in poverty reduction are higher for the postreform period, but the difference between the pre- and Datt and Ravallion 173 post-1991 trends are statistically signi�cant only for the headcount index and then only at about the 8 percent level.34 Alternatively, the trend could be de�ned by the level of the poverty measure or mean consumption/income rather than by its log. Doing so con�rms the �nding of an acceleration of growth (in mean income and consumption) in the post-1991 period but yields no evidence of a parallel acceleration in poverty reduction. (Details are in the supplemental appendix.) Growth and poverty trends in urban and rural areas are similar to those at the national level described above. While the (survey-based) mean consumption growth rates were higher (nearly twice as high) in the post-1991 period than in the pre-1991 period in both rural and urban areas, only the acceleration in urban growth was statistically signi�cant. There are some indications of a faster poverty decline after 1991, more notably in rural areas, but the increase was often not statistically signi�cant. For instance, there was no signi�cant acceleration in the trend decline in the poverty gap or the squared poverty gap Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 in either rural or urban areas. Only for the headcount index is the increase in the trend rate of poverty decline signi�cant—at the 10 percent level in rural areas and at the 3 percent level in urban areas. Part of the reason that the faster postreform growth has not yielded corre- spondingly higher rates of poverty reduction is that rising inequality has accompanied the higher overall growth. As in many developing countries, the gap between urban and rural living standards is an important dimension of overall inequality. The urban mean has risen faster than the rural mean in India. The trend rate of growth in mean consumption based on the NSS since 1958 has been 0.87 percent a year (standard error of 0.10 percent) for urban areas and 0.65 percent (0.14 percent) for rural areas.35 So inequality between urban and rural areas increased. What has happened to inequality within urban and rural areas? The Gini indices calculated from the relevant NSS rounds, but without adjusting for the difference between the uniform and the mixed-recall period, suggest that in rural areas inequality declined, whereas in urban areas it declined until about 1980 and tended to increase thereafter. However, this changes after controlling for the mixed-recall periods of the several NSS rounds since the 1990s, which have a dampening effect on measured inequality (as already noted). Figure 2, which gives the predicted values after controlling for the differences in recall periods between surveys, shows evidence of a clear rising trend in inequality within both rural and urban areas after 1991. The next subsection looks at whether the rising inequality in the postreform period, both between and within urban and rural areas, attenuated the impact of growth on poverty. 34. The supplemental appendix provides a complete set of statistical tests. 35. The rural mean was rising relative to the urban mean during most of the 1950s. This period is excluded from the calculation because it is so unusual. 174 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. Trends in Urban and Rural Inequality in India Controlling for Changes in Survey Reference Periods Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Note: The lines show predicted Gini indices after controlling for the effect of mixed-recall period rounds (as distinct from the actual values plotted, which are naturally without controls). Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. Growth Elasticities of Poverty Reduction Elasticities of the three poverty measures are estimated by regressing the log poverty measure on log mean consumption per person from the NSS, consump- tion per person as estimated by the NAS and population census, and net domestic product (income, for short) per person, also from the NAS and census (table 1). In addition, an "adjusted" estimate adds a control variable for the �rst difference of the log of the ratio of the consumer price index for agri- cultural laborers to the national income deflator (that is, the difference in the rate of inflation implied by the two deflators). This allows for possible bias in estimating the growth elasticity due to the difference in the deflator used for the NAS data and that used for the poverty lines. For 1958–2006 as a whole, the national poverty measures responded signi�- cantly to economic growth by all three measures. This also holds when the IV estimator is used to reduce the potential for spurious correlation arising from common survey measurement errors. The (absolute) elasticities are higher when using NSS consumption rather than NAS consumption. The elasticities are lowest for per capita income. This may be due to intertemporal consump- tion smoothing, which may make poverty (in terms of consumption) less T A B L E 1 . Elasticities of National Poverty Measures to Growth in India, 1958–2006 Elasticity of poverty measure with respect to: Mean consumption from Mean private consumption National Sample Surveys from national accounts Mean net domestic product Ordinary Instrumental Poverty measure Period least squares variable Unadjusted Adjusted Unadjusted Adjusted Headcount index Whole period – 1.62( –26.0) – 1.60(– 61.4) – 0.90(– 9.57) – 0.50(– 9.76) – 0.65(– 9.20) – 0.35( – 9.27) Up to 1991 – 1.58( –27.8) – 1.57(– 75.2) – 0.98(– 6.77) – 0.51(– 7.35) – 0.73(– 6.07) – 0.36( – 6.35) After 1991 – 2.07( –21.4) – 2.07(– 22.9) – 0.70(– 5.10) – 0.62(– 2.99) – 0.49(– 4.13) – 0.42( – 2.70) Ho: pre-1991 F(1,34 or 32)Prob. 16.08(0.00) 24.91(0.00) 1.50(0.23) 0.25(0.62) 1.43(0.24) 0.12(0.73) elasticity ¼ post-1991 elasticity Poverty gap index Whole period – 2.66( –21.8) – 2.68(– 35.5) – 1.53(– 10.6) – 0.95(– 11.5) – 1.11(– 10.3) – 0.68( – 11.5) Up to 1991 – 2.63( –20.3) – 2.66(– 33.5) – 1.75(– 8.74) – 1.09(– 10.6) – 1.31(– 7.97) – 0.80( – 9.88) After 1991 – 2.94( –12.2) – 2.78(– 11.5) – 0.97(– 4.94) – 0.80(– 2.43) – 0.69(– 4.17) – 0.56( – 2.24) Ho: pre-1991 F(1,34 or 32)Prob. 1.10(0.30) 0.19(0.66) 5.96(0.02) 0.67(0.42) 5.21(0.03) 0.67(0.42) elasticity ¼ post-1991 elasticity (Continued ) Datt and Ravallion 175 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 176 TABLE 1. Continued Elasticity of poverty measure with respect to: Mean consumption from Mean private consumption National Sample Surveys from national accounts Mean net domestic product Ordinary Instrumental Poverty measure Period least squares variable Unadjusted Adjusted Unadjusted Adjusted Squared poverty gap index Whole period – 3.48( –19.7) – 3.48(– 31.8) – 2.03(– 10.7) – 1.31(– 10.7) – 1.48(– 10.5) – 0.94( – 10.9) Up to 1991 – 3.48( –18.0) – 3.52(– 26.3) – 2.37(– 9.63) – 1.58(– 10.6) – 1.79(– 8.86) – 1.16( – 10.3) THE WORLD BANK ECONOMIC REVIEW After 1991 – 3.49( –8.20) – 3.28(– 7.73) – 1.17(– 4.74) – 0.95(– 2.20) – 0.84(– 4.17) – 0.69( – 2.10) Ho: pre-1991 F(1,34 or 32) Prob. 0.00(0.99) 0.26(0.61) 9.51(0.00) 1.78(0.19) 8.36 (0.01) 1.56(0.22) elasticity ¼ post-1991 elasticity Note: Numbers in parentheses are t-ratios based on heteroskedasticity and autocorrelation-consistent standard errors. Results are based on regressions of log poverty measures against log consumption or net product per person using 37 surveys spanning 1958– 2006. All regressions include a control for surveys that used a mixed-recall period. The “adjusted� estimates control for the difference in the rates of inflation implied by the rural consumer price index and the national income deflator (Ravallion and Datt 1996). The instrumental variables for the survey mean regressions included lagged survey means (split urban and rural), current and lagged mean consumption from the national accounts, current and lagged rural and urban consumer price indices, current and lagged rural population shares, interval between mid-points of survey periods, and a time trend. The regressions also incorporate a kink at survey round 47 (July– December 1991) so that there is no discontinuity in the predicted values of log poverty measures between the pre- and post-1991 periods. Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Datt and Ravallion 177 responsive in the short term to income growth than to consumption growth. Imperfect matching of the time periods between the NSS and the NAS could also be attenuating the elasticities using NAS growth rates. However, the more important reason for lower (absolute) elasticities with NAS consumption or income is likely the divergence between NSS and NAS growth rates of mean consumption or income. Note that: d ln P d ln P d ln m ð6Þ ¼ : : d ln C d ln m d ln C An elasticity of m with regard to C (NAS consumption per capita) of around 0.5 (section II) would yield a poverty elasticity with regard to m that is about double that with regard to C—roughly in accord with table 1. When the period is split at 1991, the (absolute) elasticity of the headcount index with respect to the survey mean is appreciably higher in the post-1991 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 period (2.07) than in the pre-1971 period (1.58), and the difference is statisti- cally signi�cant.36 However, for the poverty gap measures, the difference in the elasticities for the two periods (2.63 and 2.94) is much smaller and is not statistically signi�cant. Finally, for the squared poverty gap measure, the elasti- cities are the same for the two periods (about 3.48). The pattern is similar using the IV method to control for correlated measurement errors, although the difference between the two periods is narrower and for the squared poverty gap measure the post-1991 elasticity (3.28) is lower than the pre-1991 elasticity (3.52). The vanishing difference in post- and pre-1991 elasticities for the higher order measures of poverty is consistent with the increase in inequality during the postreform period, given that the higher order poverty measures will tend to be more responsive to rising inequality. In contrast to the growth rates based on the survey means, both NAS-based growth rates indicate lower (absolute) elasticities in the post-1991 period, although the difference between the two periods is generally not statistically signi�cant. Exceptions are for the “unadjusted� elasticities of poverty gap and squared poverty gap, which are signi�cantly lower in the postreform period. It is notable, however, how much difference there is in the elasticity based on the NSS consumption growth rates and those based on the NAS rates for the post-1991 period. The much lower NAS elasticities reflect the much faster NAS-based growth than NSS-based growth. Since this growth divergence is more pronounced in the period after 1991, for the poverty gap and squared poverty gap measures it yields even lower (absolute) elasticities for this period relative to the pre-1991 period. 36. See Table 3 later in the article. These results are based on regressions of log poverty measures on the log survey mean interacted with dummy variables for pre- and post-1991 periods and a dummy variable for mixed-recall period surveys. The regressions also incorporate a kink at NSS round 47 (July–December 1991) such that there is no discontinuity in the predicted values of log poverty measures between the pre- and post-1991 periods. 178 THE WORLD BANK ECONOMIC REVIEW The estimated semi-elasticities, from the regression of Pt on ln mt, show a lower poverty impact of growth in the survey mean in the post-1991 period for the headcount index (–0.73, t ¼ –45.8), the poverty gap ( –0.34, t ¼ –32.3), and the squared poverty gap (–0.17, t ¼ –25.3) than in the pre-1991 period (–0.63, t ¼ –15.7; –0.20, t ¼ –9.82; and –0.08, t ¼ –7.24). This is to be expected; if elasticities are similar between the two periods, but poverty has fallen, absolute rates of decline will be lower in the later period. To summarize: the proportionate response of poverty to economic growth when measured from the NSS data remained roughly the same across the pre- and postreform periods, though with a slightly higher elasticity for the head- count index. However, there are signs that the responsiveness to growth measured through the NAS has declined during the postreform period. Urban–Rural Composition of Consumption Growth Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Table 2 summarizes the results of testing the poverty impact of the urban-rural composition of consumption growth.37 Table 3 presents the test statistics on whether the urban-rural composition of growth matters and whether the pop- ulation shift effect is signi�cant. These results on the relative effects of urban and rural growth are presented for national poverty measures and separately for urban and rural areas. IMPACT ON NATIONAL POVERTY. For the pre-1991 period, the hypothesis that only the overall rate of growth matters for poverty reduction is strongly rejected (table 3). The weaker hypothesis of uniform poverty effects of urban and rural growth is also strongly rejected. This echoes the results from Ravallion and Datt (1996) that the growth effects on poverty before1991 are attributable largely to rural consumption growth, with virtually no contri- bution from urban growth and only a limited contribution from the Kuznets process. However, there is a signi�cant structural shift between the pre-1991 and post-1991 periods. The hypothesis that growth effects are the same during the two periods is rejected (at the 8 percent level of signi�cance or better; see table 2). In the post-1991 period, the rural growth rate remains signi�cant for poverty reduction (with the possible exception of the squared poverty gap index), though the growth effects are smaller in absolute terms. Unlike in the pre-1991 period, rural growth does not appear to be the prime driver of national poverty reduction. The most notable change is that the (share- weighted) urban growth variable is now highly signi�cant. The null hypothesis that only the overall growth rate matters for poverty reduction in the post-1991 period can also be largely rejected (see table 3), although the evidence for a Kuznets effect is weaker during this period and limited to the headcount index. 37. Table 2 uses mean consumption from the surveys, since the NAS data do not permit an urban-rural breakdown. T A B L E 2 . Impacts on Poverty of the Urban-rural Composition of Growth: 1951–2006 National poverty Urban poverty Rural poverty Poverty measure Period Variable Coef�cient t-ratio Coef�cient t-ratio Coef�cient t-ratio Headcount index Up to 1991 Urban growth – 0.38 – 1.03 – 0.64 – 11.42 0.46 1.47 Rural growth – 1.45 – 21.79 – 0.08 – 4.16 – 1.38 – 34.32 After 1991 Urban growth – 3.73 – 2.40 – 0.94 – 3.47 – 2.96 – 2.07 Rural growth – 0.98 – 3.88 – 0.03 – 0.25 – 1.01 – 5.18 Ho: Pre-1991 coef�cient ¼ Post-1991 coef�cient F(2,34) (prob.) 2.787 (0.08) 0.52 (0.60) 3.44 (0.04) Ho: All Pre-1991 coef�cients ¼ Post-1991 coef�cients F(3,34) (prob.) 2.16 (0.11) 0.97 (0.42) 2.96 (0.05) Poverty gap index Up to 1991 Urban growth 0.21 0.27 – 0.67 – 4.54 0.90 1.16 Rural growth – 2.19 – 26.32 – 0.14 – 4.04 – 2.06 – 16.82 After 1991 Urban growth – 8.19 – 2.79 – 2.24 – 3.84 – 5.31 – 1.88 Rural growth – 1.59 – 3.72 0.00 0.03 – 1.59 – 3.27 Ho: Pre-1991 coef�cient ¼ Post-1991 coef�cient F(2,34)(prob.) 4.12(0.02) 3.25(0.05) 2.13(0.13) Ho: All Pre-1991 coef�cients ¼ Post-1991 coef�cients F(3,34)(prob.) 2.79(0.06) 4.50(0.01) 1.47(0.24) Squared poverty gap index Up to 1991 Urban growth 0.47 0.44 – 0.58 – 3.55 1.51 1.48 Rural growth – 2.69 – 15.27 – 0.17 – 4.13 – 2.54 – 11.80 After 1991 Urban growth – 11.64 – 2.33 – 3.95 – 4.77 – 7.45 – 1.70 Rural growth – 1.66 – 1.54 – 0.33 – 1.27 – 1.19 – 1.35 Ho: Pre-1991 coef�cient ¼ Post-1991 coef�cient F(2,34)(prob.) 2.73(0.08) 11.03(0.00) 2.33(0.11) Ho: All Pre-1991 coef�cients ¼ Post-1991 coef�cients F(3,34) (prob.) 1.86(0.15) 7.42(0.00) 1.56(0.22) Note: These are the p coef�cients in the regressions in equations (3) and (5) rather than elasticities. All regressions include a control for surveys that used a mixed-recall period (by adding the change between surveys in a dummy variable for such surveys). The regressions are estimated using a 2-stage GMM estimator. The instruments for the urban and rural growth variables included lagged survey means (split urban and rural), current and lagged mean consumption from the national accounts, current and lagged rural and urban consumer price indices, current and lagged rural population shares, Datt and Ravallion interval between mid-points of survey periods and a time trend. The t-ratios are based on heteroskedasticity and autocorrelation-consistent standard errors. Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. 179 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 180 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Test Statistics on the Signi�cance of the Pattern of Growth and the Kuznets Effect Pattern of growth Pattern of growth matters matters Ho: Kuznets effect Ho: piu ¼ pr pu ¼ pr ¼ pn ¼ p Ho: pn ¼ 0 Poverty measure Sector F(1,34) Prob. F(2,34) Prob. t ratio Prob. Headcount index Pre-1991 National 7.55 0.01 7.31 0.00 – 2.18 0.04 Urban 63.05 0.00 32.36 0.00 – 1.32 0.20 Rural 32.17 0.00 22.27 0.00 – 1.76 0.09 Post-1991 National 2.55 0.12 4.06 0.03 – 1.76 0.09 Urban 7.71 0.01 4.25 0.02 0.47 0.64 Rural 1.60 0.21 4.85 0.01 – 1.77 0.09 Poverty gap index Pre-1991 National 7.77 0.01 12.76 0.00 – 3.94 0.00 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Urban 9.38 0.00 5.63 0.01 – 1.69 0.10 Rural 11.74 0.00 12.04 0.00 – 3.33 0.00 Post-1991 National 4.72 0.04 2.78 0.08 0.28 0.78 Urban 10.84 0.00 9.10 0.00 1.62 0.12 Rural 1.56 0.22 1.24 0.30 0.25 0.81 Squared poverty gap index Pre-1991 National 6.98 0.01 8.49 0.00 – 3.08 0.00 Urban 4.37 0.04 8.52 0.00 – 3.48 0.00 Rural 11.74 0.00 9,74 0.00 – 2.72 0.01 Post-1991 National 3.54 0.07 1.82 0.18 0.31 0.76 Urban 13.56 0.00 10.09 0.00 1.68 0.01 Rural 1.81 0.19 1.38 0.27 – 0.31 0.76 Note: See equations (4), (5.1), (5.2) and discussion in text. Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. The emergence of a signi�cant effect of urban growth on national poverty is the most striking feature of these results. Table 4 reports the elasticities of national headcount, poverty gap, and squared poverty gap measures with respect to urban and rural growth. The contrast between the pre-1991 and post-1991 periods is compelling. While urban growth did not seem to matter for national poverty reduction before 1991, after 1991 not only did a sig- ni�cant urban growth effect emerge, but the urban growth elasticities of all three national poverty measures were higher (in absolute terms) than the corre- sponding rural growth elasticities. IMPACTS ON RURAL AND URBAN POVERTY. The urban-rural decomposition reveals something about the source of these differences between the pre- and post-reform periods. The hypothesis of no structural change is rejected for measures of the depth and severity of poverty in urban areas, but only for the headcount index in rural areas. However, for the rural depth and severity of Datt and Ravallion 181 T A B L E 4 . Elasticities of Poverty with Respect to Urban and Rural Growth: 1951–2006 Poverty measure Period National poverty Urban poverty Rural poverty Headcount index Urban growth Pre-1991 – 0.09 – 0.85 0.13 Rural growth Pre-1991 – 1.11 – 0.35 – 1.29 Urban growth Post-1991 – 1.21 – 1.26 – 1.26 Rural growth Post-1991 – 0.66 – 0.08 – 0.90 Poverty gap index Urban growth Pre-1991 0.05 – 0.89 0.25 Rural growth Pre-1991 – 1.68 – 0.61 – 1.91 Urban growth Post-1991 – 2.65 – 2.79 – 2.32 Rural growth Post-1991 – 1.08 0.01 – 1.46 Squared poverty gap index Urban growth Pre-1991 0.11 – 0.78 0.43 Rural growth Pre-1991 – 2.07 – 0.77 – 2.36 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Urban growth Post-1991 – 3.77 – 4.73 – 3.31 Rural growth Post-1991 – 1.12 – 0.83 – 1.11 Note: Elasticities are evaluated at means for the pre- and post-1991 periods using the par- ameter estimates reported in table 1. Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. poverty, too, the hypothesis of similar effects of urban growth for the two sub- periods is rejected. For the pre-1991 period, urban growth reduced urban poverty (see table 2), but so did rural growth, which had a signi�cant impact on poverty in both urban and rural areas for all three poverty measures. Indeed, for the squared poverty gap, the (absolute) elasticity of urban poverty to rural growth (0.77) is virtually the same as to urban growth (0.78; see table 4). The effect of urban growth, which for the pre-1991 period is con�ned to urban poverty, appears to be too small to be detected in the national average poverty measures in this period. The data for the post-1991 period look very different. Urban economic growth not only reduced urban poverty (as it did before), but had positive feed- back effects on rural poverty, especially the rural headcount index. Indeed, the estimated elasticities of rural poverty measures to urban growth are even higher than to rural growth. On the other hand, rural economic growth remains important to rural poverty reduction (in particular, for the incidence and depth of rural poverty), although there are signs that rural consumption growth has been somewhat less effective (in elasticity terms) against rural poverty in the post-1991 period. Also, the spillover effect to the urban poor has become con- siderably weaker in the post-1991 period for the headcount index and the poverty gap, though it remains strong for the squared poverty gap, suggestive of a continuing ( propoor) distributional effect in urban areas of rural economic expansion (see table 4). 182 THE WORLD BANK ECONOMIC REVIEW Figure 3 shows the estimated impact of urban economic growth for the periods before and after 1991. For each period, the �gure plots the change in log national headcount index that remains unexplained by rural growth against the change in log urban mean consumption. There was no signi�cant poverty-reducing effect of growth in mean urban consumption in the pre-1991 period, but a signi�cant impact emerges after 1991. The qualitative results are generally robust to the choice of poverty measure. As in Ravallion and Datt (1996), the growth elasticities tend to be highest (in absolute value) for the squared poverty gap and higher for the poverty gap than for the headcount index. As in Ravallion and Datt, the higher growth elasticity of the poverty gap than the headcount index implies that growth also reduces the depth of poverty (as measured by the mean poverty gap relative to the poverty line). Similarly, the even higher elasticity of the squared poverty gap implies that growth reduces inequality among the poor (as measured by the coef�cient of variation). Thus, the impacts of growth within and between Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 sectors are not con�ned to households in a neighborhood of the poverty line. There are two notable exceptions. The �rst is in the pre-1991 data for urban areas, where a slightly lower elasticity is found for the squared poverty gap than for the poverty gap in the effects of urban growth on urban poverty (see table 4). This suggest an underlying adverse distributional effect among the poor in the urban economic growth process of the prereform period. The second exception is in the impacts of rural economic growth on rural poverty in the post-1991 period, for the elasticity is lower for the squared poverty gap than for the poverty gap in the post-1991 period (see table 4). It appears that an adverse dis- tributional effect among the rural poor has emerged in the rural growth process of the prereform period. Compared with the earlier �ndings in Ravallion and Datt (1996), the most striking new result is the evidence that the urban economic growth process since 1991 has been appreciably more effective in reducing rural (and national) poverty. Since the regressions for rural poverty include rural mean consumption, the urban growth effect can be interpreted as a distributional effect. Supportive evidence is provided by the following regression of changes in the rural log Gini index (G r) of inequality on the (share-weighted) urban and rural growth rates:38 91 m 91 m D ln Gr t ¼ 1:54ð1 À dt ÞsutÀ1 D ln mut À 3:64 dt sutÀ1 D ln mut ð1:75Þ ðÀ1:68Þ 91 m 91 m ð7Þ À 0:20 ð1 À dt ÞsrtÀ1 D ln mrt þ 1:48 dt srtÀ1 D ln mrt ðÀ1:13Þ ð2:50Þ 1t R2 ¼ 0:32; n ¼ 41 À 0:08 DMRPt þ ^ ðÀ1:67Þ 38. Population shift effects were included (as in equation 5.2), but they were insigni�cant and are not reported. The share-weighted urban and rural growth terms are instrumented, as in table 3. Datt and Ravallion 183 F I G U R E 3. Poverty Impacts of Urban Economic Growth in India Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Note: The shaded area shows the 95 percent con�dence interval. Source: Authors’ calculations based on consumption data from 47 National Sample Surveys and on private �nal consumption expenditure and net domestic product data from national accounts and the population census; see text for details. 184 THE WORLD BANK ECONOMIC REVIEW where d91 t ¼ 1 for the post-reform period. From equation (7) it can be seen that, unlike in the pre-1991 period, higher growth rates of mean urban consumption since 1991 have reduced inequality in rural areas (signi�cant at the 10 percent level). Rural consumption growth, on the other hand, has had the opposite effect. Implications of Measurement Errors Concerns about underestimation of consumption in the NSS have implications for assessing how the urban-rural composition of growth has affected poverty. The proportionate bias in the NSS estimates of mean consumption may well be greater in India’s urban areas, where (as noted) it is widely thought that the NSS does not fully capture the consumption of the rich (notably for consumer durables and celebrations). Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Even so, the direction of any net bias in these estimates of the growth elas- ticity of poverty reduction is unclear a priori. There are three sources of poten- tial bias. First, greater measurement error in the log of mean consumption in urban areas than in rural areas would imply greater attenuation bias in the esti- mate of the impact of urban economic growth on poverty, leading to underesti- mation of the true elasticity, meaning that the true elasticity is more negative. Second, to the extent that the NSS is not fully capturing the growth in con- sumption by the relatively rich, the measured mean consumption growth rate from the surveys may be lower than the true rate.39 Call this the “growth-rate bias.� This will partly or even fully offset the attenuation bias; indeed, if the effect is strong enough, the measurement error in the mean may lead to an overestimate of the true elasticity, meaning that the true elasticity is less nega- tive. Third, some of the bias in estimating mean consumption will be passed onto the poverty measures—also pushing toward overestimation of the elas- ticity. This can be called the “spillover bias.� The net effect of these three potential sources of bias is unclear. Nor is it clear how much all of this would matter to the comparison of elas- ticities between the pre-and post-1991 periods. Since the balance of these effects cannot be determined on theoretical grounds, the conclusion that urban economic growth has become more poverty reducing may not be robust to cor- recting for measurement error in the NSS. The spillover bias is unlikely to be strong, since it is consumption by the urban nonpoor that tends to be underes- timated by the NSS; correcting for this bias would not have much effect on the poverty measures. However, by the same logic, the growth rate bias could be large, and so there can be no presumption that the attenuation bias would dominate. 39. In more technical terms, the measurement error in the NSS mean is not just a simple additive error in the log mean, as in the standard formulation of the attenuation bias in a regression coef�cient due to additive measurement error in the regressor. Datt and Ravallion 185 It could be argued that measurement error in the NSS has become a bigger problem in more recent years. That conjecture is at least consistent with the increasing divergence between the NSS mean and the NAS consumption aggre- gates, although this divergence could also stem from a rising share of the com- ponents of consumption included in the NAS aggregates that are not included in the NSS (including measurement errors in the NAS). Evidence is found of a lower elasticity of NSS consumption to NAS consumption in the postreform period, although the difference is small and not statistically signi�cant.40 However, this would presumably strengthen both the attenuation bias and the growth rate bias, leaving the net effect indeterminate. I V. C O N C L U S I O N S While progress against poverty has been uneven, the long-run trend has been a Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 decline in all three poverty measures based on a new time series of survey- based poverty measures for urban and rural India spanning 50 years, including 15 years after economic reforms started in earnest in the early 1990s. Exponential ( proportionate) trends are higher for the poverty gap and squared poverty gap indices than for the headcount index, reflecting gains to those living well below the poverty line. Both urban and rural poverty measures have shown a trend decline; rural poverty measures have historically been higher than urban measures, though the two have been converging over time. Progress against poverty has been maintained in the postreform period. Indeed, there was a higher proportionate rate of progress against poverty after 1991, although the difference in trend rates of change between the two periods is statistically signi�cant only for the headcount index. The linear trend—the annual percentage point reduction in the poverty measures—remained about the same in the postreform period. The responsiveness of poverty to growth in the survey mean—the growth elasticity of poverty reduction—has also gener- ally remained the same between the two periods; only for the headcount index is there a signi�cant increase in the absolute growth elasticity in the postreform period. When growth as measured in the NAS is used, there are signs that the postreform growth process has become less propoor in the sense of attaining a lower proportionate rate of poverty reduction from a given rate of growth. This seems to be the result largely of the faster postreform growth not being fully reflected in the surveys, and of the increase in inequality during the postreform period. The data do not make a robust case for saying that the growth elasticity of poverty reduction has risen (or fallen) since the reforms began. 40. The elasticities obtained by regressing consumption growth from the NSS on consumption growth from the NAS (with controls for changes in whether the round used a mixed-recall period and for changes in the log ratio of the rural price index to the NAS deflator) indicate that the elasticity is lower in the post-1991 period, declining from 0.57 (4.47) in the pre-1991 period to 0.45 (t ¼ 3.29). However, the null hypothesis that the elasticities for the two subperiods are the same cannot be rejected. 186 THE WORLD BANK ECONOMIC REVIEW Recognizing that the fortunes of the poor in urban and rural areas are linked in various ways—through trade, migration, and transfers—this study also revisited earlier ( prereform) �ndings on the relative importance of growth in urban and rural areas to poverty reduction in both areas and nationally (Ravallion and Datt 1996). Like that 1996 study, this one �nds that the pattern of growth matters for poverty reduction. But it also �nds a striking change in the relative importance of urban and rural economic growth in the postreform period. The 1996 study found that urban economic growth helped reduce urban poverty but brought little or no overall bene�t to the rural poor; the main driving force for overall poverty reduction was rural economic growth. This study con�rms that �nding for the data up to 1991, but the picture changes after 1991. As before, urban growth reduced urban poverty, and rural growth reduced rural poverty. But there is much stronger evidence of a feed- back effect from urban economic growth to rural poverty reduction in the post-1991 data than was found in the pre-1991 data. There are also signs that Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 the post-1991 rural growth has been less poverty reducing in rural areas. The relatively weak performance of India’s agricultural sector and the widen- ing disparities between urban and rural living standards remain important con- cerns, including for India’s poor. However, it is encouraging that rising overall living standards in India’s urban areas in the postreform period appear to have had signi�cant distributional effects favoring the country’s rural poor. While the attribution of this effect to the reforms is hardly conclusive—since there can be no comparison group for India after 1991 without the reforms—these �ndings are consistent with the view that with India’s efforts to create a more open and productive market economy has come a reversal in the historical pattern of weak feedback effects of urban economic growth on rural living standards. This may be a surprising conclusion considering that sectors that rely on skilled labor have been the most dynamic. However, the more relevant obser- vation is that the nonfarm sectors that use unskilled labor more intensively— notably trade, construction, and the “unorganized� manufacturing sectors—have seen higher employment growth in the postreform period. This is plausibly the main reason behind the stronger spillover effect of urban economic growth on the rural distribution of levels of living since 1991. This encouraging �nding comes with a warning, however. While the rural poor have bene�ted more from urban economic growth in the postreform economy, it can also be expected that they will be more vulnerable in the future to urban-based economic shocks. REFERENCES Ahluwalia, Montek S. 1978. “Rural Poverty and Agricultural Performance in India.� Journal of Development Studies 14 (3): 298–323. ——— 1985. “Rural Poverty, Agricultural Production, and Prices: A Reexamination.� In Agricultural Change and Rural Poverty: Variations on a Theme by Dharam Narain, ed. J.W. Mellor, and G.M. Desai. Baltimore, MD: Johns Hopkins University Press. Datt and Ravallion 187 ——— 2002. “Economic Reforms in India: A Decade of Gradualism.� Journal of Economic Perspectives 16 (2): 67– 88. Banerjee, Abhijit, and Thomas Piketty. 2005. “Top Indian Incomes: 1956–2000.� World Bank Economic Review 19 (1):1–20. Bell, Clive, and R. Rich. 1994. “Rural Poverty and Agricultural Performance in Post- Independence India.� Oxford Bulletin of Economics and Statistics 56 (2): 111–33. Bhagwati, Jagdish. 1993. India in Transition. Oxford: Clarendon Paperbacks. Bhattacharya, B., and S. Sakthivel. 2004. “Regional Growth and Disparity in India: Comparison of Pre- and Post-Reform Decades.� Economic and Political Weekly 39 (10): 1071– 77. Bhattacharya, N., D. Coondoo, and R. Mukherjee. 1991. Poverty, Inequality and Prices in Rural India. New Delhi: Sage Publications. Boyce, James K. 1986. “Kinked Exponential Models for Growth Rate Estimation.� Oxford Bulletin of Economics and Statistics 48 (4): 385– 91. Bruno, Michael, Martin Ravallion, and Lyn Squire. 1998. “Equity and Growth in Developing Countries: Old and New Perspectives on the Policy Issues.� In Income Distribution and High-Quality Growth, ed. Vito Tanzi, and Ke-young Chu. Cambridge, MA: MIT Press. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Chakravarty, Sukhamoy. 1987. Development Planning: The Indian Experience. Delhi: Oxford University Press. Chaudhuri, Shubham, and Martin Ravallion. 2006. “Partially Awakened Giants: Uneven Growth in China and India.� In Dancing with Giants: China, India, and the Global Economy, ed. L. Alan WintersShahid Yusuf, Washington, DC: World Bank. Datt, Gaurav. 1997. “Poverty in India 1951–1994: Trends and Decompositions.� International Food Policy Research Institute, Washington, DC. Datt, Gaurav, and Martin Ravallion. 1992. “Growth and Redistribution Components of Changes in Poverty: A Decomposition with Application to Brazil and India in the 1980s.� Journal of Development Economics 38 (2): 275–95. ———. 1997. “Macroeconomic Crises and Poverty Monitoring: A Case Study for India.� Review of Development Economics 1 (2): 135–52. ———. 1998. “Farm Productivity and Rural Poverty in India.� Journal of Development Studies 34 (4): 62– 85. ———. 2002. “Has India’s Post-Reform Economic Growth Left the Poor Behind?� Journal of Economic Perspectives 16 (3): 89– 108. Deaton, Angus. 2003. “Prices and Poverty in India, 1987–2000.� Economic and Political Weekly. January 25. ———. 2005. “Measuring Poverty in a Growing World (or Measuring Growth in a Poor World).� Review of Economics and Statistics 87 (1): 1– 19. ———. 2008. “Price Trends in India and their Implications for Measuring Poverty.� Economic and Political Weekly. February 9. Deaton, Angus, and Jean Dre ` ze. 2002. “Poverty and Inequality in India: A Re-Examination.� Economic and Political Weekly. September 7. Deaton, Angus, and A. Tarozzi. 2005. “Prices and Poverty in India.� In The Great Indian Poverty Debate, ed. Angus Deaton, and Valerie Kozel. New Delhi: MacMillan. ` ze, Jean, and Amartya Sen. 1995. India: Economic Development and Social Opportunity. Delhi: Dre Oxford University Press. Eswaran, Mukesh, and Ashok Kotwal. 1994. Why Poverty Persists in India. Delhi: Oxford University Press. Foster, Andrew D., and Mark R. Rosenzweig. 2004a. “Agricultural Development, Industrialization and Rural Inequality.� Brown University and Harvard University, Providence, RI and Cambridge, MA. ———. 2004b. “Agricultural Productivity Growth, Rural Economic Diversity, and Economic Reforms: India, 1970–2000.� Economic Development and Cultural Change 52 (3): 509–42, 188 THE WORLD BANK ECONOMIC REVIEW Foster, James, J. Greer, and Erik Thorbecke. 1984. “A Class of Decomposable Poverty Measures.� Econometrica 52 (3): 761–65. Gaiha, Raghav. 1995. “Does Agricultural Growth Matter to Poverty Alleviation?� Development and Change 26 (2): 285–304. Heltberg, Rasmus. 2004. “The Growth Elasticity of Poverty.� In Growth, Inequality and Poverty, ed. Anthony Shorrocks, and Rolph Van Der Hoeven. Oxford: Oxford University Press. India, Planning Commission. 1979. Report of the Task Force on Projections of Minimum Needs and Effective Consumption. New Delhi. ———. 1993. Report of the Expert Group on Estimation of Proportion and Number of Poor. New Delhi. ———. 2009. Report of the Expert Group to Review the Methodology for Estimation of Poverty. New Delhi. Jacoby, Hanan, Mariano Rabassa, and Emmanuel Skou�as. 2010. “Distributional Implications of Climate Change in India.� Development Research Group, World Bank, Washington, DC. Jha, Raghbendra. 2000. “Growth, Inequality and Poverty in India. Spatial and Temporal Characteristics.� Economic and Political Weekly. March 11. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Joshi, Vijay, and Ian Little. 1996. India’s Economic Reforms 1991–2001. Oxford: Clarendon Press. ˆ te D’Ivoire.� Review of Kakwani, Nanak. 1993. “Poverty and Economic Growth with Application to Co Income and Wealth 39 (2): 121– 39. Korinek, Anton, Johan Mistiaen, and Martin Ravallion. 2006. “Survey Nonresponse and the Distribution of Income.� Journal of Economic Inequality 4 (2): 33 –55. Kotwal, Askok, Bharat Ramaswami, and Wilma Wadhwa. 2009. “Economic Liberalization and India Economic Growth: What’s the Evidence?� University of British Columbia, Vancouver, BC. Kuznets, Simon. 1955. “Economic Growth and Income Inequality.� American Economic Review 45 (1):1– 28. Lanjouw, Peter, and Rinku Murgai. 2009. “Poverty Decline, Agricultural Wages and Nonfarm Employment in Rural India, 1983–2004.� Agricultural Economics 40 (2): 243– 64. Lipton, Michael, and Martin Ravallion. 1995. “Poverty and Policy.� In Handbook of Development Economics Vol 3, ed. Jere Behrman, and T.N. Srinivasan. Amsterdam: North-Holland. Panagariya, Arvind. 2008. India: The Emerging Giant. Oxford: Oxford University Press. Pur�eld, Catriona. 2006. “Mind the Gap: Is Economic Growth in India Leaving Some States Behind?� IMF Working Paper 06/103. International Monetary Fund, Washington, DC. Ravallion, Martin. 2000. “Should Poverty Measures be Anchored to the National Accounts?� Economic and Political Weekly. August 26. ———. 2003. “Measuring Aggregate Economic Welfare in Developing Countries: How Well do National Accounts and Surveys Agree?� Review of Economics and Statistics 85: 645–52. ———. 2007. “Inequality is Bad for the Poor.� In Inequality and Poverty Re-Examined, ed. John Micklewright and Steven Jenkins. Oxford: Oxford University Press. Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2007. “New Evidence on the Urbanization of Global Poverty.� Population and Development Review 33 (4): 667–702. Ravallion, Martin, and Gaurav Datt. 1996. “How Important to India’s Poor is the Sectoral Composition of Economic Growth?� World Bank Economic Review 10 (1): 1– 26. ———. 2002. “Why Has Economic Growth Been More Pro-Poor in Some States of India than Others?� Journal of Development Economics 68 (2): 381–400. Saith, Ashwani. 1981. “Production, Prices and Poverty in Rural India. Journal of Development Studies 17: 196–213. Sen, Abhijit. 2005. “Estimates of Consumer Expenditure and its Distribution. Statistical Priorities after the NSS 55th Round.� In The Great Indian Poverty Debate, ed. Angus Deaton, and Valerie Kozel. Delhi: Macmillan Press. Datt and Ravallion 189 Sen, Abhijit, and Hiamnshu. 2004a. “Poverty and Inequality in India 1.� Economic and Political Weekly. September 18. ———. 2004b. “Poverty and Inequality in India 2: Widening Disparities During the 1990s.� Economic and Political Weekly, September 25. Sen, Amartya. 1976. “Poverty: An Ordinal Approach to Measurement.� Econometrica 46 (2): 437–46. Sen, Kunal. 2009. “International Trade and Manufacturing Employment: Is India Following the Footsteps of Asia or Africa?� Review of Development Economics 13 (4): 765–77. Sundaram, K., and Suresh D. Tendulkar. 2001. “NAS-NSS Estimates of Private Consumption for Poverty Estimation: A Disaggregated Comparison for 1993–94.� Economic and Political Weekly January 13. Vakil, C.N., and P.R. Brahmanand. 1956. Planning for an Expanding Economy. Bombay: Vora and Company. van de Walle, Dominique. 1985. “Population Growth and Poverty: Another Look at the Indian Time Series Data.� Journal of Development Studies 21 (3): 429– 39. World Bank. 2005. World Development Report 2006: Equity and Development, Washington, DC: World Bank. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 ———. Forthcoming. Perspectives on Poverty in India: Stylized Facts from Survey Data. Washington, DC: World Bank. Are The Poverty Effects of Trade Policies Invisible? Monika Verma, Thomas W. Hertel, and Ernesto Valenzuela Beginning with the WTO’s Doha Development Agenda and establishment of the Millennium Development Goal of reducing poverty by 50 percent by 2015, poverty impacts of trade reforms have become central to the global development agenda. This has been particularly true of agricultural trade reforms due to the importance of grains in the diets of the poor, presence of relatively higher protection in agriculture, as well as heavy concentration of global poverty in rural areas where agriculture is the main source of income. Yet some in this debate have argued that, given the extreme vola- tility in agricultural commodity markets, the additional price and therefore poverty impacts due to trade liberalization might well be indiscernible. This paper formally tests the “invisibility hypothesis� using the method of stochastic simulation in a trade- poverty modeling framework. The hypothesis test is based on the comparison of two samples of price and poverty distributions. The �rst originates solely from the inherent variability in global staple grains markets, while the second combines the effects of inherent market variability with those of trade reform in these same markets. Results, at the national and stratum level indicate that the short-run poverty impacts of full trade liberalization in staple grains trade worldwide, are distinguishable in only four of the �fteen countries, suggesting that impacts of more modest agricultural trade reforms are indeed likely to be invisible in short run. Countries that show statistically signi�cant short run impacts are the ones characterized by high staple grains tariffs and/or a moderate degree of grain markets variability. Within each country, results are heterogeneous. In two thirds of the sample countries, agriculturally self-employed poor experience statistically signi�cant poverty impacts from trade liberalization. However, this �gure is under a third for all the other strata. Agricultural trade reform, computable general equilibrium, poverty headcount, volatility, stochastic simulation, hypothesis testing. JEL codes: C12, C68, F17, I32, Q17, R20 Monika Verma (corresponding author: verma3@purdue.edu) is Post-Doctoral Research Associate, Center for Global Trade Analysis, Department of Agricultural Economics, Purdue University. Thomas Hertel (hertel@purdue.edu) is Distinguished Professor, Department of Agricultural Economics, and Executive Director, Center for Global Trade Analysis, Purdue University. Ernesto Valenzuela (ernesto. valenzuela@adelaide.edu.au) is Senior Lecturer and Executive Director, Centre for International Economics, University of Adelaide. The authors wish to thank Antoine Bouet, Paul Preckel, William Masters, Alain de Janvry and three anonymous referees for helpful suggestions which led to signi�cant improvements of this manuscript. A supplemental appendix to this article is available at https://www.gtap.agecon.purdue.edu/resources/ res_display.asp?RecordID=3386 THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 190– 211 doi:10.1093/wber/lhr014 Advance Access Publication May 17, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 190 Verma, Hertel, and Valenzuela 191 World Trade Organization’s (WTO) Doha Development Agenda, and the Millennium Development Goal to reduce poverty by 50 percent by the year 2015, served to bring the poverty impacts of trade reforms into central focus for global policy makers. This has been particularly true of agricultural trade reforms due to the importance of grains in the diets of the poor, relatively higher protection in agriculture, as well as the heavy concentration of global poverty in rural areas where agriculture is the main source of income. Three quarters of the world’s poor reside in rural areas (World Development Report 2008), mostly depending for their livelihoods on agriculture; it is therefore hardly surprising that changes in primary commodity prices have been identi�ed as one of the most important linkages between international trade and poverty (Winters 2000). Agricultural commodity prices are of course inherently volatile, due to the combination of inelastic demand and supply, high perishability, high transport costs, and exposure to random weather shocks. The recent 2007/2008 food price spike,in fact, has been esti- mated to have thrown more than one hundred million people temporarily into poverty (Ivanic and Martin 2008). Given this background volatility in agricultural prices and poverty, some have argued that the additional poverty impacts due to trade liberalization might well be indiscernible. Indeed, in a critique of an early draft of Cline ´s (2004) book on trade policy and poverty, Rodrik (2003) made the point that the impacts of reforming agricultural protection in developed economies on world prices are likely to be dwarfed by the inherent volatility of agricultural markets. Similar sentiments surfaced in the context of the debate over the poverty impacts of trade liberalization under the Doha Development Agenda (Hertel and Winters 2006). This paper terms this assertion, the ‘invisibility hypothesis’. The goal of this paper is to formally test the invisibility hypothesis using a model of global trade, linked to poverty modules for �fteen developing countries. It is important to point out up front that statistically failing to reject the invisibility hypothesis by no means implies that agricultural trade reform is economically irrelevant. Even in cases where the long run impacts of agricul- tural trade reform are large, and of lasting importance, the short run impacts of such reforms on poverty might be statistically indiscernible due to the extreme volatility in international agricultural markets. As witnessed in recent years, commodity price swings of more than one hundred percent within a given year are not uncommon. These swings can themselves have a devastating effect on the poor – and they can also bene�t those households which are net sellers of agricultural products. Given the signi�cance of such commodity market volatility for the poor, it is important to couch agricultural trade reforms in this context. Also, the fact is that such reforms do not take place in a vacuum, and the presence of extreme market volatility will shape the way the world perceives them. It is important that those advocating agricultural trade 192 THE WORLD BANK ECONOMIC REVIEW reforms not overstate the near term impacts, which may indeed be dominated by other factors. On the other hand, it is also important to consider that, while the poverty changes induced by trade reforms may in some cases be smaller than those swings caused by inherent commodity market volatility, the gains from trade policy reforms represent permanent changes and are therefore likely to be of greater economic signi�cance than the transitory changes induced by annual market volatility. Previous literature on poverty impacts of trade reforms in the presense of inherent price variability is limited (Valenzuela 2009). Bourguignon et al. (2004) developed a stylized framework to assess the impact of export price variability on household income volatility. The related topic of the impact of higher food prices on poverty has also drawn attention (de Janvry and Sadoulet 2010; Ivanic and Martin 2008) as have the impacts of trade reforms on income distribution (Robbins 1996; Lunati and O ´ Connor 1999). However, none of these authors have offered a formal test of the invisibility hypothesis. The con- tribution of this paper is to provide such a test. The invisibility hypothesis is formulated as follows: Due to the high degree of volatility inherent in agricul- tural commodity markets, the incremental impact of agricultural trade liberali- zation on agricultural prices and the ensuing poverty impacts will be statistically invisible. The focus is on a subset of commodities – staple grains – which are often subject to high levels of protection, and which also represent a large share of the budget for the poorest households. Volatility in staple grains production is modeled by sampling from a distribution of productivity shocks derived from a time series analysis of Food and Agriculture Organization (FAO) production data. This supply-side volatility is implemented in a Computable General Equilibrium (CGE) framework – the agriculture-speci�c GTAP-AGR model (Keeney and Hertel 2005). The general equilibrium approach permits us to capture the implications of changes in national commodity and factor prices, resulting from changesin global trade policies as well as uncertainty in world grain yields, while retaining economy-wide consistency. Our analysis concen- trates on the implications of these earnings and price changes, for the utility of households in the neighborhood of the poverty line, asking whether they might fall below this poverty line or be lifted out of poverty as a result of these com- modity market shocks. By aggregating across the diverse socio-economic groups within the economy, a conclusion about the change in national poverty headcount can be inferred for each draw from the agricultural productivity dis- tribution. The resulting distribution of poverty headcounts is contrasted with the same distribution when trade reforms are implemented in combination with the inherent commodity productivity volatility. The �rst set of results, stemming from the inherent variability in global staple-grains markets, is referred to as the stochastic baseline scenario, while the combined effects of the inherent market variability and trade reforms is referred to as the stochastic policy reform scenario. While the model is general equilibrium in nature, price Verma, Hertel, and Valenzuela 193 volatility only in the staple grains markets is considered and therefore, to be consistent, the trade reforms are also only implemented in the staple grains sector. A further quali�cation stems from the fact that this is a static approach. Clearly a dynamic stochastic model would be preferred. This would permit us to distinguish permanent from transient shocks, with important implications for agents’ responses to these different types of shocks. However, this would introduce additional complexities that exceed the scope of this paper. In order to get an adequately broad representation of the diverse economies and circumstances in which the world’s poor live, this analysis is undertaken for �fteen developing countries in South Asia, Latin America and Sub-Saharan Africa. The remainder of this study is organized as follows. The methodology is described next (Section I). Section II presents the results for the moments of distributions for variables driving poverty headcounts changes before formally testing the invisibility hypothesis. It also provides a discussion of sensitivity of our results to the assumption of exogenous trade policy changes. Caveats, con- clusions and policy implications are discussed last (in Section III). I. ME T H O D O LO GY One approach to testing the invisibility hypothesis would be to develop a single country trade/poverty model in great detail and test this hypothesis in the context of that particular country. This is attractive, as it would allow development of the poverty component in considerable detail (see Hertel and Winters 2006, for ten country case studies undertaken to assess the national impacts of WTO reforms). However, there are several problems with this approach. Firstly, using a national model makes it dif�cult to generate stochas- tic global price shocks in a consistent manner. Secondly, WTO agricultural reforms typically entail signi�cant liberalization in developed markets, so without a global framework it is problematic to accurately assess the poverty impacts of such reforms on developing countries. Finally, readers would very likely argue that the results were speci�c to the country under investigation, if such tests of the invisibility hypothesis were undertaken only for an individual country. Therefore, a multi-country approach to testing the invisibility hypoth- esis is adopted. The cost of doing so is that the poverty analysis is necessarily rather simple and symmetric across countries. Poverty Headcount Analysis The analysis here relies on the trade/poverty approach outlined in Hertel et al. (2009). Those authors focus on poverty headcount changes in diverse house- hold population strata across a range of developing countries. A �rst order approximation to such poverty headcount changes may be written as follows 194 THE WORLD BANK ECONOMIC REVIEW (the hats denote percentage changes in the associated variables): X ^ rs ¼ À1rs � ^ H yprs ¼ À1rs � ap rsj ðw ^ rj À C ^ p Þ: r ð1Þ j Here, the index r denotes region (focus country), s the population stratum 1 and the superscript p signi�es that the variable in question is associated with earnings and consumption patterns at the poverty level of utility. Any shock to the economy that alters the after-tax returns to factor j (wrj ) and/or the prices of consumption goods, will affect the poverty level of income ( yp rs), the cost of p Phouseholds ^ living for poor (Cr ) and therefore strata poverty headcounts (Hrs). The term j ap rsj ð w ^ rj À C p r Þ in equation (1) is the percentage change in real factor income in stratum s of region r, taking into account the cost of living changes for households at the poverty line in stratum s of region r. The change in cost of living at the poverty line in region r, denoted C ^ p , is the change in r household expenditure required to keep utility constant at its poverty level, once a new set of prices is obtained. This change is derived by solving the household expenditure minimization problem at the new prices, while keeping utility �xed at the poverty level. Thus households are permitted to alter their optimal consumption bundle in response to the new commodity prices. Apart from the “driver� variables (after-tax factor earnings and commodity prices), two more elements play an important role in determining poverty head- count impacts. Coef�cient ap rsj is the share of factor earning j in total income for households at the poverty line, in stratum s of region r. For a given increase in factor earnings (e.g., unskilled agricultural labor), a stratum that obtains 90 percent of its income from this concerned factor, will experience a greater income rise than one with only 10 percent of its income attributable to that factor. Since these are shares, P the summation over factorpearnings types for any given stratum equals one ( j ap rsj ¼ 1). The values for arsj in our sample of 15 countries are obtained from household surveys and range from 0 to 0.99 (as shown in Appendix Table A1, available at https://www.gtap.agecon.purdue .edu/resources/res_display.asp?RecordID=3386). The second coef�cient of interest in equation (1) is 1rs, the poverty elasticity with respect to income in stratum s of region r. The higher the poverty elasticity, the greater the head- count reduction from a given increase in income for households at the poverty line in that particular stratum. Estimates of 1rs range from 0.0006 to 8.9 (Appendix Table A2), and vary widely by stratum and country.2 The change in total poverty headcount in a region is obtained by summing over stratum headcounts; therefore, the percentage change in national head- count can be written as share weighted sum of percentage headcounts changes 1. There are seven strata: Agriculturally self-employed, non-agriculturally self-employed, rural wage labor, urban wage labor, rural diversi�ed, urban diversi�ed and transfer stratum. 2. More details on the elasticities can be found in Verma et al. (2011). Verma, Hertel, and Valenzuela 195 at the stratum level: X ^r ¼ H ^ rs ; brs �H ð2Þ s where the shares (brs) are the share of stratum s in total poverty in the region r. brs plays an important role in determining how the stratum headcount changes get translated into the aggregate regional headcount.3 The initial equilibrium values for all of these coef�cients are estimated from household survey data for the 15 focus countries (Hertel et al. 2004) and are reported in Appendix Table A2. Substituting equation (1) in (2) gives the regional headcount in terms of its driving factors X X p ^r ¼ À H brs � 1rs � arsj ðw ^ p Þ; ^ rj À C ð3Þ r s j (3) can be further decomposed into changes due to pre-tax factor earnings ^m (w ^ r ), income tax changes (T ^ r ) designed to ensure revenue neutrality ^ rj þ T rj ¼ w of policy and the cost of living changes due to changed consumption prices, evaluated relative to the change in net national income: X X p  H^r ¼ À brs � 1rs � ^m arsj w À y ^r ^ r þ 1r ð C ^p À ^ þ 1r � T yr Þ : ð4Þ rj r s j The �rst term in equation (4) can be termed the earnings effect and involves the changes in factor earnings of the poor relative to national income. The second term is the tax effect and the last term identi�es the effect of changes in cost of living relative to net national income. The term 1r is the regional poverty elasticity andP is de�ned as the poverty share-weighted sum of strata poverty elasticities ( s brs � 1rs ). Any increase in taxes or relative cost of living raises poverty headcount in a region while increased relative factor incomes work towards poverty reduction. Overall, the poverty headcount in stratum s of country r falls when real income increases; the amount by which it falls depends on the density of the population in the neighborhood of the poverty line. Equation (4) offers a useful framework for analyzing the poverty impacts of trade and commodity market volatility. There are, however, some important limitations to its use which deserve a mention. Foremost among these is the 3. Consider for expository purposes that the poverty headcount for the rural diverse stratum for both Brazil and Uganda fell by 50 percent and other strata were unaffected (H ^ rs ¼ 0 8s = ruraldiverseÞ, then regional poverty headcount in Brazil would fall by a mere 1.5 (0.03 x 50) percent while in Uganda by a 37.5 (0.75 x 50) percent. The results are so diverse due to the big difference (0.03 versus 0.75) in the share of poverty population concentrated in the rural diverse stratum in the two countries, as can be seen from Appendix Table A2. 196 THE WORLD BANK ECONOMIC REVIEW static composition of the strata as the earnings specialization of households isn’t allowed to change; large shocks may induce a household to switch employment (e.g. moving from agriculture to non-agriculture), although this is less likely in the short run. In addition, the focus is only on changes in the poverty headcount; ignoring higher order measures, such as the poverty gap. The virtue of this simple approach is that it can be readily implemented across a wide range of household strata and countries, thereby permitting us to gener- alize our �ndings. Global General Equilibrium Model To calculate the impact of trade policy reforms on poverty headcount as per equation (4), one must �rst determine the impact of trade policy reforms on the poverty “drivers�, wrj and Cp r . The inability of partial equilibrium frame- works to predict the changes in economy-wide factor returns, which play a very prominent role in the analysis, forces us to use a CGE model in our analy- sis. One of the main criticisms of CGE models is the absence of validation (Kehoe et al. 1995). Accordingly, special attention is devoted to validating the model with respect to staple grains markets. This study employs the GTAP-AGR model of Keeney and Hertel (2005) which is explicitly designed to focus on issues of agricultural trade liberaliza- tion. (See Appendix I for details on the model structure and data sources used). A short-run factor market speci�cation is used such that land is commodity- speci�c, capital is sector-speci�c and labor is imperfectly mobile between agri- culture and non-agriculture sectors. The degree of inter-sector mobility is deter- mined by the choice of relevant parameters in the model. These are set, based on evidence on labor mobility from the OECD (2001). In addition, the model is modi�ed to accommodate the replacement of lost tax revenue from trade reforms, in the form of a non-distorting uniform ad valorem tax on income, making each scenario �scally neutral. Stochastic Simulation and Model Validation The credibility of any simulation model hinges very much on whether the model can produce reliable predictions for key endogenous variables, based on historical shocks. In practice, there are very few natural experiments involving trade policy reforms. WTO rounds are typically concluded once every decade or two, and their implementation is gradual and fraught with controversy. National reforms are sometimes more clear-cut; however, their effects are often confounded with other signi�cant events (e.g., a �nancial crisis, or a recession, etc.). Therefore a different type of natural experiment which is somewhat unique to the staple grains markets, is used for validation – the focus is on how well the model captures the economic impacts of random historic vola- tility in agricultural productivity, largely induced by weather-related shocks. Given the relative stability of demand for subsistence goods such as staple grains, demand-side volatility is ignored here; characterizing only supply side Verma, Hertel, and Valenzuela 197 volatility in the staple grains markets and thereupon asking whether the model is capable of reproducing observed price volatility in these same markets. If the model can accurately characterize inter-annual price volatility in response to supply side shocks then it is also a valid tool for looking at the short run impacts of tariff shocks in these same markets. The validation approach involves using production shocks derived from the residuals of time series models of FAO grains production data. By sampling from the derived distribution, the stochastic simulation seeks to mimic the ran- domness inherent in these markets. Solving the CGE model repeatedly, each time with a different set of productivity draws, produces the resulting distri- bution of price changes for each region. The validation then involves compar- ing the model results for grain price variation, with FAO observed price variation in each region. With the aim of improving the CGE replication of observed FAO price variability, the model’s consumer demand elasticities were adjusted for a few regions; details of the approach are given in the next sub- section. After ensuring the historic price variation is faithfully replicated, one can concentrate on contrasting the poverty headcount distributions associated with the stochastic baseline and stochastic policy reform scenarios and testing for statistical difference between these two sets of results. Characterizing Volatility. Tyers and Anderson (1992) characterize uncertainty in global food markets by sampling from a distribution of supply shocks. Valenzuela et al. (2007) use this approach to validate a model of global wheat trade. The same approach has been used here. Autoregressive Moving Average (ARMA) models are used to characterize systematic changes in staple grains pro- duction, using the ARMA residuals to de�ne the distributions of productivity shocks. This speci�cation is appealing in modeling grain crops production because past values appear to carry a great deal of information about current values and prediction errors arise largely from weather-related shocks to pro- duction. Staple grains production data from the FAO for the period 1991 to 2006 (FAOSTAT)4 is used to calculate the productivity shocks for aggregate regions.5 The 15 focus countries inherit the shocks from their respective parent region. The model selection is guided by the signi�cance of the AR and MA com- ponents, the Akaike Information Criteria (AIC), and autocorrelation in residuals for alternative model speci�cations. The normalized standard devi- ations of the production residuals from the estimated time series models are used to create a distribution reflecting random regional productivity variation. 4. While paddy rice and wheat are the same across GTAP and FAOSTAT terminology, the Coarse-grains category under GTAP covers barley, maize, mop corn, rye, oats, millet, sorghum, buckwheat, quinoa, fonio, triticale, canary seed, mixed grain and cereals nes. reported in FAO data. 5. Calculations using FAOSTAT data show that measures of observed volatility in output vary considerably depending what aggregation of crops and regions is used. Generally speaking, the higher the level of aggregation, the lower is the volatility that the CGE model is adjusted to replicate. The aggregation scheme for regions is provided on Appendix Table A4. 198 THE WORLD BANK ECONOMIC REVIEW The greatest production volatility is seen in Russia6, Sub-Saharan Africa and Eastern Europe. With the assumption that productivity follows a symmetric tri- pffiffiffi the end points of this distribution are determined by the angular distribution, formula: mean + 6  productivity standard deviation. This estimated distri- bution of productivity shocks for each region provides the basis for implement- ing the stochastic baseline scenario. The methodology involves sampling from this distribution of productivity shocks and solving the CGE model repeatedly. The results for each solve of the model are stored and the means and standard deviations of the stored results for all endogenous variables are calculated. The sampling is done by means of Gaussian Quadrature (GQ), a numerical integration technique developed as an alternative to Monte-Carlo simulations, and implemented for GTAP models by Pearson and Arndt (2000). The GQ technique is chosen instead of the more traditional Monte Carlo approach, as it signi�cantly reduces the number of simulations while still preserving the accuracy of the resulting means and stan- dard deviations for endogenous variables (DeVuyst and Preckel 1997). The validity of evaluating the impacts of trade liberalization in the context of a volatile grains market environment critically depends on the capability of the CGE system to replicate historical price variability. This capacity of the model is assessed by comparing the model simulated volatility for staple-grains prices, to FAO-observed volatility (Table 1). Since staple grains represent a composite of many commodities, a range of historic price volatilities from the FAO data base is reported in the �rst column of this Table. For example, in Bangladesh, price volatility of rice, wheat and coarse-grains, as measured by the normalized stan- dard deviation of ARMA residuals, ranges from 5% to 12%. In the Philippines, this is a smaller range (10% to 13%). Initial results indicated that the model overstated price volatility for Philippines, Bangladesh, Colombia, Peru, Venezuela, Malawi and Mozambique; while it understated the same for Thailand. Aiming to replicate the price volatility for these regions more closely, the consumer demand elasticities in these regions were re-calibrated.7 Speci�cally, demand elasticities were increased for regions where price volatility was over-predicted by the model, while they were reduced for Thailand. Elasticities were also increased for all the rest of Sub-Saharan African regions, as the model predicted unusually high price volatility for these countries. The price volatility results, after adjusting the elasticities, are reported for comparison (Table 1). This calibration process enables the CGE model to replicate the FAO data price variation in most cases (with the exceptions of Thailand, Colombia and Venezuela). For Colombia and Venezuela the model over-states price volatility. This could be due to the Andean Price Band 6. This region includes Russia and all the constituent states of the former Soviet Union. 7. More details and justi�cation for this approach is provided in Verma (2010). Verma, Hertel, and Valenzuela 199 T A B L E 1 . Historic versus Model Generated Price Volatility and Associated Percentage Changes in Poverty Headcount Historic Model Generated Mean Percent Volatility Range Volatility Results Change (stochastic baseline)** Bangladesh 5-12 11 0.19 Indonesia 9-19* 11 0.07 Philippines 10-13* 13 2 0.10 Thailand 11-14 7 0.02 Vietnam 7 0.18 Brazil 11-20 12 0.09 Chile 7-21 11 0.03 Colombia 4-10 14 0.07 Mexico 7-9 9 0.12 Peru 6-15 15 0.08 Venezuela 6-11 18 0.12 Malawi 21-30 23 2 0.01 Mozambique 16-20 19 0.12 Uganda 22 2 0.07 Zambia 19 0.12 Source: FAO Price-Stat Data 1991-2006, Model generated price variation results and Authors Calculations using Model Simulation Results * FAO Price data on wheat is not available for Indonesia and Philippines; so the range reflects the price volatility of rice and coarse grains only. FAO Price data on none of the crops is available for Vietnam, Uganda and Zambia. ** These changes in Poverty Headcount in absence of any policy shock arise as a result of inherent changes in agricultural commodity prices. System policy which was implemented in 1995 and involved variable tariffs in Peru8, Colombia, Venezuela and Ecuador aimed at restricting price fluctu- ations in these markets (Villoria et al. 2002). The model used here does not reflect these country speci�c policies and therefore misses these effects.9 For Thailand the model under-predicts price volatility – a problem similar to that faced by Valenzuela et al. (2007) who found that the same type of global CGE model under-predicted price variations for most exporters (Thailand is a major exporter of rice). The base case scenario here does not incorporate the endogenous response of border policies to changes in global market condition such as the export bans and import policy changes which arose in the context of the 2007/2008 food crisis. These policies tend to exacerbate price volatility – particularly for exporters (see Valenzuela et al. for more details). Implications of such policy endogeneity are briefly explored in Section III below. 8. In the case of Peru, the model generated price variation reaches the upper limit of the observed price variation range. 9. In principle it would be desirable to model these policies explicitly. However owing to the diverse range and complexity of policies across countries, such an endeavor is better-suited to a country case study approach. 200 THE WORLD BANK ECONOMIC REVIEW With mean zero agricultural productivity shocks under the stochastic base- line, mean zero outcomes for most model variables are to be expected. The last column of Table 1 in fact shows that the mean changes for the poverty head- counts10 are less than 1 percent. Modeling Staples Trade reforms. The year 2001 is adopted as the bench- mark, as it is the base year for Doha proposals on tariff cuts and also the base year for the GTAP version 6.1 data (Dimaranan 2006). The �rst column of Table 2 shows tariffs in the staple grains sector for all of the 15 focus countries. Mexico has the highest import tariffs for staple grains, followed by Thailand and Peru. Overall, the focus countries have much lower tariffs on staple grains than do the non-focus countries (rest of world). This study con- siders a scenario of trade liberalization which involves the complete removal of tariffs in all focus, as well as non-focus countries.11 Though our simulations focus on full liberalization in all countries, under realistic trade negotiations different countries may undertake different levels of agricultural tariff reductions. To be consistent with the stochastic baseline simulations (variabil- ity is restricted to staple grains production), trade reforms in other sectors of the economy are not considered. Thus, our stochastic policy reform scenario is the combined effect of inherent market variability and the complete elimination of effectively applied tariffs in staple-grains market. II . R ES ULT S Elimination of the import tariffs for staple grains is expected to result in lower consumer prices in countries with high initial tariffs (see Table 2). Since the average import tariff in the focus regions is about 11 percent, countries with higher than 11 percent tariffs are expected to experience greater consumer price reductions, and therefore potential greater poverty reductions (abstracting from the earnings side of the poverty story). We focus on the impacts of trade reform in the context of the stochastic simulations.12 A good starting point – before focusing on poverty headcount distributions, at the aggregate regional as well as the disaggregated stratum levels – is the 10. Any big numbers in thousands of units can be explained by the presence of a big poverty base (Appendix Table A5). Note that as the percent change in poverty headcounts is the average percentage change in the variable across 22 simulations, the decomposition of results though along the lines of deterministic setup is not as straightforward. Most of the analysis therefore focuses not on what is driving the means but on a more relevant question that the stochastic framework can answer: whether the distributions with and without reforms are different. 11. The focus is on tariffs-only policy reform as data on domestic support and export subsidies is not available on a consistent global basis. Croser and Anderson (2010) using a partial equilibrium framework and a recent World Bank comprehensive set of indicators of distortions to agricultural incentives (Anderson and Valenzuela 2008) found that border measures in agricultural markets account for more than 85 percent of global loss of welfare. 12. Readers interested in a detailed deterministic analysis of the impacts of tariff reform alone are referred to Appendix II. Verma, Hertel, and Valenzuela 201 T A B L E 2 . Import Tariffs on Staples and Mean and Standard Deviations for Staples Prices (Percentage Change), Before and After Tariff Liberalization Under a Stochastic Scenario Country(s) Stochastic Baseline Scenario Stochastic Policy Scenario ts (Tariffs) Mean Standard-Deviation Mean Standard-Deviation Bangladesh 4.54 0.7 6.0 0.0 5.7 Indonesia 1.47 0.9 7.7 2 4.0 6.3 Philippines 6.19 1.1 6.1 2 12.1 4.7 Thailand 20.42 1.1 4.4 25.0 5.7 Vietnam 2.76 2.6 7.2 6.9 5.8 Brazil 0.14 1.1 4.9 1.6 3.9 Chile 6.98 1.0 9.4 0.7 9.2 Colombia 12.77 0.5 3.8 2 2.8 3.6 Mexico 23.94 0.9 8.6 2 10.9 7.2 Peru 16.46 0.5 4.9 2 5.1 4.4 Venezuela 12.10 1.2 9.7 2 1.1 9.4 Malawi 0.08 4.0 20.6 3.1 19.9 Mozambique 2.11 2.8 16.8 2 1.5 16.0 Uganda 0.72 2.4 14.5 0.8 14.1 Zambia 2.90 2.9 17.5 2.2 17.0 Poverty Regions’ Average Tariff 11.39 Average Tariff for other Regions 34.89 World Average Tariff 30.65 Source: Calculations using GTAP Database version 6 for tariffs and using Model Simulation Results for others The average applied import tariffs are calculated as P "  #! X VIWirs X  VIWirs P Pr P ts ¼ � Ãtirs i r i VIWirs r r VIWirs where the Value of Imports (VIWirs) at World prices by commodity(i), source (r) and destination(- s);and the tariff rates (tirs) come from GTAP version 6 database. comparison of pre and post reform distributions of endogenous variables that “drive� the poverty headcount results. As indicated in equation (1), these are the consumption prices and factor earnings. Distributions of Driver Variables The comparison of the mean and standard deviations of driving factors – staple grains consumption prices (affecting the cost of living) and real after tax factor earnings (affecting income) – across the stochastic baseline and stochas- tic policy scenarios should provide insights into the results of formal test of the poverty headcount distributions under the same scenarios. If the moments of distributions for these variables differ little across the two scenarios, then results for poverty headcounts will also very likely not be distinguishable. 202 THE WORLD BANK ECONOMIC REVIEW Tables 2 and 3 present the results for these driver variables – staple con- sumption prices and factor incomes respectively – for all 15 countries. For example the mean staple grains consumption price in Mexico increases by 0.9 percent under stochastic baseline while it falls by 10.9 percent under the sto- chastic policy scenario; so there is a difference of 11.9 percentage points between the two scenarios for Mexico. The same �gure for Thailand is 25.0 – 1.1 ¼ 23.9 percentage points (Table 2). For Mexico, most of the change is driven by the reduction in prices as a result of removing high import tariffs in the country. The increase in staples price in Thailand is driven by the increased price of rice owing to increase in rice export demand (Appendix II). The reforms seem to bene�t consumers in Latin America; as the mean staple prices are lower (except for Brazil13) and no more volatile under the stochastic policy scenario as compared to the baseline scenario (Table 2). Also it is interesting to see that, while mean outcomes (especially for staple grains) show some differ- ence, the standard deviations across the scenarios are very similar, this we suspect, is partly due to the omission of endogenous trade policy responses to world market price variability in this analysis. (This issue is explored in more detail below.) A similar comparison of after-tax real factor earnings is offered for the focus countries (Table 3). The �rst panel in the table gives the percentage point differences in means while the bottom panel gives the same for standard devi- ations. A positive number indicates that the post liberalization mean or stan- dard deviation for the factor endowment in a given country is higher than under the baseline (no liberalization) scenario. The changes in factor returns to land and agricultural capital are greater than those for other factors due to their sector speci�city and more limited factor mobility respectively. Also, as with the results in Table 2, the changes in standard deviations are generally modest. Small changes in the standard deviation compared to the mean suggest that the Kolmogorov-Smirnov (henceforth KS) two sample test14 can be used as a formal test of difference in distributions of consumption prices and factor earn- ings. The KS test answers the question: are the observations under one scenario systematically larger or smaller than under the other scenario? Test results for staples consumption prices and factor earnings are provided in Appendix III. The main conclusion of the test is that, while some factors earnings are statisti- cally different post liberalization, this �nding is by no means universal across factors and countries. Given that strata are de�ned based on income specializ- ation, it is therefore likely that the results will show within-country, across- stratum differences in the visibility of poverty headcounts. The next section 13. The deterministic results in Appendix II show the reason for increased post liberalization prices in Brazil on account of increased demand for coarse-grain exports from the country. 14. This test in comparison to other non-parametric tests performs better for cases where there is not much difference in variance (Baumgartner et al. 1998). T A B L E 3 . Differences in Mean and Standard Deviations of Real After-Tax Factor earnings between Stochastic Policy and Stochastic Baseline Scenarios (percent changes) Bangladesh Indonesia Philippines Thailand Vietnam Brazil Chile Colombia Mexico Peru Venezuela Malawi Mozambique Uganda Zambia Mean Land 2 1.0 2 2.5 2 3.9 15.4 4.5 1.4 2 1.7 2 0.7 2 2.1 2 2.5 2 0.7 2 1.7 2 1.4 2 0.6 2 1.3 Ag-Unskilled 0.0 2 0.4 0.2 1.4 0.5 0.0 2 0.1 0.2 0.4 0.0 0.0 2 0.2 0.3 0.0 0.0 Ag-Skilled 0.0 2 0.1 0.9 0.9 0.4 0.0 2 0.1 0.2 0.5 0.0 0.0 2 0.1 0.4 0.1 0.1 NonAg-Unskilled 0.2 0.7 1.9 2 2.1 2 0.2 0.0 0.1 0.3 0.8 0.4 0.1 0.0 0.5 0.2 0.2 NonAg-Skilled 0.2 0.8 2.1 2 2.2 2 0.1 0.0 0.1 0.3 0.8 0.4 0.1 0.0 0.5 0.2 0.2 Wage-Unskilled 0.2 0.5 1.2 2 1.5 2 0.1 0.0 0.1 0.3 0.8 0.3 0.1 2 0.1 0.4 0.1 0.1 Wage-Skilled 0.2 0.8 2.1 2 2.2 2 0.1 0.0 0.1 0.3 0.8 0.4 0.1 0.0 0.5 0.2 0.2 Ag-Capital 2 1.0 2 2.5 2 3.9 15.4 4.5 1.4 2 1.7 2 0.7 2 0.6 2 2.6 2 0.7 2 1.7 2 1.4 2 0.6 2 1.4 Nag-Capital 0.2 0.8 2.1 2 2.4 2 0.1 2 0.1 0.1 0.4 0.9 0.6 0.1 0.0 0.5 0.2 0.2 Transfers 0.1 0.5 1.2 2 1.0 0.2 0.0 0.0 0.3 0.8 0.3 0.1 2 0.3 0.3 0.0 0.1 Standard Deviation Land 2 0.7 2 0.7 2 1.2 1.8 2 2.4 0.0 2 0.1 0.1 2 0.2 2 0.8 2 0.1 0.2 2 0.3 2 0.5 2 0.3 Ag-Unskilled 0.0 0.0 0.3 0.1 0.0 0.0 0.0 0.0 0.0 2 0.1 0.0 0.6 0.1 0.0 0.1 Ag-Skilled 2 0.1 0.0 2 0.3 0.1 0.0 0.0 0.0 0.0 0.0 2 0.1 0.0 0.5 0.0 2 0.1 0.2 NonAg-Unskilled 2 0.1 2 0.2 2 0.5 0.0 2 0.3 0.0 0.0 0.0 0.0 2 0.1 0.0 0.4 0.0 2 0.1 0.1 NonAg-Skilled 2 0.2 2 0.2 2 0.6 0.0 2 0.3 0.0 0.0 0.0 0.0 2 0.1 0.0 0.4 0.0 2 0.1 0.1 Wage-Unskilled 2 0.1 2 0.1 2 0.4 0.0 2 0.2 0.0 0.0 0.0 0.0 2 0.1 0.0 0.5 0.0 0.0 0.1 Wage-Skilled 2 0.2 2 0.2 2 0.6 0.0 2 0.3 0.0 0.0 0.0 0.0 2 0.1 0.0 0.4 0.0 2 0.1 0.1 Ag-Capital 2 0.7 2 0.7 2 1.2 1.8 2 2.4 0.0 2 0.1 0.1 0.0 2 0.8 2 0.1 0.2 2 0.3 2 0.5 2 0.3 Nag-Capital 2 0.1 2 0.2 2 0.4 0.0 2 0.3 0.0 0.0 0.0 0.0 2 0.1 0.0 0.4 0.0 2 0.1 0.0 Transfers 2 0.1 2 0.1 2 0.2 2 0.1 2 0.1 0.0 0.0 0.0 0.0 2 0.1 0.0 0.1 0.0 0.0 0.0 Source: Authors’ calculations Using Model Simulation Results Verma, Hertel, and Valenzuela 203 204 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . K-S Test Statistics, P-Values and Moments of Distributions Across the Baseline and Policy Scenarios for Poverty Headcount Changes Trade Liberalization Stochastic Baseline in Stochastic Scenario Framework (in thousands) Calculated KS Standard Standard Test Statistic P-value Mean deviation Mean deviation Bangladesh 0.18 0.84 83 598 14 553 Indonesia 0.41 0.04 10 41 25 31 Philippines 0.27 0.39 2 11 277 98 262 Thailand 0.18 0.63 0 7 1 8 Vietnam 0.14 0.87 3 13 11 0 Brazil 0.14 0.92 21 113 27 106 Chile 0.36 0.06 0 3 2 3 Colombia 0.32 0.20 3 16 26 16 Mexico 0.64 0.00 11 104 2 128 96 Peru 0.36 0.06 3 24 29 20 Venezuela 0.27 0.39 4 24 21 23 Malawi 0.32 0.20 0 17 8 26 Mozambique 0.23 0.57 7 46 26 48 Uganda 0.27 0.39 2 12 14 25 9 Zambia 0.14 0.92 7 59 2 62 Source: Authors’ calculations using Model Simulation Results The negative numbers under the mean columns are to be interpreted as a reduction in poverty headcount. offers a formal test of differences between poverty headcount distributions at both the country and stratum levels. Distribution of Poverty Headcounts The KS test is implemented to formally compare the two distributions of poverty headcounts, resulting from stochastic baseline and stochastic policy reform scenarios. The null hypothesis is that the two distributions are not stat- istically different and are therefore hard to tell apart. Calculated KS test stat- istic values, along with the associated P-values are reported for all the focus countries (Table 4). The table shows that poverty headcount changes following trade liberalization are statistically perceptible at a 10 percent level of signi�- cance, in just four countries: Indonesia, Chile, Peru and Mexico. Figure 1 shows what the results look like graphically for two cases – Bangladesh and Mexico – one where the two distributions are not statistically distinct and the other where they are clearly differentiated. The lines in the �gure are the sample cumulative density functions (CDFs) for poverty head- count changes in the two countries under alternative scenarios. The sample Verma, Hertel, and Valenzuela 205 F I G U R E 1. Empirical Cumulative Distributions of Poverty Headcount Changes in Mexico and Bangladesh Source: Model simulation results observations refer to the pooled samples generated by the repeated model simu- lations under the two stochastic scenarios (see Appendix III for technical details). The maximum vertical distance between the two lines is the KS test statistic. For Bangladesh, these sample CDFs lie very close together while they lie farther apart and do not overlap for Mexico. The samples in question are those generated under the stochastic baseline and under stochastic trade liberal- ization. The �gure brings out the nature of our results very clearly – the effects are visibly distinct in one case while not in the other. This point is further underscored via a set of diagnostic plots reported for all the sample countries in Appendix IV. 206 THE WORLD BANK ECONOMIC REVIEW The results of the KS test of the invisibility hypothesis for all seven strata in each of the 15 focus countries are provided in Table 5. At the 10 percent level of signi�cance ( p-value less than 0.1), for only 30 of the 105 country- stratum pairs the results turn out to be statistically distinct. Note that in Indonesia the headcount changes are statistically visible only in the agricultural stratum; however because this stratum has a 42 percent share in the national poverty headcount (Appendix Table A2), its statistical signi�cance carries over to the overall national-level invisibility hypothesis test results. Conversely, while poverty headcount changes are statistically signi�cant for all but the rural and urban diverse strata in Philippines, the changes at the national level are not sig- ni�cant. This stems from the fact that the two diversi�ed strata comprise over 70 percent of Philippines’s poverty population (23 percent for urban diverse and 49 percent in rural diverse, as shown in Appendix Table A2). Two other cases that stand out in these results are Chile and Thailand. For Chile, while the national level impacts of trade reform are visible, only the agricultural stratum is statistically signi�cant and it makes up only 26 percent (Appendix Table A2) of the country’s poverty headcount. In Thailand, while all strata show signi�cantly perceptible poverty headcount results, the same does not hold for the national level results due to the fact that the agricultural head- count reductions are offset by the urban increases. The lower panel of Table 5 sheds further light on this conundrum. For Thailand while the stratum head- counts scenarios differ signi�cantly, the total headcount does not vary much between the two scenarios. The opposite is true for Chile, where poverty rises for most strata. The broad �ndings are that short-run poverty changes resulting from liberal- izing staples sectors are large enough to be discernable in four of the �fteen focus countries: Peru, Mexico Chile and Indonesia. The P-values for the KS test (Table 4) suggest that the impacts are most likely to be visible in Latin America and least likely to be so in Africa and Asia. The visibility of impacts depends on initial level of tariffs, degree of inherent volatility and magnitude of policy shocks. Mexican poverty reduction bene�ts from reduction in a rela- tively high staple grains import tariff (Table 2), whereas most of the Sub-Saharan African countries fail to get visible impacts due to their highly volatile domestic markets. Also, even though the results are not statistically visible at the country level for some cases, a look at a more disaggregated (stratum) level can reveal a different result. In a cross country comparison, while the national level results for Peru and Malawi look very different (Table 4); the change in agricultural stratum poverty in both countries is equally visible (Table 5). Similarly poverty changes in rural diverse stratum in Peru and Venezuela are invisible to the same degree (Table 5). What If Trade Policy Changes Are Endogenous? So far the analysis assumed that trade policy changes are exogenous and are not subject to short term manipulation. However as seen in recent years, T A B L E 5 . P-Values for KS test of “Invisibility Hypothesis� at Stratum Level Agriculturally Non-agriculturally Rural Urban Transfer Rural Urban self-employed self-employed Labor Labor Income Diverse Diverse Bangladesh 0.84 0.84 0.84 0.84 0.84 0.84 0.84 Indonesia 0.01 0.20 0.25 0.25 0.39 0.87 0.63 Philippines 0.06 0.01 0.00 0.00 0.01 0.39 0.84 Thailand 0.00 0.00 0.00 0.00 0.00 0.02 0.04 Vietnam 0.001 0.84 1 1 0.63 0.84 0.04 Brazil 0.25 0.63 0.63 0.63 0.92 1 1 Chile 0.06 0.84 0.84 0.84 0.92 0.57 0.20 Colombia 0.06 0.22 0.20 0.20 0.20 0.20 0.20 Mexico 0.00 0.001 0.001 0.001 0.00 0.00 0.00 Peru 0.00 0.04 0.04 0.04 0.02 0.39 0.22 Venezuela 0.39 0.39 0.39 0.39 0.39 0.39 0.39 Malawi 0.01 0.84 0.84 0.84 0.63 0.25 0.25 Mozambique 0.57 0.57 0.57 0.57 0.57 0.57 0.57 Uganda 0.84 0.92 0.92 0.92 1 0.57 0.84 Zambia 0.87 0.92 0.92 0.92 0.92 0.92 0.92 Mean Poverty Headcount Changes across Scenarios and Strata (in ‘000) Agriculturally Non-agriculturally Rural Urban Transfer Rural Urban Scenario self-employed self-employed Labor Labor Income Diverse Diverse Total Thailand Reform 2 5.1 1.3 2.8 0.2 4.3 2 3.3 0.5 0.7 No-reform 2 0.3 0.1 0.2 0.0 0.3 2 0.1 0.0 0.2 Chile Reform 1.1 0.0 0.0 0.0 0.1 0.2 0.3 1.7 No-reform 2 0.2 0.0 0.0 0.1 0.2 0.0 0.0 0.1 Verma, Hertel, and Valenzuela Source: Authors’ calculations using Model Simulation Results 207 208 THE WORLD BANK ECONOMIC REVIEW exporters and importers frequently resort to imposing export taxes and lower- ing import tariffs respectively, when world market prices rise sharply. Martin and Anderson (2012) argue that such policy actions contributed as much as a third of observed world price changes in the recent commodity price boom. Because the analysis thus far has ignored this possibility, it has potentially understated the bene�ts of altogether eliminating trade barriers for staple grains. The logic behind this argument is as follows. If Martin and Anderson are correct, and endogenously varying trade policies serve to amplify world price changes in the face of production shortfalls, then fully eliminating such policies should have a stabilizing influence on world prices. Furthermore, by reducing price variations, trade liberalization would be more likely to result in statistically discernable changes in poverty. To explore this possibility, a robustness check is undertaken to compare the liberalization scenario to a new baseline scenario wherein export taxes and import tariffs respond endogenously to changes in commodities’ export and import prices faced by a country. Lacking information on country-speci�c approaches to insulation and seeking to obtain an outer bound on our results, the results reported in this section pertain to the case where all countries seek to insulate their domestic markets from world price changes. This is done by manipulating border taxes to eliminate half of the deviation in border prices (relative to a global trade price index).15 The results indicate that when all exporters and importers resort to such responses, world prices under the new baseline scenario increase by a factor of about two in comparison to the case when trade shocks are treated as exogenous. The standard deviation of inter- national prices as well becomes twice as large; while not much is achieved on moderating domestic price movements, due to the greater international price volatility under the new baseline. In effect, when every country attempts to export its price variability, no country is able to stabilize its prices. This con- �rms the theoretical arguments presented by Martin and Anderson. Turning to the invisibility hypothesis; because domestic commodity and factor prices are the driving forces behind poverty headcount results, and they aren’t affected greatly since the attempt to insulate is frustrated by greater world price vola- tility, one should not expect to see a big difference in the poverty results. Applying the KS test for the new regional headcount numbers, still leads to the invisibility hypothesis being rejected for the same four countries (Appendix Table A11); however, at the stratum level results do look somewhat different, with 5 more cases gaining signi�cance (Appendix Table A12) in the presence of endogenous policies. 15. This value of 0.5 is not entirely random. Anderson and Nelgen (2010) provide estimates suggesting that only about half the movement in international prices is transmitted to domestic markets for the period spanning 1985-2007. Verma, Hertel, and Valenzuela 209 III. CONCLUSIONS This study developed a framework to address the question about the relative size of trade policy induced poverty changes versus those induced by the inherent volatility in agricultural markets; this is a different question and should not be confused with the more familiar question of quantifying the poverty impacts of trade reforms. Even if trade reforms are economically relevant, it is entirely poss- ible that trade policy reform induced changes in a country’s poverty headcounts are large but invisible, due to the high degree of commodity market volatility, as seen in the case of the Philippines. Conversely, modest impacts from grains liber- alization may be visible in markets like Chile. The differences in visibility results can be explained by the differences in initial level of tariffs, degree of inherent volatility and magnitude of policy shocks faced by a country. Overall, at the national level, the short-run poverty impacts of full liberaliza- tion of grains’ trade are statistically distinguishable in less than a quarter of our sample countries. Even though policies do affect poverty headcounts in the remaining 11 countries, the changes are masked by the price changes due to the volatile nature of grains markets. So, broadly speaking, this study fails to reject the hypothesis that the short run national poverty impacts of trade pol- icies are in fact invisible in the presence of volatile commodity markets. It is therefore important for the advocates of agricultural trade liberalization to not overstate the near term impacts. However, the results vary by stratum within countries, and the results for individual strata can be very different from the country level results. An extreme example is given by the case of Thailand where, even though poverty headcounts are visible at the stratum level, the invisibility hypothesis at the national level cannot be rejected. Also, not surprisingly, the visibility is highest for poverty changes amongst the agriculturally self-employed (Table 5). Results for this stratum are found to be signi�cant for 9 of the 15 sample countries. Therefore, the answer to the invisibility question depends on the level (national or stratum) at which the question is asked. Certainly the impacts of trade reforms on agriculture-specialized households in countries with relatively stable commodity markets are quite likely to be visible. APPENDICES Appendices can be found on Internet at:https://www.gtap.agecon.purdue.edu/ resources/res_display.asp?RecordID=3386. REFERENCES Anderson, K., and S. Nelgen. 2010. “Trade Barrier Volatility and Agricultural Price Stabilization.� CEPR Discussion Paper 8102, London, November. 210 THE WORLD BANK ECONOMIC REVIEW Anderson, K., and E. Valenzuela. 2008. “Estimates of Global Distortions to Agricultural Incentives, 1955 to 2007.� Database available at www.worldbank.org/agdistortions. Baumgartner, W., P. Weiß, and H. Schindler. 1998. “A Nonparametric Test for the General Two-Sample Problem.� Biometrics, Vol. 54 (3), 1129–35. Bourguignon, F., S. Lambert, and A. Suwa-Eisenmann. 2004. “Trade exposure and income volatility in cash-crop exporting developing countries.� European Review of Agricultural Economics, Vol 31 (3), 369 –87. Cline, W.R. 2004. Trade Policy and Global Poverty. Washington, D.C. Institute for International Economics. Croser, J., and K. Anderson. 2010. “Changing contributions of different agricultural policy instruments to global reductions in trade and welfare.� Policy Research Working Paper Series 5345, The World Bank. de Janvry, A., and E. Sadoulet. 2010. “The Global Food Crisis and Guatemala: What Crisis and for Whom?� World Development, 38 (9), 1328–1339. DeVuyst, E.A., and P.V. Preckel. 1997. “Sensitivity Analysis Revisited: A Quadrature Based Approach.� Journal of Policy Modeling, 19 (2), 175–85. B.V. Dimaranan (editor). 2006. “Global Trade, Assistance, and Production: The GTAP 6 Data Base.� Center for Global Trade Analysis, Purdue University. FAOSTAT (Food and Agriculture Organization Statistical Database). Internet site: www.fao.org. Hertel, T.W., M. Ivanic, P.V. Preckel, and J.A.L. Cran�eld. 2004. “The Earnings Effects of Multilateral Trade Liberalization: Implications for Poverty.� The World Bank Economic Review, Vol. 18 (2), 205 –36. Hertel, T.W., and L.A. Winters (eds.). 2006. “Putting Development Back Into the Doha Agenda: Poverty Impacts of a WTO Agreement.� New York: Palgrave Macmillan, co-published with the World Bank. Hertel, T.W., R. Keeney, M. Ivanic, and L.A. Winters. 2009. “Why Isn’t the Doha Development Agenda more Poverty Friendly?� Review of Development Economics, Volume 13 (4), 543–59. Ivanic, M., and W. Martin. 2008. “Implications of higher global food prices for poverty in low-income countries�, Agricultural Economics, Volume 39 (0), 405–16. Keeney, R., and T.W. Hertel. 2005. “GTAP-AGR: A Framework for Assessing the Implications of Multilateral Changes in Agricultural Policies.� GTAP Technical Paper Series No.24. Purdue University, West Lafayette, Indiana, U.S.A. Kehoe, T.J., C. Polo, and F. Sancho. 1995. “An Evaluation of the Performance of an Applied General Equilibrium Model of the Spanish Economy.� Economic Theory, Vol. 6 (1), 115–41. Lunati, M.R., and D. O’Connor. 1999. “Economic opening and the demand for skills in developing countries: a review of theory and evidence�. OCDE/GD(96)182. Paris: OECD. Martin, W., and K. Anderson. 2012. “Export Restrictions and Price Insulation During Commodity Price Booms.� American Journal of Agricultural Economics 94 (1), (forthcoming). Organization for Economic Cooperation and Development. 2001. Market Effects of Crop Support Measures. OECD Publications, Paris, France. Pearson, K., and C. Arndt. 2000. “Implementing Systematic Sensitivity Analysis Using GEMPACK.� GTAP Technical Paper No. 3. Internet site: www.gtap.org Robbins, D. J. 1996. “Evidence on trade and wages in the developing world.� OCDE/GD(96)182. Paris: OECD. Rodrik, D. 2003. Trade Liberalization and Poverty: Comments, presented at the Center for Global Development/Global Development Network Conference on Quantifying the Impact of Rich Countries’ Policies on Poor Countries, Institute for International Economics, Washington, D.C., October 23 –24, 2003. Tyers, R., and K. Anderson. 1992. Disarray in World Food Markets. Cambridge: Cambridge University Press. Verma, Hertel, and Valenzuela 211 Valenzuela, E., T.W. Hertel, R. Keeney, and J. Reimer. 2007. “Assessing Global Computable General Equilibrium Model Validity Using Agricultural Price Volatility.� American Journal of Agricultural Economics, May 2007, Vol. 89 Issue 2, p383– 397. Valenzuela, E. 2009. “Poverty Vulnerability and Trade Policy: General Equilibrium Modelling Issues.� VDM Verlag Dr. Mu ¨ ller. Verma, M. 2010. Assessing The Poverty Impacts When Commodity Prices Are Volatile. Ph.D. Dissertation, Department of Agricultural Economics, Purdue University. Verma, M., T.W. Hertel, A. Rios, and M. Ivanic.2011. “GTAP-POV: A Framework for Assessing the National Poverty Impacts of Global Economic and Environmental Policies.� Center for Global Trade Analysis, Purdue University. Mimeo. Villoria, N., and D.R. Lee. 2002. “The Andean Price Band System: Effects on Prices, Protection and Producer Welfare.� Paper presented at American Agricultural Economics Association, 2002 Annual meeting, July 28–31, Long Beach, CA. Winters, L.A. 2000. Trade and Poverty: Is There a Connection? in Trade, Income Disparity and Poverty. Ben David, D.; H. Nordstrom and L. A. Winters, eds. Special Study 5, Geneva: WTO. World Development Report. 2008.Agriculture For Development. Washington, D.C.: World Bank. Corruption and Con�dence in Public Institutions: Evidence from a Global Survey Bianca Clausen, Aart Kraay, and Zsolt Nyiri 1 Well-functioning institutions matter for economic development. In order to operate effectively, public institutions must also inspire con�dence in those they serve. We use data from the Gallup World Poll, a unique and very large global household survey, to document a quantitatively large and statistically signi�cant negative correlation between corruption and con�dence in public institutions. This suggests an important indirect channel through which corruption can inhibit development: by eroding con�- dence in public institutions. This correlation is robust to the inclusion of a large set of controls for country and respondent-level characteristics. Moreover we show how it Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 can plausibly be interpreted as reflecting at least in part a causal effect from corrup- tion to con�dence. Finally, we provide evidence that individuals with low con�dence in institutions exhibit low levels of political participation, show increased tolerance for violent means to achieve political ends, and have a greater desire to “vote with their feet� through emigration. JEL classi�cation: D73, O12, O17 Despite considerable debate over de�nitions, measurement, and methodology, it is widely-accepted among academics and policymakers that well-functioning public institutions play an important role in economic development. In turn, a key ingredient in the effectiveness of public institutions is the con�dence that they inspire among those whom they serve. For example, households or �rms who do not have con�dence in the police or the courts are unlikely to avail themselves of their services, and may resort to other informal means of prop- erty protection or dispute resolution. Similarly, if individuals lack con�dence in the honesty of the electoral process they are unlikely to vote, leading to low 1. Aart Kraay (corresponding author), The World Bank, Development Economics Research Group, akraay@worldbank.org, Bianca Clausen, The World Bank, Development Economics Research Group, bclausen@worldbank.org, Zsolt Nyiri, German Marshall Fund, znyiri@gmfus.org. We would like to thank the Knowledge for Change (KCP) Program of the World Bank for �nancial support, and Alfonso Astudillo, Michael Clemens, Mary Hallward-Driemeier, Claudio Raddatz, and seminar participants at Gallup, the World Bank, and Stanford University. We are also grateful to the Gallup Organization for enabling this project by providing access to the Gallup World Poll data, and especially to Gale Muller, Vice Chairman and General Manager of the Gallup World Poll. The views expressed here do not reflect those of the Gallup Organization, the German Marshall Fund of the United States, the World Bank, its Executive Directors, or the countries they represent. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 212– 249 doi:10.1093/wber/lhr018 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 212 Clausen, Kraay, and Nyiri 213 turnout rates that cast doubt on elected of�cials’ popular mandates and their ability to carry out their agendas. These effects of corruption on con�dence have not been lost on policymakers. A recent quotation from Kai Eide, UN Special Representative of the Secretary-General for Afghanistan, neatly encap- sulates this view: “..[Corruption] pushes people away from the state and under- mines our joint efforts to build peace, stability and progress for Afghanistan’s peoples.� 2 In this paper we empirically investigate the role of corruption in undermin- ing con�dence in public institutions. We document a quantitatively large and statistically signi�cant partial correlation between measures of corruption and con�dence in public institutions using a unique dataset. The Gallup World Poll (GWP) is a large cross-country household survey, interviewing more than 100,000 households in over 150 countries, annually or biennially in most countries since 2006. We use questions from the 2008/2009 wave of the GWP, covering over 78,000 respondents in a single cross-section of 103 countries to Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 study the links between corruption and con�dence in public institutions in both developed and developing countries. Not surprisingly, in countries where respondents report a high incidence of personal experiences with corruption, and in which corruption is perceived to be widespread, con�dence in public institutions is also low. Much more interestingly, we show that this pattern also holds across individuals within countries: individuals who experience cor- ruption and who report that corruption is widespread also tend to have lower con�dence in public institutions. We show that this correlation is robust to the inclusion of a large set of variables to control for respondent-level character- istics, including a number of proxies intended to capture the respondent’s ten- dency to complain or report more negatively on corruption and con�dence than might otherwise be objectively warranted. Our goal in this paper is not to develop new theoretical understandings of the links between corruption and con�dence in public institutions. Rather, our much more modest objective is to signi�cantly improve the quality of the exist- ing empirical evidence on the relationship between the two. Relative to the existing empirical literature on this topic (which we discuss in more detail below), we offer three important contributions. First and most basic, our study covers a much larger set of countries and respondents than any previous work, which due to data limitations typically has been focused on small, usually regionally-focused samples of countries. Second, several features of the GWP allow us to include a very rich set of respondent-level control variables, impor- tantly including proxies for respondents’ unobserved propensity to respond negatively to both questions about corruption and con�dence that might arti�- cially bias our results towards �nding a strong effect of corruption on con�dence. 2. UNAMA Press Release, United Nations Assistance Mission, August 20, 2008. 214 THE WORLD BANK ECONOMIC REVIEW Third and perhaps most important, we offer a more serious treatment of a key identi�cation problem that has largely been ignored by the existing litera- ture. Simply documenting that survey respondents answer “yes� to a question like “is corruption a problem in your country� and “no� to a question like “are you con�dent in your national government�, as most of the previous lit- erature has done, does little to identify the direction of causation between the two. Perhaps respondents’ perceptions of the prevalence of corruption drive their low con�dence in institutions, but just as plausibly the opposite could be true: individuals who lack con�dence in public institutions might as a result express the view that corruption is widespread. We address concerns about endogeneity in two ways. The �rst is to exploit the difference in responses to two questions asked in the GWP. As we discuss in more detail below, the GWP asks both a generalized perceptions of corrup- tion question, as well as a very speci�c experiential question which asks whether the respondent has been asked for a bribe in the past 12 months. The Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 advantage of the latter question is that it is much more plausibly exogenous to respondents’ con�dence in public institutions since it in large part reflects the decision of a public of�cial to solicit a bribe from the respondent, rather than the respondent’s own characteristics. Consistent with this view, we �nd that the estimated effect of the experienced corruption question is substantially smaller and less statistically signi�cant than the corresponding estimated effect using the generalized perceptions question. However it remains strongly signi�- cant and quantitatively large, supporting our claim of an important and plausi- bly causal effect running from corruption to con�dence in public institutions. Second, as we argue in more detail below, even the partial correlation between corruption experiences and con�dence might reflect some degree of reverse cau- sation from con�dence to corruption experiences. To assess this concern, we also perform a bounds analysis which shows that such reverse causation is unli- kely to fully overturn our �nding of a signi�cant causal effect of corruption on con�dence. The rest of this paper proceeds as follows. In the next section we review the related literature. Section II describes the main features of the Gallup World Poll and compares our key corruption variables with other widely used ones. Section III contains our main empirical results linking corruption to con�dence in public institutions. In Section IV we explore a number of robustness checks for this partial correlation, and in Section V we discuss in detail the identi�- cation problem and potential solutions. In Section VI we briefly document some consequences of the corruption-induced loss of con�dence in public insti- tutions, showing that individuals with low con�dence in public institutions are less likely to engage in the political process, are more likely to condone vio- lence as a means to further political ends, and are more likely to “vote with their feet� by emigrating. Section VII concludes. Clausen, Kraay, and Nyiri 215 I . RE L AT E D L I T E R AT U R E It is widely accepted by scholars and policymakers that well-functioning insti- tutions are important for development. This view has been informed by a wide range of historical analysis, case studies, and cross-country empirical analysis. A few examples from this very large literature include North (1990), Knack and Keefer (1995, 1997), Kaufmann et al. (1999), Acemoglu et al. (2001), and Rodrik et al. (2004). The idea that a lack of con�dence in public institutions undermines their effectiveness has also been widely studied. A few examples of this literature include Easton (1965, 1975), Gibson and Caldeira (1995), Putnam (2000), Uslaner (2002), Gibson et al. (2003), and Mishler and Rose (2005). There is also a large literature on the direct economic consequences of corruption for growth and investment, including Mauro (1995), Knack and Keefer (1995), Mo (2001), Pellegrini and Gerlagh (2004), and Me ´ on and Sekkat (2005), and reviewed by Me ´ on and Sekkat (2004) and Lambsdorff Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (2007). Our contribution is to the small but growing literature on the effects of cor- ruption that operate through con�dence in public institutions. As our contri- bution in this paper is primarily empirical, we focus in this review only on the empirical aspects of previous papers that have studied the links between cor- ruption and con�dence in public institutions. We refer the interested reader to some of these other papers, most notably Anderson and Tverdova (2003), for an extensive discussion of the various theoretical channels through which cor- ruption might impact con�dence in public institutions. A number of early papers in this literature exploit only the country-level variation in perceptions of corruption and con�dence in public institutions. These include Pharr (2000) who looks at aggregate data over time for one country (Japan); Della Porta (2000) who provides a narrative discussion of country-level averages of both corruption and con�dence for just three countries; and Anderson and Tverdova (2003) who combine country-level data on corruption perceptions from the Transparency International Corruption Perceptions Index with household survey data on con�dence from 16 mostly developed countries. The major drawback of such studies is the possibility that excluded country characteristics (or year effects in the case of Pharr (2000)) may be confounding the observed relationship between corruption and con�- dence in public institutions. A second set of papers improves on these by relying on household-level variation in survey responses to questions about corruption and con�dence to estimate the correlation between the two. These include Rose, Mishler, and Haerpfer (1998), Mishler and Rose (2001), Catterberg and Moreno (2005), and Chang and Chu (2006), who all document a negative partial correlation between perceptions of corruption and con�dence in public institutions in small and regionally-focused samples of countries. These papers however do not address the identi�cation problem to which we have referred in the 216 THE WORLD BANK ECONOMIC REVIEW introduction: it is unclear from the partial correlations documented by these authors whether respondents’ perceptions of corruption drive their con�dence in public institutions, or the converse. Also in this category is a related paper by Hellman and Kaufmann (2004), who investigate how an alternative measure of corruption perceptions influ- ences �rms’ con�dence in, and use of, public institutions. They use data from the World Bank’s Business Environment and Enterprise Performance Survey of 6500 �rms in transition economies in 2002 to construct a measure of perceived ‘crony bias’ as the difference between �rms’ perceptions of their own influence and the influence of other �rms they view as having strong political connec- tions. They show that �rms who perceive a great deal of crony bias in policy- making have less con�dence in the judiciary, are less likely to use courts, are more likely to pay bribes, and are more likely to cheat on their taxes. Here as well, however, the direction of causation between corruption perceptions and con�dence in institutions is unclear. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Four more recent papers improve on the ones discussed so far by relying on respondent-level data on personal experiences with corruption (and not simply perceptions of corruption) to study the effects on con�dence in public insti- tutions. Seligson (2002) uses survey data for four Latin American countries to test the effects of corruption experiences on perceptions of the legitimacy of the political system at the individual level. He �nds that exposure to corruption erodes belief in the political system and reduces interpersonal trust. Bratton (2007) uses survey data from 18 African countries to document that percep- tions of corruption are negatively correlated with respondents’ satisfaction with public services, but somewhat surprisingly, personal experience with bribery is positively associated with user satisfaction. These papers share with ours the advantage of relying on corruption experiences questions which plausibly are more exogenous than corruption perceptions. However, these papers do not consider the further possibility we address later in the paper, that even responses to the corruption experiences question may be endogenous to responses to the con�dence questions.3 Finally, Cho and Kirwin (2007) and Lavalle ´ e, Raza�ndrakoto and Roubaud (2008) also use a set of African countries covered by the Afrobarometer survey to investigate directly the links between con�dence in public institutions and both corruption perceptions and corruption experiences questions using respondent-level variation. Unlike the rest of the literature surveyed so far, these papers are the only ones to explicitly acknowledge the potential for reverse causality and seek to address it. Cho and Kirwin (2007) in particular emphasize the possibility of vicious circles: corruption undermines con�dence 3. The identi�cation problem is compounded by the fact that, despite having record-level data for many countries, Bratton (2007) does not appear to include country �xed effects in his speci�cations. This opens the possibility that unobserved country-level effects are confounding the relationship between corruption and satisfaction with public services that he studies. Clausen, Kraay, and Nyiri 217 in public institutions, and this in turn increases the acceptability of offering bribes to obtain public services, increasing the prevalence of corruption.4 Both papers propose using instrumental variables drawn from the same survey in order to address this identi�cation problem. However, as we explain in more detail below in our discussion of identi�cation, this strategy depends on the validity of – in our view implausible – exclusion restrictions that the authors fail to adequately justify. In summary, the existing literature on the effect of corruption on con�dence in public institutions has been based on small samples of countries, and has for the most part failed to recognize or address the dif�culty of isolating the direc- tion of causation between corruption and con�dence. In the remainder of this paper we show how we can use the very large sample size and the richness of the GWP core questionnaire to make progress on these issues. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 II. CORRUPTION AND CONFIDENCE IN INSTITUTIONS IN THE GALLUP WORLD POLL The Gallup World Poll (GWP) has been �elded annually or biennially since 2006 in over 150 countries representing 95% of the world’s adult population, and asks questions on a wide range of topics. This makes it the largest (in terms of country coverage) annual multi-country household survey in the world. The surveys are based on a standard methodology and considerable effort goes into ensuring comparability across countries. The surveys are designed to be nationally representative of people who are 15 years old or older and great efforts are made to interview households in rural areas, as well as politically unstable and insecure areas. The surveys are in-depth face-to-face interviews in all countries except the most developed countries such as Western Europe or Australia where a shorter version of the survey is �elded by phone.5 The majority of the core questions on the Gallup World Poll are not political in nature. Instead they concern individuals’ well-being, asking about their every- day lives, level of happiness, life-satisfaction, expectations about their future, daily experiences of stress, etc.6 This tends to build a higher level of trust between the interviewer and respondent than a more technical-sounding 4. A related point is made by Sacks (2011) who argues theoretically and empirically that it is dif�cult for governments to embark on public sector reform programs absent some measure of public trust in the government. If in turn poor public sector management leads to corruption which undermines trust in government, there is the possibility of a “trap� where governments viewed as corrupt do not have the legitimacy required to carry out reforms that might actually reduce corruption. 5. For documentation of the GWP survey methodology refer to http://www.gallup.com/consulting/ worldpoll/108079/Methodological-Design.aspx 6. In this context, we note that a number of recent scholarly papers have used the GWP data for empirical research. Examples include Deaton (2008, 2009), Helliwell (2008), Ng et al. (2008), Stevenson and Wolfers (2008), Deaton et al. (2009), Gandelman and Herna ´ ndez-Murillo (2009), Helliwell et al. (2009), Krueger and Malec ´ (2009), and Pelham et al. (2009). The majority of these ˇ kova focus on GWP questions related to subjective assessments of personal well-being. 218 THE WORLD BANK ECONOMIC REVIEW government-use questionnaire. Together with an explicit statement by the enumerator regarding the con�dentiality of responses, this likely helps to improve respondent candor on some of the more sensitive questions in the survey.7 We combine countries from the 2008 and early 2009 waves of the GWP into a single cross-section of countries8. As our key measure of corruption we use the following speci�c question about the respondent’s personal experience with corruption: “Sometimes people have to give a bribe or present in order to solve their problems. In the last 12 months, were you, personally, faced with this kind of situation, or not (regardless of whether you gave a bribe/present)?� This question, which we will refer to this as the “corruption experiences� ques- tion was a new addition to the core GWP questionnaire in the 2008 wave of surveys. However, for reasons of timing and questionnaire space, it was asked in only 115 of the 124 countries covered in our sample of the GWP in 2008 and early 2009. This question was asked in most high-income OECD, Latin Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 American, Asian and African countries, but coverage of Eastern Europe is scarcer. Nevertheless the breadth of GWP data still allows us to study the effects of corruption experiences in a much larger sample of countries and respondents than any previous work. The GWP also asks a more generic question about the corruption percep- tions of respondents that we will use alongside the experience question in this paper: “Is corruption widespread throughout the government in this country, or not?� We refer to this as the “corruption perceptions� question. It was asked in 112 of the 124 countries in our sample. However, as the samples of countries in which the corruption experience and perception questions were �elded do not match perfectly, the sample in which both questions were asked comprises 103 countries. Appendix Table A contains a full description of all questions from the GWP used in the paper, and the �nal sample of 103 countries is listed in Appendix Table B. There are substantial conceptual and practical differences between the cor- ruption experiences and corruption perceptions question. The former asks about a respondent’s personal experiences with corruption, while the latter solicits the respondent’s views about the prevalence of corruption, regardless of whether the respondent has witnessed or experienced any corrupt acts himself. We note �rst that one would naturally expect to see differences between the responses to the two questions. The corruption experiences question is poten- tially a good gauge of “petty� or administrative corruption that individuals might be likely to experience in their everyday lives: a policeman asking for a 7. However, one should not conclude that all respondents are fully candid in their responses to all questions. For approaches to identifying reticent respondent biases and applications, see Azfar and Murrell (2009) and Clausen, Kraay and Murrell (2010). 8. At the time of our access to the data, the relevant corruption questions had been asked only once in each country, and so we are unfortunately not able to exploit any within-country over-time variation in the data. Clausen, Kraay, and Nyiri 219 bribe instead of issuing a ticket, or a bureaucrat soliciting an irregular payment for a permit. On the other hand, the corruption perceptions question can potentially capture the prevalence of broader forms of corruption, particularly at higher levels of government. The downside of this latter question of course is that it does not draw on the respondent’s personal experience, but rather is informed by the respondent’s exposure to second-hand information about corrupt activities.9 As we argue in more detail in Section 4, a crucial advantage of the corruption experiences question is that it is less likely to suffer from reverse causality, in the sense that individuals’ con�dence in institutions affects their corruption experiences. This will be very important for our interpretation of the empirical results that follow. Figures 1 and 2 illustrate the country-level variation in these two measures of corruption from the GWP. Figure 1 plots country average corruption percep- tions versus corruption experiences. All countries in the sample fall above the 45-degree line, indicating that on average, respondents are more likely to Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 answer “yes� to the corruption perceptions question than to the corruption experiences question, in all countries. In some countries, this gap is large: for example, Japan and Italy have low rates of personal experience with corrup- tion, but nevertheless strong perceptions of widespread corruption in govern- ment. One interpretation is that this suggests low rates of petty or administrative corruption but a greater incidence of high-level or political cor- ruption. In Figure 2 we plot the two corruption questions from the GWP against a broad perceptions-based measure of corruption, the Worldwide Governance Indicators ‘Control of Corruption’ variable (Kaufmann et al., 2008). Both corruption questions display a fairly strong negative correlation with the Control of Corruption measure. However, this correlation is far from perfect, in part due to the fact that the Control of Corruption measure aggre- gates information from a large number of different data sources. Our main objective in this paper is to empirically document the links between corruption and con�dence in public institutions. We measure the latter using another question in the GWP, which asks respondents about their con�dence in a variety of institutions at the national level. Speci�cally, the GWP asks “Do you have con�dence in each of the following?: (a) the military, (b) judicial system and courts, (c) national government, (d) health care or medical systems, (e) �nancial institutions or banks, (f ) religious organizations, 9. In fact, this second-hand information or “hearsay� effect might very well arti�cially amplify the relationship between perceived corruption and con�dence in public institutions. If a person who was solicited for a bribe tells all his/her friends about the experience, the experience of a single corrupt act may raise perceptions of the prevalence of corruption and lower con�dence in institutions among all his/ her friends. Consistent with this we do in fact �nd that (a) typically a substantially larger fraction of respondents state that corruption is widespread than those who respond to having personally experienced a bribe situation, and (b) the correlation between corruption perceptions and con�dence is stronger than the correlation between corruption experiences and corruption. We are grateful to an anonymous referee for pointing this out to us. 220 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. GWP Corruption Perceptions and Experiences Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (g) quality and integrity of the media, and (h) honesty of elections. In this paper we are primarily interested in con�dence in public institutions, and so in our core speci�cations we sum together the responses to (a), (b), (c) and (h) to obtain an index of con�dence in public institutions that ranges from 0 (respon- dents who report no con�dence in any of the four institutions) to 4 (respon- dents who report con�dence in all four institutions). We do not include (d), (e), (f ), and (g) as these questions do not refer to purely public institutions. Consistent with this interpretation, we �nd that responses to the four questions on public institutions are more strongly correlated with each other (with a median pairwise correlation of 0.42) than they are with responses to the ques- tions about other institutions that are not necessarily public (with a median pairwise correlation of 0.28). We will discuss in more detail below the extent to which this strong correlation of responses regarding the two types of institutions is attributable to unobserved individual-speci�c effects that might subsequently bias our estimates of the effects of corruption on con�dence. 10 10. We note also that the con�dence questions refer to national-level public institutions, whereas the corruption experiences question might in part reflect respondents’ interactions with local, rather than national-level, public of�cials. To the extent that respondents entertain different views about different levels of government this would work against us by weakening the correlation between the corruption and con�dence responses. As a robustness check however we have veri�ed that our main speci�cations also deliver similar results when we use two questions about con�dence in local institutions also in the GWP: (i) “In the city or area where you live, do you have con�dence in the local police force?� and (ii) “Do you approve of the leadership of the city or area where you live?�. Responses to these questions on local institutions are strongly correlated with responses to questions on national-level institutions, with a median pairwise correlation of 0.44, which is similar to the median pairwise correlation among national-level responses. Clausen, Kraay, and Nyiri 221 F I G U R E 2. Correlation of GWP Corruption Experiences and Perceptions Questions with Worldwide Governance Indicators (WGI) ‘Control of Corruption’ Variable Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Figure 3 documents how this measure of con�dence in institutions from the GWP compares with the most closely-related variables on con�dence in institutions taken from the World Values Survey.11 While the two measures are highly correlated in the common sample of countries for which both 11. The WVS asks about respondents’ con�dence in a variety of institutions. We aimed to match this con�dence index as closely as possible to our GWP index and therefore aggregated the answers to the following four questions into an index ranging from 0 to 4: “I am going to name a number of organizations. For each one, could you tell me how much con�dence you have in them [. . .]: a) the armed forces, b) the courts, c) the government (in your nation’s capital), d) parliament.� 222 THE WORLD BANK ECONOMIC REVIEW F I G U R E 3. Comparing Con�dence in Institutions: Country Average Values of GWP and WVS Indices Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 measures are available (a correlation of 0.81), it is worth noting the signi�- cantly smaller country coverage of the WVS. The circles in the graph rep- resent countries that are present in our sample of the GWP but not in the most recent wave of the WVS. Using the GWP index therefore signi�cantly increases the available cross-country sample to study effects of corruption on con�dence in institutions. We note however that this very large increase in country coverage offered by the GWP comes at the cost of a smaller number of respondents per country. Our sample size varies with availability of explanatory variables, but ranges from around 500 to 750 respondents per country, depending on the set of vari- ables considered. In contrast, the WVS survey used to construct Figure 3 fea- tures on average 1419 respondents per country. And the Afrobarometer Surveys used in several papers in this literature feature on average more than 1000 respondents by country-year (see for example Table 4 in Lavalle ´ e, Raza�ndrakoto and Roubaud (2008)), although in a much smaller cross- section of just 18 countries in two waves).12 Finally, Figure 4 documents the relationship between the corruption ques- tions and the con�dence in institutions index at the country level. The top panel plots corruption perceptions against con�dence in institutions and the 12. A further distinction of the GWP relative to the Afrobarometer Surveys is that it provides respondents only with binary response options to the corruption and con�dence questions (Yes/No), whereas the Afrobarometer Surveys offer more graduated responses (for example, “never�, “once or twice�, “a few times�, or “often� are possible responses to the corruption experiences question). There are advantages and disadvantages to both approaches. While the more graduated response in principle offers more detail, this detail can be dif�cult to interpret absent clear evidence on how respondents “anchor� the distinction between categories such as “a few times� and “once or twice�. Clausen, Kraay, and Nyiri 223 F I G U R E 4. Con�dence in Institutions and Corruption Experiences/Perceptions bottom panel plots corruption experiences against con�dence. Both graphs Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 display a negative relationship between corruption and con�dence although this is much more pronounced for corruption perceptions. Here, all countries with very low average corruption perceptions score high on con�dence in institutions. Scandinavian countries are the ones with the lowest perceived corruption and the highest con�dence in institutions. Turning to corruption experiences, we see that in general countries with a higher share of people that have experienced corruption report lower con�dence in institutions. However, there are a number of countries that have low levels of experienced corruption but still report low con�dence. In this group we �nd particularly Latin American and Caribbean countries such as Panama, Argentina, Peru, and Trinidad and Tobago. 224 THE WORLD BANK ECONOMIC REVIEW I II . MA IN R E SU LTS : R E S PO N D E N T -LE V E L E V I D E N C E ON CORRUPTION AND CONFIDENCE IN INSTITUTIONS While the cross-country relationship between corruption and con�dence in insti- tutions described above is suggestive of a link between the two, it is also far from convincing. A major concern here is that there may be many country- speci�c factors driving both variables. For example, some countries may simply have dysfunctional governments. On the one hand this will lead to high levels of corruption, and on the other hand public institutions naturally do not inspire con�dence in such an environment. Any correlation between our two variables would simply reflect the omitted variable of government quality that is driving both corruption and con�dence in public institutions. Another related possibility has to do with frame-of-reference issues in the survey responses themselves. It is plausible for example that citizens of rich countries have greater expectations of the quality and extent of public services provided by the government than do Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 citizens in poor countries. In this case small departures from these high stan- dards might result in lower reported con�dence in rich countries. Similarly there might be greater tolerance of corruption in poor countries than in rich countries, resulting in lower reported corruption perceptions or experiences in poor countries. Thus cross-country differences in expectations of corruption and public service quality might also spuriously contribute to the cross-country cor- relation between measured corruption and con�dence.13 To address this �rst concern, we primarily focus on the respondent-level variation within countries to study the relationship between corruption and con�dence in institutions. Doing so allows us to control for any omitted country-level characteristics that might be driving the cross-country correlation. Table 1 documents the distinction between the within- and between-country results. Columns 1 and 3 reflect the between-country variation, showing coef�- cients of cross-country linear regressions of con�dence in institutions on the two corruption measures, using country-averaged data. In contrast columns 2 and 4 capture the within-country variation, reporting estimates of the corre- sponding regressions including country �xed effects.14 In all cases we �nd a negative correlation between corruption and con�dence in institutions that is highly statistically signi�cant. In the cross-country variation, the estimated coef�cients imply that a one-standard-deviation increase (across countries) in either of the two corruption measures reduces con�dence in institutions by 13. A better approach to dealing with this problem of frame-of-reference issues is at the survey design stage, for example through the introduction of anchoring vignettes to provide common context to respondents’ qualitative responses. This option is unfortunately not available to us in the GWP which did not �eld such vignettes. 14. It would technically be more appropriate to estimate an ordered probit model because of the discrete and ordered nature of our dependent variable. Doing this does not change the sign or level of signi�cance of the coef�cients. However, because of the dif�culties involved with interpreting ordered probit coef�cients as marginal effects, we chose to present linear regression results throughout the paper. Clausen, Kraay, and Nyiri 225 T A B L E 1 . Bivariate Cross Country and Fixed Effects Regressions on the Relationship between Con�dence in Institutions and Corruption (1) (2) (3) (4) Con�dence in Con�dence in Con�dence in Con�dence in institutions institutions institutions institutions cross-country �xed effects cross-country �xed effects Corruption experiences -2.318*** -0.287*** (-3.10) (-8.94) Corruption perceptions -1.620*** -0.854*** (-5.17) (-21.32) _cons 2.118*** 2.904*** (17.73) (13.15) N 103 78063 103 78063 No. of countries 103 103 103 103 R-sq 0.098 0.230 0.233 0.271 Estimation in columns (1) and (3) is by ordinary least squares on country-level averages of all Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 variables, with heteroskedasticity-consistent standard errors. Estimation in columns (2) and (4) is by weighted least squares using sampling weights provided by Gallup, and heteroskedasiticity- consistent standard errors are clustered at the country level. T-statistics in parentheses: * p , 0.10, ** p , 0.05, *** p , 0.01 Source: Authors’ analysis with data from Gallup World Poll between 0.2 and 0.3 points on a 0-4 scale.15 Within countries, the relationship between corruption and con�dence is also very strong. Here a one standard deviation increases of either corruption variable within a country leads to a reduction of con�dence in institutions of between 0.1 and 0.3 points.16 Throughout the paper, we assess the signi�cance of the within-country results using standard errors that are clustered at the country level, and observations are weighted using sampling weights provided by Gallup. Anticipating our later discussion of endogeneity problems, we note that the estimated effect of the corruption perceptions question is nearly three times as large as the effect of the corruption experiences question. This is consistent with our view that the former is much more likely to be endogenous to respon- dents’ con�dence in public institutions, and that the latter much more plausibly identi�es a causal effect running from corruption to con�dence. While the esti- mated coef�cients are statistically signi�cant and quantitatively large, we note that the explanatory power of corruption for the con�dence question is limited. In particular, in the �xed-effects regressions, the bulk of the R-squared is due to the country dummies. In contrast, the within R-squared net of the country �xed effects is 0.01 for the corruption experiences question, and 0.06 for the corruption perceptions question. Although within-country regressions in Table 1 control for country-level omitted variables, a possible objection is that there may also be a variety of 15. Note that the cross-country standard deviations of corruption experiences and perceptions are 0.083 and 0.184, respectively. 16. Within-country standard deviation of corruption experiences is 0.360 and of corruption perceptions 0.368. 226 THE WORLD BANK ECONOMIC REVIEW individual-speci�c characteristics that influence both (i) respondents’ con�dence in institutions; and (ii) the likelihood that they view corruption as prevalent, or that they report having been solicited for a bribe.17 For example, richer, older, and more educated people might have more interactions with the state and so be more likely to �nd themselves exposed to corruption, and might also be more likely to have a cynical world view that precludes expressing con�dence in public institutions. To control for this we introduce a set of core control variables that we have found to be correlated with the corruption questions, and that also tend to be signi�cant predictors of con�dence in institutions. These include respondent age, gender, marital status, education, and the logarithm of self-reported income. We also introduce as basic control variables whether the household in which the respondent lives has access to the internet and a television. Access to such media may have ambiguous effects on individual’s opinions about and experiences with corruption and institutions. On the one hand, of�cials might have a harder Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 time extracting bribes from more informed citizens that have had the chance to obtain information about laws and regulations concerning their dealings with government. On the other hand, coverage of corruption cases in the media might influence corruption perceptions of individuals and may therefore have a direct effect on the answers to the perceptions question used in the GWP. Table 2 presents the results adding these basic control variables. We note �rst that missing data presents a problem when introducing our set of core control variables. In particular, data availability for education and income is incomplete, and this decreases our sample to about 57,000 individuals in 94 countries. To aid in comparison with the previous results, we �rst repeat the results with no controls from Table 1 in the smaller sample for which the control variables are available, and then report results with controls. Reducing the size of the sample in this way makes little difference for the effect of corruption on con�dence in public institutions: the results without control variables in columns (1) and (3) of Table 2 are essentially identical to those in columns (2) and (4) of Table 1. Second, we note that while the additional control variables featured in Table 2 do show some correlation with both the 17. Of course, controlling for country-level �xed effects will not address concerns about variations in quality of subnational governments. It could for example be the case that within a country, some local governments are corrupt and deliver low-quality public services, and as a result respondents have low con�dence in local government. It could then be that some of our observed within-country correlation reflects heterogeneity in government performance across local governments. It is dif�cult to control for this directly as the GWP does not contain much information attitudes towards local governments. As an imperfect proxy for this, we average responses to the question “In the city or area where you live, are you satis�ed or dissatis�ed with: (a) The public transporation systems, (b) the roads and highways, and (c) the educational system or the schools. This question does clearly ask respondents about public services in their locality, however it is not clear whether these are provided by local or by national-level governments. Despite this ambiguity, we include this as a control variables (results not reported but available from authors on request). Doing so has minimal effects on the size and signi�cance of our estimated effects of corruption on con�dence. We are grateful to an anonymous referee for pointing out this possible interpretation. Clausen, Kraay, and Nyiri 227 T A B L E 2 . Fixed Effects Regressions Including Control Variables (1) (2) (3) (4) Con�dence in Con�dence in Con�dence in Con�dence in institutions institutions institutions institutions Corruption experiences -0.298*** -0.282*** (-7.73) (-7.41) Corruption perceptions -0.870*** -0.865*** (-20.71) (-20.54) Male 0.00637 -0.00831 (0.30) (-0.41) Age -0.0167*** -0.0149*** (-6.51) (-5.80) Age2 0.000210*** 0.000190*** (7.39) (6.69) Married 0.0933*** 0.0775*** (4.57) (3.76) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Secondary education -0.121*** -0.115*** (-3.94) (-4.02) Tertiary education -0.0853 -0.108** (-1.65) (-2.45) Income -0.000873 -0.0104 (-0.07) (-0.82) Internet access -0.0437 -0.0600** (-1.38) (-2.08) TV -0.0321 -0.0228 (-0.66) (-0.48) N 57095 57095 57095 57095 No. of countries 94 94 94 94 R-sq 0.226 0.230 0.271 0.275 Estimation is by weighted least squares using sampling weights provided by Gallup, and heteroskedasiticity-consistent standard errors are clustered at the country level. T-statistics in parentheses: * p , 0.10, ** p , 0.05, *** p , 0.01 Source: Authors’ analysis with data from Gallup World Poll corruption and con�dence variables, we �nd that the estimated coef�cients on the corruption variables change very little, declining just slightly in absolute value. Finally, we note that the control variables all enter with expected signs and are generally signi�cant. Older individuals seem to have a lower degree of con�dence in institutions although this relationship is not linear. Also, married respondents express higher con�dence than single ones. Higher income and education as well as access to internet and TV appear to reduce con�dence although these latter effects are not statistically signi�cant in all cases. While the results in Table 2 are suggestive of a strong relationship between con�dence in institutions on the one hand, and corruption perceptions and experiences on the other, one might nevertheless reasonably worry that this correlation is driven by other unobserved respondent-speci�c characteristics.18 18. Of course a preferred way of dealing with this type of heterogeneity is to identify our effects using individual-level over time variation in responses to the corruption and con�dence questions. Unfortunately this option is not available to us in our single cross-section of countries and respondents. 228 THE WORLD BANK ECONOMIC REVIEW A leading possibility is that, conditional on the basic control variables described above, some individuals may simply have a negative outlook or worldview which makes them more likely to think that corruption is widespread, and at the same time drives their lack of con�dence in public institutions. Kaufmann and Wei (2000) coin this as a "kvetch" effect, after the Yiddish word for habitual complaining. To the extent that this drives the observed correlation between cor- ruption and con�dence in public institutions, we cannot interpret it as a causal link from the former to the latter. At �rst glance, one might think that this potential problem of kvetch is less severe for the corruption experiences question than for the corruption percep- tions question. However, while the former ostensibly is an objective question about the respondent’s experience, there are nevertheless ways in which kvetch might creep into responses to this question as well. First, respondents prone to kvetch might simply falsely claim that they had been solicited for a bribe. They might also be more likely to interpret ambiguous interactions with a public of�- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 cial as a request for a bribe. Therefore, respondents who in general tend to com- plain a lot might also be more likely to report interactions with public of�cials as involving a request for a bribe. Second, the question about experiences with bribery follows a battery of other questions about corruption, one of which is the corruption perceptions questions described above. It is possible that respon- dents prone to kvetch might want to reinforce their point of stating that govern- ment corruption is a problem by subsequently answering that they personally have found themselves in a bribe situation, even if this is not the case. Our strategy for dealing with this potential problem is to introduce control variables that we think may be good proxies for the propensity to kvetch. We consider three sets of such proxies.19 The �rst set relies on questions in the survey that focus on individuals’ self-reported well-being. For example, the GWP asks respondents whether they are satis�ed with their living standards, and which rung on the ladder of life that they �nd themselves. The GWP also asks respondents whether they have felt a variety of emotions such as worry, stress, or happiness in the previous day. These variables are plausibly correlated with individual respondents’ predisposition to complain. Second, the GWP asks respondents their opinions about a number of country-level variables including whether the economy is doing well or poorly, whether the economic outlook is favorable, and whether corruption is getting better or worse. Since our regressions include country �xed effects that soak up all national-level vari- ation, variation in individuals’ responses to these questions can be interpreted as capturing their idiosyncratic perceptions of the same national-level reality, and as such will also plausibly be correlated with kvetch. As a �nal control for kvetch, we note that the battery of questions from which our "con�dence in institutions" variables are drawn includes a further question about con�dence in religious organizations. It seems plausible to us 19. See Appendix Table A for a detailed description of the kvetch proxies and the speci�c GWP questions used in their construction. Clausen, Kraay, and Nyiri 229 that corruption perceptions or experiences are likely to have little direct impact on con�dence in religious organizations. However there might be an indirect effect through kvetch: individuals more likely to complain in general might also report less con�dence in religious organizations purely because of their propen- sity to kvetch. This suggests using a kind of differencing strategy to control for kvetch. In particular, one might ask whether corruption reduces the difference in con�dence in public institutions and con�dence in religious organizations. Alternatively and more flexibly, we can simply introduce con�dence in religious organizations directly into our main speci�cation as a control for kvetch. Table 3 documents the results controlling for these proxies for kvetch. Since not all of the kvetch variables are available for all observations, our sample shrinks further to 49,019 respondents in 90 countries. As in Table 2, we �rst document that our main results with basic respondent-level controls do not change as we move to this smaller sample (compare columns (1) and (3) in Tables 2 and 3). More interesting is how our results on the effects of corruption Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 perceptions and experiences on con�dence in institutions change when we control for kvetch. We �nd that the estimated impact of corruption on con�dence falls by about 34 percent (for the corruption experiences question) and by 40 percent (for the corruption perceptions question). This is a good indication that kvetch effects are present in the data and are at least partially addressed by the controls that we introduce. Interestingly, while both the corruption perceptions and corruption experiences questions might be subject to kvetch, we think it is plausible that kvetch effects are stronger for the former. The results in Table 3 are consistent with this: the coef�cient on the corruption perceptions falls rela- tively more after the introduction of the kvetch controls. We note also that the kvetch controls are all highly-signi�cant predictors of the con�dence responses, and collectively contribute to a substantial increase in the explanatory power of the regressions (the R-squared increases from 0.22 to 0.38 in the case of corrup- tion experiences, and from 0.27 to 0.39 in the case of corruption perceptions. However, even after introducing these very rigorous controls for kvetch, the negative relationship between corruption and con�dence remains highly signi�- cant and the magnitude of both corruption coef�cients remains large.20 20. A closely-related interpretation of these results is that individuals vary in their extent of “generalized trust�, which could be thought of as the opposite of ‘kvetch’. In the extreme, one could very well interpret responses to the corruption perceptions question and responses to questions about con�dence in public institutions as both simply serving as proxies for individuals’ “generalized trust�. Our strategy for dealing with this problem would be the same as our strategy for dealing with ‘kvetch’ as the two are quite similar. The �rst is to introduce controls that might serve as proxies for ‘kvetch’ or “generalized trust� (although we note that Newton and Norris (2000) examined the question if trust and con�dence is a feature of basic personality types but found little evidence to support this hypothesis.) In this respect our strategy of controlling for con�dence in non-public institutions is particularly helpful because it directly controls for individuals’ con�dence and focuses only on the differential degree of con�dence in public relative to non-public institutions. The second is to emphasize the corruption experiences question, which as we have argued is less likely to be tainted by either “kvetch� or “generalized trust�. 230 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 : Fixed Effects Regressions Controlling for Kvetch (1) (2) (3) (4) Con�dence in Con�dence in Con�dence in Con�dence in institutions institutions institutions institutions Corruption 2 0.280*** 2 0.185*** experiences ( 2 6.67) ( 2 5.62) Corruption 2 0.873*** 2 0.518*** perceptions ( 2 20.05) ( 2 16.37) Male 0.0123 2 0.00647 2 0.00512 2 0.0144 (0.52) ( 2 0.34) ( 2 0.23) ( 2 0.78) Age 2 0.0157*** 2 0.00135 2 0.0141*** 2 0.00118 ( 2 5.90) ( 2 0.54) ( 2 5.21) ( 2 0.47) Age2 0.000198*** 0.0000456* 0.000180*** 0.0000425 (6.76) (1.66) (6.02) (1.53) Married 0.0890*** 0.0399** 0.0739*** 0.0342* (4.05) (2.13) (3.37) (1.80) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Secondary 2 0.131*** 2 0.118*** 2 0.124*** 2 0.116*** education ( 2 3.94) ( 2 4.79) ( 2 4.04) ( 2 4.81) Tertiary education 2 0.0822 2 0.0791* 2 0.102** 2 0.0916** ( 2 1.49) ( 2 1.74) ( 2 2.18) ( 2 2.21) Income 2 0.000267 2 0.0470*** 2 0.00943 2 0.0488*** ( 2 0.02) ( 2 3.68) ( 2 0.68) ( 2 3.76) Internet access 2 0.0612* 2 0.0784*** 2 0.0817** 2 0.0872*** ( 2 1.71) ( 2 3.17) ( 2 2.55) ( 2 3.68) TV 2 0.0327 2 0.101*** 2 0.0234 2 0.0943** ( 2 0.67) ( 2 2.86) ( 2 0.49) ( 2 2.62) Ladder of life 0.0151*** 0.0139** (2.74) (2.61) Standard of living 0.228*** 0.220*** (8.30) (8.02) Emotions 0.0533*** 0.0519*** (4.90) (4.77) Economy good/bad 0.530*** 0.489*** (17.52) (17.20) Economic outlook 2 0.190*** 2 0.183*** ( 2 11.61) ( 2 11.70) Corruption trend 2 0.272*** 2 0.207*** ( 2 15.54) ( 2 12.61) Religious 0.705*** 0.689*** organizations (19.25) (18.81) N 49019 49019 49019 49019 No. of countries 90 90 90 90 R-sq 0.218 0.378 0.264 0.392 Estimation is by weighted least squares using sampling weights provided by Gallup, and heteroskedasiticity-consistent standard errors are clustered at the country level. T-statistics in par- entheses: * p , 0.10, ** p , 0.05, *** p , 0.01 Source: Authors’ analysis with data from Gallup World Poll Clausen, Kraay, and Nyiri 231 I V. R O B U S T N E S S OF THE MA IN R E SU LTS Thus far we have seen that there is a large and statistically signi�cant partial correlation between measures of corruption and con�dence in public insti- tutions, and that this result is robust to the addition of (a) country �xed effects, (b) a set of respondent-level controls, and (c) a set of proxies for ‘kvetch’. In this section we subject these main results to a variety of further robustness checks. We �rst disaggregate the con�dence in institutions measure into its four components and investigate how the effects of corruption vary across these components. We then also estimate our main speci�cation country-by-country, and document how the estimated coef�cients on the corruption questions vary by country, by level of corruption, and by level of development. In Table 4 we disaggregate the con�dence in institutions measure into its four components: con�dence in the military, judiciary, national government, Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 and in the honesty of elections. In the �rst four columns we report results for our core speci�cation, using each of these components of the overall con�dence measure separately as the dependent variable.21 We do this for both the corrup- tion experiences (top panel) and corruption perceptions measure (bottom panel). In all cases, we include, but do not report estimated coef�cients for, the full set of control variables used in Table 3. For the corruption experiences question, we �nd only modest differences across components in terms of the magnitude of the estimated partial correlation between corruption and con�- dence. This effect is largest for con�dence in the judiciary at 0.06, and smallest for con�dence in the honesty of elections, at 0.04. There is somewhat more variation across the various con�dence measures for the corruption perceptions question. The estimated effect of corruption is much lower for con�dence in the military, at 0.06, than it is for the other three measures, which range from 0.13 to 0.17. Thus far we have assumed that the slope of the relationship between corrup- tion and con�dence in public institutions is the same in all countries, at all income levels, and at all levels of corruption. We now relax this assumption and re-estimate our main speci�cation from Table 3, country-by-country, so that we can investigate how this slope varies across countries. We note �rst that the means of the country-by-country estimates in Table 5 are slightly smaller than the pooled estimates in Table 3 (at -0.13 and -0.47 for the corrup- tion experiences and perceptions questions, respectively). The sign of the esti- mated coef�cient is also fairly consistently negative across countries, with 67 percent (91 percent) of country estimates being negative for the corruption experiences ( perceptions) question. However, and not surprisingly, in many 21. Since the dependent variable for the individual con�dence in institutions regressions is a binary variable, a probit speci�cation would be more appropriate than the linear probability model. However, to improve comparability with previous results we report estimates from linear probability models here. We have also estimated the speci�cations in Table 4 using a probit model and �nd a similar pattern of relative magnitudes of the effect of corruption on the different con�dence in institutions variables. 232 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Disaggregation of “Con�dence in Institutions� Index (1) (2) (3) (4) Military Judiciary National Gov. Elections linear linear linear linear Dependent variable is -0.0431*** -0.0576*** -0.0480*** -0.0367*** Corruption experiences (-4.74) (-5.02) (-4.70) (-3.48) R-sq 0.232 0.234 0.278 0.271 Dependent variable is -0.0623*** -0.133*** -0.166*** -0.157*** Corruption perceptions (-6.82) (-12.18) (-13.54) (-13.99) R-sq 0.234 0.241 0.291 0.282 N 49019 49019 49019 49019 No. of countries 90 90 90 90 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Estimation is by weighted least squares using sampling weights provided by Gallup, and heteroskedasiticity-consistent standard errors are clustered at the country level. T-statistics in parentheses: * p , 0.10, ** p , 0.05, *** p , 0.01 countries the estimated effects are not statistically signi�cant, given the much smaller sample of observations on which to base inference in each country. In fact, the mean number of observations per country for the regressions in Table 5 is just 594, as opposed to 49,019 in the pooled regressions of Table 3. We next examine how these estimated coef�cients vary across regions (using the standard World Bank regional classi�cation). While it is evident that cor- ruption experiences as well as perceptions affect con�dence negatively in all regions on average, the magnitude and strength of the relationship varies widely across regions, from -0.06 to -0.32 in the case of corruption experi- ences, and from -0.10 to -1.00 in the case of corruption perceptions. In the case of corruption experiences, the largest mean estimated effect is for the South Asia region. The relationship between corruption and con�dence in insti- tutions is also the strongest in this region with 60 percent of countries report- ing a statistically signi�cant negative relationship. At the same time however, while South Asia showed the largest coef�cient of corruption experiences, its perceptions coef�cient is the smallest among the regions in our sample. In the remaining panels of Table 5 we document how the estimated corre- lation between corruption and con�dence varies with the average level of cor- ruption, and the level of development, of the country. To do this, we divide countries into three equal groups according to their country-level average score on the corruption question, and also their level of GDP per capita. We then report the mean (across countries) of the estimated slope coef�cient on corrup- tion from the country-by-country regressions, for each group. In the case of corruption experiences, there is a pronounced non-linear relationship in countries’ overall level of corruption. In countries where reported corruption experiences are on average either very low or very high, the estimated effect of corruption experiences on con�dence in institutions is small (at 0.07 and 0.09 T A B L E 5 . Disaggregation into Subgroups Depending on Geographic Region, Level of Corruption, and Income CORRUPTION EXPERIENCES CORRUPTION PERCEPTIONS Mean Standard Proportion of Mean Standard Proportion of No. of estimated deviation of Proportion of negative and estimated deviation of Proportion of negative and countries slope slope negative sign�cant slope slope negative sign�cant in group coef�cient coef�cient coef�cients coef�cients* coef�cient coef�cient coef�cients coef�cients* Full sample 90 -0.134 0.255 0.678 0.222 -0.466 0.376 0.911 0.633 Europe & 11 -0.172 0.200 0.727 0.272 -0.312 0.419 0.818 0.454 Central Asia Middle-East & 4 -0.153 0.282 0.750 0.250 -0.730 0.250 1.000 1.000 North Africa East Asia & 9 -0.060 0.175 0.667 0.222 -0.224 0.297 0.778 0.333 Paci�c South Asia 5 -0.319 0.265 0.800 0.600 -0.104 0.259 0.600 0.400 Latin America & 19 -0.194 0.271 0.684 0.369 -0.490 0.191 1.000 0.737 Caribbean Sub-Saharan 18 -0.061 0.278 0.611 0.222 -0.472 0.521 0.889 0.556 Africa High income: 20 -0.113 0.281 0.650 0.000 -0.561 0.211 1.000 0.800 OECD High income: 4 -0.094 0.143 0.750 0.000 -1.006 0.385 1.000 0.750 non-OECD Low level of 30 -0.066 0.241 0.600 0.000 -0.485 0.407 0.933 0.633 corruption experiences/ perceptions Clausen, Kraay, and Nyiri (Continued ) 233 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 TABLE 5. Continued 234 CORRUPTION EXPERIENCES CORRUPTION PERCEPTIONS Mean Standard Proportion of Mean Standard Proportion of No. of estimated deviation of Proportion of negative and estimated deviation of Proportion of negative and countries slope slope negative sign�cant slope slope negative sign�cant in group coef�cient coef�cient coef�cients coef�cients* coef�cient coef�cient coef�cients coef�cients* Medium level of 30 -0.243 0.238 0.833 0.367 -0.447 0.345 0.867 0.667 corruption experiences/ perceptions High level of 30 -0.093 0.256 0.600 0.300 -0.465 0.384 0.933 0.600 corruption experiences/ THE WORLD BANK ECONOMIC REVIEW perceptions Low income 30 -0.129 0.272 0.600 0.300 -0.370 0.437 0.867 0.400 Medium income 30 -0.153 0.244 0.733 0.300 -0.449 0.406 0.867 0.733 High income 30 -0.121 0.256 0.700 0.067 -0.579 0.234 1.000 0.767 * statistically signi�cant coef�cients at at least the 5 percent level were included Each row provides summary statistics on indicated slope coef�cients estimated country-by-country within the indicated country groups. Country-level estimation is by weighted least squares using sampling weights provided by Gallup. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Clausen, Kraay, and Nyiri 235 respectively). In contrast, for intermediate-corruption countries, the adverse effect of corruption on con�dence is much larger. This suggests that in countries where corruption is rare, a respondent’s iso- lated experience with having been solicited for a bribe will not be enough to sub- stantially undermine his or her faith in overall public institutions. And similarly, in countries where corruption is widespread, personal experiences with or per- ceptions of corruption might also not change con�dence in public institutions because this con�dence is very low to begin with. In contrast, for countries with a moderate prevalence of corruption, personal experiences with corruption have a stronger adverse impact on con�dence in public institutions. Interestingly, however, this pattern is not present in the corruption perceptions question, nor is it present when countries are divided into groups according to income levels. V. C O N C E R N S A B O U T E N D O G E N E I T Y Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 We now discuss the extent to which the partial correlation between corruption and con�dence in public institutions can be interpreted as a causal effect from the former to the latter. As noted in the introduction, there is an important identi�- cation problem: corruption might lead to a loss of con�dence in public institutions as we emphasize here, but at the same time, respondents who report low con�- dence in public institutions might as a result hold the belief that corruption is widespread as well. This point is also noticed by Cho and Kirwin (2007) who argue that individuals who do not trust public institutions might be more likely to resort to bribery to advance their interests, or to believe that corruption is wide- spread. This can lead to vicious circles where corruption and a lack of con�dence in public institutions feed off each other. This potential for bi-directional causa- tion complicates the interpretation of the partial correlation between corruption and con�dence in institutions that we have documented. This is the classic identi�- cation problem: the observed correlation between corruption and con�dence might reflect causal effects from corruption to con�dence that we emphasize. But it could also reflect causation in the opposite direction. We note �rst that a particular strength of the corruption experiences ques- tion is that it is much less likely to be prone to reverse causation than the cor- ruption perceptions question. To see why, recall that the experience question asks respondents whether they have been solicited for a bribe during the past 12 months. To the extent that the decision to solicit a bribe originates with the public of�cial with whom the respondent is interacting, there should be no pro- blems of reverse causation: it seems unlikely that a public of�cial would even know the respondent’s con�dence in public institutions, let alone base his decision to solicit a bribe on it.22 This stands in contrast with the corruption 22. Indeed, the pattern of reverse causation might go against our results, if individuals with low con�dence in public institutions choose not to interact with government agencies and so are less likely to report having been asked for a bribe. 236 THE WORLD BANK ECONOMIC REVIEW perceptions question, where there is a more plausible channel of causation in the opposite direction: individuals who have low con�dence in public insti- tutions may precisely for this reason also believe that corruption is widespread in government. This potential endogeneity bias may in part account for the fact that in most of our speci�cations thus far, the estimated slope of the relationship between corruption perceptions and con�dence is larger in absol- ute value, and typically is also much more signi�cant, than in the regressions using the corruption experiences question. Thus we argue that our results using the corruption experiences question provide a more plausible estimate of the causal effect of corruption on con�dence in public institutions than do our results with the corruption perceptions question. At the same time, we acknowledge that there may still be such endogeneity bias, although to a lesser extent, even in the corruption experiences question. This would occur if respondents expressing a low con�dence in public insti- tutions are more likely to interpret an ambiguous interaction with a public of�- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 cial as a request for a bribe than other respondents with higher con�dence in public institutions. Such potential endogeneity bias is extremely dif�cult to correct using purely cross-sectional observational data such as what we have in the GWP. This is because the usual strategy with observational data of identify- ing instruments (variables that plausibly affect only corruption, but not con�- dence in institutions, and vice versa) is very dif�cult to implement since it is hard to make a compelling case for the requisite exclusion restrictions. In particular, we �nd it hard to make a convincing case that there are vari- ables in the GWP that predict corruption at the individual level, but do not have direct predictive power for con�dence in institutions, that we could then use as instruments for corruption. To illustrate why we think this approach is not promising, consider the identifying assumptions implicit in the few papers in the literature that have attempted this instrumental variables strategy. Cho and Kirwin (2007) make the identifying assumption that variables such as respondents’ overall trust in others, and their perceptions of the political influ- ence of ethnic groups, matter only for corruption and have no direct effect on con�dence in institutions (see the exclusion restrictions implicit in their Table 1). Lavalle ´ e, Raza�ndrakoto, and Roubaud (2008) claim with little justi- �cation that a dummy variable indicating that the respondent is head of the household, and a variable capturing the respondent’s views on the acceptability of paying a bribe, matter only for corruption and have no direct effect on con�dence. We do not �nd such exclusion restrictions to be convincing. One might easily imagine that any of these variables are directly correlated with con�- dence in public institutions: for example respondents’ might believe that paying a bribe is acceptable precisely because they have no con�dence in public insti- tutions. It is also striking that in both papers, the instrumented estimates of the effects of corruption on con�dence are vastly larger in absolute value than the uninstrumented estimates, while the feedback problem these authors seek to Clausen, Kraay, and Nyiri 237 correct would suggest that the true effects of corruption on con�dence should be much smaller in absolute value than the corresponding OLS estimates (see columns (1) and (2) of Table 1 in Cho and Kirwin (2007) and Table 4 in Lavalle´ e, Raza�ndrakoto and Roubaud (2008)). These counterintuitive results likely are due to a failure of the exclusion restrictions required to justify the instrumental variables estimator.23 24 In contrast, we have consistently found that the magnitude of the effect of the more exogenous corruption experiences question on con�dence is always substantially smaller than the effect of the corruption perceptions question, consistent with the view that the former is less tainted by reverse causation. Absent compelling instruments, we use an argument based on Leamer (1981) to provide a rough bound on the extent to which our estimates might reflect reverse causation. To make this concrete let y denote the portion of con- �dence that is orthogonal to all of the control variables, including the country �xed effects, in columns 2 and 4 of Table 3, and let x denote the same orthog- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 onal component of corruption. The possibility of causal effects in both direc- tions between corruption and con�dence can be captured by the assumption that y and x are generated by the following system of two equations: y ¼ bx þ 1 ð1Þ x ¼ gy þ y We are primarily interested in the slope coef�cient b which captures the effect of corruption on con�dence. However, we cannot identify this effect absent some instrument that shifts corruption without at the same time affecting con�- dence, i.e. we need to �nd a variable that is included in the second equation but excluded from the �rst. Absent such an instrument, the problem is simply that there are four unknown parameters in this system (b, g, and the two variances of the error terms), while there are just three moments in the data (V(x), V(y), and COV(x,y)).25 However, we can still make progress by exploring how our 23. Lavalle´ e, Raza�ndrakoto and Robaud (2008) claim support for their identi�cation strategy in the fact that tests of overidentifying restrictions fail to reject the null of instrument validity. Here they fall into the (unfortunately common) pitfall of failing to recognize that such tests are valid only if at least one instrument is indeed valid. We think it is very dif�cult to make such a case even for just one instrument in this context. 24. An alternative approach sometimes used with survey data is to use the average of the corruption question across all observations within a pre-speci�ed group, for example all respondents in the same city, as an instrument for corruption. This is plausible as an identi�cation strategy only to the extent that we think that the unexplained portion of con�dence is uncorrelated across respondents within a group. This assumption is dif�cult to justify in practice. 25. In fact things might be even more complicated, as we have assumed for simplicity that the covariance between the two structural errors is zero as well. We justify this simplifying assumption by observing that in Table 3 we have already controlled for a large set of variables that might simultaneously be driving corruption and con�dence. Thus it is more plausible that the errors in the orthogonalized system here are independent. 238 THE WORLD BANK ECONOMIC REVIEW estimate of b would change given differing assumptions on the strength of the reverse causation captured by g. To do this, express the three observable data moments in terms of the four unknown parameters, and then solve for b con- ditional on a value of g. Then by varying g we can explore the robustness of our conclusions about b to alternative assumptions regarding the strength of the reverse causation. Some simple algebra delivers this very natural estimator for b as a function of g: ^ ¼ COV ðx; yÞ À gV ðyÞ b ð2Þ V ðxÞ À gCOV ðx; yÞ Note that when g ¼ 0 we retrieve the OLS estimator, i.e. ^ ¼ COV ðx; yÞ=V ðxÞ, since in this case there is no feedback from con�dence b to corruption, and so OLS is valid. On the other hand, note that b ^ ¼ 0 when g ¼ COV(x,y)/V( y) which is simply the OLS estimate of the feedback effect in Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 the second equation. This is because if there is in fact no causal effect running from corruption to con�dence, then the second equation can be estimated by OLS.26 Moreover, the range from g ¼ 0 to g ¼ COV(x,y)/V( y) seems to us to be a reasonable prior bound for the magnitude of reverse causation. It seems reasonable to assume that g , 0, i.e. less con�dence implies more corruption. However, the magnitude of this effect is likely to be less (in absolute value) than g ¼ COV(x,y)/V( y). If it were not, then the data would imply that b . 0, i.e. that corruption raises con�dence in public institutions, which seems implausible. We plot this estimate of b (on the vertical axis) as a function of g (on the horizontal axis) in Figure 5, using this prior plausible range of values for the magnitude of reverse causation. The top panel refers to the corruption experi- ences question, and the bottom to the corruption perceptions question. In both panels, when g ¼ 0 we retrieve the OLS estimates of b on the horizontal axis corresponding to those in Columns (2) and (4) of Table 3. As we allow for the possibility of more and more reverse causation, i.e. as g becomes more and more negative capturing a stronger effect of con�dence on corruption, our esti- mate of the main effect of interest, b, becomes closer and closer to zero. We also report 95 percent con�dence intervals for b, and these suggest that our estimate of b would be insigni�cantly different from zero only if g were very 26. While rarely used, it is interesting to note that the basic argument here is nearly 80 years old! Leamer (1981) credits Leontief (1929) with �rst performing this basic calculation. A very recent and growing literature on instrumental variables estimation with imperfect instruments can be thought of as resurrecting some of these basic insights as well (see for example Kraay (forthcoming), Conley, Hansen and Rossi (forthcoming), and Nevo and Rosen (forthcoming). In the case of OLS which we consider here where the corruption variable serves as its own instrument, the unobserved feedback parameter g governs the strength of the correlation between the instrument and the error term. The approaches in these three more recent papers can be thought of as a more formal way of exploring how the IV estimator varies with alternative assumptions about the strength of the correlations of the instrument with the error term. Clausen, Kraay, and Nyiri 239 F I G U R E 5. Robustness of Main Results to Reverse Causation Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 large (in absolute value). In particular, we note that the 95 percent con�dence interval for b includes zero only when g , -0.01 in the case of the corruption experiences question, and when g , -0.04 for the corruption perceptions ques- tion. This represents less than one-quarter of the plausible range for g indicated on the horizontal axis in each �gure. We conclude from this that it is a priori very plausible that there may be causal effects running in both directions between corruption and con�dence in public institutions. In this paper we are concerned primarily with the channel from corruption to con�dence. While we are unable to formally isolate this channel using credible instruments given data limitations, we nevertheless argue that there are at least two reasons why the results we show are at least partially interpretable as a causal effect from corruption to con�dence. The 240 THE WORLD BANK ECONOMIC REVIEW �rst is that, as we have discussed, it is much more dif�cult to see the channel for potential reverse causation in the results using the corruption experiences question. The second is that, even if reverse causation were present, it would need to be extremely strong in order to undermine our conclusion of a statisti- cally signi�cant effect of corruption on con�dence. VI. WHY DOES THE ADVERSE EFFECT OF CORRUPTION O N CO N F I D E N C E I N I N S T I T U T I O N S M AT T E R ? Thus far we have documented a strong negative relationship between corrup- tion and con�dence in public institutions. We conclude by using a small number of variables available in the GWP to investigate some direct conse- quences of this loss of con�dence. We do so in an effort to shed some light on what might be some of the mechanisms through which corruption-induced lack of con�dence in public institutions could undermine the functioning of Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 those institutions. In particular, we �nd some evidence that reduced con�dence in public institutions leads to a reduction in political participation, raises support for violent means of political expression, and increases the desire of respondents to vote with their feet through emigratation. We interpret each of these as a signal of respondents’ likelihood to “opt out� of participation in public institutions in a country. This in turn is suggestive of how lack of con�- dence in public institutions undermines their effectiveness, but it is of course far from the �nal word. We draw on a number of questions from the GWP to measure these conse- quences of corruption-induced losses in con�dence. To measure political par- ticipation, we use the GWP question which asks “In the past month, have you voiced your opinion to a public of�cial?" As a measure of support for violent forms of protest, we use a question from the GWP which asks: “Do you think groups that are oppressed and are suffering from injustice can improve their situation by peaceful means alone?� And �nally, the desire to emigrate is captured by response to the question "Ideally, if you had the opportunity, would you like to move permanently to another country, or would you prefer to continue living in this country?" In Table 6 we document the relationship between corruption, con�dence, and these three outcomes. In the �rst column, we report the simple bivariate relationship between the con�dence variable and the three outcome variables of interest, and in the second column we introduce the full set of control vari- ables from Table 3. We �nd strong evidence that a lack of con�dence in public institutions raises sympathy for violent protest, raises the desire to migrate, and reduces political participation. We next investigate the extent to which this reflects the effect of corruption perceptions and corruption experiences. In columns three and four we estimate regressions of the three outcome variables on the two corruption variables alone (but still controlling for the full set of control variables from Table 3). Here we �nd evidence that those individuals T A B L E 6 . Why Do Adverse Effects of Corruption Matter? (1) (2) (3) (4) (5) (6) Achieve change by Achieve change by Achieve change by Achieve change by Achieve change by Achieve change by peaceful means peaceful means peaceful means peaceful means peaceful means peaceful means Con�dence in 0.0958*** 0.0776*** 0.0764*** 0.0767*** institutions (8.70) (6.48) (6.41) (6.61) Corruption 2 0.0833*** 2 0.0690** experiences ( 2 2.84) ( 2 2.45) Corruption 2 0.0572* 2 0.0174 perceptions ( 2 1.76) ( 2 0.58) N 46249 46249 46249 46249 46249 46249 No. of countries 89 89 89 89 89 89 Controls no yes yes yes yes yes (1) (2) (3) (4) (5) (6) Like to move to Like to move to Like to move to Like to move to Like to move to Like to move to other country? other country? other country? other country? other country? other country? Con�dence in 2 0.127*** 2 0.0676*** 2 0.0629*** 2 0.0622*** institutions ( 2 8.45) ( 2 4.63) ( 2 4.27) ( 2 4.26) Corruption 0.249*** 0.236*** experiences (7.15) (6.63) Corruption 0.130*** 0.0964*** perceptions (4.35) (3.22) Clausen, Kraay, and Nyiri N 34184 34184 34184 34184 34184 34184 (Continued ) 241 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 TABLE 6. Continued 242 (1) (2) (3) (4) (5) (6) Achieve change by Achieve change by Achieve change by Achieve change by Achieve change by Achieve change by peaceful means peaceful means peaceful means peaceful means peaceful means peaceful means No. of countries 69 69 69 69 69 69 Controls no yes yes yes yes yes (1) (2) (3) (4) (5) (6) Voiced opinion to Voiced opinion to Voiced opinion to Voiced opinion to Voiced opinion to Voiced opinion to of�cial of�cial of�cial of�cial of�cial of�cial Con�dence in 0.0259*** 0.0127 0.0183* 0.0127 institutions (4.25) (1.26) (1.80) (1.25) Corruption 0.301*** 0.305*** experiences (9.00) (9.09) THE WORLD BANK ECONOMIC REVIEW Corruption 2 0.00537 0.00135 perceptions ( 2 0.19) (0.05) N 48774 48774 48774 48774 48774 48774 No. of countries 90 90 90 90 90 90 Controls no yes yes yes yes yes Estimation is by weighted least squares using sampling weights provided by Gallup, and heteroskedasiticity-consistent standard errors are clustered at the country level. T-statistics in parentheses: * p , 0.10, ** p , 0.05, *** p , 0.01 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Clausen, Kraay, and Nyiri 243 who have experienced corruption or who perceive corruption to be high in their country show support for violent protest and express increased desire to permanently leave their country. In contrast however, having had a corruption experience raises the likelihood of individuals voicing their opinion to public of�cials. While the GWP does not ask about the nature of this interaction with a public of�cial, it is possible that this positive correlation reflects precisely respondents complaining to public of�cials about their experience with corruption. Finally, we introduce both corruption measures together with con�dence in institutions as explanatory variables. Doing so sheds light on whether the effects of corruption on these outcomes operate only through con�dence in institutions (in which case the corruption variables would not enter signi�- cantly), or whether there are direct effects of corruption (in which case they would enter signi�cantly even after controlling for con�dence in public insti- tutions). In the case of corruption experiences, there seems to be fairly clear Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 evidence of both direct and indirect effects, as both the corruption and con�- dence variables enter signi�cantly. In the case of corruption perceptions however we �nd evidence of a direct effect only for the emigration question. Overall these �ndings provide some support to the �ndings of Putnam (2000) and Uslaner (2002) that institutional trust contributes to citizen’s involvement in the political process. VII. CONCLUSIONS In this paper we have used data from the Gallup World Poll, a unique and very large global household survey, to document a quantitatively large and statisti- cally signi�cant negative effect of corruption on con�dence in public insti- tutions. This highlights an important, but relatively under-examined, channel through which corruption can inhibit development. Our �ndings are robust to the inclusion of a large set of controls for country and respondent-level charac- teristics. In addition to considering a much larger sample of countries and a more thorough set of control variables, our main contribution relative to the existing literature is our treatment of potential endogeneity biases. We have argued that a key advantage of speci�c experiential questions about corruption is that they are much more plausibly exogenous to respondents’ reported con�- dence in public institutions. As a result, the partial correlation between such questions and con�dence can much more plausibly be interpreted as a causal effect from the former to the latter. 244 THE WORLD BANK ECONOMIC REVIEW A P P E N D I X T A B L E A . Variable Descriptions Variable Wording of Question in GWP De�nition Con�dence in Index composed of four subcategories of this question: scale of 0 to 4 with 4 institutions "In this country, do you have con�dence in each of indicating highest the following, or not? How about the military? con�dence Judicial system and courts? National government? Honesty of elections?" Corruption "Sometimes people have to give a bribe or a present in dummy: 1 indicating experiences order to solve their problems. In the last 12 months, exposure to bribery were you, personally, faced with this kind of situation, or not (regardless of whether you have the bribe/present or not)?" Corruption "Is corruption widespread throughout the government dummy: 1 indicating perceptions in this country, or not?" corruption is widespread Male Share of male Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 respondents Age Age in years Married "What is your current marital status?"; responses of dummy: 1 indicating "married" as well as "domestic partner" were married/domestic aggregated to form the "Married" variable partner Secondary "What is your highest level of education?" dummy: 1 indicating education highest level is tertiary education Tertiary "What is your highest level of education?" dummy: 1 indicating education highest level is secondary education Income "What is your total monthly household income, before Income in US dollars taxes? Please include income from wages and salaries, remittances from family member living elsewhere, farming and all other sources." Internet access "Does your home have access to the internet?" dummy: 1 indicating yes TV "Does your home have a television?" dummy: 1 indicating yes Ladder of life “Imagine a ladder numbered from zero at the bottom scale of 0 to 10 with to ten at the top. Suppose we say that the top of the 10 being best life ladder represents the best possible life for you, and the bottom of the ladder represents the worst possible life for you. On which step of the ladder would you say you personally feel you stand at this time, assuming that the higher the step the better you feel about your life, and the lower the step the worse you feel about it? Which step comes closest to the way you fell?" Standard of “Are you satis�ed or dissatis�ed with your standard of (0 or 1) with 1 living living, all the things you can buy and do?�. indicating satis�ed (Continued ) Clausen, Kraay, and Nyiri 245 APPENDIX TABLE A. Continued Variable Wording of Question in GWP De�nition Emotions Index composed of three subcategories of this question: scale of 0 to 3 with 3 “Did you experience the following feelings during a indicating yes to all lot of the day yesterday? How about Worry? Stress? 3 questions Happiness?� Economy good/ "Do you believe the current economic conditions in dummy: 1 indicating bad this country are good, or not?" good Economic "Right now, do you think the economic conditions in dummy: 1 indicating outlook this country as a whole, are getting better or getting better worse?" Corruption "Do you think the level of corruption in this country is dummy: 1 indicating trend lower, about the same, or higher than it was 5 years corruption is higher ago?" Religious "In this country, do you have con�dence in each of the dummy: 1 indicating organizations following, or not? How about religious con�dence Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 organizations (churches, mosques, temples etc.)?" Voiced opinion "Have you done any of the following in the past dummy: 1 indicating to public month? How about voiced your opinion to a public "yes" of�cial of�cial?" Achieve change "Some people believe that groups that are oppressed dummy: 1 indicating by peaceful and are suffering from injustice can improve their "peaceful means means situations by peaceful means alone. Other do not alone will work" believe that peaceful means alone will work to improve the situation for such oppressed groups. Which do you believe?" Like to move to "Ideally, if you had the opportunity, would you like to dummy: 1 indicating other country move permanently to another country, or would you "would like to prefer to continue living in this country?" move" 246 A P P E N D I X T A B L E B . Countries in Core Sample by Geographical Region Europe & East Asia & Latin America & Sub-Saharan Middle-East & High income: High income: Central Asia Paci�c South Asia Caribbean Africa North Africa OECD non-OECD Armenia Cambodia Bangladesh Argentina Botswana Algeria Australia Estonia Azerbaijan Indonesia India Bolivia Burkina Faso Djibouti Austria Israel Belarus Laos Nepal Brazil Burundi Iran Belgium Malta Hungary Malaysia Pakistan Chile Cameroon Lebanon Canada Trinidad & Tobago Latvia Mongolia Sri Lanka Colombia Chad Denmark Lithuania Philippines Costa Rica Ethiopia Finland Moldova Taiwan Dominican Rep. Ghana France THE WORLD BANK ECONOMIC REVIEW Poland Thailand Ecuador Kenya Germany Russia Vietnam El Salvador Liberia Ireland Turkey Guatemala Madagascar Italy Ukraine Haiti Mauritania Japan Honduras Niger Luxembourg Mexico Senegal Netherlands Nicaragua Sierra Leone New Zealand Panama Tanzania Norway Paraguay Togo Portugal Peru Uganda South Korea Uruguay Zambia Spain Venezuela Sweden United Kingdom Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Clausen, Kraay, and Nyiri 247 REFERENCES Acemoglu, Daron, Simon Johnson, and James A. Robinson 2001. “The Colonial Origins of Comparative Development: An Empirical Investigation.� American Economic Review, 91(5), 1369–1401. Anderson, Christopher J., and Yuliya V. Tverdova 2003. “Corruption, Political Allegiances, and Attitudes toward Government in Contemporary Democracies.� American Journal of Political Science, 47(1), 91– 109. Azfar, Omar, and Peter Murrell 2009. Identifying Reticent Respondents: Assessing the Quality of Survey Data on Corruption and Values, Economic Development and Cultural Change, University of Chicago Press, Vol. 57(2), 387– 411. Bratton, Michael 2007. Are You Being Served? Popular Satisfaction with Health and Education Services in Africa, Afrobarometer, Working Paper No. 65. Catterberg, Gabriela, and Alejandro Moreno 2005. “The Individual Bases of Political Trust: Trends in New and Established Democracies.� International Journal of Public Opinion Research, 18(1), 408–443. Chang, Eric C.C., and Yun-han Chu 2006. “Corruption and Trust: Exceptionalism in Asian Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Democracies?.� The Journal of Politics, 68(2), 259 –271. Cho, Wonbin, and Matthew F. Kirwin 2007. A Vicious Cycle of Corruption and Mistrust in Institutions in Sub-Saharan Africa: A Micro-Level Analysis, Afrobarometer Working Paper, No. 71, Michigan State University. Clausen, Bianca, Aart Kraay, and Peter Murrell 2010. Does Respondent Reticence Affect the Results of Corruption Surveys? Evidence from the World Bank Enterprise Survey for Nigeria. World Bank Policy Research Department Working Paper No. 5415. Conley, Timothy, Christian Hansen, and Peter Rossi forthcoming. “Plausibly Exogenous.� Review of Economics and Statistics. Deaton, Angus 2008. “Income, Health, and Wellbeing Around the World: Evidence from the Gallup World Poll.� Journal of Economic Perspectives, 22(2), 53– 72. ——— 2009. Aging, Religion, and Health, National Bureau of Economic Research Working Paper No. 15271. Deaton, Angus, Jane Fortson, and Robert Tortora 2009. Life (Evaluation), HIV/AIDS, and Death in Africa, National Bureau of Economic Research Working Paper No. 14637. Della Porta, Donatella 2000. “Social Capital, Beliefs in Government, and Political Corruption,� in Susan J. Pharr, and Robert D. Putnam eds.: Disaffected Democracies: What’s Troubling the Trilateral Countries?, Princeton, NJ: Princeton University Press. Easton, David 1965. A Systems Analysis of Political Life, New York: John Wiley. ——— 1975. “A Re-Assessment of the Concept of Political Support.� British Journal of Political Science, 5(4), 435– 457. Gandelman, Ne ´ n Herna ´ stor, and Rube ´ ndez-Murillo 2009. “The Impact of Inflation and Unemployment on Subjective Personal and Country Evaluations.� Federal Reserve Bank of St. Louis Review, 91(3), 107–126. Gibson, James L., and Gregory A. Caldeira 1995. “The Legitimacy of Transnational Legal Institutions: Compliance, Support, and the European Court of Justice.� American Journal of Political Science, 39(2), 459–89. Gibson, James L., Gregory A. Caldeira, and Lester K. Spence 2003. “Measuring Attitudes toward the United States Supreme Court.� American Journal of Political Science, 39(2), 354–367. Helliwell, John F. 2008. Life Satisfaction and Quality of Development, National Bureau of Economic Research Working Paper No. 14507. 248 THE WORLD BANK ECONOMIC REVIEW Helliwell, John F., Christopher P. Barrington-Leigh, Anthony Harris, and Haifang Huang 2009. International Evidence on the Social Context of Well-Being, National Bureau of Economic Research Working Paper No. 14720. Hellman, Joel, and Daniel Kaufmann 2004. “The Inequality of Influence�. Available at SSRN: http ://ssrn.com/abstract=386901 or doi:10.2139/ssrn.386901. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi 1999. Governance Matters, World Bank Policy Research Department Working Paper No. 2196. Kaufmann, Daniel, and Shang-Jin Wei 2000. “Does “Grease Money� Speed Up the Wheels of Commerce?.� IMF Working Papers, 00/64, Washington DC: International Monetary Fund. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi 2008. “Governance Matters VII: Aggregate and Individual Governance Indicators 1996-2007.� Policy Research Working Paper, 4654, Washington DC: The World Bank. Kaufmann, Daniel, and Joel Hellman 2004. “Political Inequality and the Subversion of Institutions in Transition Economies�, in Janos Kornai, and Susan Rose-Ackerman (eds). “Building a Trustworthy State in Post-Socialist Transition�, Palgrave-MacMillan. Knack, Stephen, and Philip Keefer 1995. “Institutions and Economic Performance: Cross-Country Tests Using Alternative Institutional Measures.� Economics and Politics, 7(3), 207– 227. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 ——— 1997. “Why Don’t Poor Countries Catch Up? A Cross-National Test for an Institutional Explanation.� Economic Inquiry, 35(3), 590– 602. Kraay, Aart forthcoming. “Instrumental Variables Estimation with Uncertain Exclusion Restrictions: A Bayesian Approach.� Journal of Applied Econometrics. Krueger, Alan B., and Jitka Malec ´ 2009. “Attitudes and Action: Public Opinion and the ˇ kova Occurrence of International Terrorism.� Science, 235, 1534– 1536. Lambsdorff, Johann Graf 2007. The Institutional Economics of Corruption and Reform: Theory, Evidence, and Policy, Cambridge: Cambridge University Press. ´ e, Emmanuelle, Mireille Raza�ndrakoto, and Francois Roubaud 2008: Corruption and Trust in Lavalle Political Institutions in Sub-Saharan Africa, Afrobarometer, Working Paper No. 102. Leamer, Edward E. 1981. “Is it A Supply Curve or Is it a Demand Curve? Partial Identi�cation Through Inequality Constraints.� Review of Economics and Statistics, 63(3), 319–327. Leontief, Wassily 1929. “Ein Versuch Zur Statistischen Analyse von Angebot und Nachfrage (An Inquiry on the Statistical Analysis of Supply and Demand).� Weltwirtschaftliches Archiv, 30(1), 1– 53. Mauro, Paolo 1995. “Corruption and Growth.� The Quarterly Journal of Economics, 110(3), 681 –712. ´ on, Pierre-Guillaume, and Khalid Sekkat 2004. “Does the Quality of Institutions Limit the MENA’s Me Integration in the World Economy.� The World Economy, 27(9), 1475–1498. ´ on, Pierre-Guillaume, and Khalid Sekkat 2005. “Does Corruption Grease or Sand the Wheels of Me Growth?.� Public Choice, 122(1), 69– 79. Mishler, William, and Richard Rose 2001. “What are the Origins of Political Trust? Testing Institutional and Cultural Theories in Post-Communist Societies.� Comparative Political Studies, Vol. 34(1), 30 –62. ——— 2005. “What Are the Consequences of Trust? A Test of Cultural and Institutional Theories in Russia.� Comparative Political Studies, 38(9), 1050–1078. Mo, Pak Hung 2001. “Corruption and Growth.� Journal of Comparative Economics, 29(1), 66– 79. Nevo, Aviv, and Adam Rosen forthcoming. “Inference with Imperfect Instrumental Variables.� Review of Economics and Statistics. Newton, Kenneth, and Pippa Norris 2000. “Con�dence in Public Institutions: Faith, Culture, or Performance?,� in Susan J. Pharr, and Robert D. Putnam: Disaffected Democracies: What’s Troubling the Trilateral Economies, Princeton, NJ: Princeton University Press. Clausen, Kraay, and Nyiri 249 Ng, Weiting, Ed Diener, Raksha Aurora, and James Harter 2008. “Affluence, Feelings of Stress, and Well-Being.� Social Indicators Research, doi:10.1007/s11205-008-9422-5. North, Douglass C. 1990. Institutions, Institutional Change, and Economic Performance, New York: Cambridge University Press. Pelham, Brett, Steve Crabtree, and Zsolt Nyiri 2009. “Technology and Education.� Harvard International Review, Summer, 74 –76. Pellegrini, Lorenzo, and Reyer Gerlagh 2004. “Corruption’s Effects on Growth and its Transmission Channels.� Kyklos, 57(3), 429–456. Pharr, Susan 2000. “Of�cials’ Misconduct and Public Distrust: Japan and the Trilateral Democracies,“ in Susan J. Pharr, and Robert D. Putnam eds.: Disaffected Democracies: What’s Troubling the Trilateral Countries?, Princeton, NJ: Princeton University Press. Putnam, Robert D. 2000. Bowling Alone, New York: Simon & Schuster. Rodrik, Dani, Arvind Subramanian, and Francesco Trebbi 2004. “Institutions Rule: The Primacy of Institutions over Geography and Integration in Economic Development.� Journal of Economic Growth, 9(2), 131 –165. Rose, Richard, William Mishler, and Christian Haerpfer 1998. Democracy and its Alternatives: Understanding Postcommunist Societies, Baltimore: The Johns Hopkins University Press. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Sacks, Audrey 2011. “The Antecedents of Approval of the Incumbent Government and Trust in Government in sub-Saharan Africa, Latin America, and Six Arab Countries�. Manuscript, the World Bank Institute. Seligson, Mitchell A. 2002. “The Impact of Corruption on Regime Legitimacy: A Comparative Study of Four Latin American Countries.� The Journal of Politics, 64(2), 408–433. Stevenson, Betsey, and Justin Wolfers 2008. “Economic Growth and Subjective Well-Being: Reassessing the Easterlin Paradox.� Brookings Papers on Economic Activity, Spring, 1– 87. Uslaner, Eric M. 2002. The Moral Foundations of Trust, Cambridge: Cambridge University Press. Agricultural Distortions in Sub-Saharan Africa: Trade and Welfare Indicators, 1961 to 2004 Johanna L. Croser and Kym Anderson For decades, agricultural price and trade policies in Sub-Saharan Africa hampered farmers’ contributions to economic growth and poverty reduction. This paper draws on a modi�cation of so-called trade restrictiveness indexes to provide theoretically precise partial-equilibrium indicators of the trade and welfare effects of agricultural policy distortions to producer and consumer prices in 19 African countries since 1961. Annual time series estimates are provided not only by country but also, for the region, by commodity and by policy instrument. The �ndings reveal the considerable extent of policy reform over the past two decades, especially through reducing export Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 taxation; but they also reveal that national policies continue to reduce trade and econ- omic welfare much more in Sub-Saharan Africa than in Asia or Latin America. JEL classi�cations: F13, F14, F15, N57, Q17, Q18 In the 1960s and 1970s, governments of many Sub-Saharan African countries adopted macroeconomic, sectoral, trade and exchange rate policies that directly or indirectly taxed farm household earnings, particularly from export commod- ities. These anti-agricultural, anti-trade, welfare-reducing policies, which were also prevalent in numerous other developing country regions up to the early 1980s (Krueger, Schiff and Valdes 1988), have since been subject to major reform. How far has that reform effort gone in altering the trade- and welfare-reducing characteristics of farm and food policies in Sub-Saharan Africa? This matters greatly for economic development and poverty alleviation, because 60 percent of Sub-Saharan Africa’s workforce is still employed in agri- culture, nearly 40 percent of the population is earning less than $1/day, and more than 80 percent of the region’s poorest households depend directly or indirectly on farming for their livelihoods (World Bank 2007, Chen and Kym Anderson (corresponding author, kym.anderson@adelaide.edu.au) is George Gollin Professor of Economics at the University of Adelaide in Australia, and former Lead Economist (Trade Policy) in the Development Research Group of the World Bank in Washington DC. Johanna Croser ( johlou@ gmail.com) at the time of preparing this paper was a PhD candidate in the School of Economics at the University of Adelaide, and is currently a commercial lawyer with the law �rm Johnson Winter and Slattery in Sydney, Australia. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 250– 277 doi:10.1093/wber/lhr012 Advance Access Publication May 23, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 250 Croser and Anderson 251 Ravallion 2010). Furthermore, because Africa is the focus of several new major agricultural development assistance programs, there is an on-going need to monitor the extent of changes over time in market-distorting policy interven- tions by national governments. The present paper serves two purposes. First, it briefly outlines a method- ology appropriate for both assessing trends and fluctuations in past policy choices and monitoring annual changes in those policies as soon as data become available each year. And second, it provides estimates for the past half- century which indicate the changing extent of government intervention in the region’s agricultural markets. Those indicators also reveal the contributions of different countries, commodities and policy instruments to the region’s overall reform of agricultural and food policies. The indicators of price distortions draw on the family of trade restrictiveness indexes, which in turn draw on – but go beyond – the type used by the OECD Secretariat for monitoring agricultural and food policies of high-income Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 countries ( producer and consumer support estimates, see OECD 2010). More speci�cally, they indicate what trade tax, if applied equally to all farm products for a country, would generate the same trade- (or welfare-)reducing outcome as the actual national structure of producer and consumer price distortions in place in any year. In doing so, a methodological advance is made by incorpor- ating nontradable products in our estimates of the indexes, which turns out to be important in the African agricultural policy context. Economy-wide computable general equilibrium models also are able to provide estimates of the trade and welfare effects of policies for a point in time. However, for lack of econometric estimates such models typically depend on myriad assumptions about parameter values. Furthermore, they apply to just one particular previous year and, being data intensive, tend to be updated infrequently and with a long delay. They are thus unable to provide annual revisions of time series trends and fluctuations on the regular, comparable, and timely basis desired by the policy community. Data for construction of the indexes reported below come from the World Bank’s Distortions to Agricultural Incentives database (Anderson and Valenzuela 2008). The database gives consistent measures of price-distorting policies for 75 countries for the period 1955 to 2007. The data for the 21 African countries in that database are discussed comprehensively in Anderson and Masters (2009), but that study did not include estimates of the indexes reported below. In this paper we focus on 19 of those African countries, leaving aside Egypt and South Africa because they are large and far more afflu- ent than the rest of the sample. The sub-sample comprises �ve countries of eastern Africa (Ethiopia, Kenya, Sudan, Tanzania, and Uganda), four in southern Africa (Madagascar, Mozambique, Zambia, and Zimbabwe), �ve large economies in Africa’s western coast (Cameroon, Co ˆ te d’Ivoire, Ghana, Nigeria, and Senegal), and �ve smaller economies of West and Central Africa for which cotton is a crucial export (Benin, Burkina Faso, Chad, Mali, and 252 THE WORLD BANK ECONOMIC REVIEW Togo). We concentrate on 1961 to 2004, since those are the years for which the African data are most complete. The paper is structured as follows: the next section summarizes the method- ology to be used. This is followed by a discussion of the data in the World Bank’s Agricultural Distortions database. We then report estimates of the series of indexes, before drawing conclusions. I . ME T H O D O LO GY There is a growing literature that identi�es ways to estimate indicators of the trade- and welfare-reducing effects of international trade-related policies as scalar index numbers. This literature serves a key purpose: it overcomes aggre- gation problems (across different intervention measures and across industries) by using a theoretically sound aggregation procedure to answer precise ques- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 tions regarding the trade or welfare reductions imposed by each country’s trade and trade-related policies. These measures represent a substantial improvement on commonly used measures. The usual tools for summarizing price-distorting policy trends in a country or region (see, e.g., Anderson and Masters 2009) are measures of the unweighted or weighted mean nominal rate of assistance (NRA) and consumer tax equivalent (CTE), the standard deviation of industry NRAs for the sector, and in a few instances the weighted mean NRA for exportable versus import- competing covered products.1 Authors often need to report more than one measure to gain an appreciation of the nature of the policy regime. For example, indicators of dispersion of NRAs are a reminder that there are additional welfare losses from greater variation of NRAs across industries within the sector (Lloyd 1974). Further, if import-competing and exportable sub-sectors have NRAs of opposite sign, they need to be reported separately because they would offset each other in calculating the aggregate sectoral NRA. While those various indicators are useful as a set, policy makers would �nd it more helpful to have a single indicator to capture the overall trade or welfare effect of an individual country’s regime of agricultural price distortions in place at any time, and to trace its path over time and make cross-country com- parisons. To that end, the scalar index literature has been developed. The pio- neering theoretical work is by Anderson and Neary (summarized in their 2005 book), with an important partial equilibrium contribution by Feenstra (1995). The theory de�nes an ad valorem trade tax rate which, if applied uniformly across all tradable agricultural commodities in a country will generate the same 1. The OECD (2010) measures similar indicators to the NRA and CTE, called producer and consumer support estimates (PSEs and CSEs). The main difference from an African viewpoint, apart from the CSE having the opposite sign to the CTE, is that the NRA and CTE are expressed as a percentage divergence from undistorted (e.g., border) prices whereas the PSEs/CSEs relate to the divergence from actual (distorted) prices. Croser and Anderson 253 reduction in sectoral trade, or in economic welfare, as the actual cross-product structure of distortions.2 In recent years, several empirical papers have provided series of estimates of scalar index numbers for individual countries. Irwin (2010) uses detailed import tariff data to calculate the Anderson-Neary Trade Restrictiveness Index for the United States in 1859 and annually from 1867 to 1961; Kee, Nicita and Olarreaga (2009) estimate indexes for 78 developing and developed countries for a single point in time (the mid-2000s); and Lloyd, Croser and Anderson (2010) estimate indexes for 75 developed and developing countries using the World Bank’s Distortions to Agricultural Incentives database for the period 1955 to 2007. In addition to being useful in summarizing the agricultural and food policy regime in an individual country, the Anderson-Neary scalar index measures can be adapted to reveal two other aspects of agricultural policy: the relative contributions of different policy instruments to reductions in trade or welfare Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (Croser and Anderson 2011), and the trade- and welfare-reducing effects of policy in a single global or regional commodity market (Croser, Lloyd and Anderson 2010). In this paper we utilise the methodology to estimate all three types of indexes. In doing so, we extend the theory and analysis to include non- tradables, which have not been addressed in previous studies but which are of practical signi�cance in poorer African countries where nontradables account for a non-trivial share of the gross value of agricultural production. Country level trade- and welfare-reduction indexes To capture distortions imposed by each country’s border and domestic policies on its trade volume and economic welfare, we adopt the methodology from Lloyd, Croser and Anderson (2010). Those authors de�ne a Trade Reduction Index (TRI) and a Welfare Reduction Index (WRI) and estimate them by con- sidering separately the distortions to the producer and consumer sides of the agricultural sector (which can differ when there are domestic measures in place in addition to or instead of trade measures). As their names suggest, the two indexes respectively provide a single indicator the (partial equilibrium) of the trade- or welfare-reducing effects of all distortions to consumer and producer prices of farm products from all agricultural and food policy measures in place. The TRI and WRI thus go somewhat closer to what a computable general equi- librium (CGE) can provide in the way of estimates of the trade and welfare (and other) effects of price distortions, while having the advantage of providing an annual time series. Fortuitously, estimates of the actual price distortions are 2. Other indexes de�ne an ad valorem trade tax rate which, if applied uniformly across all tradable products, will generate the same government revenue (Bach and Martin 2001), or the same real national income and general equilibrium structure of the economy (Anderson 2009a), as the actual cross-product structure of distortions. 254 THE WORLD BANK ECONOMIC REVIEW available in the NRAs and CTEs of the World Bank’s Distortions to Agricultural Incentives database. The derivation of the two indexes for n import-competing industries leads to the expressions for the TRI and WRI for the import-competing sector of a country shown in Box 1. BOX 1: Expressions for the TRI and WRI TRI WRI T ¼ {Ra þ Sb}, with W ¼ fR02 a þ S02 bg1=2 , with n  n  n 1 n 1 P P P 2 2 P 2 2 R¼ ri ui and S ¼ s i vi R0 ¼ ri ui and S0 ¼ si vi i¼1 i¼1 i ¼1 i ¼1 2 C P Ã2 P where ui ¼ pà i ðdxi =dpi Þ= pi ðdxi =dpC à i Þ ¼ ri ð p i x i Þ = ri ð p à i xi Þ Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 P i P i Ã2 P Ã2 P à à vi ¼ pi ðdyi =dpi Þ= pi ðdyi =dpi Þ ¼ si ð pi yi Þ= si ð pi yi Þ, P Ã2 iP Pi P Ã2 C 2 2 pP a¼ pi dxi =d pi = pÃi dmi =dpi , and b ¼ À pÃi dyi =d i = pi dmi =dpi . i i i i Variable de�nitions: T — Trade Reduction Index; W — Welfare Reduction Index; R — index of average consumer price distortions; S —index of average producer price distortions; R0 — Consumer Distortion Index; S0 — Producer Distortion Index; si — the rate of distortion of the producer price in pro- portional terms; ri — rate of distortion of the consumer price in proportional terms; ui — weight for each commodity in R and R 0 , which is proportional to the marginal response of domestic consumption to changes in international free-trade prices and can be written as a function of the domestic price elasticity (at the protected trade situation) of demand (ri); vi — weight for each commodity in S and S’, which is proportional to the marginal response of domestic production to changes in international free-trade prices and can be written as a func- tion of the domestic price elasticity (at the protected trade situation) of supply, (si); pà i — border price; pP à C à i ¼ pi ð1 þ si Þ — distorted domestic price; pi ¼ pi ð1 þ ri Þ — distorted domestic consumer price; xi ¼ xi ð pC i Þ — quantity of good i demanded (as a function of own domestic price); yi ¼ yi ð pPi Þ — quantity of good i supplied (as a function of own domestic price); a (b) — weight of consumption ( production) in the WRI or TRI, which is proportional to the ratio of the marginal response of domestic demand (supply) to a price change relative to the mar- ginal response of imports to a price change. Source: Synthesized from Lloyd, Croser and Anderson (2010) Essentially the import-competing TRI and WRI are constructed from appropri- ately weighted averages of the level of distortions of consumer and producer prices. The TRI is a mean of order one, and the WRI a mean of order two, but they use the same weights. Because the WRI is a mean of order two, it better reflects the welfare cost of agricultural price-distorting policies because it recog- nizes that the welfare cost of a government-imposed price distortion is related to the square of the price wedge. It thus captures the disproportionately higher welfare costs of peak levels of assistance or taxation, and is positive regardless of whether the government’s agricultural policy is favouring or hurting farmers. Croser and Anderson 255 The TRI and WRI can each be extended so as to add the exportable and nontradable sub-sectors of agriculture (see Appendix). Distortions to exporta- ble industries are inputted into the TRI as negative values because a positive (negative) price distortion in an exporting industry has a trade-expanding (- reducing) effect, and thus decreases (increases) the TRI. Distortions to nontrad- able industries are inputted into the TRI as zero values because a domestic price distortion in a nontradable industry by de�nition has neither a trade-expanding nor trade-reducing effect for that industry because of its assumed prohibitively high trade costs.3 The expressions for the TRI and WRI weights above show that estimates of own-price elasticities are required to compute the indexes. In line with Lloyd, Croser and Anderson (2010), and in the absence of reliable elasticity estimates for all countries and periods, we adopt some simplifying assumptions in this paper. We assume that a country’s domestic price elasticities of supply are equal across commodities within a sub-sector, and likewise for domestic price Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 elasticities of demand. This powerful simplifying assumption allows us (in the empirical section below) to �nd appropriately weighted aggregates of distor- tions on the production and consumption sides of the market simply by aggre- gating the change in consumer ( producer) prices across commodities and using as weights the sectoral share of each commodity’s domestic value of consump- tion ( production) at undistorted prices. We expect these simplifying elasticity assumptions (which still allow for differences across countries and between the demand and supply elasticities of each product within each country) to have only a small impact on the reported indexes. This is because elasticities appear in both the numerator and denomi- nator of the weight expressions, and therefore cancel each other out to a con- siderable extent. Further, Kee, Nicita and Olarreaga (2008, p. 677) show that the TRI and WRI can be decomposed into three components (namely, the mean and the variance of the distortion rates and the co-variance between the square of the distortion rate and the appropriate price elasticity). Since the elas- ticity enters into only the third component (see Appendix), and since in prac- tice that component tends to be small relative to the other two components (as noted by Anderson and Neary (2005) and found empirically by Kee, Nicita and Olarreaga (2009) and Irwin (2010)), errors from adopting these simplify- ing elasticity assumptions are unlikely to be a major problem. These assump- tions also make practical sense in the context of computing time series of indexes for Africa, where there is a dearth of reliable and consistent estimates of price elasticities of demand and supply for different time periods over the 3. It is conceivable that a distortion to the price of a nontradable could have an indirect trade consequence because of non-zero cross-price elasticities of demand or supply between tradables and nontradables. However, as with estimates of NRAs and CTEs, our estimates of TRIs and WRIs assume those cross-price elasticities (and also those between tradable products) are zero. We make this assumption not only to simplify greatly the algebra but also because reliable estimates of all the relevant cross elasticities for Africa over the 45-year period under review are unavailable. 256 THE WORLD BANK ECONOMIC REVIEW past half-century for each of the covered agricultural products in each of our focus countries.4 Policy instrument trade and welfare reduction indexes The above country-level TRI and WRI measures are the aggregate of the trade- or welfare-reducing indicators of all the policy measures in place. The variables si and ri, as domestic-to-border price ratios, can theoretically encompass distor- tions provided by all trade tax/subsidy and non-tariff trade measures, plus domestic price support measures ( positive or negative), plus direct interven- tions affecting farm input prices. Furthermore, where multiple exchange rates operate, the measures can encompass an estimate of the import or export tax equivalent of that distortionary regime too. Whilst it is desirable to have such an aggregated country level indicator that is so encompassing, agricultural policy analysts are sometimes interested also in Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 indicators of the relative contribution of different policy instruments to reductions in trade or welfare. To provide this insight, it is possible to use the Anderson-Neary framework to construct indexes of policy distortions at the instrument level to facilitate this comparison.5 To capture distortions imposed by each African country’s different policy instruments on its trade volume and its economic welfare, we adopt the meth- odology from Croser and Anderson (2011). These authors de�ne an Instrument Trade Reduction Index (ITRI) and an Instrument Welfare Reduction Index (IWRI), which can be estimated by considering the distortion from a single policy instrument to the producer and consumer sides of the market. The methodology in Croser and Anderson (2011) identi�es four types of border distortions (import taxes and subsidies, and export taxes and subsidies), for which individual ITRI and IWRI series can be estimated. In addition to the border measures, the series for domestic distortions in the form of production, consumption and input taxes and subsidies can be estimated. To estimate the trade-reducing effect of an individual instrument, those authors derive expressions for the change in import volume from the individual policy measures, which are used as the basis for deriving ITRIs. To estimate the welfare-reducing effect of individual instruments, those authors make an assumption about the allocation of the total welfare loss from the combination of individual policy instruments. They assume that border measures are applied �rst, and that this may be supplemented by additional domestic price 4. Sensitivity analysis by Croser, Lloyd and Anderson (2010) shows little difference in the overall TRI and WRI estimates for commodities globally when differentiated elasticity estimates from Tyers and Anderson (1992) were used in place of common ones in each country. 5. Note that most of the series of TRI and WRI indicators in the literature are for single instruments anyway. For example, Irwin (2010) uses only import tariffs, and Kee, Nicita and Olarreaga (2009) report two series of indexes — one based on tariffs only, the other on tariffs plus non-tariff import barriers. Croser and Anderson 257 distortions (which, in practice for Africa, are relatively minor). Thus the dom- estic distortion’s welfare reduction is the residual from subtracting the border measures’ effects from the total welfare reduction of all policy measures. This allocation assumption provides a lower-bound on welfare losses from border measures and an upper-bound on welfare losses from domestic measures. The derivation of the ITRI and IWRI follows essentially the same steps as those for the country-level indexes which encompass all forms of price distor- tion. The difference in the algebraic methodology is to specify separate indexes for the nine different types of policy instrument (for details see Croser and Anderson 2011). Simplifying assumptions can be made in the absence of reliable price elasticity estimates, and again these assumptions have a minimal effect on the estimates. Commodity market trade and welfare reduction indexes In addition to constructing country-level and instrument-speci�c indexes, this Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 paper makes use of another methodology within the Anderson-Neary frame- work to analyse a different aspect of agricultural policy in Sub-Saharan Africa. We construct indexes that show the extent to which African markets for indi- vidual farm commodities are distorted relative to others. We employ the meth- odology in Croser, Lloyd and Anderson (2010) for this purpose. This methodology is novel because whereas all previous work within the trade restrictiveness indexes literature has focused on constructing index numbers of distortions from the perspective of a single country, this methodology instead takes a regional view of individual commodity markets. The regional commodity TRI (WRI) is equal to the uniform trade tax that has the same effect on regional trade volume (welfare) as the existing set of dis- tortions in the region’s national commodity markets. The measures are con- structed in the same way as those for individual country indexes, except that instead of summing across distortions in different industries for a single country, the measures are constructed by summing across distortions in differ- ent countries for a single commodity. The indexes are computed using data on the domestic production and consumption sides of the region’s national com- modity markets, and the measures account for all forms of border and dom- estic price distortions in each country for the commodity market of interest, as well as incorporating import-competing and exportable countries into the measure. II. DISTORTIONS TO A G R I C U LT U R A L I N C E N T I V E S D ATA B A S E This study makes use of the World Bank’s Distortions to Agricultural Incentives database (Anderson and Valenzuela 2008). The database is an output from a global research project seeking to improve the understanding of agricultural policy interventions and reforms in Asia, Europe’s transition econ- omies, Latin America and the Caribbean as well as Africa. The database 258 THE WORLD BANK ECONOMIC REVIEW contains annual estimates of nominal rates of assistance (NRA) (positive or negative) for key farm products in 75 countries that together account for between 90 and 96 percent of the world’s population, farmers, agricultural GDP, and total GDP. There are 21 African countries in the database. We concentrate on the sample of 19 Sub-Saharan African countries listed in the introduction, excluding relatively affluent Egypt and South Africa. For those 19 African focus countries, the database contains around 6000 consistent estimates of annual NRAs to the agricultural sector between 1955 and 2004 or 2005, and the same number of CTEs. Country coverage up to 1960 is much less than from 1961, so the series of estimates presented in this paper begins in that latter year. The estimates of NRA and CTE in the database are at the commodity level and cover a subset of 41 agricultural products in Africa. These so-called covered products account for around 70 percent of each country’s total agricul- tural production over the period studied. The data identi�es each year whether Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 each commodity in each country is considered an importable, exportable or nontradable, a status that may change over time. In the 19 African focus countries, tradable products account for between 40 and 55 percent of the gross value of production of all covered agricultural products (last column of Table 1). The range of policy measures included in the NRA estimates in the Distortions to Agricultural Incentives database is wide. By calculating domestic-to-border price ratios, the estimates include assistance provided by all tariff and nontariff trade measures, plus any domestic price support measures ( positive or negative), plus an adjustment for the output-price equivalent of direct interventions affecting prices of farm inputs. Where multiple exchange rates operate, estimates of the import or export tax equivalents of that distor- tion are included as well. The range of measures included in the CTE estimates include both domestic consumer taxes and subsidies and trade and exchange rate policies, all of which drive a wedge between the price that consumers pay for each commodity and the international price at the border. Anderson and Masters (2009) note several patterns that emerge in the distor- tions to agricultural incentives in the focus countries. In the 1960s and 1970s, many African governments had macroeconomic, sectoral and trade policies that increasingly favored the urban sector at the expense of farm households, and favored production of import-competing farm goods at the expense of exportables. The policy regime was characterized as pro-urban (anti- agricultural) and pro-self-suf�ciency (anti-agricultural trade). Since the 1980s, Africa has reduced its anti-agricultural and anti-trade biases, but many distor- tions still remain. For the countries in this study, Table 1 and Figure 1 illustrate those patterns. The weighted average NRA for the 19 countries is almost always below zero, indicating that agricultural price, trade and exchange rate policies together have reduced the earnings of farmers in these countries. The average rate of Croser and Anderson 259 T A B L E 1 . Nominal Rates of Assistance for import-competing, exportable and nontradable covered agricultural products, 19 African focus countries,a 1961 to 2004 (percent) NRA, agricultural productsa Tradables share (%) of value of All All Standard all covered Covered Covered covered Covered covered deviation agricultural exportables importables tradablesb nontradables products of NRAsb production 1961– 64 2 30 123 3 0 21 34 49 1965– 69 2 39 62 2 15 0 2 11 33 55 1970– 74 2 47 30 2 27 0 2 17 31 55 1975– 79 2 52 22 2 30 21 2 23 37 54 1980– 84 2 47 4 2 28 21 2 18 35 46 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 1985– 89 2 50 49 2 26 22 2 15 33 46 1990– 94 2 49 5 2 27 22 2 16 31 41 1995– 99 2 32 3 2 15 23 2 10 25 39 2000– 04 2 32 7 2 16 23 2 10 26 43 Source: Anderson and Valenzuela (2008) a. Nominal rates of assistance for the 19 African focus countries are weighted by the gross value of production at undistorted prices for the relevant sub-sector. b. The simple average of the 19 focus countries’ standard deviation of NRA around its weighted mean. F I G U R E 1. Nominal Rates of Assistance for import-competing, exportable and all covered agricultural products, 19 African countries, 1961 to 2004 Source: Anderson and Valenzuela (2008). 260 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Nominal rates of assistance, all covered agricultural products, 19 African focus countries, 1961 to 2004 (percent) 1961– 1965– 1970– 1975– 1980– 1985– 1990– 1995– 2000– 64 69 74 79 84 89 94 99 04 Africa -1 -11 -17 -23 -18 -15 -16 -10 -10 Benin na na 23 21 21 21 24 24 21 Burkina Faso na na 22 23 24 21 23 23 0 Cameroon 24 28 2 12 2 25 2 19 25 24 24 21 Chad na na 2 12 2 11 28 21 23 23 21 Coˆ te d’Ivoire 2 29 2 35 2 33 2 40 2 40 2 28 2 22 2 22 2 28 Ethiopia na na na na 2 12 2 15 2 17 2 10 27 Ghana 2 15 2 28 2 23 2 41 2 32 28 23 25 22 Kenya 13 22 2 24 2 15 2 30 28 2 30 24 4 Madagascar 2 19 2 23 2 20 2 38 2 51 2 26 27 24 2 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Mali na na 26 28 27 23 25 27 0 Mozambique na na na 2 56 2 42 2 51 24 5 14 Nigeria 21 12 7 5 8 15 4 0 25 Senegal 2 15 2 12 2 33 2 34 2 30 5 7 2 10 2 12 Sudan 2 26 2 37 2 48 2 28 2 33 2 39 2 54 2 29 2 15 Tanzania na na na 2 50 2 60 2 52 2 30 2 29 2 17 Togo na na 21 21 22 22 24 23 21 Uganda 23 25 2 12 2 24 2 12 2 14 21 1 1 Zambia 2 24 2 32 2 42 2 57 2 26 2 68 2 53 2 34 2 36 Zimbabwe 2 36 2 36 2 44 2 54 2 47 2 43 2 45 2 38 2 73 Source: Anderson and Valenzuela (2008) direct taxation (negative NRA) of African farmers rose until the late 1970s before declining by more than half over the next 25 years. Table 2 reports the country-level NRAs for covered products for each of the 19 countries in this sample. It reveals the considerable diversity within the sample. In some countries — such as Cameroon, Ghana, Senegal, Uganda, Tanzania, and Madagascar —a reduction in the taxing of farmers is evident following the regional peak in 1975–79, while in other countries — such as Cote d’Ivoire, Zambia, and Zimbabwe — high levels of agricultural taxation appear to have persisted. The country level aggregate measures hide the degree of variation in NRA estimates within countries. Column 6 of Table 1 suggests the standard devi- ations around the weighted mean NRA for covered products in each country has been high, but has declined somewhat since the mid-1990s. An indication of the extent of variation between groups of products is seen even when com- paring the average NRAs for import-competing and exportable product groups (Figure 1). The gap between those two groups’ average NRAs has tended to narrow over the period shown, suggesting there has been a decline also in the anti-trade bias in Africa’s agricultural policies since the mid-1990s. Croser and Anderson 261 F I G U R E 2. Trade and Welfare Reduction Indexes and Nominal Rate of Assistance for all covered agricultural products, 19 African countries, 1961 to 2004 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Anderson and Croser (2009). Notwithstanding their valuable contribution, sectoral averages of NRAs can be misleading as indicators of the aggregate extent of price distortion within the sector. They can also be misleading when compared across countries that have varying degrees of dispersion in their NRAs (and CTEs) for farm pro- ducts. To see why, we now turn to the TRI and WRI estimates. III. TRADE A N D W E L FA R E R E D U C T I O N I N D E X E S T I M AT E S The regional aggregate TRI for the 19 African focus countries for all covered products is positive and large over the period under analysis (middle line in Figure 2). The positive TRI indicates that overall agricultural policy in African countries reduced trade. The extent of that has decreased over time, however, with the �ve-year TRI averages of between 20 and 25 percent in the �rst two decades of data falling to around 10 percent in the most recent decade. The TRI has the opposite sign to the NRA (see bottom line in Figure 2) because the TRI correctly aggregates the effect of all policies that reduce trade volume, regardless of whether they make a positive or negative contribution to the NRA. The importance of the difference in these aggregations of the trade-reducing effect of policies can be seen in the early 1960s, for example, when the average NRA was around zero but the TRI was quite high (the latter capturing the trade-reducing effect of both import taxes and export taxes, which offset one another in the NRA estimate). Similarly, in the late 1980s the 262 THE WORLD BANK ECONOMIC REVIEW NRA changes from around 2 15 percent to 2 10 percent at a time when the TRI increases from 20 to 30 percent: the aggregate NRA gives the impression that policies are becoming less distorted in this period but, because the upward trend in the NRA is caused by an increase in import taxes, the TRI correctly reveals that agricultural policies are in fact becoming more trade-restrictive in this time period. The WRI series for all covered products is necessarily positive and every- where lies above the TRI series (compare middle and upper lines in Figure 2). The WRI series correctly demonstrates the negative welfare consequences that flow from both negative and positive price distortions. Furthermore, the WRI series provides a better indicator of the welfare cost of distortions than the average level of assistance or taxation, due to the inclusion in the WRI of the ‘power of two’. A weighted arithmetic mean does not fully reflect the welfare effects of agricultural distortions because the dispersion of that support or taxation across products has been ignored. By contrast, the WRI captures the Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 higher welfare costs of peak levels of assistance or taxation. The aggregate African results mask country-level diversity in the TRI and WRI series. Some countries — such as Cote d’Ivoire, Ethiopia, Sudan, Tanzania and Zimbabwe — persistently restricted trade (in aggregate) through- out the period under analysis (Table 3). Other countries — such as Kenya, Mozambique and Zambia — have had periods in which policies in aggregate have expanded agricultural trade slightly (via import subsidies). In terms of the WRI, there is less diversity across countries, since WRI measures are all necess- arily positive (Table 4). The extent to which agricultural policy reduced aggre- gate welfare does differ across countries, however. Some countries have low reductions in welfare, including Uganda and most cotton-exporting countries. Figure 3 provides a snapshot for 2000–04 of the diversity in the WRI and TRI for each of the 19 countries, with the weighted African average in the middle and close to Kenya. A useful way of understanding the overall welfare reduction for Africa from agricultural policy is to compute the country contributions to the WRI for the 19 African focus countries as a whole. Contributions can be found by comput- ing dollar values of the welfare reduction index for each country (by multiply- ing the WRI percent by the average of the gross value of production and consumption at undistorted prices). Table 5 shows that Nigeria, Sudan and Ethiopia dominate the region’s contributions. The last column of Table 5 reports country contributions to the decline in the regional WRI from its value of 44 percent in 1975–79 to its value of 27 percent in 2000–04. Nigeria and Sudan dominate that overall reduction, together accounting for around 80 percent of the fall in the WRI. However, Cameroon, Madagascar, Senegal and Uganda have slightly offsetting effects on the regional fall in the WRI over that period. It is worth noting that the TRI and WRI for all covered products is signi�- cantly lower than that for just tradables. This is because nontradables account Croser and Anderson 263 T A B L E 3 . Trade Reduction Index, all covered agricultural productsa, 19 African focus countries, 1961 to 2004 (percent) 1961– 1965– 1970– 1975– 1980– 1985– 1990– 1995– 2000– 64 69 74 79 84 89 94 99 04 Africa 24 22 20 21 15 24 14 9 10 Benin na na 2 1 1 0 2 3 1 Burkina Faso na na 2 3 4 1 3 3 0 Cameroon 2 5 6 14 12 3 2 2 1 Chad na na 12 11 8 1 3 3 1 Coˆ te d’Ivoire 13 13 24 27 19 17 12 15 22 Ethiopia na na na na 14 16 19 11 9 Ghana 6 11 10 22 20 15 7 3 7 Kenya 2 16 2 19 24 12 21 19 27 9 11 Madagascar 20 15 2 13 6 21 17 3 3 8 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Mali na na 4 7 6 3 5 7 0 Mozambique na na na 27 26 2 14 1 6 24 Nigeria 39 38 31 18 11 19 7 8 1 Senegal 14 10 30 36 28 25 26 7 12 Sudan 29 28 29 29 22 56 40 17 31 Tanzania na na na 16 18 34 30 16 17 Togo na na 0 1 2 1 4 3 1 Uganda 2 4 8 14 9 10 2 2 1 Zambia 21 1 1 36 2 11 2 45 2 27 27 29 Zimbabwe 33 38 43 51 29 37 19 10 12 Source: Anderson and Croser (2009) based on NRA and CTE data in Anderson and Valenzuela (2008). a. Includes all import-competing, exportable and nontradable products, with nontradable sectors assumed to have a zero level of distortion on the volume of trade. for a large share of the gross value of production and consumption (see �nal column of Table 1). The TRI estimates for all covered products are roughly half, and WRI estimates are roughly two-thirds, what there are with nontrad- ables excluded. It is useful to compare the TRI and WRI results for the African focus countries with those for other developing country regions, which are reported in Lloyd, Croser and Anderson (2010). The African focus countries’ policies have been, and remain, the most trade- and welfare-reducing. However, all three regions have shown a trend towards less damaging agricultural policies in recent years (Figure 4). Policy instrument results We now turn to the national decompositions of the TRI and WRI to the policy instrument level. Figure 5 provides a summary of the estimates of the contri- bution to the weighted average WRI series for the 19 African focus countries of four different border measures: taxes and subsidies on both imports and 264 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Welfare Reduction Index, all covered agricultural products, 19 African focus countries, 1961 to 2004 (percent) 1961– 1965– 1970– 1975– 1980– 1985– 1990– 1995– 2000– 64 69 74 79 84 89 94 99 04 Africa 49 46 45 44 39 45 40 28 27 Benin na na 9 6 7 4 8 7 4 Burkina Faso na na 9 13 14 5 9 9 9 Cameroon 9 14 17 29 22 12 11 10 4 Chad na na 24 23 20 5 9 8 6 Coˆ te d’Ivoire 28 36 36 40 38 30 25 25 31 Ethiopia na na na na 22 24 27 20 16 Ghana 17 30 28 44 49 36 17 11 15 Kenya 35 39 29 34 38 28 35 26 29 Madagascar 23 27 26 43 55 37 21 11 13 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Mali na na 16 20 18 8 13 14 9 Mozambique na na na 63 52 63 18 18 41 Nigeria 87 78 68 54 45 63 48 36 31 Senegal 17 15 38 41 36 50 55 11 16 Sudan 36 40 51 40 40 65 79 42 44 Tanzania na na na 58 65 62 53 46 38 Togo na na 4 5 9 5 10 8 5 Uganda 6 9 20 35 24 24 4 4 4 Zambia 26 41 47 57 31 69 58 39 42 Zimbabwe 39 45 50 56 46 42 46 40 72 Source: Anderson and Croser (2009) based on NRA and CTE data in Anderson and Valenzuela (2008). exports. The �gure demonstrates the very substantial role that export taxes have played in the reduction of welfare in the region. On average, more than half the welfare reductions has come from anti-agricultural export taxing pol- icies over the period studied, but the decline in them has contributed most to reform in recent decades: the gross contribution of export taxes to the reduction in the WRI over the period 1985–89 to 2000–04 is 93 percent. The remaining 7 percent is made up of a 34 percent gross contribution from import tax cuts offset by a 2 28 percent contribution from export subsidies ( 2 13 percent) and import subsidies ( 2 15 percent). The contributions to TRI and WRI estimates for African countries from domestic distortions are small, never accounting for more than 5 percent of the overall regional TRI or WRI. This can be seen in Table 6. That table also reveals the far greater dominance of export taxation in Africa as compared with developing Asia and Latin America, particularly in the 2000-04 period. Commodity TRI and WRI results The TRI and WRI estimates for individual regional commodity markets provide a different perspective on the level of distortion in the focus countries Croser and Anderson 265 F I G U R E 3. Trade and Welfare Reduction Indexes, all covered agricultural products, 19 African countries and regional averagea, 2000-04 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Anderson and Croser (2009) a. To get the African regional average, the national indexes are weighted by the average of the gross value of production and consumption at undistorted prices. over the period under analysis. Table 7 reveals considerable diversity in the dis- tortions in different commodity markets in Africa. Fruit and vegetable com- modity markets, which tend to have a high share of nontradable production, have low WRI estimates on average, whereas traded commodities such as tropi- cal crops, oilseeds and livestock tend to have more welfare-reducing policies in place. Grains, which comprise a mixture of tradable and nontradable products, had highly-distortionary policies in the 1960s on average, but these have been reduced over time. As of 2000–04, the indexes suggest sugar and cotton markets continue to have highly distorted policies in terms of both the trade and welfare effects of their policies. I V. C O N C L U S I O N S Reform of agricultural policy in Africa is topical at present. Recently announced international aid and investment programs, domestic policy reforms, and the negotiation of international and regional trade agreements are on the agenda, not to mention climate change. Measurement of intervention levels is required to assess policy initiatives in each of these areas. Certainly 266 THE WORLD BANK ECONOMIC REVIEW T A B L E 5 . Country contributionsa to the regional Welfare Reduction Index for African focus countries,b all covered agricultural products, 1961 to 2004, and to its fall from 1975-79 to 2000-04 (percent) Contribution to fall in WRI between 1975– 79 and 1961– 64 1970– 74 1980– 84 1990– 94 2000– 04 2000– 04 Africa WRI 49 45 39 40 27 Cameroon 2 3 2 1 0 24 Cote d’Ivoire 3 5 6 4 5 1 Ethiopia - - 10 10 9 na Ghana 2 3 4 2 3 1 Kenya 2 2 3 3 2 1 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Madagascar 1 3 4 1 1 22 Mozambique - - 2 1 2 2 Nigeria 74 51 37 38 35 34 Senegal 1 2 2 2 1 23 Sudan 10 21 18 30 27 44 Tanzania - - 7 3 6 1 Uganda 1 4 4 0 1 28 Zambia 1 2 1 1 2 1 Zimbabwe 3 4 3 3 5 7 Africa 100 100 103 100 100 100 Source: Authors’ calculations from data in Anderson and Croser (2009). a. Country contributions are computed by converting national percentage WRIs to dollar values by multiplying by the average of the gross value of production and consumption at undis- torted prices. b. Benin, Burkina Faso, Chad, Mali, and Togo are not shown as each of their contributions was less than 0.5 percent. economy-wide models can measure the welfare and trade (and other) effects of policy in a particular country or market, and do it better than can partial equi- librium analysis where there are potentially offsetting policies such as import taxes and import subsidies. Such models require, however, reliable data on the structure of the economy and sound econometric estimates of myriad par- ameters, neither of which are easily found for the poorer countries of Africa. Even where economy-wide models are available, they are calibrated to a par- ticular year (typically 5 þ years ago) and are incapable of providing easily updatable time series indicators of the national and regional effects of distor- tional policies. Scalar index measures, by contrast, can provide meaningful partial equili- brium indicators of the welfare and trade effects of policy interventions in agri- culture in poorer countries. As demonstrated above, these indexes can be Croser and Anderson 267 F I G U R E 4. Trade- and Welfare-Reduction Indexes, Sub-Saharan Africa, Asia and Latin America, all covered tradable agricultural products, 1960a to 2004 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Generated from estimates in Anderson and Croser (2009) a. The �rst period is 1961-64 for African countries. estimated using no more than already available price and quantity data used to generate NRAs and CTEs (or PSEs and CSEs), and so are relatively inexpensive to generate and update annually for timely policy monitoring.6 6. This is indeed what is being planned in a new FAO/OECD project called Monitoring African Food and Agricultural Policies, funded by the Bill and Melinda Gates Foundation (see www.fao.org/ mafap). These indexes are also being considered by USAID for inclusion among the policy indicators to be estimated to monitor future policy developments as they affect global hunger and food security (see www.state.gov/s/globalfoodsecurity). 268 THE WORLD BANK ECONOMIC REVIEW F I G U R E 5. Decomposition of the Welfare Reduction Index due to border measures, by policy instrument, 19 focus African countries, 1961 to 2004 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Croser and Anderson (2011). The scalar index numbers reported in this paper are thus a major sup- plement to the widely-used price distortion measures such as NRAs or PSEs, because they correctly aggregate offsetting policies and because the WRI prop- erly captures the much higher welfare costs associated with the largest price distortions. True, the indexes measured in this study (like NRAs or PSEs) do not make use of price elasticity estimates, but if and when reliable estimates become available for the many agricultural products of the region, they can be incorporated to revise our estimates. Meanwhile, both theory and other recent empirical studies (see, e.g, Croser, Lloyd and Anderson 2010) provide comfort in suggesting the use of differentiated elasticity estimates across commodities would not make much difference to the results. The methodology in the paper adopts the standard partial equilibrium approach still presented in most textbooks on trade policy or welfare econ- omics. In particular, it is based on the benchmark of competitive markets. The methodology ignores the existence of divergences such as externalities and gov- ernance problems, including administrative costs. The trade and welfare reduction indexes reported above may be over- or under-stated to the extent that such problems exist. For example, in some cases where there is market failure, we know from second-best theory that policies that increase assistance to a lightly protected sector may increase rather than decrease national econ- omic welfare. Even so, the series reported in this paper almost certainly give a better indication of trade and welfare effects of policies than the NRA/CTE measures from which they are built. Croser and Anderson 269 T A B L E 6 . Contributions from different policy instruments on the production side to the TRI and WRI for covered products by different policy instruments,a by developing country region,b 1980–84 and 2000–04 (percent) (a) TRI 1980-84 2000-04 Africa Asia Latin Amer. Africa Asia Latin Amer. All measures 24 37 21 16 10 7 Border measures 22 33 20 13 9 7 Export tax 24 27 21 15 2 6 Export subsidy 24 21 21 23 21 23 Import tax 10 9 6 7 10 5 Import subsidy 28 22 25 26 21 22 Domestic taxes & subsidies 2 4 1 3 1 0 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Production tax 2 3 0 3 1 0 Production subsidy 0 0 1 0 0 0 (b)WRI 1980-84 2000-04 Africa Asia Latin Amer. Africa Asia Latin Amer. All measures 54 61 46 38 20 25 Border measures 48 44 38 33 17 18 Export tax 25 29 23 16 2 7 Export subsidy 4 1 1 3 3 3 Import tax 11 12 7 8 12 7 Import subsidy 8 2 6 6 1 2 Domestic taxes & subsidies 6 17 8 5 3 7 Production tax 5 15 1 4 1 2 Production subsidy 1 1 8 0 2 5 Source: Croser and Anderson (2011). a. Each instrument share is computed in the following two steps: (1) indices are converted to constant 2000 $US by multiplying the index by the average value of production or consumption for that instrument group at the country level; (2) each instrument dollar amount index is divided by the country average value of production or consumption. The measures in the table — which are like a weighted average of an overall regional index — therefore reflect both the absolute size of the index for each policy instrument and the relative importance of that policy instrument in the region. b. Africa includes Egypt and South Africa (unlike in previous tables); Asia excludes Japan; Latin America includes the Caribbean. Notwithstanding those caveats, two clear conclusions can be drawn from the empirical estimates presented in this paper. One is that they con�rm that there has been very substantial policy reform in African agriculture over recent decades, especially in phasing out export taxation. The other is that they reveal there is still a long way to go before that reform process is complete, since the trade- and 270 THE WORLD BANK ECONOMIC REVIEW T A B L E 7 . Commodity Welfare Reduction Index, African regional market of 19 focus countries, 31 covered agricultural products, 1961–64 to 2000–04 (percent) 1961-64 1965-69 1970-74 1975-79 1980-84 1985-89 1990-94 1995-99 2000-04 Grains 59 50 44 34 28 33 26 20 18 Cassava 0 0 1 1 3 1 1 4 3 Maize 114 73 63 71 54 67 40 38 35 Millet 18 18 11 5 10 13 16 18 8 Rice 31 30 40 36 48 60 38 16 18 Sorghum 153 144 118 95 83 95 80 52 49 Wheat 17 37 40 30 14 16 35 16 16 Oilseeds 28 42 54 49 47 40 72 43 36 Cashew na na na 80 80 85 61 13 11 Groundnut 27 43 54 51 50 35 60 41 47 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Oilseed na na na na 47 52 61 56 42 Palmoil 25 31 45 26 28 44 132 50 13 Sesame 50 60 62 65 56 44 47 45 38 Soybean na 14 34 44 45 44 56 52 64 Sunflower 0 0 0 0 0 0 0 0 0 Tropical 36 41 45 61 54 49 53 44 51 crops Cocoa 31 51 46 62 54 41 37 37 38 Coffee 39 41 46 64 56 48 47 35 21 Cotton 42 35 44 57 59 59 71 59 64 Sugar 22 35 47 49 43 38 45 45 87 Tea 12 8 24 56 52 47 51 50 49 Tobacco 39 38 48 56 50 50 40 39 58 Fruit & 0 0 0 4 5 5 2 5 5 vegetables Banana 2 4 0 2 2 1 5 5 2 Bean 7 10 3 48 62 73 35 42 40 Roots & 0 0 0 0 0 0 0 0 0 tubers Pepper na 42 9 39 47 80 30 62 27 Plantain 0 0 0 0 0 0 0 0 0 Potato na na na 0 0 0 0 0 0 Sweet 0 0 0 0 0 0 0 0 0 potato Yam 0 0 0 1 2 1 1 4 4 Livestock 30 36 52 35 33 68 66 40 38 Beef 34 42 58 29 29 60 73 43 42 Camel 38 60 34 38 34 68 84 49 99 Milk 19 16 41 36 29 79 40 30 29 Sheepmeat 42 48 61 46 38 59 70 54 33 Source: Authors’ calculations. Croser and Anderson 271 welfare-reduction indexes associated with the present decade’s policies are still substantial and reveal large differences across countries and commodities. FUNDING This work is a product of a World Bank research project on Distortions to Agricultural Incentives (Project P093895, see www.worldbank.org/ agdistortions) which was �nancially supported by the governments of the Netherlands (BNPP), the United Kingdom (DfID) and Ireland; and by the Australian Research Council (DP0880565). AC K N OW L E D G E M E N T S The authors are grateful for helpful referee comments and for the distortion estimates provided by the authors of the various African country case studies, Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 reported in Anderson, K. and W. Masters (eds.), Distortions to Agricultural Incentives in Africa, Washington DC: World Bank, 2009. The views expressed are the authors’ alone and not necessarily those of the World Bank and its Executive Directors, nor the countries they represent, nor of the institutions providing the project research funds. A P P E N D I X : D E R I VA T I O N OF TRADE- AND WELFARE-REDUCTION INDEXES Lloyd, Croser and Anderson (2010) outline a methodology for computing indexes which accurately capture the state of trade policy regime in an individual country in a theoretically meaningful way. Their methodology, which draws heavily on the Anderson and Neary (2005) methodology, de�nes partial equilibrium indexes which aggregate the production and consumption sides of the economy separately (instead of trade data as is more commonly done with trade restrictiveness indexes). This form of index is well-suited to agricultural distortions research, where data are available for production and consumption of individual farm com- modities. This Appendix briefly outlines that theory for the import-competing sector of a small open economy (with further details and extensions available in Lloyd, Croser and Anderson 2010 and Croser, Lloyd and Anderson 2010). Consider an individual country and assume it has a small, open economy in which all markets are competitive. The market for an import good may be dis- torted by a tariff and other nontariff border measures or by behind-the-border measures such as domestic subsidies and price controls. The effect of a coun- try’s distortions on its import volume is captured by the Trade Reduction Index (TRI), de�ned as the uniform tariff rate which, if applied to all goods in the place of all actual border and behind-the-border price distortions, would result in the same reduction in the volume of imports (summed across products by valuing them at the undistorted border price) as the actual distortions. 272 THE WORLD BANK ECONOMIC REVIEW Suppose the market for one good, good i, is distorted by a combination of measures that distort its consumer and producer prices. For the producers of the good, the distorted domestic producer price, pP i , is related to the à P à border price, pi , by the relation, pi ¼ pi ð1 þ si Þ where si is the rate of distor- tion of the producer price in proportional terms. For the consumers of the good, the distorted domestic consumer price, pC i , is related to the border price by the relation, pCi ¼ pà i ð1 þ r i Þ where r i is the rate of distortion of the consumer price in proportional terms. In general, ri = si . Using these relations, the change in the value of imports in the market for good i is given by: DMi ¼ pà à i Dxi À pi Dyi ð1Þ 2 pC Ã2 pP ¼ pà i dxi =d i ri À pi dyi =d i si where the quantities of good i demanded and supplied, xi and yi, are Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 functions just of their own domestic price: xi ¼ xi ðpC pP i Þ and yi ¼ yi ð i Þ. Strictly speaking, this result holds only for small distortions. In reality rates of distortion may not be small. If, however, the demand and supply functions are linear over the relevant price range, the effect on imports is given by equation (1) with constant slopes of the demand and supply curves (dxi / dpC i and dyi / dpP i , respectively). If the functions are not linear, this expression pro- vides an approximation to the loss. With n importable goods subject to different levels of distortions, the aggre- gate reduction in imports, in the absence of cross-price effects in all markets, is given by: X n X n 2 pC 2 pP DM ¼ pà i dxi =d i ri À pà i dyi =d i si ð2Þ i ¼1 i¼1 Setting the result equal to the reduction in imports from a uniform tariff, T, gives: X n X n X n 2 pC 2 pP 2 pà i dxi =d i ri À pà i dyi =d i si ¼ pà i dmi =dpi T i¼1 i ¼1 i¼1 Solving for T, gives " # T ¼ fRa þ Sbg ð3aÞ X n X 2 pC 2 pC where R ¼ ri ui with ui ¼ pà i dxi =d i = pà i dxi =d i ; ð3bÞ i¼1 i " # X n X 2 pP 2 pP S¼ si vi with vi ¼ pà i dyi =d i = pà i dyi =d i ; and ð3cÞ i¼1 i Croser and Anderson 273 X X 2 pC 2 a¼ pà i dxi =d i = pà i dmi =dpi ; and i i X X ð3dÞ 2 pP 2 b¼À pà i dyi =d i = pà i dmi =dpi i i Evidently, the uniform tariff T can be written as a weighted average of the level of distortions of consumer and producer prices, R and S (the Consumer and Producer Assistance Indexes, respectively). An important advantage of using this decomposition of the index into producer and consumer effects is that it treats correctly the effects of non-tariff measures and domestic distor- tions that affect the two sides of the market differently. In equation 3c (equation 3b), the weights for each commodity are pro- portional to the marginal response of domestic production (consumption) to changes in international free-trade prices. These weights can be written as, Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 among other things, functions of the domestic price elasticities (at the protected trade situation) of supply and demand (si and ri, respectively):7 X n X n ui ¼ ri ð pà i xi Þ= ri ð pà à i xi Þ and vi ¼ si ð pi yi Þ= si ð pà i yi Þ ð4Þ i i The other index de�ned in Lloyd, Croser and Anderson (2010), the Welfare Reduction Index (WRI), measures the effect of a country’s distortions on its economic welfare. The derivation follows the same steps as in the derivation of the TRI except that instead of starting from the loss in trade volume from a policy, one starts from a loss of consumer and producer surplus (a welfare loss, Li). With n importable goods subject to different levels of distortions, the aggregate welfare loss, in the absence of cross-price effects in all markets, is given by: ( ) 1 X n Xn L¼ ð pà si Þ2 dyi =d pP i À ð pà 2 pC i ri Þ dxi =d i ð5Þ 2 i¼1 i i¼1 The uniform tariff rate, W, that generates an aggregate deadweight loss iden- tical with that of the differentiated set of tariffs is determined by the following equation: X n 2 X n 2 X n 2 ð pà pP i si Þ dyi =d i À ð pà pC i ri Þ dxi =d i ¼ À ð pà i W Þ dmi =dpi ð6Þ i¼1 i¼1 i¼1 W is thus the uniform tariff which, if applied to all goods in the place of all actual tariffs and NTMs and other distortions, would result in the same 7. These expressions can also be written as functions of, among other things, the domestic price elasticities at the free trade points. 274 THE WORLD BANK ECONOMIC REVIEW aggregate loss of welfare as the actual distortions. Solving for W, we have: W ¼ fR02 a þ S02 bg1=2 ð7aÞ " #1 Xn 2 0 2 where R ¼ ri ui ð7bÞ i¼1 " #1 X n 2 S0 ¼ s2 i vi ð7cÞ i¼1 with ui, vi, a and b as de�ned for equation 3 above. W is the desired Welfare Reduction Index, while R0 and S0 are the contributions to W from consumer and producer price distortions, respectively. They, like their appropriately weighted average W, are means of order two. As with the index T, we can deal with, and analyse, the production and consumption sides of the sector Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 separately. Extension to exportable sectors Lloyd, Croser and Anderson (2010) show how the indexes can each be extended to include the exportables sub-sector. This is facilitated by way of aggregating the import-competing and exportables sub-indexes where the weights for each sub-sector are the share of the sub-sectors’ value of production (consumption) in the total value of production (consumption). The resulting measure is the import tax/export subsidy which, if applied uniformly to all pro- ducts in the sector, would give the same loss of welfare as the combination of measures distorting consumer and producer prices in the import-competing and exportable sub-sectors. In the case of the TRI it is important to keep separate track of the subsets of import-competing and exportable goods because the sign of an NRA in the exportable sub-sector ( positive or negative) has the opposite effect on the TRI. That is, while an export subsidy in the exportable sub-sector reduces welfare in the same way as an import tax in the import-competing sub-sector, the export subsidy will increase trade and the import tariff reduces trade. Extension to nontradables sectors In this paper we extend indexes to include nontradable sectors. Because non- tradables generally have low or zero distortions, an index that does not take into account these sectors will tend to overstate the trade- and welfare-reducing effect of overall agricultural policy. To include nontradables, we keep separate track of three sub-sectors of the economy: import-competing, exportable and nontradable sub-sectors. We gen- erate sub-sector-speci�c TRI and WRI indexes (as we previously did for each of the import-competing and exportable sub-sectors). The three sub-sector Croser and Anderson 275 indexes are then aggregated using as weights each sub-sectors’ share of value of production (consumption) in the total value of production (consumption). For the WRI, because distortions in the nontradable sub-sector can cause welfare distortions, we proceed as expected and si and ri values in equations 7b and 7c are the actual level of distortion in the nontradable sub-sector. For the TRI, however, we make an assumption that si and ri values in equations 3b and 3b are zero. This assumption means distortions to nontrad- able products do not alter the sector’s trade volume, and that the contribution of nontradables to the TRI is only through their share in the sector’s total value of production (consumption). Why elasticities are of minor importance To assess how important is the simplifying assumption in this paper that the domestic price elasticities of supply are equal across commodities within a Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 country, and likewise for elasticities of demand, consider the standard form of the Producer Assistance Index (PAI) from equation (7c): " #1 X n 2 X S0 ¼ s2 i vi with vi ¼ pÃ2 pP i dyi =d i = pà 2 pP i dyi =d i i¼1 i This partial equilibrium measure can be broken down into three parts:8 1 s2 þ V2 S0 ¼ ½ s þ rs Š : 2 The three parts are: P † production-weighted average producer distortions,  s¼ si hi , where hi is the production share of good i; i † production-weighted P variance of producer distortions, 2 V2s ¼ ð si À s Þ hi ; and i † the covariance between each producer distortion and its elasticity of output supply scaled by the production weighted average output supply elasticity,P  ; s2 Þ , where si is the elasticity of output supply rs ¼ covðsi =s and s ¼ si hi . i The formula makes explicit that an increase in the dispersion of producer distortions increases the partial equilibrium index relative to production- weighted average producer distortion. In addition, the partial equilibrium dis- tortion index will be larger than the production-weighted average producer dis- tortion when the covariance between supply elasticities and producer distortion 8. Kee, Nicita and Olarreaga (2008, p. 677) show the decomposition for the usual Anderson and Neary index, which is based on import volumes, import demand elasticities and trade distortion measures. 276 THE WORLD BANK ECONOMIC REVIEW measures is positive. An analogous decomposition can be derived for the Consumer Assistance Index (CAI). In the absence of elasticity data across time and countries, it is possible to estimate PAIs, CAIs, TRIs and WRIs with the simplifying assumption that domestic price elasticities of supply are equal across commodities within a country, and likewise for elasticities of demand. The simplifying assumption equates to a computation of the PAI in which the third component of the decomposition shown above is zero. Anderson and Neary (2005, p. 293) observe that elasticities are ‘not very influential’ in affecting trade restrictiveness indices because elasticities appear in both the numerator and denominator of the indices. In the PAI expression in equation 7c, for example, elasticities appear in both the numerator and denominator of the vi expression. In the third term of the PAI decomposition above, the elasticity for good i is scaled by the production weighted average elasticity for all goods. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 In empirical work, Kee, Nicita and Olarreaga (2008) note that the contri- bution of the covariance term to their estimates trade restrictiveness indexes is very small in practice. Irwin (2010), in his historical study of US trade policy, similarly shows that empirically the covariance is a very small factor relative to the average tariff and variance of the tariff. His estimated indexes depend almost entirely on the mean and variance of tariff rates, which are independent of elasticities. Thus both theory and recent empirical analyses suggest reasonable approxi- mations of the PAI, CAI, TRI and WRI can be obtained even when elasticity estimates are unavailable. REFERENCES Anderson, J.E. (2009a), ‘Consistent Trade Policy Aggregation’, International Economic Review 50(3): 903 –27. Anderson, J.E., and J.P. Neary (2005), Measuring the Restrictiveness of International Trade Policy, Cambridge MA: MIT Press. Anderson, K. (2009b), ‘Five Decades of Distortions to Agricultural Incentives’, chapter 1 in K. Anderson (ed.), Distortions to Agricultural Incentives: A Global Perspective, 1955-2007, London: Palgrave Macmillan and Washington DC: World Bank. Anderson, K., and J.L. Croser (2009), National and Global Agricultural Trade and Welfare Reduction Indexes, 1955 to 2007, database available at www.worldbank.org/agdistortions K. Anderson, and W. Masters (eds.) (2009), Distortions to Agricultural Incentives in Africa, Washington DC: World Bank. Anderson, K., and E. Valenzuela (2008), Global Estimates of Distortions to Agricultural Incentives, 1955 to 2007, database available at www.worldbank.org/agdistortions Bach, C., and W. Martin, (2001), ‘Would the Right Tariff Aggregator for Policy Analysis Please Stand Up?’ Journal of Policy Modeling 23: 621–35. Chen, S, and M. Ravallion (2010), ‘The Developing World is Poorer Than We Thought, But No Less Successful in the Fight Against Poverty’, Quarterly Journal of Economics 125(4): 1577–1625, November. Croser and Anderson 277 Croser, J., and K. Anderson (2011), ‘Changing Contributions of Different Agricultural Policy Instruments to Global Reductions in Trade and Welfare’, World Trade Review 10, 2011 (forthcoming). Croser, J.L., P.J. Lloyd, and K. Anderson (2010), ‘How do Agricultural Policy Restrictions to Global Trade and Welfare Differ Across Commodities?’ American Journal of Agricultural Economics 92(3): 698–712, April. Feenstra, R. (1995). ‘Estimating the Effects of Trade Policy’, in G. Grossman and K. Rogoff (eds.), Handbook of International Economics, Vol. 3, Amsterdam: Elsevier. Irwin, D. (2010), ‘Trade Restrictiveness and Deadweight Losses from U.S. Tariffs, 1859-1961’, American Economic Journal: Economic Policy 2: 111–33, August. Kee, H.L., A. Nicita, and M. Olerreaga (2008), ‘Import Demand Elasticities and Trade Distortions’, Review of Economics and Statistics 90(4): 666– 82, November. ——— (2009), ‘Estimating Trade Restrictiveness Indexes’, Economic Journal 119(534): 172 –99, January. Lloyd, P.J. (1974), ‘A More General Theory of Price Distortions in an Open Economy’, Journal of International Economics 4(4): 365– 86, November. Lloyd, P.J., J.L. Croser, and K. Anderson (2010), ‘Global Distortions to Agricultural Markets: New Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Indicators of Trade and Welfare Impacts, 1960 to 2007’, Review of Development Economics 14(2): 141–60, May. OECD (2010), Agricultural Policies in OECD Countries: Monitoring and Evaluation 2010, Paris: Organization for Economic Cooperation and Development. Tyers, R., and K. Anderson (1992), Disarray in World Food Markets: A Quantitative Assessment, Cambridge and New York: Cambridge University Press. World Bank (2007), World Development Report 2008: Agriculture for Development, Washington DC: World Bank. Thresholds in the Finance-Growth Nexus: A Cross-Country Analysis Hakan Yilmazkuday Thresholds of inflation, government size, trade openness, and per capita income for the �nance-growth nexus are investigated using �ve-year averages of standard vari- ables for 84 countries from 1965 to 2004. The results suggest that (i) high inflation crowds out positive effects of �nancial depth on long-run growth, (ii) small govern- ment sizes hurt the �nance-growth nexus in low-income countries, while large govern- ment sizes hurt high-income countries, (iii) low levels of trade openness are suf�cient for �nance-growth nexus in high-income countries, but low-income countries need higher levels of trade openness for similar magnitudes of the �nance-growth nexus, (iv) catch-up effects through the �nance-growth nexus are higher for moderate per capita income levels. Financial development, Economic growth, Thresholds3Cross- country analysis JEL Classi�cation: E31, E44, F36, O16, O47 In a seminal study, Lucas (1985) argues that the bene�ts obtained by individ- uals from eliminating the whole macroeconomic instability in a given economy are almost certain to be negligibly small, when compared with those that can be obtained with more growth.1 Therefore, even the global �nancial crisis that has started at the end of 2007, considered to be the biggest one since the Great Depression by most economists, should not matter from a welfare analysis point of view, and countries, especially the developing ones, should still focus on the long-run growth. In this context, the impact of �nancial development on the long-run growth is of particular interest: A healthy �nancial system not only encourages savings, but also improves the allocation of such savings to ef�cient investment projects; this, in turn, encourages an ef�cient and high level of capital formation to promote growth. However, what are the necessary economic conditions and/or environments to achieve such a healthy �nance- growth nexus? Does high inflation lead �nancial depth to show its negative impacts on growth or does it only eliminate the positive effects? Is there any Hakan Yilmazkuday is an assistant professor in the Department of Economics at Florida International University, Miami, FL 33199; his e-mail address is skuday@gmail.com. The author thanks Elisabeth Sadoulet and two anonymous referees for their helpful comments and suggestions. The usual disclaimer applies. 1. See Imrohoroglu (2008) and the discussion therein. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 278– 295 doi:10.1093/wber/lhr011 Advance Access Publication May 18, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 278 Hakan Yilmazkuday 279 optimal level of trade openness or government size for the development of �nance-growth nexus in low-income and high-income countries? Who bene�ts most from the catch-up (convergence) effects through the �nance-growth nexus? Is the �nance-growth nexus stable through time? All these questions are sought to be answered here by investigating the historical experiences of 84 countries from 1965 to 2004 and considering the nonlinearities in the �nance- growth nexus through a continuous threshold analysis. The effect of inflation on growth is found to be negative, especially in the lit- erature on empirical growth. This is attributed to increasing uncertainties, mostly because of increasing relative price variability, increasing dif�culties in planning, or increasing expectations of disinflation (see Fischer, 1993, Barro, 1996, Temple, 2000, for various arguments and surveys of empirical litera- ture). While measuring such effects, Bruno and Easterly (1998) show that growth falls sharply when the inflation rate crosses the threshold of 40 percent per year. In the context of �nance-growth nexus, uncertainty due to high inflation can be through the flow of information about the investment projects and returns, used by intermediaries. Rousseau and Wachtel (2002) show that the impact of �nancial depth on growth disappears for inflation rates above 6.5 or 13.4 percent, depending on the �nancial-depth measure used. In the very same context, by using a slightly different method, this study �nds that high inflation crowds out the positive effects of �nancial depth on long-run growth; however, the threshold inflation rate estimated by this study is about 8 percent, independent of the �nancial-depth measure used. The government expenditure can promote growth through the provision of public goods, such as property rights, national defense, legal system, and police protection; however, large public expenditures would tend to crowd out poten- tially productive private investments. The empirical evidence is in line with this claim suggesting that the effects of government size on growth are mixed: Landau (1983) claims that the growth of government size hurts growth, while Kormendi and Meguire (1985) �nd no connection between government size and growth. Furthermore, Ram (1986) �nds that government size has positive effect on growth, while Levine and Renelt (1992) show that there is a fragile statistical relationship between growth and the growth of government size. Karras (1996) reports that there is an optimal government size, and, on an average, it is about 23 percent of the GDP. Demetriades and Rousseau (2010) contend that government expenditure has positive effect on �nancial development of countries that are in the midrange of economic development, and a strongly negative effect on the wealthiest countries, but little effect on poor countries. In the context of the �nance-growth nexus, this study shows that small government sizes hurt low-income countries (e.g., owing to the lack of suf�cient public goods, such as infrastructure or prop- erty rights, to have an effective �nancial system), while large government sizes hurt high-income countries (e.g., owing to the crowding-out effect described earlier); thus, the optimal government size, on an average, is found to be between 11 and 19 percent, which is lower than that suggested by Karras (1996). 280 THE WORLD BANK ECONOMIC REVIEW Trade openness can endorse growth through providing access to large and high-income markets, together with low-cost intermediate inputs and technol- ogies; however, it can also lead to more vulnerability through international shocks (either trade or �nance). Such effects of trade openness on growth have been studied extensively (see Yanikkaya, 2003, for a comprehensive survey). Although relatively recent works by Dollar (1992), Sachs and Warner (1995), Edwards (1998), Frankel and Romer (1999), and Dufrenot et al. (2010) assign an important role for trade openness in economic growth, considerable skepti- cism does exist about this relationship, as summarized by Rodriguez and Rodrik (2000). They show that low levels of trade openness are suf�cient for the �nance-growth nexus in high-income countries, because they already have their high-income (and mostly large) national markets and �nancial intermedi- aries who can help in this process. On the contrary, low-income countries need higher levels of trade openness for similar magnitudes of �nance-growth nexus, because they can bene�t from larger, high-technology and high-income markets only through high levels of openness. Starting with Gerschenkron (1952), the argument that low-income countries can grow faster than high-income countries has been studied extensively. According to Gerschenkron, the so-called "catch-up effect" is due to the low costs of industrialization in low-income countries through imitating already- developed technologies in high-income countries. Barro and Sala-i-Martin (1995) connect this story to the neoclassical theory of diminishing returns to physical capital, which should cause more advanced countries to grow more slowly than the less advanced countries. However, in empirical terms, the evi- dence is mixed: Besides many others, Baumol (1986) �nds evidence for the catch-up effect in some OECD countries, while DeLong (1988) could �nd no evidence in the historical data of over a century. In the context of �nance- growth nexus, using ad hoc measures of development, Rousseau and Yilmazkuday (2009) claim that �nancial depth has higher effects on low- income countries than on high-income countries. However, as �nancial devel- opment is costly and dif�cult, one would expect that catch-up effects would start manifesting only after the income crosses a certain threshold value. Considering all possible income levels, this study shows that the catch-up effect, through the �nance-growth nexus, does not start until a country reaches the threshold per capita income level of about $665 (in constant 1995 U.S. dollars), and that it would not work effectively until that income level reaches about $1,636 (in constant 1995 U.S. dollars). The �nance-growth nexus has been studied extensively, especially after the classic studies by Hildebrand (1864), Schumpeter (1911), and Sombart (1916, 1927), among others, who emphasized the proactive role of �nancial services in promoting growth and development. Goldsmith (1969), McKinnon (1973), and Shaw (1973) carried out theoretical studies stressing the connection between a country’s �nancial superstructure and its real infrastructure. While Goldsmith focuses on the effect of economy’s �nancial superstructure on the Hakan Yilmazkuday 281 acceleration of economic growth to the extent of relating economic perform- ance to migration of funds to the best projects available, McKinnon and Shaw emphasize that government restrictions, such as interest-rate ceilings, high reserve requirements, and directed credit programs encumber �nancial develop- ment and ultimately reduce growth. Similar conclusions were drawn by other economists who developed models of endogenous growth theories in which growth and �nancial structure are explicitly de�ned. In particular, the works by Durlauf et al. (2005), Levine (2005), and Khan et al. (2006) provide useful survey of literature on this aspect. Recent literature on empirical growth analysis, following Barro (1991) and Levine and Renelt (1992), focuses on growth equations, including a standard set of explanatory variables that provide robust and widely accepted proxies for growth determinants. King and Levine (1993) extend this empirical frame- work by including measures of �nancial development. Most of the recent studies have moved toward threshold analysis to capture possible nonlinearities in these growth equations. They split the cross-country data based on the countries’ �nancial development levels (e.g., low, intermediate, and high �nan- cial development of Rioja and Valev, 2004, and Rousseau and Wachtel, 2011, or deviations from optimal �nancial development as reported by Graff and Karmann, 2006), inflation rates (e.g., below or above optimal threshold inflation as reported by Fischer, 1993, Bruno and Easterly, 1998, Khan and Senhadji, 2001, Khan et al., 2006, and Rousseau and Yilmazkuday, 2009), or development status (e.g., ‘developed’ vs. ‘developing’ status as reported by Rousseau and Yilmazkuday, 2009, and Rousseau and Wachtel, 2011). The split-up was achieved mostly through discrete measures that may suppress the actual nonlinear relation between growth and other variables.2 An exception here is the study by Rousseau and Wachtel (2002), who use a rolling-regression framework by ordering the data according to 5-year inflation rate averages, which can be thought as a continuous (rather than a discrete) analysis. However, they could not obtain any information from rolling-regression by ranking countries according to other variables, such as the initial per capita income, openness, or government size, among many others. Another drawback of rolling-regression technique is that sequential regressions have different sample sizes: Rousseau and Wachtel (2002) used 50 observations to start with, and then added one observation at a time until the full sample was included. A potential problem with this technique was that the estimated coef�cients might not be comparable owing to the changes in the power of the estimation through the Law of large numbers. Another exception is the study by Rousseau and Wachtel (2011), who also employed the rolling-regression framework by ordering the data according to �nancial development of countries. Nevertheless, their study also lacked any information that can be obtained from rolling-regression by ranking countries according to other threshold 2. See Hansen (1999, 2000) for recent econometric techniques to determine discrete thresholds. 282 THE WORLD BANK ECONOMIC REVIEW variables mentioned earlier. Another drawback of the rolling-regression tech- nique in their study is that they used 20-country windows in each regression, which may be problematic owing to small sample size (i.e., the signi�cance of the coef�cient estimates may not be reliable because of the Law of large numbers). In contrast, this study considers the thresholds in several possible explanatory variables in the �nance-growth nexus through rolling-window two-stage least squares regressions with constant and large sample sizes, to capture all possible nonlinearities. Technically speaking, this approach general- izes the threshold frameworks used in earlier studies (mentioned above) to �gure out how nonlinear growth estimates and their signi�cance change if all the observations are ordered by a variable of interest (e.g., inflation, govern- ment size, trade openness, or initial per capita income). I I . D ATA A N D B A S E L I N E G R O W T H R E G R E S S I O N S The data set was constructed for 84 countries covering the period 1965–2004 as a panel of country observations from the World Bank’s World Development Indicators.3 The list of countries is given in the note under Table 1. Following Barro (1991) and Levine and Renelt (1992), the baseline growth equations included a standard set of explanatory variables that provide robust and widely accepted proxies for growth determinants. The dependent variable was the growth rate of real per capita output averaged over 5-year periods from 1965 to 2004. The regression analysis included standard explanatory variables, such as log initial per capita GDP, log initial secondary enrollment rate (SEC), the ratio of liquid liabilities (i.e., M3) to GDP, the ratio of M3 less M1 to GDP, inflation rate, openness, and government size. The log of initial per capita GDP for each 5-year period in constant 1995 U.S. dollars is expected to have a negative coef- �cient because of convergence (i.e., the tendency for countries with lower start- ing levels of GDP to “catch up� with countries of higher GDP). The log of the initial secondary school enrollment rate for each 5-year period (i.e., the percen- tage of the high school aged population actually enrolled) is expected to have a positive coef�cient to reflect a country’s commitment to the development of human capital; school enrollment rates are more widely available than other more precise measures of human capital. Two measures of �nancial sector depth, each averaged for individual 5-year periods, were used: (i) the ratio of liquid liabilities (i.e., M3) to GDP and (ii) the ratio of M3 2 M1 to GDP. The broad money supply M3 included all deposit-type assets and was presumed to relate to the extent and intensity of intermediary activity; M3 2 M1 took the pure transactions assets out of the ratio to reflect more closely the 3. Original raw data set covers the period 1960–2004. But, considering that the missing observations in all possible variables will have a consistent analysis across different model speci�cations, the data set was reduced to cover only the period of 1965–2004. T A B L E 1 . Descriptive Statistics, 1965–2004, 84 Countries Per capita income Per capita Initial Government Trade Inflation M3 M3-M1 Variable growth (%) initial GDP SEC (%) (% GDP) Openness (%) (% GDP) (% GDP) Mean 1.77 5961 50.33 14.69 60.80 15.25 45.48 28.49 Maximum 11.66 45888 146.32 40.59 212.49 351.97 184.03 156.44 Minimum 2 9.27 145 1.00 4.36 8.92 0.49 4.15 2 13.14 Standard deviation 2.67 8498 30.43 5.52 31.73 26.37 28.21 24.28 Coef�cient of variation 1.50 1.43 0.61 0.38 0.52 1.73 0.62 0.85 Correlations Per capita income growth (%) 1.00 Per capita initial GDP 0.11 1.00 Initial SEC (%) 0.16 0.75 1.00 Government (% GDP) 2 0.02 0.45 0.41 1.00 Trade openness 0.11 2 0.06 0.11 0.33 1.00 Inflation (%) 2 0.18 2 0.12 2 0.06 2 0.04 2 0.13 1.00 M3 (% GDP) 0.25 0.50 0.55 0.33 0.29 2 0.16 1.00 M3-M1 (% GDP) 0.26 0.49 0.54 0.26 0.22 2 0.12 0.83 1.00 Note: The list of 84 countries is as follows: Algeria, Argentina, Australia, Austria, Bangladesh, Barbados, Belgium, Bolivia, Brazil, Cameroon, Canada, Central African Republic, Chile, Colombia, Costa Rica, Cote d’Ivoire, Denmark, Dominican Republic, Ecuador, Egypt, Arab Rep., El Salvador, Fiji, Finland, France, Gambia, The, Ghana, Greece, Guatemala, Guyana, Haiti, Honduras, Iceland, India, Indonesia, Iran, Islamic Rep., Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kenya, Korea, Rep., Lesotho, Luxembourg, Malawi, Malaysia, Malta, Mauritius, Mexico, Morocco, Nepal, Netherlands, New Zealand, Nicaragua, Niger, Nigeria, Norway, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Portugal, Rwanda, Senegal, Sierra Leone, South Africa, Spain, Sri Lanka, Sudan, Sweden, Switzerland, Syrian Arab Republic, Thailand, Togo, Trinidad and Tobago, Turkey, United Kingdom, United States, Uruguay, Venezuela, RB, Zimbabwe. Source: Author’s analysis based on data sources discussed in the text. Hakan Yilmazkuday 283 284 THE WORLD BANK ECONOMIC REVIEW intermediation activities of the depository institutions. The inflation rate was measured as the average annual growth rate of the consumer price index (CPI) in each 5-year period, where deflationary episodes were �ltered. This allowed explicit examination of the direct effects of price inflation on growth, and a negative coef�cient is expected. The total government expenditure, in terms of the percentage of GDP, and the international trade openness averaged for each 5-year period served as additional control variables. Although the role of gov- ernment expenditure is weak, large public expenditures would tend to crowd out potentially more productive private investments, especially in higher- income countries. To control for any country-size and income-level effects on openness, international trade openness was measured as residuals from a regression of international trade (the sum of exports and imports) as a percen- tage of GDP on country size (measured by log GDP) and income level (measured by log per capita GDP). To control for scale effects in the interpret- ation of the empirical results, minimum international trade openness (measured by residuals) was scaled up to the minimum value of the international trade as a percentage of GDP, because that minimum value was least affected by the country size and income level. In a growth regression, this adjustment will have no effect on the coef�cient estimates, because it will be captured by the inter- cept. This adjusted trade openness is expected to have a positive effect on growth. The descriptive statistics of the data set (averaged over 5-year periods from 1965 to 2004) are provided in Table 1. It is evident from these statistics that the annual per capita income growth rates ranged between 2 9 and 12 percent, the per capita initial GDP levels between $145 and $46,000, the initial SEC between 1 and 146 percent, the government expenditure between 4 and 41 percent, the adjusted trade openness between 9 and 212 percent, the inflation rate between 0 and 352 percent, M3 (% of GDP) between 4 and 184 percent, and M3 2 M1 (% of GDP) between 2 13 and 156 percent. These wide ranges warrant a threshold analysis per se. The coef�cients of variation (a normalized measure of dispersion of a probability distribution, calculated as the standard error over the mean) show that the dispersions of per capita income growth, per capita initial GDP, and inflation rate are high across the countries, while those of the government expenditure and trade are low. Therefore, one might expect to have relatively higher threshold effects from per capita income growth, per capita initial GDP, and inflation rate. The correlations across vari- ables are also depicted in the lower part of Table 1. The expected signs of cor- relation coef�cients between growth and explanatory variables are consistent with the foregoing discussion. Almost all variables are positively correlated with each other, except for inflation, which is negatively correlated with all the variables, implying possible distortionary effects of positive price changes in all the transmission channels in the economy. Estimation was carried out by instrumental variables (i.e., two-stage least squares) with initial values of �nancial depth, inflation, government Hakan Yilmazkuday 285 expenditure, and trade for each 5-year period serving as instruments in the �rst stage. Fixed effects for the 5-year periods were also included, because global business cycle conditions often involved shocks with common growth effects across the countries.4 Table 2 presents the results that replicate the linear regression analysis of Rousseau and Yilmazkuday (2009); the only difference is that this study has employed an adjusted trade openness measure for openness, as described earlier. Column 1 contains the baseline growth model where the coef�cient for initial GDP is negative and is thus consistent with the theory of conditional convergence but is not statistically signi�cant, while the coef�cient on the initial SEC is positive and signi�cant at 1-percent level. As the baseline speci�cation is expanded in the remaining columns of the Table, the coef�cient on the initial GDP remains negative throughout and is statistically signi�cant in 6 of the 12 regressions. The initial secondary enrollment retains its positive and statistically signi�cant coef�cient throughout. Column 2 of Table 2 includes trade openness and government expenditure as controls to form an extended baseline. Openness is positively and signi�- cantly related to growth in this speci�cation and all others in which it appears, while the coef�cients on government expenditure are negative and statistically signi�cant throughout. These �ndings are consistent with the priors for these controls. When inflation is included to the baseline model in Column 3 and to the extended baseline in Column 4, the coef�cients on inflation become negative, but statistically signi�cant at the 5-percent level only in Column 4; this �nding is consistent with that of earlier studies. When any of the two �nancial vari- ables are included to the baseline and extended baseline in Columns 5 –8, both the measures become positively and signi�cantly related to the growth at 1-percent level.5 Finally, when both �nancial depth and inflation are included in the remaining columns of Table 2, although the effects of the �nancial vari- ables remain, the statistical signi�cance of the inflation coef�cients falls to 10-percent level without additional controls (Columns 9 and 11) and when the full conditioning set was included (Columns 10 and 12), the inflation coef�- cients are no longer signi�cant. The dampening of the effect of log initial GDP and inflation on growth, combined with �nancial development, calls for an explanation. Why does the effect of log initial GDP disappear when it is combined with log initial SEC, 4. For robustness, country-�xed effects were also included in the regressions, but the results were not at all affected by this inclusion; the only effect was on the explanatory power of the regression, which shifted up when the country-�xed effects were included. These results of additional sensitivity analysis can be obtained using the published Matlab codes. 5. The ratio of total domestic credit to GDP was also experimented as a measure of �nancial development that would bring non-depository intermediaries into the analysis; however, it was found that this variable was not statistically signi�cant in any of the speci�cations. This echoes the results (i.e., covering the period 1960 to 2004) recently obtained by Rousseau and Wachtel (2011). Therefore, the analysis is limited to the two �nancial measures as described earlier. 286 T A B L E 2 . Instrumental variables growth regressions, 1965–2004, 84 Countries Dependent Variable: Growth of Per Capita Income (%) Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Log of initial GDP 2 0.124 2 0.023 2 0.154 2 0.031 2 0.277* 2 0.143 2 0.299* 2 0.169 2 0.284* 2 0.153 2 0.309** 2 0.181 (0.107) (0.119) (0.106) (0.119) (0.110) (0.121) (0.110) (0.123) (0.111) (0.121) (0.110) (0.123) Log of initial SEC (%) 1.101** 1.033** 1.154** 1.082** 0.912** 0.883** 0.923** 0.896** 0.951** 0.917** 0.971** 0.940** (0.251) (0.212) (0.252) (0.213) (0.248) (0.211) (0.252) (0.210) (0.250) (0.215) (0.253) (0.213) Government (% GDP) 2 0.062** 2 0.061* 2 0.068** 2 0.059* 2 0.067** 2 0.058* (0.025) (0.024) (0.024) (0.024) (0.024) (0.024) Trade openness 0.014** 0.013** 0.010** 0.011** 0.009** 0.010** THE WORLD BANK ECONOMIC REVIEW (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Inflation (%) 2 0.012* 2 0.010 2 0.006† 2 0.005 2 0.008† 2 0.007 (0.005) (0.007) (0.004) (0.007) (0.004) (0.007) M3 (% GDP) 0.023** 0.021** 0.022** 0.020** (0.005) (0.005) (0.005) (0.005) M3 2 M1 (% GDP) 0.029** 0.025** 0.027** 0.024** (0.006) (0.006) (0.005) (0.006) R-bar sqd. 0.18 0.21 0.20 0.23 0.22 0.24 0.22 0.24 0.23 0.25 0.24 0.25 Note: †, * and ** indicate signi�cance at the 10 percent, 5 percent, and 1 percent levels, respectively. Standard errors are in parentheses. Growth rates are �ve-year averages. Estimation is by two-stage least squares. The initial values of government, trade, inflation, M3, and M3-M1 in each �ve-year period are used as instruments for the corresponding �ve-year averages. All equations include �xed effects for time periods that are not shown. The sample size in each equation is 485. Source: Author’s analysis based on data sources discussed in the text. Hakan Yilmazkuday 287 government expenditure, or trade? Is the direct effect of inflation on growth as important as the one suggested by the regressions in Columns 3 and 4 of Table 1? Or, does inflation inhibit growth primarily through its effects on the smooth operation of the �nancial sector, as indicated by the regressions in Columns 9–12 of Table 2? Is there a continuum of combinations of inflation rates and levels of �nancial development that are associated with a given rate of growth? If a continuum exists, linear regression analysis seems ineffective in showing it clearly, especially given the negative correlation between inflation and �nancial depth ( 2 0.16 for M3 and 2 0.12 for M3 2 M1 in Table 1); nevertheless, a continuous threshold analysis can shed more light on under- standing nonlinearities in growth regressions. I I I . T E C H N I C A L A N A LY S I S For a continuous threshold analysis, rolling-window two-stage least squares regressions were employed with a constant window size of 120 after ordering the data according to the threshold variable. For instance, if the inflation thresholds were of interest, all the observations (i.e., the pooled sample of 5-year average data from all the countries) were sorted in the order of the lowest to the highest inflation rates; the �rst regression was run with the �rst 120 observations of the sorted data set, the second regression by moving the 120 window toward higher inflation rates by one observation, and so on. The selection of a constant window size was important for comparison of coef�- cient estimates across the windows, while the selection of a window size of 120 was important to ensure a fair distribution across the power of the regressions and the degree of nonlinearity. Nevertheless, the results of this study are robust to the selection of the window size; the results obtained under different poss- ible window sizes are almost the same as those that will be discussed below.6 For a consistent inference across linear and nonlinear frameworks, the rolling- window regressions used the speci�cations in Columns 10 and 12 of Table 2, depending on the �nancial-depth measure used. The corresponding results are given in Figs 1–2, where the x-axes show the median of the threshold variable in 120 sample windows (i.e., the variable according to which all the observations have been sorted). The y-axes of the �gures in the left panel of Figs 1 –2 show the coef�cient estimates of the �nance variable (either M3 or M3 2 M1 as a percentage of GDP). The bold solid lines show the coef�cient estimates and the dashed lines the 10-percent con�dence intervals. For the sake of robustness, Fig. 1 considers the �nance variable of M3 as a percentage of GDP, and Fig. 2 M3 2 M1. The results are similar in terms of the signi�cance of the estimated parameters, but slightly different in terms of the coef�cient estimates. 6. Although these sensitivity analyses were skipped to save space, they can easily be obtained using the published Matlab codes. 288 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Thresholds (with M3 as the �nance variable) Note: The dashed lines in the �gures of left panel show the 10 percent con�dence intervals, while the dashed lines in the �gures of right panel show the mean of R-bar squared values. Source: Author’s analysis based on data sources discussed in the text. The top rows of Figs 1 –2 replicate the inflation-threshold analysis of Rousseau and Wachtel (2002), this time by using a rolling-window regression with a constant window size of 120 (Rousseau and Wachtel [2002] used a rolling-regression analysis, where they started with 50 observations and included one more observation for each additional estimation). The purpose of this exercise is to investigate the effects of inflation on the �nance-growth Hakan Yilmazkuday 289 F I G U R E 2. Thresholds (with M3-M1 as the �nance variable) Note: The dashed lines in the �gures of left panel show the 10 percent con�dence intervals, while the dashed lines in the �gures of right panel show the mean of R-bar squared values. Source: Author’s analysis based on data sources discussed in the text. nexus. It is evident that the coef�cient estimates of �nancial depth are signi�- cant only when the inflation rate is below approximately 8 percent, indepen- dent of the �nancial-depth measure used. Financial depth appears to need a reasonably low inflation environment to promote long-run growth effectively; otherwise, as shown in Figs 1–2, the �nancial-depth effects on growth approach zero as inflation increases. This threshold value is in line with the 290 THE WORLD BANK ECONOMIC REVIEW values suggested by Rousseau and Wachtel (2002), which vary between 6.5 and 13.4 percent, depending on the �nancial-depth measure used. Within this picture, the signi�cance may not be an indisputable guidance, because with a high number of observations in a panel framework and a large number of regressions, the signi�cance at conventional levels may imply 5 or 10 percent of type-1 errors (rejecting the null when it should be maintained). At the same time, failure to meet conventional signi�cance levels does not imply the cer- tainty that the null is true (type-2 error).7 Therefore, it is also worth focusing on the coef�cient estimates without considering their signi�cance. The coef�- cient estimates of �nancial depth are non-negative for almost any level of inflation. This contrasts with the result reported by Rousseau and Wachtel (2002), who show that �nancial depth has a negative coef�cient estimate for inflation rates above 13.4 or 15.9 percent, depending on the �nancial-depth measure used. In sum, according to the present study, the worst-case scenario with high inflation rates is to have an ineffective �nancial-depth effect on the long-run growth. Besides Rousseau and Wachtel (2002), this study considered the thresholds in variables other than inflation. First, the second rows of Figs 1–2 analyze the effects of government-expenditure thresholds on the �nance-growth nexus. Independent of the �nancial-depth measure used, the coef�cient estimates of the �nancial depth were found to be signi�cant for countries with government expenditures of approximately between 11 and 19 percent of their GDPs. Although the �nance-coef�cient estimates were found to be non-negative for almost any government size, consistent with their signi�cance levels, their effects on growth were found to be lower for government sizes lower than 11 percent or higher than 19 percent. Thus, the historical cross-country data shows that the government size must be optimal for signi�cant and positive effects of �nancial depth on the long-run growth. This may be in line with the expectations of stimulus effects of government expenditures on low-income countries in promoting productive private investments through the �nancial system, and distortionary effects of government expenditures in high-income countries through crowding out potentially more productive private invest- ments. To test this claim, the income levels of countries were checked with gov- ernment sizes below 11 percent, between 11 and 19 percent, and above 19 percent. The results supported the claim by showing that, on an average, countries with government sizes below 11 percent have per capita income levels of about $1,053, those between 11 and 19 percent about $2,148, and those above 19 percent about $6,628. Second, the third rows of Figs 1 –2 analyze the effects of trade-openness thresholds on the �nance-growth nexus. It is evident that the coef�cient esti- mates of �nancial depth become signi�cant for countries that have adjusted 7. The author would like to thank the anonymous referee for pointing out this issue for the signi�cance of the parameters. Hakan Yilmazkuday 291 trade openness lower than about 35 percent or higher than about 75 percent of their GDPs. Coef�cient estimates of �nancial depth are found to be non- negative for almost any degree of trade openness. Therefore, according to his- torical cross-country data, there is also evidence that optimal trade openness has signi�cant and positive effects of �nancial depth on the long-run growth. While investigating the possible economic reasons behind this result, it was observed that countries with trade openness between 35 and 75 percent have an average per capita income level of about $1,481, those with lower than 35 percent openness about $2,945 and those with higher than 75 percent openness about $2,143. This suggests that higher-income countries bene�t from �nancial depth mostly through their national markets. However, for lower-income countries, the story is different: to bene�t from �nancial depth, lower-income countries should have a trade openness of at least about 75 percent, failing which the effect of �nancial depth on the long-run growth is almost none. This can be linked to the market shares of the countries: if a country has access to high-income markets, through their national markets or international trade, then �nancial depth helps growth; otherwise, �nancial depth remains ineffec- tive, and the country suffers from a disconnected �nance-growth nexus. Third, the fourth rows of Figs 1–2 extend the analyses carried out by Rousseau and Yilmazkuday (2009) and Rousseau and Wachtel (2011) by con- sidering the continuous log per capita initial GDP thresholds. They both con- sidered ad hoc splits of countries in terms of their developments, simply by setting a threshold of a certain amount of per capita real income (e.g., countries with per capita income of less than US $3,000 a year in 1995 are ‘developing’ and those with higher income ‘developed’). This approach sup- presses changes in the development of countries through time, because the development of a country is measured based on its performance during the base year; moreover, there may be many other categories of development in a more continuous sense. The present study used a robust measure of develop- ment, namely the per capita initial income for the time period considered in the pooled sample. It is evident that the effect of �nancial development on growth was signi�cant for countries with a per capita income of more than about $665 ( ¼ exp[6.5]). Although the signi�cant effect of �nancial depth on growth was found to increase until the per capita income reached about $1,636 ( ¼ exp[7.4]), it started decreasing above this level. Yet, the coef�cient estimates of �nancial depth were non-negative for almost any degree of initial GDP level. The decreasing effects of �nancial depth on growth for countries with per capita income levels higher than $1,636 were consistent with the catch-up effect, which suggest that low-income countries have the potential to grow at a faster rate than high-income countries, because the diminishing returns (in particular, to physical capital) are not as strong as those in countries with high levels of capital. Furthermore, low-income countries can replicate production methods, technologies, and institutions currently used in developed countries, and combine them with their cheap labor opportunities. Therefore, 292 THE WORLD BANK ECONOMIC REVIEW on an average, as the income of the countries increases, the effect of �nancial depth on growth goes down. Finally, the results for time thresholds are depicted at the bottom of Figs 1 –2, where 120-sample-size windows were not used; instead, the 5-year periods were used as thresholds through which 5-year averages were taken. The y-axes of the �gures in the left panels of Figs 1 –2 again show the coef�cient estimates of �nancial depth, while the x-axes show the median of each 5-year period con- sidered. It is evident that, consistent with the �ndings of Rousseau and Wachtel (2011), the effects of �nancial development on growth are decreasing through time. Nevertheless, �nancial-depth effects on growth are found to be positive at almost all times. The economic reasoning behind this can be the diminishing returns to capital: as the countries get richer through time, because of diminish- ing returns, �nancial depth becomes less effective on growth. However, as all the countries were employed for each 5-year period, this may also reflect the effect of �nancial depth on growth for a subset of countries. Thus, according to the foregoing discussions, this does not rule out the scope for future �nance-growth nexus. Although the coef�cient estimates in the left panels of Figs 1–2 depict the �nance-growth nexus with the thresholds in inflation, government size, trade openness, and initial GDP, they do not provide any information on the relative importance of these thresholds. In particular, as each of these thresholds can substitute for the other, it is not clear which threshold is more important stat- istically. To answer this question, explanatory powers of the rolling regressions with each threshold (in terms of R-bar squared values) are provided in the right panels of Figs 1 –2. While the y-axes of the �gures in the right panel show the R-bar squared values, the dashed lines show the mean R-bar squared values. Using M3 (M3 2 M1) as the percentage of GDP as the �nance variable, the mean R-bar squared values are 0.18 (0.18), 0.25 (0.25), 0.15 (0.16), 0.29 (0.29), and 0.15 (0.16) for the threshold variables of inflation, government size, trade openness, initial GDP, and time, respectively. Hence, statistically, the initial GDP seems to be the most important threshold variable. This recon- �rms the importance of the catch-up effects on the �nance-growth nexus that start only after a country reaches a particular threshold value of income. I V. C O N C L U D I N G R E M A R K S A N D D I S C U S S I O N This research paper has generalized the empirical studies on the �nance-growth nexus by considering the thresholds in several explanatory variables. Following are the suggestions that emerged from this study: (i) Inflation rates above 8 percent eliminate the positive effects of �nancial depth on the long-run growth. (ii) Optimal government size (% GDP) for the �nance-growth nexus is between 11 and 19 percent; government sizes below 11 percent hurt the low-income countries, and those above 19 percent hurt the high-income countries. (iii) Optimal trade openness for the �nance-growth nexus is below about 35 Hakan Yilmazkuday 293 percent for high-income countries, and above about 75 percent for low-income countries. (iv) The catch-up effect through �nance-growth nexus starts when a country passes the threshold per capita income level of about $665; it has its highest impact when the per capita income is about $1,636; its impact decreases as the per capita income increases. (v) There is evidence to show that �nancial-depth effects on growth decrease through time. (vi) The thresholds in the initial per capita income seem to be more important than other thresholds. However, this study is not without caveats. First, the �nancial-depth measures used here may not fully reflect the actual �nancial development, especially during crisis periods with high inflation rates or low levels of per capita income, although the averaging of the variables across 5-year periods amended some of these extreme cases. Second, despite strong evidence in favor of the nexus between �nance and growth, the exact causality between �nance and growth is still a subject of debate (see Demetriades and Hussein, 1996; Arestis and Demetriades, 1997; Andrianova and Demetriades, 2004, 2008); therefore, the results and policy implications of this paper should be quali�ed with respect to the certain causality assumptions and the estimation method- ology employed. Third, the results reflect mostly the average historical experi- ences of the countries in the sample, rather than providing strong policy implications for future development of a country. Overall, in line with the sug- gestions of Durlauf and Johnson (1995), Liu and Stengos (1999), and Durlauf (2001), this study offers one important message: The typical cross-country growth regressions are inadequate, because the �nance-growth nexus is shown to be nonlinear. Although this study has focused mostly on the �nance-growth nexus, it can easily be extended to investigate other determinants/channels of growth, such as institutions, debt, law, private investment, foreign direct invest- ment, inequality, volatility/uncertainty, culture, beliefs, �nancial aid, central bank independence, and so on. And, these could be the possible topics for future research. REFERENCES Andrianova, S., and P. Demetriades. 2004. “Finance and Growth: What We Know and What We Need to Know�, (with), in C.A.E. Goodhart (ed.), Financial Development and Growth: Explaining the Links, Palgrave Macmillan, 38–65. ———. 2008. “Sources and Effectiveness of Financial Development: What We Know and What We Need to Know�, in B. Guha-Khasnobis, and G. Mavrotas (eds.), Financial Development, Institutions, Growth and Poverty Reduction, Studies in Development Economics and Policy Series, Palgrave Macmillan, 10 –34. Arestis, P., and P. Demetriades. 1997. “Financial Development and Economic Growth: Assessing the Evidence�, The Economic Journal, 107: 783–799 Barro, R.J. 1991. “Economic growth in a cross section of countries.� Quarterly Journal of Economics 106: 407–443. ———. 1996. “Inflation and growth.� Review, Federal Reserve Bank of St. Louis 78: 153– 169. Barro, R.J., and X. Sala-i-Martin. 1995. Economic Growth, McGraw-Hill, Inc. U.S.A. 294 THE WORLD BANK ECONOMIC REVIEW Baumol, W. 1986. “Productivity growth, convergence, and welfare.� American Economic Review 76: 1072– 1085. Bruno, M., and W. Easterly. 1998. “Inflation crises and long-run growth.� Journal of Monetary Economics 41: 3 –26. Chen, S-T., and C-C. Lee. 2005. “Government size and economic growth in Taiwan: A threshold regression approach.� Journal of Policy Modeling 27 (9): 1051–1066. DeLong, J.B. 1988. “Productivity growth, convergence and welfare: Comment. American Economic Review.� 78: 1138– 1154. Demetriades, P., and P.L. Rousseau. 2010. “Government, Openness and Finance: Past and Present�, NBER Working Paper 16462. Dollar, D. 1992. “Outward-oriented developing countries really do grow more rapidly: evidence from 95 LDCs, 1976– 85.� Economic Development and Cultural Change 40(3): 523–44. Dufrenot, G., V. Mignon, and C. Tsangarides. 2010. “The trade-growth nexus in the developing countries: a quantile regression approach.� Review of World Economics 146 (4): 731–761. Durlauf, S. 2001. “Manifesto for a growth econometrics.� Journal of Econometrics 100: 65– 69. Durlauf, S., and J. Johnson. 1995. “Multiple regimes and cross-country growth behavior.� Journal of Applied Econometrics 10: 365 –384. Durlauf, S., P.A. Johnson, and J. Temple. 2005. “Growth Econometrics.� Handbook of Economic Growth, in: Philippe Aghion, and Steven Durlauf (ed.), Handbook of Economic Growth, edition 1,volume 1, chapter 8, pages 555–677. Edwards, S. 1998. “Openness, productivity and growth: what do we really know?� Economic Journal 108: 383– 398. El Khoury, A.C., and A. Savvides. 2006. “Openness in services trade and economic growth.� Economics Letters 92(2): 277–283. Fisher, S. 1993. “The Role of Macroeconomic Factors in Growth.� Journal of Monetary Economics 32: 485 –512. Frankel, J.A., and D. Romer. 1999. “Does trade cause growth?� American Economic Review 89 (3): 379 –399. Gerschenkron, A. 1952. “Economic backwardness in historical perspective.� In: B. Hoselitz, Editor, The progress of underdeveloped areas, University of Chicago Press, Chicago, pp. 5– 30. Goldsmith, R. 1969. Financial Structure and Development, Yale University Press, New Haven, CT. Graff, M., and A. Karmann. 2006. “What Determines the Finance-growth Nexus? Empirical Evidence for Threshold Models.� Journal of Economics 87(2): 127–157. Hansen, B. 1999. “Threshold Effects in Non-dynamic Panels: Estimation, Testing and Inference.� Journal of Econometrics 81: 594– 607. ———. 2000. “Sample Splitting and Threshold Estimation.� Econometrica 68: 575–603. ¨ cher fu Hildebrand, B. 1864. “Natural-, Geld- und Kreditwirtschaft.� Jahrbu ¨ konomie und ¨ r Nationalo Statistik 2: 1–24. Imrohoroglu, A. 2008. “Welfare Costs of Business Cycles.� The New Palgrave Dictionary of Economics, Second Edition. Kalaitzidakis, P., T.P. Mamuneas, A. Savvides, and T. Stengos. 2001. “Measures of Human Capital and Nonlinearities in Economic Growth.� Journal of Economic Growth 6 (3): 229–54. Karras, G. 1996. “The optimal government size: further international evidence on the productivity of government services.� Economic Inquiry 34 (2): 193–203. Khan, M.S., and A.S. Senhadji. 2001. “Threshold effects in the relationship between inflation and growth.� IMF Staff Papers 48, 1– 21. Khan, M.S., A.S. Senhadji, and B.D. Smith. 2006 “Inflation and Financial Depth.� Macroeconomic Dynamics 10 (2): 165– 182. Hakan Yilmazkuday 295 King, R.G., and R. Levine. 1993. “Finance and growth: Schumpeter might be right.� Quarterly Journal of Economics 108 (3): 717 –737. Kormendi, R.C., and P.G. Meguire. 1985. “Macroeconomic determinants of growth.� Journal of Monetary Economics 16: 141–163. Landau, D. 1983. “Government expenditure and economic growth: A cross-country study.� Southern Economic Journal 49: 783–792. Levine, R. 2005. “Finance and growth: theory and evidence.� In P. Aghion, and S. N. Durlauf, eds., Handbook of Economic Growth, Volume 1A. Elsevier North Holland, Amsterdam, 865– 934. Levine, R., and D. Renelt. 1992. “A sensitivity analysis of cross-country growth regressions.� American Economic Review 82: 942– 963. Liu, Z., and T. Stengos. 1999. “Non-linearities in cross-country growth regressions: a semiparametric approach.� Journal of Applied Econometrics 14: 537–538. Lucas, R. 1985. Models of Business Cycles. Yrjo Jahnsson Lectures. Oxford: Basil Blackwell. McKinnon, R. 1973. Money and Capital in Economic Development, The Brookings Institution, Washington, DC. Ram, R. 1986. “Government size and economic growth: A new framework and some evidence from cross-section and time-series data.� American Economic Review 76: 191– 203. Rioja, F., and N. Valev. 2004. “Does One Size Fit All? A Reexamination of the Finance and Growth Relationship.� Journal of Development Economics 74: 429–47. Rodriguez, F., and D. Rodrik. 2000. “Trade policy and economic growth: a skeptic’s guide to the cross- national evidence�, in (B. Bernanke, and K. Rogoff, eds.), Macroeconomics Annual 2000, Cambridge MA: MIT Press for NBER Rousseau, P.L., and P. Wachtel. 2002. “Inflation Thresholds and the Finance-Growth Nexus.� Journal of International Money and Finance 21: 777–793. ———. 2011. “What is happening to the impact of �nancial deepening on economic growth?� Economic Inquiry 49 (1): 276– 288. Rousseau, P.L., and H. Yilmazkuday. 2009. “Inflation, Financial Development and Growth: A Trilateral Analysis.� Economic Systems 33 (4): 310– 324. Sachs, J.D., and A. Warner. 1995. “Economic reform and the process of global integration’.� Brookings Papers on Economic Activity 1: 1 –118. Schumpeter, J.A. 1911. The Theory of Economic Development, Harvard university Press, Cambridge, MA. Shaw, E.S. 1973. Financial Deepening in Economic Development, Oxford Univ. Press, New York. Sombart, W. 1916. Der moderne Kapitalismus, Band I: Die Vorkapitalistische Wirtschaft; Band II: Das ¨ ische Wirtschaftsleben im Zeitalter des Fru europa ¨ nchen: Duncker and ¨ hkapitalismus, 2. Aufl., Mu Humblot. ———. 1927. Der moderne Kapitalismus, Band III: Das Wirtschaftsleben im Zeitalter des ¨ nchen: Duncker and Humblot. Hochkapitalismus, Mu Temple, J. 2000. “Inflation and growth: stories short and tall.� Journal of Economic Surveys 14: 395– 426. Yanikkaya, H. 2003. “Trade openness and economic growth: a cross-country empirical investigation.� Journal of Development Economics 72 (1): 57– 89. The Value of Vocational Education: High School Type and Labor Market Outcomes in Indonesia David Newhouse and Daniel Suryadarma This paper examines the relationship between the type of senior high school attended by Indonesian youth and their subsequent labor market outcomes. This topic is timely in light of a recent policy shift that aims to dramatically expand vocational education. The analysis controls for an unusually rich set of predetermined characteristics, and exploits longitudinal data spanning fourteen years to separately identify cohort and age effects. There are four main �ndings. First, the estimated wage premium for voca- tional graduates, relative to general graduates, is greater for women than men. Second, the returns to public vocational school for men have plummeted for the most recent cohort, and male vocational graduates now face a large wage penalty. Third, the generally favorable outcomes of public school graduates can be partly explained by non-random sorting of students with higher test scores and better-educated parents into public schools. Finally, these peer effects appear to be particularly important for students with above-average test scores, as men with high scores earn a surprisingly small premium from graduating from vocational or private general school. These small returns for high-scoring men, as well as the dramatic fall in the earnings premium for all male vocational graduates, raise important concerns about the current expansion of public vocational education and the relevance of the male vocational curriculum in an increasingly service-oriented economy. JEL Classi�cations: I21, J24, O15 Expanding access to vocational education can be an attractive option for pol- icymakers in developing countries seeking to improve labor market outcomes. For example, Tanzania prioritized vocational education in the late 1960s David Newhouse (corresponding author) is a labor economist in the Social Protection and Labor unit at the World Bank; his email address is dnewhouse@worldbank.org. Daniel Suryadarma is a research fellow at the Arndt-Corden Department of Economics, Australian National University; his email address is daniel.suryadarma@anu.edu.au. This work was supported by a grant from the Netherlands Ministry of Development Cooperation. The views expressed here are personal and do not implicate the World Bank, its management, or executive board. This study is a background paper for the Indonesia Jobs Report. We thank Vivi Alatas and Andrew Leigh for their support and encouragement, and Dandan Chen, Wendy Cunningham, Eric Edmonds, Edgar Janz, Daan Pattinasarany, and seminar participants at World Bank Of�ce Jakarta and the Institute for the Study of Labor for helpful comments. All remaining errors are our own. A supplemental appendix to this article is available at http://wber.oxfordjournals.org. THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 296– 322 doi:10.1093/wber/lhr010 Advance Access Publication May 18, 2011 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 296 Newhouse and Suryadarma 297 (Kahyarara and Teal, 2008), and South Korea followed suit thirty years later, both in response to a perceived shortage of skilled workers. In both cases, the expansion policy failed, as parents continued to prefer general to vocational education and refused to send their children to vocational schools (KRIVET, 2008).1 The Korean and Tanzanian experiences have not deterred the Indonesian Ministry of Education from enthusiastically embracing vocational education. The government, aiming to reduce high unemployment rates among educated youth, pledged to reverse the current ratio of high school graduates, from 70 percent general to 70 percent vocational, by 2015 (Ministry of National Education, 2006). Although this target is likely infeasible, the ministry has frozen the construction of new public general high schools and converted selected general schools to vocational schools, despite scant evidence that voca- tional education improves labor market outcomes. Worldwide, empirical evidence on the merits of vocational education is mixed. Vocational graduates earn a wage premium in Egypt (El-Hamidi, 2006), Israel (Neuman and Ziderman, 1991), and Thailand (Moenjak and Worswick, 2003). In contrast, general graduates earn a higher wage in Suriname (Horowitz and Schenzler, 1999) and, for students that continue on to university, in Tanzania (Kahyarara and Teal, 2008). Finally, Lechner (2000), KRIVET (2008), and Malamud and Pop-Eleches (2008) �nd no signi�cant differences in labor market outcomes between the two educational tracks in East Germany, South Korea, and Romania respectively. There is one study that we know of that examines the outcomes of voca- tion high school graduates in Indonesia (Chen, 2009). This study follows a single cohort of students three years after graduation and �nds that vocational school graduates, compared with general school graduates, experience similar wage and employment outcomes. Unfortunately, this study suffers from several limitations. First, the sample is restricted to recent high school gradu- ates aged 18 to 21, and therefore only measures very short-run impacts. In addition, two thirds of this young sample is not working, and the econo- metric technique used to correct for this relies on dubious assumptions.2 Because of the small sample size, the estimated effects of vocational education are insuf�ciently precise to rule out large returns.3 Finally, the analysis does not distinguish between men and women, despite important gender differ- ences in both the nature of the vocational education curriculum and labor market participation. 1. Some studies use the term academic education. In this paper, we use the term general education. 2. The Heckman selection equation is identi�ed by excluding parental education, previous household income, and junior high test score from the earnings equation. 3. In the OLS estimates, the 95 percent con�dence interval ranges from 0 to 60 percent of average earnings, while in the IV estimates, the 95 percent con�dence interval ranges from -50 to 150 percent of average earnings. 298 THE WORLD BANK ECONOMIC REVIEW The mixed conclusions of past studies have contributed to a contentious debate on the validity of standard regression estimates, given that selection of students into vocational and general tracks is not random. Attributes that could influence whether a student chooses one track over the other include scholastic ability, parental education, and location of residence. Failure to control for these variables may confound estimates of the returns to vocational education. In developing countries, access to data on these attributes is rare. Although many studies attempt to correct for non-random selection into work, we know of only two studies that address the role of unobserved determinants of school type.4 In this paper, we use a rich longitudinal household survey from Indonesia to evaluate the outcome of vocational high school graduates relative to general school graduates along four dimensions: earnings, labor market participation, risk of unemployment, and job quality. This study does not directly address potential bias due to omitted unobserved characteristics that may confound the estimates. Nevertheless, the data contain a rich set of control variables that allow us to control for non-random selection more carefully than the vast majority of existing studies.5 The set of control variables include the district where a person graduated from junior high school, whether they lived in a city, town, or village at age twelve, grade repetition and outside employment during elementary and junior high school, adult height, and the level of parental edu- cation. Junior high exit exam scores are not included as a control variable, because they are only available for the youngest cohort. Evidence from this cohort indicates that the omission of test scores has minor effects on the esti- mated effects of school type. Our paper makes three main contributions to the literature. The �rst is dis- tinguishing between public and private schools when assessing vocational edu- cation. While there has been a resurgence of interest in the ef�cacy of public versus private schooling in developing countries, this is the �rst research to our knowledge that explicitly distinguishes between public and private voca- tional education at the high school level.6 The second main contribution is 4. The only study that uses a plausibly exogenous source of variation in vocational school attendance is Malamud and Pop-Eleches (2008), which employs a regression discontinuity design to evaluate a 1973 policy that promoted general education in Romania. Chen (2009) uses the proportion of schools reported by village households that are vocational as an instrument for school type. Other studies control for observables (Kahyarara and Teal (2008), and Lechner (2000)), or model selection into work rather than school type (El-Hamidi (2006) and Moenjak and Worswick (2003)). In a review of several prominent studies between 1980s and 1990s, Bennell (1996) criticizes many studies’ failure to correct for bias due to choice of school type and participation in work. 5. Of course, this does not imply that we have controlled for the full set of potentially confounding variables. For example, student motivation and aspirations, and parental income and occupation, are not observed. 6. Newhouse and Beegle (2006) �nd that public junior secondary school students in Indonesia perform better than private school students in national examinations. In contrast, Jimenez, Lockheed, and Paqueo (1991) and World Bank (2007) �nd that private primary school students outperform public school students in several other developing countries. Newhouse and Suryadarma 299 estimating heterogeneous effects of school type, across scholastic ability, age, and family background. The �nal main contribution is the use of a household panel, covering fourteen years, to distinguish between age and cohort effects and assess changes in the returns to vocational education over time. To the extent that bias due to confounding unobserved characteristics remains con- stant over time, this provides an accurate estimate in the changes in returns over time. There are four main �ndings. First, the estimated return to vocational edu- cation, relative to general education, is greater for women than men. Female public vocational graduates enjoy a particularly large wage premium over female graduates of other types of schools, while males bene�t from attending public school, whether general or vocational. Second, the returns to public vocational school for men have plummeted for the most recent cohort. and male vocational graduates now face a large wage penalty. In contrast, returns to public vocational school have, if anything, improved for the most recent cohort of women. This decline for men cannot be explained by an increase in supply, as the probability that both men and women graduated from public vocational school has declined over time. Third, the favorable outcomes of public school graduates partly results from the non-random sorting of students with higher test scores and better- educated parents into public schools. In the most recent cohort, public vocational and general schools attracted the highest-scoring students. Finally, the peer effects created by this sorting are particularly important for students with above- average test scores. The estimated wage premium for public general graduates is noticeably larger for high-scoring students than low-scoring students, particu- larly for men. For males with high entering test scores, the estimated wage premium for high school graduates, compared to to non-graduates, is less than ten percent for public vocational school and negative for private schools. The remainder of this paper is organized as follows. The next section pro- vides background on the Indonesian education system and the mix of voca- tional versus general education. Section II describes the data. Section III analyses school choice patterns. Section IV investigates the effects of different school types of labor market outcomes. Sections V to VII explore heterogeneity in the effects across different types of people. The �nal section concludes and provides policy recommendations. I. SECONDARY EDU CATION IN INDONESIA The secondary education system in Indonesia is divided into junior and senior secondary school, which each take three years to complete. The country has two different school systems, secular and Islamic, and in this paper we focus exclusively on the former.7 In the secular school system, children graduating 7. In 2007, the National Socioeconomic Survey (Susenas) shows that only 8.4 percent school-age children are enrolled in the Islamic system. 300 THE WORLD BANK ECONOMIC REVIEW from junior high school, usually at around 15 years of age, must choose whether to enroll in a vocational or general high school.8 These school types are distinct. Only a small portion of the curriculum used in general and vocational schools overlap, mostly in the subjects of English and Indonesian. General schools offer three majors: natural science, social science, and language. On the other hand, the vocational stream provides a choice between many majors. Each vocational school usually focuses on just one or two majors. The available vocations are business management; technical, which includes machinery and information technology; agriculture and for- estry; community welfare; tourism; arts and handicraft, and health care. In addition, there are very specialized vocational high schools that focus on avia- tion and shipbuilding. Of all these choices, the most popular are the �rst two, business management and technical.9 The public cost of providing vocational education is at least as high as general education. Ghozali (2006) �nds that a public vocational student costs the public 28 percent more annually than a public general student.10 Meanwhile, the amount of per student public funds spent in private schools is lower—about 40 percent and 20 percent lower in the vocational and general streams, respectively—and private vocational schools receive the same amount of public funds as private general schools. Households, meanwhile, face higher out of pocket costs expenses in private schools. Comparing the four school types, households report that private general schools are the most expensive, followed by private and public vocational schools respectively, with public general schools being the least expensive.11 Vocational school expansion plan In 2006, the Ministry of National Education began expanding vocational schools. According to their strategic plan (Ministry of National Education, 2006), the main reason for this policy is to increase the size of the labor force that is ready-to-work, especially among those who do not continue to tertiary education. In addition, the Ministry argues that because the unemployment rate of vocational graduates is lower than general graduates, increasing the share of vocational graduates in the mix would result in a lower overall unemployment rate. The policy’s target is to achieve a 50:50 vocational to general student ratio by 2010, and a 70:30 ratio by 2015. As Figure 1 shows below, the ratio was 24:76 in 2007. In order to achieve this target, the ministry has recommended a moratorium on building new general schools. Instead, the government will 8. Better senior secondary schools also select students based on their test scores. 9. This information is taken from the National Labor Force Survey (Sakernas). We cannot separate the labor market effects of different vocational choices in our dataset. Despite this limitation, the dataset of our choice has many more advantages, such as those we list in the introduction. 10. Public cost is de�ned as the amount of government spending on each school type. 11. Figure S1.1 in the supplemental appendix (available at http://wber.oxfordjournals.org) compares school costs between vocational and general schools. Newhouse and Suryadarma 301 F I G U R E 1. Vocational School Enrollment, 1992–2007 Note: �gures calculated from the National Socioeconomic Survey (Susenas), various years construct new vocational schools and convert some general schools into vocational schools. Enrollment trends Enrollment in vocational high school has been steadily declining, as the number of vocational students has declined from about 1.6 million in 1999 to about 1.2 million in 2006 (Figure 1). Over the same period, the proportion of high school students in vocational schools declined from 27 percent to just 20 percent, as more students choose general education over vocational education. The share attending vocational school jumped in 2007, as the vocational school expansion policy took effect. In light of the historical trend, it appears extremely unlikely that the ministry will meet either the 50:50 target in 2010 or the 70:30 goal �ve years later. I I . D ATA The primary data source for this study is the Indonesia Family Life Survey (IFLS), a longitudinal household survey that began in 1993. Three full follow-up waves were conducted, in 1997, 2000, and 2007. The �rst wave represented about 83 percent of Indonesia’s 1993 population, and covered 13 of the nation’s 27 provinces. This initial round interviewed roughly 7,200 households. By 2007, the number of households had grown to 13,000 as the survey attempts to re-interview many members of the original sample that form or join new house- holds. Household attrition is quite low, as around 5 percent of households are lost each wave. Overall, 87.6 percent of households that participated in IFLS1 are interviewed in each of the subsequent three waves (Strauss et al., 2009). 302 THE WORLD BANK ECONOMIC REVIEW The sample is constructed as follows. We began with respondents who were interviewed at least once between the ages of 18 and 50, as a detailed edu- cation history is only available for respondents aged 50 or younger. Next, we limited our sample to individuals who were born between 1940 and 1980. We then dropped individuals who were never interviewed after they graduated junior secondary, as well as those who were full-time students when inter- viewed. We then dropped observations that did not report complete school information. Finally, to avoid identi�cation based on functional form assump- tions, we restrict the sample to the region of common support (Heckman and Vytlacil, 2001; Tobias, 2003). To do this, we estimated the probability that each person either leaves the schooling system without graduating from senior secondary or attends each of the four school types using a multinomial logit model, and dropped observations for which the estimated probability of attending public general school falls outside the range of all public general graduates. Finally, we dropped reported wages from the bottom and top per- centile from wage regressions to avoid distorted results due to outliers. Table S2.1 in Appendix S2 shows the number of observations that were dropped during each stage of this process. After dropping observations outside the region of common support, the �nal sample consists of 17,485 total labor market observations on 7,607 individ- uals. These individuals are divided into three cohorts. The oldest cohort con- sists of those born from 1940 to 1963, the middle cohort covers those born from 1964 to 1972, and the youngest cohort contains those born from 1973 to 1980. The IFLS survey asks the youngest cohort to report their performance in the junior secondary �nal examination.12 Hence, for this most recent cohort, a direct measure of scholastic ability is available. Descriptive statistics for all variables are given in Table S2.2 in Appendix S2. All estimates are separated by sex, because men and women exhibit different labor market participation patterns and they select different education majors. According to the 2006 National Labor Force Survey (Sakernas), 64 percent of men choose a technical or industrial major, while 56 percent and 29 percent of women are enrolled in business management and tourism majors, respectively. I I I . U N D E R S TA N D I N G S C H O O L C H O I C E To better understand the determinants of an individual’s school choice, we esti- mate the following multinomial logit regression: Ti ¼ aZ Zi þ ai Pi þ ad Pd þ 1i ð1Þ where Ti is a �ve-category variable indicating senior secondary school type or 12. The examination is designed to be nationally comparable by the Ministry of National Education. We standardize the scores by year of junior secondary graduation to take into account possible quality changes in the exam over time. Newhouse and Suryadarma 303 non-graduation, Zi is a vector of predetermined characteristics, Pi is parental education, and Pd is district-level parental education shares. Table 1 provides the estimated marginal effects of selected independent vari- ables. It shows that the reduction in vocational enrollment observed in Figure 1 resulted �rst in an increase in the probability of attending private school, and then decreases in high school attendance. The top rows of the third column show that men in the middle and recent cohorts were 9.6 and 11.4 percentage points less likely to enroll in public vocational schools than those in the oldest cohorts. Men in the middle cohort were more likely to attend general school, by 9.6 percentage points, but private vocational school has become more popular for men in the youngest cohort. Girls have also increasingly turned away from public vocational education. The probability of attending general and vocational private schools both increased by 8.7 and 6.5 percentage points respectively for the middle cohort. This increase in private general persisted for the youngest cohort. The table also shows that a higher percentage of men in the recent cohort left without completing senior secondary education, compared to men in the old and middle cohort. To a certain extent, we �nd a similar pattern among the recent cohort of women. We believe that there are two plausible expla- nations for this pattern. First, the composition of junior high school graduates changed, partly as a result of a nine-year compulsory education program that was enacted by the government in the early 1990s. This program caused some students to graduate from junior high school that would have dropped out in the older cohorts, and these new junior high graduates were less likely to con- tinue on to high school.13 Second, the rapid increase in the supply of high school graduates eroded the returns to completing high school. Turning to par- ental education, the children of highly educated parents are more likely to attend general schools. Increased paternal education raises the probability of attending private general school the most, followed by public general schools. The pattern is similarly strong among women. The Effect of Test Scores on School Choice Test score data is available for the most recent cohort (those born between 1973 and 1980). For this cohort, we examine how test scores relate to school choice, and whether including test scores alters the estimated effect of the other independent variables, especially parental education. Table S3.1 in Appendix S3 provides the estimation results for men, while Table S3.2 shows the results for women. 13. This generational shift from dropping out after completing primary school to dropping out after junior secondary school would have been clearer had our sample consisted of all individuals, not just those who completed junior secondary school. However, since our main interest is to describe the choice of senior secondary education, we choose to continue using our current sample. 304 T A B L E 1 . Determinants of School Enrollment: selected variables, full sample Men Women No senior Public Public Private Private No senior Public Public Private Private secondary general vocational general vocational secondary general vocational general vocational Personal characteristics Middle Cohort 0.6 2.9* 2 9.6*** 6.7*** 2 0.7 2 9.7*** 2.4 2 7.9*** 8.7*** 6.5*** THE WORLD BANK ECONOMIC REVIEW (2.1) (1.7) (1.7) (1.7) (1.5) (2.4) (2.2) (2.0) (2.0) (1.7) Recent cohort 12.8*** 2 4.5*** 2 11.4*** 2 0.8 4.0** 3.9 0.6 2 12.2*** 2.9* 4.7*** (2.2) (1.7) (1.6) (1.7) (1.6) (2.5) (1.8) (1.8) (1.8) (1.5) Repeated grade in junior secondary 0.1 2 3.7 2 3.4 6.6 0.3 5.5 2 1.4 2 5.7 2 0.3 1.9 (4.4) (4.2) (2.9) (4.8) (3.7) (8.2) (6.1) (4.8) (8.0) (5.1) Lived in small town at age 12 2 5.1** 2.6 0.4 2 0.3 2.4 2 6.0*** 4.6** 2 0.6 0.4 1.6 (2.1) (1.7) (1.5) (1.6) (1.7) (2.1) (1.9) (1.6) (1.7) (1.5) Lived in big city at age 12 2 6.5** 7.2*** 2 0.2 2 3.2* 2.7 2 5.4** 5.8** 1.1 4.2** 2 5.6*** (2.6) (2.2) (1.8) (1.8) (2.1) (2.7) (2.3) (2.1) (2.1) (1.6) Height 2 0.3** 0.1 2 0.0 0.3** 2 0.1 2 0.5*** 0.2 0.1 0.3** 2 0.1 (0.1) (0.1) (0.1) (0.1) (0.1) (0.2) (0.1) (0.1) (0.1) (0.1) Parental education Father graduated elementary 2 13.7*** 6.1** 1.6 5.7*** 0.3 2 15.4*** 2.3 1.1 6.8*** 5.2* (3.5) (2.4) (2.3) (2.1) (2.6) (4.1) (3.1) (2.8) (2.3) (2.7) Father graduated junior secondary 2 24.6*** 10.1*** 1.6 12.1*** 0.9 2 28.7*** 7.5** 1.3 11.3*** 8.6*** (4.0) (3.1) (2.8) (3.0) (3.3) (4.7) (3.6) (3.3) (3.0) (3.2) Father graduated senior secondary 2 24.6*** 9.5** 2 0.0 15.8*** 2 0.6 2 42.5*** 14.2*** 4.2 18.2*** 6.0 (5.7) (3.7) (3.3) (4.0) (3.9) (5.2) (4.8) (3.8) (4.1) (3.7) Father graduated university 2 27.6*** 19.9*** 2 4.6 18.1*** 2 5.8 2 40.1*** 19.4*** 1.7 13.7*** 5.3 (5.5) (5.0) (3.5) (5.1) (4.0) (6.1) (5.7) (4.6) (4.8) (5.0) Father attended vocational school 2 7.7 4.5 6.5 2 1.6 2 1.8 5.4 2 1.1 2 1.0 2 6.9** 3.6 (5.6) (3.9) (4.2) (3.3) (3.7) (5.9) (3.6) (3.0) (2.8) (3.7) Mother graduated elementary 2 2.2 1.7 2 2.7 3.9* 2 0.7 2 6.8** 5.8** 4.9*** 2 0.9 2 3.0 (2.7) (2.2) (2.0) (2.0) (2.1) (3.1) (2.3) (1.7) (2.4) (2.4) Mother graduated junior secondary 2 6.4 4.6 2 1.5 5.8* 2 2.5 2 8.8** 10.9*** 2 0.7 4.3 2 5.7** (4.3) (3.3) (2.8) (3.1) (3.0) (3.8) (3.0) (2.2) (3.3) (2.7) Mother graduated senior secondary 2 18.1*** 9.5* 2 2.7 5.1 6.1 2 10.6 9.6** 4.5 1.1 2 4.5 (5.5) (5.2) (4.3) (4.9) (5.7) (6.8) (4.4) (4.1) (5.1) (4.4) Mother graduated university 2 0.8 11.6 2 1.9 2 4.1 2 4.7 2 31.1*** 21.6** 10.5 2 1.7 0.8 (9.3) (7.9) (7.6) (4.0) (5.5) (6.6) (8.7) (8.6) (6.7) (6.9) Mother attended vocational school 2 0.5 2 0.0 4.8 2 2.5 2 1.8 2 3.9 3.2 2 1.0 2 1.6 3.4 (8.3) (4.5) (6.1) (4.2) (4.3) (7.8) (4.4) (4.1) (5.0) (5.8) Base case probability 34.8 19.7 13.3 17.9 14.4 35.9 20.0 13.1 17.5 13.5 Observations 4,040 3,567 R 2 Squared 0.103 0.124 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; �gures are marginal effects in percentage points; estimation includes province of junior secondary graduation �xed effects and all variables listed in Table S2.2; standard errors in parentheses, they are robust to heteroskedasticity and clustered at subdistrict level. Newhouse and Suryadarma 305 306 THE WORLD BANK ECONOMIC REVIEW For both sexes, students with test scores in the top tercile are far more likely to attend public schools. Moreover, private vocational schools attract the lowest scoring students. Including test scores does not alter the �nding above that highly educated parents choose general schools over vocational schools, although the effects are less precisely estimated. In sum, the probability that students enroll in public vocational schools declined substantially for the middle and youngest cohort. However, this does not seem to be caused by a decline in the quality of public vocational students, as high scoring students are still more likely to attend public schools. The more likely cause is an increase in the number of private schools, particularly private vocational schools, which have responded to the continued high demand for highly educated workers (World Bank, 2011). Two main household characteristics are associated with choice of school type: scholastic ability and parental education. With regards to the former, high scoring students choose public schools as their �rst preference, followed by private general school. For parental education, private general schools attract the sons of better-educated fathers, followed by public general schools. Private vocational schools act as a last resort; students who enroll in these schools disproportionately scores in the bottom tercile and their parents are less well educated. I V. L A B O R M A R K E T E F F E C T S OF VO CAT I O N A L E D U CAT I O N We turn from the determinants of students’ school type to their subsequent labor market experience. We examine four different outcomes: labor force par- ticipation (LFP), unemployment conditional on participation, formal sector work, and log of hourly wage.14 The reduced form model estimated is: Yit ¼ bZ Zi þ bP Pi þ bD Dd þ bY Yt þ bT Ti þ 1it ð2Þ where Yit is the labor market outcome of person i in year t. Zi and Pi, as in equation one, are de�ned as a vector predetermined individual characteristics and parental education, while Dd is a set of indicators for district of junior secondary school. Yt is a vector of interview year dummies, and Ti is a vector of categorical dummies of the senior secondary school types, including no senior secondary, with public general excluded given that our main interest is in comparing the returns to general relative to vocational.15 The equation is estimated using double robust regression, which rebalances the sample by reweighting observations according to the inverse estimated 14. The wage of self-employed individuals is calculated using their average hourly pro�t. The Statistics Indonesia urban price index is used to deflate 1993 wages, while IFLS price indices are used for subsequent years. 15. We do not control for university attendance, which is partially determined by choice of school type. Newhouse and Suryadarma 307 probability of attending the type of school that a person graduated from. While this reweighting reduces precision, it makes the estimates more robust to non-linear functional forms. A key indicator to measure the effectiveness of this reweighting procedure is the normalized difference between means of the observed control variables for different school types, compared to general public graduates (Imbens and Wooldridge, 2009). Reweighting greatly reduces the average of the normalized difference across the 42 control variables. The average normalized difference falls by 66 percent for public vocational graduates, 80 percent for private general graduates, and 95 percent for private vocational graduates.16 This indicates that the reweighting was effective. To the best of our knowledge, a plausible instrument for school choice does not exist.17 As a result, the OLS results reported would be biased to the extent that school choice is based on unobserved determinants of labor market out- comes.18 Non-random selection into employment will also bias the estimated effects of school type on formality and wages, if unobserved determinants of school type are correlated with the probability that different types of graduates choose to work. It is therefore important to control for as many pre-determined or exogenous characteristics as possible. Fortunately, the survey collects a large amount of data on individual and family characteristics. We include parental education, for both resident and non-co-resident parents; height; self-reported size of residence at age 12; grade repetition in junior high and elementary school; public lower secondary school attendance; working while attending elementary school, or lower secondary, and year of interview. In addition, the youngest cohort was asked to report their lower secondary test score, which can be used to gauge the bias due to omitting this variable. We include district of junior secondary graduation �xed effects to take into account differences in the supply of education, community characteristics, and peer effects that vary across district.19 Finally, survey year and cohort dummies are included, which captures age as well. Despite the inclusion of several observed characteristics, 16. After rebalancing, the normalized difference between public general and public vocational graduates is 0.006. For private general and vocational, the normalized difference is 0.005 and 0.001 respectively. 17. We have tried several instruments, including the share of schools of each type and the leave-out mean of enrolment in each school type in the district and year where a person graduates from junior secondary school. While the latter is a strong instrument, it is unlikely to be valid, as variation in school attendance patterns across communities is undoubtedly correlated with local labor market conditions. The best candidate instrument would be data on historical school construction, as in Duflo (2001). However, this information is unavailable, and the village censuses (Podes) show little change in the local prevalence of different types of high schools across time. Therefore, we elected to abandon the instrumental variables approach. 18. Unfortunately, it is dif�cult to speculate as to the direction of the bias, given the lack of data on unobserved characteristics such as motivation or aspirations, and the presence of several control variables in the model. 19. District of lower secondary school is highly collinear with district of secondary school, as less than a quarter of the sample attended junior and senior secondary schools in different districts. 308 THE WORLD BANK ECONOMIC REVIEW however, there are important unobserved characteristics that are omitted and we do not claim that the results are causal. Table 2 shows the estimated labor market effects of different school types relative to public general, while the full estimation results are in Table S4.2 in Appendix S4. For robustness, the fourth and �fth columns give the estimates of average and median returns.20 For men, the results show a substantial public school wage premium. Private general graduates earn nearly 25 percent less than public general graduates and private vocational students nearly 18 percent less. These penalties are substan- tial given that non-graduates earn 40 percent less. In contrast, differences between public and vocational schools are much smaller. The estimates are suf- �ciently precise to rule out a public vocational premium, relative to public general, exceeding 16 percent. Differences between public general and voca- tional graduates are more apparent, however, when examining formality.21 Graduating from a vocational school is associated with about a 6 percentage point greater chance of working in a formal job. For male graduates of private schools, the average wage penalty is similar for vocational and general graduates, although general graduates face a larger median wage penalty. Private general school graduates are also much less likely to get a formal job than private vocational graduates. Compared to public school graduates, private general graduates are 5 percentage points less likely to work in a formal job, but private vocational graduates are 5 percentage points more likely to. The results for private general graduates are particularly disappointing, since private vocational graduates tend to have lower parental education levels than private general graduates, and in the most recent cohort, lower test scores as well. Among women, private general schools are also associated with reduced labor force participation and formality rates compared with graduates of the other three school types. With regards to wages, meanwhile, public vocational graduates earn a wage premium of 16 percent. The wage estimates for females are less precise but can nonetheless rule out a public vocational wage premium that is greater than 30 percent. Private general graduates earn the least com- pared to observable similar graduates of the other three schools, although the difference is not statistically signi�cant. As with men, women with no senior secondary education earn far less than those who attend senior secondary. 20. Although median regression is more robust to outliers, it does not allow for the inclusion of district �xed effects. As a result, we included provincial rather than district effects in the median regression speci�cation. 21. A job is classi�ed as formal if the worker is a salaried employee, is self-employed with permanent workers, or is self-employed with temporary workers outside of agriculture. This de�nition, which is based on employment status and sector, is 99 percent correlated with the of�cial de�nition adopted by the Statistics Indonesia, which is based on employment status and occupation. Formal employees tend to earn higher wages and express greater job satisfaction than informal employees, particularly casual workers (World Bank, 2011). T A B L E 2 . The Effect of School Types on Labor Market Outcomes: Full sample pooled Men Women LFP Unemployment Formal Wage Wage LFP Unemployment Formal Wage Wage LPM LPM LPM OLS LAD LPM LPM LPM OLS LAD No secondary school 2 0.001 0.005 2 0.096*** 2 0.404*** 2 0.481*** 2 0.147*** 2 0.015 2 0.213*** 2 0.473*** 2 0.636*** (0.009) (0.009) (0.027) (0.053) (0.037) (0.027) (0.009) (0.031) (0.093) (0.062) Public Vocational 0.015** 2 0.007 0.062** 0.041 0.007 0.029 2 0.020* 0.087*** 0.158** 0.143*** (0.007) (0.012) (0.027) (0.057) (0.037) (0.030) (0.011) (0.028) (0.072) (0.049) Private general 0.014* 2 0.004 2 0.050* 2 0.146** 2 0.248*** 2 0.078** 0.017 2 0.084** 2 0.138 2 0.251*** (0.008) (0.009) (0.030) (0.060) (0.055) (0.031) (0.011) (0.041) (0.091) (0.079) Private vocational 0.006 0.004 0.048* 2 0.177*** 2 0.176*** 2 0.034 0.007 0.005 2 0.039 2 0.092 (0.008) (0.012) (0.028) (0.060) (0.035) (0.031) (0.012) (0.041) (0.090) (0.071) Average among public 0.971 0.053 0.712 0.698 0.043 0.734 general graduates R-squared 0.069 0.153 0.181 0.219 0.156 0.211 0.239 0.312 Observations 9,012 8,774 8,342 7,370 7,370 8,473 5,136 4,928 3,801 3,801 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; standard errors in parentheses, they are robust to heteroskedasticity and clustered at subdistrict level; LPM stands for Linear Probability Model, OLS stands for Ordinary Least Squares, and LAD for Least Absolute Deviations. In all cases, the sample is rebalanced by reweighting observations by the estimated inverse probability of attending their school type, in addition to standard individual cross-sectional weights. Robust standard errors are reported. All estimates are based on equation (2) in the text. Wage LAD estimates include provincial instead of district �xed effects. Standard errors for LAD estimates are obtained from an unweighted bootstrap procedure. Newhouse and Suryadarma 309 310 THE WORLD BANK ECONOMIC REVIEW One potential source of bias stems from the lack of a direct measure of scho- lastic ability for the entire sample. To assess the extent to which this omission generates biased estimates of the returns to different types of schools, in Table S4.1 in Appendix S4 we include test scores in the estimation results using the youngest cohort and compare them to the results when the variable is excluded. The results show, reassuringly, that the omission of test scores is a negligible source of bias. V. H E T E R O G E N E I T Y IN AGE AND COHORT Returns to vocational education may decline over time. This could occur, for example, if the speci�c skills taught in vocational schools become obsolete faster than general skills. Vocational graduates’ speci�c skills may also enable them to work immediately at a market wage after graduation, while general graduates need to be trained further by the �rms that employ them. Over time, however, general graduates may �nd it easier to upgrade their skills to cater to employers’ demands. In either case, vocational education would confer an initial advantage that would erode over time. In this section, we examine age effects for different cohorts, which enable us to separate age effects from cohort effects. There has been little research exam- ining how the returns to school type vary by age in developing counties, largely due to the lack of long-running longitudinal datasets. As we discuss in Section III, we divided the IFLS sample into three cohorts: old (those born between 1940 and 1962), middle (1963 – 1972), and young (1973 – 1980). For each cohort, we estimate the following equation: Yit ¼ bZ Zi þ bP Pi þ bD Dd þ bY Yt þ bT Ti þ bty ðTi à Yt Þ þ 1it ð3Þ In this speci�cation, bty is a 1 X 16 vector, containing the estimated effect of each school type, relative to public general, for each of the four survey waves. We discuss the outcomes in turn below, while the graphic representation and estimation results are in Appendix S5.22 We begin by looking at the effect of vocational school on labor force partici- pation, starting with men. Comparing the young and the middle cohorts, the recent cohort of men is more likely than the middle cohort to participate early in their career, although the difference is not statistically signi�cant and disap- pears by age thirty. In general, the effect of public vocational education on 22. We graph these estimated effects, separately for each cohort, on the vertical axis. The horizontal axis represents the average age of each cohort in the relevant year. Therefore, for each cohort and labor market indicator, there are four estimates of the effect, spanning fourteen years of the cohort’s life. Since the youngest cohort covers those born from 1973 to 1980, its oldest members were 20 in 1993. Since only a few members of the youngest cohort were working in 1993, these estimates are not reported. We also calculate the effect for private general and private vocational schools. However, because the vocational expansion in prioritising public vocational over public general, in this section we focus on these two school types. Newhouse and Suryadarma 311 participation is approximately one and a half percentage points, which is sizable given that only three percent of male public general graduates, on average, do not participate in the labor force. The positive effect of public vocational school on participation begins to decline at age thirty and becomes negative around the age of forty. The pattern for women, shown in Figure S5.2, is the opposite. The positive effect of public vocational on participation starts high, declines, and then recovers. The �ve percentage premium experienced by 25 year-olds decreases with age, reaches a bottom of negative percentage points in the early 30s, and then increases to ten percentage points for older women. There are no apparent cohort effects. Turning to the probability of unemployment, the difference in unemploy- ment between public general and public vocational graduates is shown in Figure S5.3 for men and Figure S5.4 for women. Men exhibit no cohort effects, as the graph is continuous across cohorts. Public vocational graduates enjoy lower unemployment from their early twenties until they turn thirty. After that, the effect of vocational education remains close to zero without becoming statistically signi�cant. For females, meanwhile, there is a sizeable cohort effect between the young and the middle cohorts. At the age of twenty-�ve, vocational graduates in the young cohort enjoy lower unemployment rates than general graduates, while vocational graduates in the middle cohort face the same unemployment rate as general graduates. At around thirty, however, the unemployment rate of voca- tional graduates in the young cohort is higher than general graduates. Looking at the age pro�le, it appears that general and vocational graduates over thirty years old have similar unemployment rates. Next, Figure S5.5 examines the effect of public vocational education on the probability of holding a formal job, conditional on being employed. For the middle cohort in 1993, when their average was 26, public vocational graduates enjoyed a large formality premium of nearly 30 percentage points, which gradually declined to 10 percentage points in 2007. That formality premium, disappeared for the younger cohort, however. In the year 2000, when their average age was 23, the formality premium for the youngest cohort was 5 per- centage points and seven years later it was essentially zero. For female public vocational graduates, unlike their male counteparts, there is no evidence of a fall in the formality premium for the youngest cohort. Figure S5.6 shows that female public vocational graduates are no more likely to work in formal jobs than public general graduates from ages 20 to 30. After age 30, public vocational graduates are slightly more likely to be in a formal job, but that formality premium gradually declines. The last labor market outcome that we examine is the reported wage, shown in Figures 2 and 3. The comparison between the middle and youngest cohort is particularly striking. In the middle cohort, public vocational graduates enjoyed an estimated 40 percent wage premium in 1993, when they were 25, which 312 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. Effect of Public Vocational on Wages, Men Note: estimation results are in Table S5.4 in Appendix S5 F I G U R E 3. Effect of Public Vocational on Wages, Women Note: estimation results are in Table S5.4 in Appendix S5 declined to essentially zero in 1997, 2000, and 2007. Male public vocational graduates in the youngest cohort, however, experienced a large wage penalty. The estimated penalty was 20 percent in 2000, when the cohort on average was 23, and 40 percent seven years later. As was the case for informality, there is no clear sign of a cohort effect for women. The public vocational wage premium for the youngest cohort, which was 60 percent for women in 2007, if anything rose compared to the middle cohort. However, the estimates are imprecise and not statistically signi�cant. Overall, the age-wage pro�le suggests a short-lived bene�t for female voca- tional graduates in their mid to late twenties, which largely disappears in their thirties before picking up again in their forties and �fties. The large public vocational premium for the oldest cohort around the age of �fty is the only estimated effect that is statistically signi�cant. This section highlights the importance of estimating both cohort and age effects and treating age effects carefully. In general, the strongest effects of Newhouse and Suryadarma 313 vocational education are experienced early in life, between the ages of 20 and 35. For example, while Table 2 shows an insigni�cant negative effect of voca- tional education on unemployment over the entire sample, results in this section show that this effect is concentrated among young graduates in their twenties. Results for graduates younger than 25, however, are contaminated by university enrollment decisions. This is because full time students are not included in the sample, and students typically do not typically graduate from university until age 25. University enrollment could explain part of the negative effect of vocational education on unemployment, for example. General second- ary school graduates are more likely to attend university than vocational gradu- ates, and university graduates are more likely to experience spells of unemployment as they search for the best job following graduation. Since the determinants of university enrollment and graduates’ job search patterns are not well understood and likely depend on unobserved factors, we focus on results for groups over 25. The results for recent graduates, particularly those between 25 and 35, suggest that the returns to public vocational school have declined sharply for men. For example, while Table 2 shows a higher formality rate among all male vocational graduates, Figure V.5 shows that the middle cohort drives this posi- tive formality rate in their youth, and that the premium has disappeared for the youngest cohort. This is consistent with the dramatic fall in the effect of voca- tional education on men’s wages. Figure 2 shows that while workers in the middle cohort enjoy a large vocational wage premium before they turn 30, individuals in the youngest cohort enjoy no such bene�t. In contrast, after enjoying a smaller wage premium at the age of 21, individuals in this cohort face an increasingly large wage penalty. Although male public vocational graduates face increasingly worse labor market outcomes, there is no sign of a similar deterioration for female public vocational graduates. One possible explanation for this decline for men relates to recent changes in the structure of the Indonesian economy. Since the �nancial crisis of 1998, the economy has increasingly relied on the service sector to generate growth. Annual growth in the industrial sector fell dramatically, from nine percent from 1990 to 1997, to four from 1999 to 2007. During the same two periods, annual service sector growth remained strong, falling slightly from seven to six. More recently, employment in the service sector has grown rapidly. From 2003 to 2007, service sector employment grew at roughly four percent per year while industrial sector employment grew at 2.5 percent per year (World Bank, 2011). The increasing prominence of the service sector could disproportio- nately affect vocationally trained males because they tend to choose technical majors. Women, on the other hand, tend to choose to study business manage- ment or tourism skills, for which demand may have remained stronger. In an increasingly service-oriented economy, there may be decreased demand for the industrial and technical majors chosen by most men in vocational schools. 314 THE WORLD BANK ECONOMIC REVIEW Another potential explanation for the recent decline in male vocational returns is deterioration in the quality of vocational training for men. For example, technical vocational training may require larger investments to remain relevant to new advances in technology. Unfortunately, it is dif�cult to investigate this further, due to the lack of data on trends in the quality of industrial education facilities. VI. HETEROGENEITY IN FA M I LY BAC K G RO U N D The second aspect of heterogeneity that we examine is family background, proxied for by father’s education. We separate the sample into two categories: those whose father has at most a junior secondary education and those whose father has at least a senior secondary education. Table 3 shows the estimation results for men. Comparing the results with the ones in Table 2, we �nd that the effects of school types on labor market outcomes are mostly limited to stu- dents from a disadvantaged background. Among these individuals, graduates of public vocational schools have a higher formality rate than public general school graduates, while private general graduates face the lowest prospects of a formal job. In addition, private school graduates face a large wage penalty rela- tive to public school graduates. In short, public schools appear to provide the most bene�t for children from disadvantaged families. The estimation results for women, shown in Table 4, give similar con- clusions. The labor market effects of school types are for the most part only signi�cant among those coming from a disadvantaged background, except the formality penalty among private general graduates. Among individuals from disadvantaged background, private general graduates fare the worst, facing a lower participation and job formality rate. In contrast, public vocational gradu- ates have the highest labor force participation and formality rate. VII. HETEROGENEITY IN ACA D E M I C AB I L I T Y The �nal aspect of heterogeneity in the labor market effects of different school types that we consider pertains to academic ability. Does higher ability mitigate or magnify the labor market effects of school types?23 Since test scores are only available for the youngest cohort, the relevant benchmarks are given in Table IV.1 in Appendix S4, which shows that recent male private general and public vocational graduates experience a substantial wage penalty. Table 5 provides the estimated effects of school type for men that scored above and below the median on their junior high exit exam. For men scoring below the mean, public vocational education is much more likely to lead to a formal job, but the average wage is much lower. Interestingly, private 23. Note that our sample is rebalanced and has common support over the test score distribution, which allows for valid comparisons across school types despite large differences in average test scores. T A B L E 3 . Estimated Effect of School Type on Employment and Job Quality, Men, by father’s education Junior secondary or below Senior secondary or above LFP Unemployment Formal Wage LFP Unemployment Formal Wage LPM LPM LPM OLS LPM LPM LPM OLS No senior secondary 2 0.001 0.008 2 0.131*** 2 0.481*** 2 0.026 0.024 2 0.008 2 0.171 (0.008) (0.008) (0.029) (0.052) (0.026) (0.042) (0.074) (0.214) Public vocational 0.002 2 0.011 0.062** 0.037 0.023 0.016 0.011 0.023 (0.006) (0.009) (0.030) (0.059) (0.015) (0.030) (0.053) (0.133) Private general 0.006 2 0.009 2 0.066** 2 0.230*** 0.045** 0.008 0.017 2 0.122 (0.008) (0.008) (0.030) (0.058) (0.018) (0.027) (0.053) (0.156) Private vocational 2 0.002 0.011 0.002 2 0.281*** 0.002 0.070* 0.049 2 0.110 (0.008) (0.013) (0.028) (0.070) (0.027) (0.039) (0.069) (0.135) Average among public general graduates 0.972 0.044 0.758 0.961 0.067 0.814 R-squared overall 0.059 0.151 0.186 0.231 0.169 0.318 0.328 0.403 Observations 6,650 6,489 6,201 5,481 1,214 1,164 1,063 928 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; standard errors in parentheses, they are robust to heteroskedasticity and clustered at subdistrict level. Newhouse and Suryadarma 315 316 T A B L E 4 . Estimated Effect of School Type on Employment and Job Quality, Women, by father’s education Junior secondary or below Senior secondary or above LFP Unemployment Formal Wage LFP Unemployment Formal Wage LPM LPM LPM OLS LPM LPM LPM OLS No senior secondary 2 0.119*** 2 0.024* 2 0.240*** 2 0.567*** 2 0.118* 2 0.021 2 0.050 2 0.048 (0.029) (0.013) (0.034) (0.098) (0.063) (0.039) (0.109) (0.151) Public vocational 0.069** 2 0.016 0.057* 0.169* 0.054 2 0.039 0.025 2 0.027 THE WORLD BANK ECONOMIC REVIEW (0.032) (0.013) (0.033) (0.090) (0.062) (0.027) (0.061) (0.132) Private general 2 0.068** 2 0.001 2 0.122*** 2 0.272** 2 0.033 0.001 2 0.125** 0.125 (0.033) (0.014) (0.044) (0.137) (0.054) (0.032) (0.062) (0.140) Private vocational 2 0.020 0.011 2 0.038 2 0.030 2 0.070 2 0.001 2 0.072 0.188 (0.037) (0.016) (0.042) (0.098) (0.086) (0.033) (0.072) (0.164) Average among public general graduates 0.640 0.050 0.711 0.700 0.073 0.815 R-squared overall 0.146 0.201 0.257 0.338 0.270 0.325 0.371 0.442 Observations 5,996 3,564 3,435 2,586 1,569 1,030 969 804 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; standard errors in parentheses, they are robust to heteroskedasticity and clustered at subdistrict level. T A B L E 5 . Estimated Effect of School Type on Employment and Job Quality, Men, by test score Low scores High scores LFP Unemployment Formal Wage LFP Unemployment Formal Wage LPM LPM LPM OLS LPM LPM LPM OLS No senior secondary 0.017 2 0.052 0.125 2 0.159 0.016 2 0.014 0.053 2 0.296* (0.045) (0.059) (0.082) (0.148) (0.041) (0.040) (0.063) (0.156) Public vocational 0.023 2 0.108 0.218** 2 0.479** 0.017 2 0.006 2 0.027 2 0.218* (0.047) (0.097) (0.102) (0.242) (0.034) (0.049) (0.066) (0.128) Private general 0.009 2 0.096 0.108 0.052 0.012 0.024 2 0.064 2 0.319** (0.046) (0.060) (0.100) (0.181) (0.030) (0.043) (0.081) (0.127) Private vocational 2 0.005 2 0.085 0.141* 2 0.024 2 0.008 0.043 2 0.001 2 0.308** (0.043) (0.066) (0.081) (0.153) (0.037) (0.055) (0.074) (0.151) Average among public general graduates 0.921 0.184 0.627 0.944 0.145 0.742 R-squared overall 0.231 0.327 0.420 0.353 0.180 0.324 0.423 0.392 Observations 1,093 1,027 896 750 1,109 1,051 916 747 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; �gures are marginal effects; standard errors in parentheses, they are robust to het- eroskedasticity and clustered at subdistrict level. Low scores are below median. Newhouse and Suryadarma 317 318 THE WORLD BANK ECONOMIC REVIEW vocational school is also associated with an increase in formality, and has no associated penalty. For men scoring above the median, the results illustrate the bene�ts of attending public general school. Public general graduates earn a 20 percent premium over public vocational students and a 30 percent premium over private school graduates. Remarkably, there appears to be no positive return to attending private school, relative to not graduating, for high scoring men. It is these high scoring men who stand the most to lose from attending vocational or lower-quality private general education in an economy that increasingly values broadly educated and cognitively skilled workers. The results for women are shown in Table 6. Most striking are the different effects of school type on wages for low and high scoring women. Unlike for men, high-scoring women face no major wage penalty for attending public vocational or private general school. The wage penalty for private vocational school, relative to public general, is 35 percent, and nearly equal to the penalty for not graduating high school. For low scoring women, there is suggestive evi- dence that vocational school helps. Low scoring women who attend public and private vocational schools earn approximately a 38 percent and 30 percent wage premium, respectively. Although not statistically signi�cant, these are large premiums. VIII. CONCLUSION This paper attempts to better understand the determinants of households’ choice of senior secondary schools in Indonesia and the labor market conse- quences of attending different types of high schools. This is the �rst paper to our knowledge from a developing country that distinguishes between public and privately provided vocational schools, to assess whether private vocational schools impart skills more relevant to a rapidly changing labor market. Another key contribution is a careful examination of heterogeneity in the effects. We examine effects separately by age, cohort, parental education, and ability. The use of longitudinal data allows for cohort effects to be distin- guished from age effects. Finally, the estimation utilizes an unusually rich set of predetermined control variables. While the possibility of bias due to unob- served characteristics cannot be dismissed, it is reassuring that for the youngest cohort, the inclusion of test scores – the most important determinant of school type – does not signi�cantly alter the results. The two most important observed determinants of school choice are test scores and parental education. Students with high test scores are most likely to attend public schools, particularly public general school. In contrast, the chil- dren of highly educated parents tend to select general schools, particularly private general, rather than vocational schools. Private vocational school is a last resort, serving students with the lowest test scores and the least educated parents. T A B L E 6 . Estimated Effect of School Type on Employment and Job Quality, Women, by test score Low scores High scores LFP Unemployment Formal Wage LFP Unemployment Formal Wage LPM LPM LPM OLS LPM LPM LPM OLS No senior secondary 2 0.078 0.051 2 0.269*** 0.051 2 0.064 2 0.101** 2 0.155** 2 0.376** (0.071) (0.050) (0.102) (0.301) (0.062) (0.050) (0.073) (0.166) Public vocational 0.008 0.070 2 0.208 0.383 2 0.016 2 0.071* 0.070 2 0.081 (0.095) (0.070) (0.133) (0.389) (0.064) (0.039) (0.086) (0.154) Private general 2 0.072 0.102 2 0.168 0.148 2 0.032 2 0.031 2 0.120 0.012 (0.087) (0.074) (0.121) (0.348) (0.081) (0.059) (0.090) (0.160) Private vocational 0.008 0.117* 2 0.122 0.301 2 0.007 0.037 0.043 2 0.351* (0.087) (0.066) (0.112) (0.292) (0.064) (0.069) (0.071) (0.191) Average among public general graduates 0.569 0.060 0.672 0.689 0.156 0.767 R-squared overall 0.239 0.402 0.484 0.548 0.324 0.381 0.434 0.506 Observations 1,219 680 624 437 1,169 774 692 525 Notes: *** 1% signi�cance, ** 5% signi�cance, * 10% signi�cance; �gures are marginal effects; standard errors in parentheses, they are robust to het- eroskedasticity and clustered at subdistrict level. Low scores are below median. Newhouse and Suryadarma 319 320 THE WORLD BANK ECONOMIC REVIEW With regard to labor market outcomes, we �nd a striking distinction between publicly and privately schooled men. Male private school graduates, compared to their public school counterparts, suffer an average wage penalty of approximately 16 percent. This large wage penalty is robust to median regression. For men with high test scores, the correlation between public general attendance and subsequent wages is particularly strong. The patterns are somewhat different for women. Public school graduates earn more than private general graduates, but there are important differences between types of graduates. Among employed public school graduates, vocational graduates have tra- ditionally fared slightly better than general graduates in the labor market, although this is no longer the case among men. In general, attending public vocational school attendance has a mild, positive, and statistically insigni�cant correlation with wages, and the estimates are suf�ciently precise to rule out wage effects greater than 16 percent for men. Public vocational schools increase the probability of obtaining a formal job, as de�ned by the Indonesian Bureau of Statistics, by six percentage points. For women, the results suggest a positive effect of public vocational education, although this effect is only clearly discernible for the oldest cohort of women. There is suggestive evidence that this positive effect of public vocational school is strongest for women with lower test scores. In contrast to men, the outcomes for female public vocational graduates in recent years have, if anything, improved. The most dramatic result, which comes from disentangling age and cohort effects, is the large drop in the wage premium for the most recent cohort of male public vocational graduates. This drop is unlikely to be explained by changes in the unobserved characteristics of vocational graduates, as there are no major changes in the observed characteristics of vocational attendance for the youngest cohort. While we cannot directly explore the underlying causes behind this drop, plausible possibilities include a fall in the educational quality of the technical and industrial majors favored by men, as well as the declining relevance of these skills in an increasingly service-oriented Indonesian economy. In sum, the results suggest that whether high schools are publicly or pri- vately administered and whether the curriculum is vocational or general are both important factors influencing graduates’ subsequent labor market out- comes. Male private school graduates earn substantially less than their publicly schooled peers. Private general school graduates perform particularly poorly, despite their parents’ higher education levels. This highlights the need for further research to investigate the importance of peer effects, curriculum, tea- chers, and reputation effects in explaining these results. The current evidence is insuf�cient to justify a recommendation to rapidly expand access to public schools. Nonetheless, given the especially strong results for men with high test scores, a logical �rst step would be ensuring access to public general schools for these high-scoring students. Newhouse and Suryadarma 321 Most importantly, the analysis provides little evidence to support the current expansion of vocational education, especially for men. The results fail to show systematic bene�ts for public vocational graduates compared to public general graduates, despite reasonably precise estimates. Furthermore, the wage penalty for male vocational graduates, in recent years, has increased dramatically. The decline has occurred as Indonesia’s industrial sector has sharply slowed and the service sector has become increasingly important to economic growth. Therefore, the general equilibrium effects of a large increase in the supply of vocational graduates should further magnify the wage penalty that we observe in this partial equilibrium investigation. This suggests that it may be worth- while to review, and possibly reform, vocational and technical education in male-dominated subjects. REFERENCES Bennell, Paul. 1996. “General versus Vocational Secondary Education in Developing Countries: A Review of the Rates of Return Evidence.� Journal of Development Studies, 33(2): 230– 247. Bennell, Paul, and Jan Segerstrom. 1998. “Vocational Education and Training in Developing Countries: Has the World Bank Got it Right?� International Journal of Educational Development, 18(4): 271–287. Chen, Dandan. 2009. Vocational Schooling, Labor Market Outcomes, and College Entry. Policy Research Working Paper 4814. Washington, DC: World Bank. Duflo, Esther. 2001. “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment�, American Economic Review, 91(4): 795– 813 El-Hamidi, Fatma. 2006. “General or Vocational Schooling? Evidence on School Choice, Returns, and ‘Sheepskin’ Effects from Egypt 1998.� Journal of Policy Reform, 9(2): 157–176. Ghozali, Abbas. 2006. Analisis Biaya Satuan Pendidikan Dasar dan Menengah. Jakarta. mimeo. Heckman, James J., and Edward Vytlacil. 2001. “Identifying the Role of Cognitive Ability in Explaining the Level of and Change in the Return of Schooling.� Review of Economics and Statistics, 83(1): 1 – 12. Horowitz, Andrew W., and Christoph Schenzler. 1999. “Returns to General, Technical and Vocational Education in Developing Countries: Recent Evidence from Suriname.� Education Economics, 7(1): 5–19. Imbens, Guido, and Jeffrey Woolridge, 2009, “Recent Developments in the Econometrics of Program Evaluation.� Journal of Economic Literature, 47(1): 5– 86 Jimenez, Emmanuel, Marlaine E. Lockheed, and Vicente Paqueo. 1991. “The Relative Ef�ciency of Private and Public Schools in Developing Countries.� World Bank Research Observer, 6(2): 205–218. Kahyarara, Godius, and Francis Teal. 2008. “The Returns to Vocational Training and Academic Education: Evidence from Tanzania.� World Development, 36(11): 2223–2242. KRIVET. 2008. Pre-employment Vocational Education and Training in Korea. mimeo. Lechner, Michael. 2000. “An Evaluation of Public-Sector-Sponsored Continuous Vocational Training Programs in East Germany.� Journal of Human Resources, 35(2): 347– 375. Malamud, Ofer, and Cristian Pop-Eleches. 2008. General Education vs. Vocational Training: Evidence from an Economy in Transition. NBER Working Paper 14155. Cambridge, MA: National Bureau of Economic Research. 322 THE WORLD BANK ECONOMIC REVIEW Ministry of National Education. 2006. Rencana Strategis Departemen Pendidikan National Tahun 2005– 2009. Jakarta: Ministry of National Education. Moenjak, Thammarak, and Christopher Worswick. 2003. “Vocational education in Thailand: a study of choice and returns.� Economics of Education Review, 22:99–107. Neuman, Shoshana, and Adrian Ziderman. 1991. “Vocational Schooling, Occupational Matching, and Labor Market Earnings in Israel.� Journal of Human Resources, 26(2): 256– 281. Newhouse, David, and Kathleen Beegle. 2006. “The Effect of School Type on Academic Achievement.� Journal of Human Resources, 41(3): 529– 557. Strauss, John, Firman Witoelar, Bondan Sikoki, and Anna Marie Wattie. 2009. The Fourth Wave of the Indonesia Family Life Survey: Overview and Field Report Volume 1. RAND Labor and Population Working Paper WR-675/1-NIA/NICHD. Santa Monica, CA: RAND. Tobias, Justin L. 2003. “Are Returns to Schooling Concentrated among the Most Ables? A Semiparametric Analysis of the Ability-earnings Relationships.� Oxford Bulletin of Economics and Statistics, 65(1): 1 –29. World Bank. 2011. Indonesia Jobs Report. Jakarta: World Bank. ———. 2007. Investing in Indonesia’s Education: Allocation, Equity and Ef�ciency of Public Expenditures. Jakarta: World Bank. ———. 1995. Training and the Labor Market in Indonesia: Policies for Productivity Gains and Employment Growth. Washington, DC: World Bank. Disability and Poverty in Vietnam Daniel Mont and Nguyen Viet Cuong Disability is signi�cantly correlated with poverty in Vietnam, according to data from the 2006 Vietnam Household Living Standards Survey, especially when the extra costs of living with a disability are taken into account. This disability-poverty link is also associated with lower educational attainment, an important factor in determining poverty and productive economic activity in general, both for household-based businesses and wage employment. Not taking into account these associations and the extra costs of disability will make some poor disabled people invisible in poverty stat- istics and impede efforts to reduce poverty. JEL codes: I12, I31, O15 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Disability and poverty are linked in developing countries (Braithwaite and Mont 2009; Fujii 2008; Mete 2008; Hoogeveen 2005; Yeo and Moore 2003; Elwan 1999). However, quantitative research in this area is usually hampered by the lack of good quality data on disability and corresponding data on consumption and other socioeconomic indicators, such as years of education. Vietnam is an exception. The 2006 Vietnam Household Living Standards Survey (VHLSS) collected high-quality data on disability that are in line with new international recommendations, along with data on consumption and other socioeconomic indicators.1 Vietnam is thus a good case study for Daniel Mont (dmont@worldbank.org; corresponding author ) is a senior poverty specialist at the World Bank in Hanoi. Nguyen Viet Cuong (c_nguyenviet@yahoo.com) is a researcher at the National Economics University in Hanoi. This work was supported by the Governance and Poverty Policy Analysis and Advice Program trust fund established by UK Department for International Development to support the World Bank’s work on poverty, governance, and statistical capacity building in Vietnam. The �ndings, interpretations, and conclusions are those of the authors and do not necessarily represent the views of the International Bank for Reconstruction and Development/ The World Bank and its af�liated organizations. The authors would like to thank Mitchell Loeb, Aleksandra Posarac, Martin Rama, Kinnon Scott, and three anonymous referees for comments on earlier drafts. 1. The 2006 VHLSS was conducted by the General Statistics Of�ce of Vietnam with technical support from the Work Bank. The survey covered 9,189 households and 39,071 individuals. The 2006 sample is representative of rural and urban areas and eight geographic regions. The disability data were collected for people ages 5 and older (36,701 people). THE WORLD BANK ECONOMIC REVIEW, VOL. 25, NO. 2, pp. 323 –359 doi:10.1093/wber/lhr019 # The Author 2011. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 323 324 THE WORLD BANK ECONOMIC REVIEW exploring the relation between poverty and disability. A previous study of poverty in Vietnam in 2006 estimated that the poverty rate was 16.4 percent for people with disabilities compared with 13.5 percent for other people (Braithwaite and Mont 2009). This gap of nearly 3 percentage points probably underestimates the impact of living with a disability. The poverty line is based on a consump- tion level that represents a minimum standard of living for the general population, but studies show that a given consumption level does not trans- late into an equivalent standard of living for people with disabilities because of their extra costs of living (Tibble 2005; Zaidi and Burchardt 2005). These costs could include additional health services, assistive devices, per- sonal assistance (whether purchased or provided by family members, with the associated opportunity costs), and additional transportation costs, among others. According to Amartya Sen’s (1985, 1993, 1999) capabilities approach, Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 poverty is not merely a shortage of material goods but also a lack of the capa- bility of combining the resources at one’s disposal to reach a minimum stan- dard of living. In the United Kingdom, once the extra costs of living with a disability were accounted for, the relation between disability and poverty rose dramatically (Kuklys 2005). In that country, 23 percent of households with people with disabilities had less than 60 percent of the median income; when the additional costs of disability were taken into account, that percentage rose to more than 47 percent. One factor usually mentioned in studies of the correlation between poverty and disability is lack of access to schooling. In a study of 11 developing countries, Filmer (2008) found that disability explained a larger part of enroll- ment de�cits than any other characteristic examined, including gender and socioeconomic status. The relation between disability and school enrollment has been found in middle income countries as well (Mete 2008; Scott and Mete 2008). This article examines the relationships between disability and educational attainment and employment and the potential impact on people’s ability to live free of poverty. The article is structured as follows. Section I presents the de�nition of disability used in this article. Section II analyzes the pattern of disability and the relation between disability and poverty, employment, and education in Vietnam. Section III notes some implications of the study. I. MEASURING DISABILITY Disability is a complex phenomenon that has been measured many ways. In crafting survey questions on disability, this study follows the recommendations of the Washington Group on Disability Statistics, established by the UN Daniel Mont and Nguyen Viet Cuong 325 Statistical Commission (www.cdc.gov/nchs/washington_group.htm). The approach is similar to the model that underlies the International Classi�cation of Functioning, Disability, and Health, which focuses on people’s ability to take particular actions in their current environment (WHO 2011). This approach is also embodied in the social model of disability (Altman 2001; Shakespeare and Watson 1997). Disability is not synonymous with having a medical condition or func- tional limitation. Rather, disabilities are the result of an environment that erects barriers that prevent people from participating fully in the economic and social life of their communities (attending school, having a job, raising a family, participating in local governance, and so on). Thus, whether a person is considered to have a disability—and how mild or severe that disability is—depends strongly on the physical, cultural, and legal environment. In recent years, measuring disability has focused on measuring the dif�- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 culties people have performing various activities (Mont 2007a, b). The Washington Group on Disability Statistics, with the involvement of at least 50 countries, has recommended using the presence of dif�culties in a core set of basic activities—sight, hearing, walking, cognition, communication, and self-care—as an operational proxy for a person having a functional limitation that puts him or her at risk of being disabled in the social model sense (CDC 2011). The importance of looking at the ability to perform actions, rather than the presence of a medical condition, is also recommended in Gertler and Gruber (2002). Simply asking people if they have a disability tends to identify only people with the most severe disabil- ities (Mont 2007a). The threshold for when having dif�culty performing activities becomes a disability is not clearly de�ned. In fact, limitations in functioning can be quite smoothly distributed (Loeb and Mont 2010). The analysis in this article uses two different thresholds to examine the sensitivity of the results to a low threshold (DISLOW) and a higher one (DISHIGH), which excludes people with lesser dif�culties. DISLOW and DISHIGH are based on answers to the Washington Group’s recommended census questions on disability that were included in the 2006 VHLSS (see box 1). Following Loeb, Eide, and Mont (2008), the low threshold is de�ned as having some dif�culty in at least two of the functional domains noted in box 1 or having considerable dif�culty in one or more domains. The high threshold is de�ned as having considerable dif�culty in at least one of the six functional domains. Thus, the cases where DISHIGH equals 1 is a subset of the observations where DISLOW equals 1. 326 THE WORLD BANK ECONOMIC REVIEW BOX 1: De�ning Disability Disability questions Introductory phrase: The next questions ask about dif�culties you may have doing certain activities because of a HEALTH PROBLEM. 1. Do you have dif�culty seeing, even if wearing glasses? a. No – no dif�culty b. Yes – some dif�culty c. Yes – a lot of dif�culty d. Cannot do at all Remaining questions have same response categories. 2. Do you have dif�culty hearing, even if using a hearing aid? 3. Do you have dif�culty walking or climbing steps? 4. Do you have dif�culty remembering or concentrating? Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 5. Do you have dif�culty (with self-care such as) washing all over or dressing? Using your usual (customary) language, do you have dif�culty communi- cating, for example, understanding or being understood? DISLOW ¼ 1 if the respondent answers “some dif�culty� to at least two of the questions, or “a lot of dif�culty� or “cannot do at all� to at least one question, otherwise DISLOW ¼ 0. DISHIGH ¼ 1 if the respondent answers “a lot of dif�culty� or “cannot do at all� to at least one of the questions, otherwise DISHIGH ¼ 0. Source: Washington Group 2008. One reason why some dif�culty in at least two areas is one threshold used for de�ning low disability relates to the survey question on vision. Minor vision dif�culties, as measured by the survey question, have been positively cor- related with consumption in a number of countries. This is unlike severe vision dif�culties and dif�culties in all the other domains, which are negatively corre- lated with consumption (Mont and Loeb 2008). This positive correlation could reflect the fact that people in jobs requiring more education (in other words, involving literacy) are quicker to notice minor dif�culties in vision. For that reason, analysis was also conducted after discarding all vision dif�culties except being unable to see. While this strengthened the relationship between disability and poverty and lowered disability prevalence, it did not qualitatively affect the results. Therefore, analyses ignoring minor vision problems are not included here. (It should be noted that minor vision problems alone are not enough to categorize someone as having a disability.) Finally, the self-reporting nature of disability is a concern. Having dif�- culty undertaking a particular activity is inherently a subjective determi- nation. Scott and Mete (2008) �nd that the negative relationship between poverty and disability weakens when a lower threshold is used because Daniel Mont and Nguyen Viet Cuong 327 richer people are more inclined to report mild dif�culties, perhaps because of higher expectations for their ability to function. (There is a similar rise in reported minor health problems moving up the income distribution.) This is another reason for using two thresholds in the analysis: the self-reporting bias is probably less for having a lot of dif�culty or being unable to do something than for having only some dif�culty. However, research also shows that answers tend to be more consistent for questions about having dif�culties undertaking particular actions than for broader questions, such as “do you have a disability� or even “do you have a particular diagnosis,� since knowledge of diagnoses is associated with access to health services (Miller, et al., 2010). I I. A N A LY S I S Functional limitations are not rare. Almost 16 percent of the Vietnamese popu- Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 lation reported at least a little dif�culty in one of the six functional domains included in the 2006 VHLSS (table 1). As in other countries, the rate of func- tional limitations increases dramatically in middle age and reaches roughly two-thirds of the population over the age of 62. For people under 40, including children, the rate is 4 –5 percent. The gender difference is not large, no doubt due at least in part to women’s longer life expectancy and thus to more age-related disabilities. The vision domain has the largest percentage of people reporting functional dif�culties; self-care has the smallest, but people with restrictions in self-care tend to have the most severe disabilities. Except for vision, poor people and those in lower expenditure quintiles are more likely to have other functional limitations than are the nonpoor and people in high expenditure quintiles. As noted, richer people may be more likely to report vision problems as they are quicker to notice them because of the nature of their work or because the type of work they do causes more eyestrain. Dif�culties across functional domains are positively correlated. The corre- lation coef�cient for having functional dif�culties in different domains ranges from 0.2 to 0.6 (table A.1 in the appendix). Overall, there is not a large differ- ence in correlation coef�cients of disabilities between poor people and nonpoor people.2 For estimating prevalence rates, the Washington Group recommends using the presence of at least some dif�culty in functioning in any of the six func- tional domains (Washington Group 2008). That yields an overall disability rate in Vietnam of 15.7 percent, which is similar to reported disability prevalence 2. People are de�ned as poor if their per capita expenditure is lower than the general poverty line, as estimated by the World Bank and the General Statistics Of�ce of Vietnam. The poverty line is equivalent to the expenditure level that allows people to meet their nutritional needs (food consumption of 2,100 calories a day) and some essential nonfood consumption, such as clothing and housing. The poverty line in 2006 was 2.56 million dong. 328 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Percentage of People Reporting Functional Limitations by Area Remembering and Any Characteristic Seeing Hearing concentrating Walking Self-care Communicating dif�culty Total 11.36 3.29 4.74 6.03 1.93 2.71 15.74 (0.24) (0.12) (0.15) (0.16) (0.08) (0.11) (0.27) Gender Male 10.16 3.06 4.12 4.67 1.84 2.40 14.49 (0.27) (0.14) (0.17) (0.17) (0.11) (0.13) (0.31) Female 12.50 3.51 5.33 7.34 2.04 3.01 16.94 (0.29) (0.15) (0.19) (0.22) (0.11) (0.14) (0.32) Age 5 –18 1.86 0.47 1.11 0.68 1.19 1.12 4.29 (0.14) (0.07) (0.11) (0.08) (0.11) (0.12) (0.21) 19– 40 2.04 0.69 1.70 1.27 0.62 1.32 5.03 (0.16) (0.08) (0.12) (0.11) (0.07) (0.12) (0.23) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 41– 62 19.75 3.01 4.96 6.97 1.35 1.86 25.31 (0.59) (0.20) (0.30) (0.34) (0.13) (0.15) (0.63) Older than 54.16 23.11 27.45 38.91 11.04 15.54 66.84 62 (1.11) (0.89) (0.98) (1.05) (0.62) (0.75) (1.03) Urban/rural Urban 14.35 3.21 4.65 6.34 1.87 2.08 18.28 (0.63) (0.24) (0.32) (0.36) (0.16) (0.19) (0.66) Rural 10.26 3.31 4.77 5.92 1.96 2.94 14.81 (0.24) (0.13) (0.16) (0.18) (0.09) (0.13) (0.28) Region Red River 10.89 3.68 4.43 6.57 2.24 2.76 15.57 Delta (0.51) (0.27) (0.29) (0.37) (0.19) (0.22) (0.58) North East 11.54 3.41 5.11 6.06 1.67 2.63 16.47 (0.58) (0.29) (0.40) (0.41) (0.20) (0.32) (0.68) North West 7.29 2.37 2.76 3.33 1.02 1.82 11.02 (0.80) (0.45) (0.51) (0.57) (0.24) (0.40) (0.90) North 9.42 3.27 4.55 5.35 2.32 3.54 13.96 Central Coast (0.59) (0.33) (0.39) (0.45) (0.27) (0.34) (0.69) South Central 10.39 2.97 3.54 5.50 2.13 2.18 14.78 Coast (0.64) (0.31) (0.37) (0.49) (0.26) (0.26) (0.74) Central 10.07 3.10 5.12 5.65 1.81 3.03 14.24 Highlands (0.80) (0.42) (0.56) (0.58) (0.28) (0.44) (0.90) South East 13.71 3.28 6.46 7.17 2.07 3.11 18.27 (0.87) (0.35) (0.54) (0.52) (0.22) (0.34) (0.92) Mekong 12.57 3.14 4.30 5.73 1.51 2.11 16.23 River Delta (0.52) (0.23) (0.28) (0.30) (0.15) (0.19) (0.56) Poverty status Nonpoor 11.98 3.25 4.56 6.05 1.84 2.43 16.11 (Continued ) Daniel Mont and Nguyen Viet Cuong 329 TABLE 1. Continued Remembering and Any Characteristic Seeing Hearing concentrating Walking Self-care Communicating dif�culty (0.27) (0.13) (0.16) (0.18) (0.09) (0.11) (0.30) Poor 7.90 3.51 5.76 5.93 2.48 4.27 13.67 (0.44) (0.29) (0.38) (0.38) (0.22) (0.36) (0.59) Expenditure quintile Poorest 8.20 3.46 5.58 5.89 2.39 4.12 13.67 (0.39) (0.26) (0.33) (0.35) (0.20) (0.31) (0.52) Near poorest 10.43 3.80 4.83 6.13 1.85 2.61 15.01 (0.42) (0.26) (0.30) (0.33) (0.17) (0.21) (0.50) Middle 11.11 3.52 4.61 5.78 2.02 2.79 15.21 (0.46) (0.26) (0.31) (0.33) (0.18) (0.22) (0.53) Near richest 12.19 2.85 4.46 6.04 1.50 2.12 16.18 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (0.52) (0.23) (0.32) (0.35) (0.15) (0.20) (0.57) Richest 14.69 2.83 4.25 6.33 1.97 1.99 18.51 (0.71) (0.25) (0.35) (0.40) (0.19) (0.20) (0.73) Note: Numbers in parentheses are robust standard errors clustered at the commune level. The sample selection of the 2006 Vietnam Households Living Standards Survey follows a method of strati�ed random cluster sampling. The survey samples households in all rural and urban pro- vinces of Vietnam (rural and urban areas of all provinces are strata). There were 64 provinces in 2006 and 128 strata. In each stratum, communes were selected randomly as a primary sampling unit. In each commune, three households were selected randomly. This study uses Stata to calcu- late the standard errors for complex survey data (svy commands in Stata). The standard error computation takes into account the effects of survey design, such as sampling weights and corre- lation between households within a primary sampling unit. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. rates in many other countries that rely on a similar approach (for example, 12.2 percent for the United Kingdom, 14.5 percent for Brazil, 18.5 percent for Canada, and 19.4 percent for the United States; Mont 2007a). This study, however, follows Loeb, Eide, and Mont (2008) in using a more restrictive threshold for DISLOW, for two reasons: to avoid the problems associated with responses to the survey question on vision, and to reduce poss- ible false positives ( people who might have a low level of functional limitation in one domain that is unlikely to have a substantial impact on their life). Disabilities related to mental health are particularly dif�cult to capture in surveys and generally require detailed instruments that are not feasible for stan- dard household surveys. Therefore, people with psychological disabilities— especially mild and moderate ones—are probably not identi�ed by the ques- tions included in the VHLSS. People with severe mental disabilities that affect their ability to care for themselves are generally identi�ed by the survey ques- tions, but some people with psychological disabilities may be left out because of the episodic nature of some mental disabilities. To the extent that people 330 THE WORLD BANK ECONOMIC REVIEW T A B L E 2 . Percentage Disabled by Disability Threshold, Gender, and Age Counting vision dif�culties only if unable to see Blind People with Characteristic DISLOW DISHIGH DISLOW DISHIGH people low vision Total 7.56 3.60 5.33 2.97 0.22 11.14 (0.18) (0.12) (0.15) (0.11) (0.03) (0.24) Gender Male 6.57 3.19 4.72 2.70 0.21 9.95 (0.21) (0.15) (0.18) (0.14) (0.04) (0.27) Female 8.50 4.00 5.91 3.23 0.23 12.28 (0.24) (0.16) (0.20) (0.15) (0.04) (0.29) Age 5 –18 1.63 1.09 1.51 1.01 0.03 1.83 (0.13) (0.11) (0.13) (0.11) (0.02) (0.14) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 19– 40 2.02 1.54 1.69 1.35 0.10 1.94 (0.14) (0.12) (0.13) (0.12) (0.03) (0.16) 41– 62 8.49 3.26 4.81 2.49 0.12 19.62 (0.37) (0.20) (0.27) (0.18) (0.04) (0.59) Older than 62 45.20 20.59 33.02 16.83 1.59 52.57 (1.05) (0.79) (0.98) (0.72) (0.24) (1.10) Urban/rural Urban 7.56 3.40 5.16 2.75 0.22 14.13 (0.40) (0.25) (0.31) (0.23) (0.06) (0.62) Rural 7.55 3.68 5.39 3.05 0.22 10.04 (0.20) (0.13) (0.17) (0.12) (0.03) (0.24) Region Red River Delta 7.74 3.79 5.39 3.16 0.18 10.71 (0.39) (0.26) (0.31) (0.24) (0.05) (0.50) North East 7.43 2.95 5.22 2.48 0.18 11.36 (0.44) (0.26) (0.37) (0.24) (0.06) (0.58) North West 4.75 2.09 3.21 2.00 0.00 7.29 (0.65) (0.38) (0.54) (0.38) (0.00) (0.80) North Central Coast 7.05 3.81 5.56 3.35 0.20 9.22 (0.48) (0.34) (0.41) (0.31) (0.07) (0.58) South Central Coast 6.53 3.81 4.73 3.14 0.24 10.16 (0.50) (0.37) (0.42) (0.33) (0.09) (0.64) Central Highlands 7.49 3.39 5.37 2.53 0.17 9.90 (0.66) (0.42) (0.52) (0.35) (0.09) (0.79) South East 8.87 4.04 6.27 3.28 0.35 13.36 (0.64) (0.38) (0.49) (0.35) (0.10) (0.85) Mekong River Delta 7.58 3.50 4.98 2.76 0.24 12.33 (0.37) (0.22) (0.29) (0.20) (0.06) (0.52) Poverty status Nonpoor 7.39 3.44 5.11 2.81 0.20 11.79 (0.20) (0.12) (0.16) (0.11) (0.03) (0.27) Poor 8.50 4.50 6.56 3.85 0.36 7.54 (0.45) (0.33) (0.39) (0.30) (0.09) (0.42) Expenditure quintile Poorest 8.13 4.19 6.32 3.59 0.29 7.90 (Continued ) Daniel Mont and Nguyen Viet Cuong 331 TABLE 2. Continued Counting vision dif�culties only if unable to see Blind People with Characteristic DISLOW DISHIGH DISLOW DISHIGH people low vision (0.40) (0.28) (0.35) (0.25) (0.07) (0.38) Near poorest 7.87 3.75 5.52 3.11 0.19 10.24 (0.38) (0.26) (0.32) (0.23) (0.06) (0.42) Middle 7.53 3.71 5.16 3.01 0.19 10.92 (0.38) (0.25) (0.31) (0.23) (0.05) (0.46) Near richest 7.01 3.10 4.80 2.50 0.15 12.04 (0.37) (0.23) (0.31) (0.21) (0.05) (0.52) Richest 7.28 3.30 4.89 2.68 0.28 14.41 (0.45) (0.26) (0.35) (0.23) (0.08) (0.69) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Note: Numbers in parentheses are robust standard errors clustered at the commune level. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. with psychological disabilities are not accounted for, the prevalence rates are underestimated, as are any negative associations with employment, education, and poverty. Table 2 reports disability prevalence rates using DISHIGH and DISLOW for different categories. Rates are reported including and excluding minor or moderate vision dif�culties. The more restrictive threshold measure, DISHIGH, yields a disability rate of 3.6 percent for the general population, which falls to 2.97 percent when the only vision dif�culty included is blind- ness. The less restrictive threshold measure, DISLOW, yields a 7.56 percent prevalence rate, which falls to 5.33 percent when blindness is the only vision dif�culty considered. More than 29 percent of the people categorized as disabled by DISLOW are disabled because of minor vision dif�culties. Many of these people appear to acquire minor vision problems in middle age. The rate of low-vision dif�culties jumps tenfold between ages 19 –40 and ages 41 –62 and more than doubles after age 62. Patterns by gender and age are not affected by which de�nition of disability is used. Girls and women have a slightly higher disability rate, and disability increases with age, reaching very high levels after age 62. Nearly half of the elderly have a disability under the low threshold measure. Poor people and people in low expenditure quintiles have slightly higher rates of disability as measured by DISLOW and DISHIGH than nonpoor people and people in high expenditure quintiles. 332 THE WORLD BANK ECONOMIC REVIEW The number of households affected by disability is much larger than the prevalence rate. Using DISLOW, 23.4 percent of households include a person with a disability. Using DISHIGH, that percentage drops to 12.4. Thus, even with a more conservative measure of disability, nearly one in eight Vietnamese live in a household that includes a person with a disability, so the costs of dis- ability are borne by a broader population than those experiencing the disability directly. Household heads account for 44.7 percent of disabled people when the DISLOW threshold is used and 33.6 percent when DISHIGH is used (table A.2 in the appendix). The rate is higher than for the general population because household heads tend to be older and age is positively correlated with disability. As expected, households that include a person with a disability are over- represented in the lower consumption quintiles. When DISHIGHNV is used to exclude any problems with minor to moderate vision dif�culties, more Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 than 13 percent of households in the bottom consumption quintile include a person with a disability compared with about 10.5 percent for the general population. The correlation between low income and people with disabilities is heightened when people with mild and moderate vision dif�culties are excluded. That group is actually slightly underrepresented in the bottom quintile. As stated earlier, this may be the result of people who lack reading skills not registering age-related mild losses in vision. However, as noted earlier, the association between disability and poverty is probably understated because it fails to account for the associated costs of living with a disability. Cost of Disability The Zaidi and Burchardt (2005) method of accounting for disability, applied in Braithwaite and Mont (2009), was used to estimate the extra costs of dis- ability, with an expanded set of assets used in the asset index. That method begins by constructing an asset index as a measure of the standard of living, S, and regresses it on per capita expenditure, Y, disability status, D, and a vector of household characteristics, X: S ¼ a lnðY Þ þ bD þ cX þ 1: ð1Þ Then, the extra cost of disability is approximately equal to - b / a. Conceptually, the idea is that, with other household characteristics held con- stant, a certain level of expenditure (or income) is associated with a certain Daniel Mont and Nguyen Viet Cuong 333 T A B L E 3 . Poverty Rates by Disability Status and Other Characteristics, with and without the Extra Costs of Disability General poverty line Adjusted poverty line Characteristic Nondisabled people Disabled people Disabled people All 15.09 17.16 22.31 (0.50) (1.01) (1.12) Gender Male 14.60 17.46 22.55 (0.51) (1.30) (1.42) Female 15.57 16.94 22.13 (0.53) (1.13) (1.26) Age 5 – 18 19.29 31.08 36.24 (0.70) (3.97) (4.11) 19 – 40 15.14 24.72 31.42 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (0.53) (3.07) (3.26) 41 – 62 9.93 11.9 15.28 (0.46) (1.35) (1.51) Older than 62 14.45 17.01 22.82 (0.99) (1.23) (1.39) Urban/rural Urban 3.61 5.53 6.63 (0.58) (1.45) (1.51) Rural 19.32 21.44 28.09 (0.64) (1.24) (1.36) Number of observations 34,007 2,694 2,694 Note: Numbers in parentheses are robust standard errors clustered at the commune-level. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. standard of living as measured by asset holdings.3 If households that include people with disabilities at that level of expenditure have lower assets, the con- clusion is that the gap in assets is caused by the presence of the disability. Braithwaite and Mont (2009) used the seven most commonly held assets in their index, but this might not have accounted for people with higher levels of 3. Equation (1) is equivalent to: lnðY Þ ¼ ð1=aÞS À ðb=aÞD À ðc=aÞX À ðc=aÞ1: The difference in log per capita expenditure between people with disabilities and those without them is equal to: D ¼ lnðYD¼1 Þ À lnðYD¼0 Þ ¼ Àb=a: Thus, the extra cost is approximately equal to: ðYD¼1 À YD¼0 Þ=ðYD¼0 Þ % Àb=a: 334 THE WORLD BANK ECONOMIC REVIEW wealth who were able to purchase other assets. Therefore, the asset list was expanded.4 The extra costs of disability were estimated to be 11.5 percent, slightly higher but of similar magnitude to the just over 9 percent in Braithwaite and Mont.5 There can be a problem of endogeneity of explanatory variables in equation (1). Thus, rather than estimate the causal effect of the explanatory variables in equation (1), the estimates are used to examine the difference in expenditures between people with and without disabilities once asset holdings and other observed variables are controlled for. Table 3 compares poverty rates for people with and without disabilities by various characteristics, both with an unadjusted poverty line and with a poverty line adjusted for the extra costs of disability. Since many households that include a person with a disability are close to the poverty line, the poverty rates increase signi�cantly when the poverty line is adjusted for the 11.5 percent extra costs of disability, especially in rural areas. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Another striking �nding is the much higher rate of poverty for households that include a child or a prime age adult with disabilities (even without accounting for the extra costs of disability). Of course, the causality may go both ways. For example, poverty may lead to disability in children because of the lack of health care and proper nutrition. It may also lead to disability in adults, but it also probably limits their ability to generate a livelihood. The difference between the poverty rates for households without disabilities and those with elderly disabled members is not that high, which brings up another point: when examining the relation between disability and poverty, it is impor- tant to account not only for the severity of the disability but also for the age when it was acquired. A child with a disability might be denied access to edu- cation and training and might experience a lifetime of discrimination. An adult who acquires a disability might already have amassed certain skills and assets. And an elderly person who acquires a disability might be beyond working age, so although their families may incur certain costs in caring for them—including foregone labor income—their own earnings potential might not be relevant. Regression Analysis The three regressions reported in table 4 therefore account not only for the severity of disability (with separate regressions for DISLOW and DISHIGH) but also the age of onset of the disability—in particular, whether the disability began in childhood or adulthood. (The tables in this section present only the estimates of variables of interest such as DISLOW and DISHIGH. Full regression results are in the appendix.) 4. The assets included in this analysis are motorbike; wardrobe; bed; tables, chairs, and sofas; television; electric fan;. cooker; flush toilet; permanent house; and tapwater. The variable Y is per capita expenditure, D is DISLOW, and X is a vector of household characteristics. 5. The estimate of - b / a is 0.115 (0.2281/1.9768), with a standard error of 0.026. The regression of the asset index is in table A.3 in the appendix. T A B L E 4 . Regressions of Natural Log of per Capita Consumption Expenditure (Log of Thousand dong) Explanatory variables Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 DISLOW acquired before age 18 – 0.2468*** – 0.1327*** –0.1285*** (0.0352) (0.0272) (0.0242) DISLOW acquired since age 18 – 0.0427** – 0.0268* –0.0353*** (0.0189) (0.0158) (0.0132) DISLOW (two or more family 0.0472 –0.0213 members, one before age 18 and one since) (0.0710) (0.0519) DISHIGH acquired before age 18 – 0.2358*** –0.1222*** – 0.1314*** (0.0407) (0.0319) (0.0267) DISHIGH acquired since age 18 – 0.0530** –0.0082 – 0.0124 (0.0246) (0.0208) (0.0165) DISHIGH (two or more family –0.0593 – 0.0559 members, one before age 18 and one since) (0.0857) (0.0783) Control variables No Yes Yes No Yes Yes District �xed effects No No Yes No Yes *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level; n ¼ 9,189. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 335 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 336 THE WORLD BANK ECONOMIC REVIEW Model 1 includes only disability variables (no control variables). Model 2 includes several control variables. Model 3 includes district �xed-effects estima- tors,6 which help remove district variables that can affect both disability and household welfare, such as epidemics and calamities. The regressions show that households that include people with disabilities are more likely to be poor, regardless of the threshold of disability or when the disability was acquired. However, the impact of the age of onset of disability is striking. The relationship is much stronger and remains statistically signi�cant after controlling for a host of household characteristics. Disability variables could have endogeneity problems; for example, house- holds might have experienced shocks that made their members more likely to acquire a disability and that also lowered income and consumption. To examine this issue, regressions included disability variables with more clearly exogenous causes (war, accident, natural calamity, and birth defects) drawn from the survey question on the cause of the disability. For regressions of per Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 capita expenditure (table A.5 in the appendix) and other the outcome variables of employment and education, these disability variables have similar effects to the overall disability variables (and thus are not reported here). The age of onset also matters in the probability that a person with a disabil- ity is employed. With DISLOW, the negative impacts of having a disability are strongly signi�cant but are somewhat mitigated if that disability was acquired as an adult (table 5). The same cannot be said of people considered to have a disability only when the more severe DISHIGH threshold is used: regardless of age of onset, a disability reduces the probability of a person working by about the same amount. These results hold whether looking at wage employment or household employment for people ages 16–60. These regressions also show that completion of primary schooling is associ- ated with a greater likelihood of working for the household business, whereas completing secondary school tends to lead more often to wage work. After controlling for education, the results show that people with disabilities work less and that their disability is associated with less education to begin with. Enrollment rates are signi�cantly lower for children with disabilities than for children without disabilities. For example, primary school enrollment for children ages 6–12 is nearly 96 percent for children without disabilities, but about 69 percent for children with mild, moderate, or severe disabilities. The gap is even higher when the more restrictive threshold is used. Given the impor- tance of education to livelihoods, this puts children with disabilities, on average, at a livelihood disadvantage from the beginning. The same result is found in developed countries (Loprest and Maag 2007). 6. A log function of per capita expenditure is used because expenditure follows a log normal distribution rather than a normal distribution. Household characteristics that affect household earning include household composition, human assets, physical assets, and regional and commune characteristics (Glewwe 1991). T A B L E 5 . Logit Regressions of Employment by Degree of Disability for People Ages 16 –60 (Odds Ratios) Dependent variable (1 if yes; 0, otherwise) Worked for a Worked for the Had done Worked for a Worked for Had done Explanatory variables wage or salary household any work wage or salary the household any work DISLOW 0.3155*** 0.2618*** 0.0498*** (0.0648) (0.0399) (0.0083) DISLOW 1.3364 2.1272*** 3.0159*** from age 18 (0.3234) (0.3740) (0.5769) DISHIGH 0.2574*** 0.2258*** 0.0390*** (0.0624) (0.0396) (0.0073) DISHIGH 0.9081 1.3102 1.4810* from age 18 (0.3020) (0.2909) (0.3427) Control variables Yes Yes Yes Yes Yes Yes District �xed effects Yes Yes Yes Yes Yes Yes *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level; n ¼ 24,710. For the odds ratios in the logit regressions, ‘sig- ni�cant’ means ‘statistically different’ from one. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 337 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 338 T A B L E 6 . Logit Regressions of Education (Odds Ratios) Explanatory variable Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 Dependent variable is school enrollment, ages 6– 17 DISLOW 0.1571*** 0.0842*** 0.0775*** (0.0292) (0.0206) (0.0189) DISHIGH 0.1179*** 0.0682*** 0.0586*** (0.0270) (0.0200) (0.0173) Control variables No Yes Yes No Yes Yes District �xed effects No No Yes No No Yes Dependent variable is primary school completion, ages 18– 62 DISLOW by aged 10 0.0631*** 0.0269*** 0.0138*** (0.0117) (0.0075) (0.0033) THE WORLD BANK ECONOMIC REVIEW DISHIGH by aged 10 0.0560*** 0.0223*** 0.0103*** (0.0116) (0.0070) (0.0028) Control variables No Yes Yes No Yes Yes District �xed effects No No Yes No No Yes Dependent variable is secondary school completion, ages 18– 62 DISLOW by age 17 0.1520*** 0.1185*** 0.1161*** (0.0426) (0.0389) (0.0332) DISHIGH by age 17 0.1812*** 0.1488*** 0.1372*** (0.0536) (0.0500) (0.0424) Control variables No Yes Yes No Yes Yes District �xed effects No No Yes No No Yes *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. Enrolled in school ¼ 1; 0, otherwise. Source: Authors’ estimation based on the 2006 Vietnam Household Living Standards Survey. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Daniel Mont and Nguyen Viet Cuong 339 A series of logit regressions were estimated to explore the relation between disability and education in Vietnam. Table 6 presents the estimates for the dis- ability variable coef�cients; the full regressions are in tables A.8–A.10 in the appendix. For both DISLOW and DISHIGH, the correlation between disabil- ity and enrollment among school-age children is statistically signi�cant at the 1 percent signi�cance level and increases as more explanatory variables are added to the model. The odds ratios for the effect of disability on enrollment using the most inclusive speci�cation (model 2) is 0.084, which for DISLOW translates into children with disabilities being nearly 0.41 times less likely to attend school, once other factors are controlled for. For DISHIGH, this rises to nearly 0.47 times. In all speci�cations, having a disability in childhood signi�cantly reduces the chances of completing school for older cohorts, regardless of the de�nition of disability or the type of school. The odds ratios using the most inclusive speci�- cation (model 2) from the school completion logits are 0.027 ( primary) and Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 0.119 (secondary) for DISLOW and 0.022 and 0.149 for DISHIGH. All are signi�cant at 1 percent level. These results show only that disability is associated with a lack of education; they do not explain the reasons behind the relationship. It could be that parents want to send their children with disabilities to school but cannot, because of barriers such as transportation dif�culties, lack of accessible schools, and lack of training or acceptance by teachers. Or it could be that parents do not want to send the children to school because the returns to edu- cation for children with disabilities do not warrant the investment, whether for reasons of inherent inability to bene�t from schooling or of barriers to employ- ment (transportation, accessibility, attitudes, and so on) that prevent people with disabilities from getting a return to the human capital they acquired in school. In addition, poor parents might have limited resources for sending chil- dren to school and so may choose to spend them on their children with the highest expected returns, perceiving their children without disabilities as being the better investment. Whatever the case, the correlation between disability and education and employment reveals that people with functional limitations have poorer out- comes than their peers without functional limitations. III. CONCLUSION Disability, whether measured by a low or a high threshold, is signi�cantly cor- related with poverty and lack of employment in Vietnam, using data from the 2006 VHLSS. After accounting for the extra costs of disability, the correlation is stronger, especially in rural areas and for households with children and prime-age adults with disabilities. That correlation is also stronger for people with more severe disabilities but is lower for people who acquire their 340 THE WORLD BANK ECONOMIC REVIEW disabilities when they are adults. Disability during childhood is signi�cantly correlated with lack of educational attainment, an important determinant of poverty. Because of the endogeneity of this system—disability causing poverty, poverty causing disability, lack of returns in the labor market affecting schooling decisions, barriers to education affecting employment opportu- nities—it is dif�cult to attribute causality. It is clear, however, that people with disabilities face more dif�cult and limited conditions and that the limited conditions are highly signi�cant when the extra costs of living with a disability are taken into account. For households with people with disabil- ities, ignoring the extra costs of disability means that poverty statistics can miss these households whose standard of living, if the higher costs were taken into account, would be equal to that of poor households without people with disabilities . Better data on disability need to be collected in conjunction with data on Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 consumption and other measures of well-being. A clearer understanding of the relationship between disability and poverty and the barriers that disabled people face in fully participating in economic life will help policymakers deter- mine where the link between disability and poverty is strongest and where the most promising and appropriate avenues are for designing interventions to weaken that link. REFERENCES Altman, B.M. 2001. “Disability De�nitions, Models, Classi�cation Schemes, and Applications.� In Handbook of Disability Studies, ed. G.L. Albrecht, K.D. Seelman, and M. Bury, 97 –122. Thousand Oaks, CA: Sage Publications. Braithwaite, J., and D. Mont. 2009. “Disability and Poverty: A Survey of World Bank Poverty Assessments and Implications.� Alter: European Journal of Disability Research 3: 219–32 CDC (Centers for Disease Control and Prevention). (2011). “Washington Group on Disability Statistics,� www.cdc.gov/nchs/washington_group.htm. Elwan, A. 1999. “Poverty and Disability: A Survey of the Literature.� SP Discussion Paper 9932, World Bank, Washington, DC. Fujii, T. 2008. “Two-Sample Estimation of Poverty Rates for Disabled People: An Application to Tanzania.� Economics and Statistics Working Paper 02-2008, Singapore Management University, Singapore. Filmer, D. 2008. “Disability, Poverty, and Schooling in Developing Countries: Results from 11 Household Surveys.� The World Bank Economic Review 22 (1):141–63. Gertler, P., and J. Gruber. 2002. “Insuring Consumption against Illness.� The American Economic Review 92 (1): 51– 70. ˆ te d’Ivoire.� Journal of Glewwe, P. 1991. “Investigating the Determinants of Household Welfare in Co Development Economics 35: 307– 37. Hoogeveen, J. 2005. “Measuring Welfare for Small but Vulnerable Groups: Poverty and Disability in Uganda.� Journal of African Economies 14 (4): 603– 31. Kuklys, W. 2005. Amartya Sen’s Capability Approach: Theoretical Insights and Empirical Applications. Studies in Choice and Welfare. New York: Springer-Verlag. Loeb, M.E., A.H. Eide, and D. Mont. 2008. “Approaching the Measurement of Disability Prevalence: The Case of Zambia.� Alter: European Journal of Disability Research 2 (1): 32– 43. Daniel Mont and Nguyen Viet Cuong 341 Loeb, M.E., and D. Mont. 2010. “A Functional Approach to Assessing Health Impacts on People with Disabilities.� Alter: European Journal of Disability Research 4 (3): 159–73. Loprest, P., and E. Maag, 2007. “The Relationship between Early Disability Onset and Education and Employment.� Journal of Vocational Rehabilitation 26 (1): 49– 62. C. Mete, ed. 2008. Economic Implications of Chronic Illness and Disease in Eastern Europe and the Former Soviet Union. Washington, DC: The World Bank. Miller, K., D. Mont, J. Madans, B. Altman, and A. Maitland 2010. “Results of a Cross-National Structured Cognitive Interviewing Protocol to Test Measures of Disability.� Quantity and Quality. 45 (4): 801–815. Mont, D. 2007a. “Measuring Disability Prevalence.� SP Discussion Paper 0706. World Bank, Washington, DC. ———. 2007b. “Measuring Health and Disability.� The Lancet 369: 1658– 63. Mont, D., and M. Loeb. 2008. “Beyond DALYs: Developing Indicators to Assess the Impact of Public Health Interventions on the Lives of People with Disabilities.� SP Discussion Paper 0815. World Bank, Washington, DC. Scott, K., and C. Mete. 2008. “Measurement of Disability and Linkages with Welfare, Employment, and Schooling: The Case of Uzbekistan.� In Economic Implications of Chronic Illness and Disease Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 in Eastern Europe and the Former Soviet Union, ed. C. Mete. Washington, DC: World Bank. Sen, A. 1985. Commodities and Capabilities. Amsterdam: North Holland. ———. 1993. “Capability and Well-being.� In The Quality of Life, ed. M. Nussbaum, and A.K. Sen. Oxford: Clarendon Press. ———. 1999. Development as Freedom. Oxford: Oxford University Press. Shakespeare, T., and N. Watson. 1997. “Defending the Social Model.� Disability and Society 12 (2): 293–300. Tibble, M. 2005. “Review of the Existing Research on the Extra Costs of Disability.� Working Paper 21. Department for Work and Pensions, Leeds, UK. Washington Group. 2008, “The Measurement of Disability: Recommendations for the 2010 Round of Censuses.�Position paper available on www.cdc.gov/nchs/washington_group.htm. WHO (World Health Organization). 2011. “International Classi�cation of Functioning, Disability and Health (ICF).� World Health Organization, http://www.who.int/classi�cations/icf/en/. Yeo, R., and K. Moore. 2003. “Including Disabled People in Poverty Reduction Work: Nothing about Us, without Us.� World Development 31 (3): 571– 90 Zaidi, A., and T. Burchardt. 2005. “Comparing Incomes When Needs Differ: Equivalization for the Extra Costs of Disability in the U.K.� Review of Income and Wealth 51 (1): 89– 114. 342 THE WORLD BANK ECONOMIC REVIEW APPENDIX T A B L E A 1 . Correlation Coef�cients of Functional Limitations for the Poor and Nonpoor Remembering Functional and limitation Seeing Hearing concentrating Walking Self-care Communicating Poor Seeing 1 Hearing 0.432*** 1 Remembering 0.381*** 0.413*** 1 and concentrating Walking 0.523*** 0.410*** 0.459*** 1 Self-care 0.204*** 0.250*** 0.363*** 0.422*** 1 Communicating 0.265*** 0.332*** 0.665*** 0.415*** 0.463*** 1 Nonpoor Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Seeing 1 Hearing 0.348*** 1 Remembering 0.363*** 0.448*** 1 and concentrating Walking 0.401*** 0.393*** 0.487*** 1 Self-care 0.198*** 0.287*** 0.377*** 0.400*** 1 Communicating 0.227*** 0.405*** 0.605*** 0.387*** 0.513*** 1 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. T A B L E A 2 . Relation between People with DISLOW and DISHIGH and Household Head Age of disabled Head Head’s spouse Head’s children Head’s parent Others DISLOW Before age 18 0.00 0.00 85.63 0.00 14.37 (0.00) (0.00) (2.83) (0.00) (2.83) 18– 30 4.00 1.56 90.33 0.00 4.12 (1.68) (0.91) (2.56) (0.00) (1.80) 31– 40 30.61 13.56 41.82 0.00 14.01 (4.38) (3.27) (5.06) (0.00) (3.93) 41– 50 53.05 30.34 8.84 0.76 7.01 (3.10) (2.91) (2.17) (0.54) (1.73) Older than age 50 51.09 22.08 0.47 23.53 2.83 (1.04) (0.88) (0.17) (1.11) (0.42) Total 44.69 20.05 13.19 17.57 4.50 (0.89) (0.75) (0.71) (0.84) (0.46) DISHIGH Before age 18 0.00 0.00 84.9 0.00 15.1 (0.00) (0.00) (3.63) (0.00) (3.63) 18– 30 3.31 0.49 91.96 0.00 4.25 (Continued ) Daniel Mont and Nguyen Viet Cuong 343 TABLE A2. Continued Age of disabled Head Head’s spouse Head’s children Head’s parent Others (1.71) (0.49) (2.68) (0.00) (2.07) 31 – 40 23.49 10.39 50.69 0.00 15.43 (4.91) (3.67) (6.38) (0.00) (5.07) 41 – 50 49.91 17.13 16.45 0.89 15.63 (5.16) (3.71) (4.59) (0.88) (3.79) Older than age 50 44.9 18.73 0.95 31.3 4.12 (1.66) (1.35) (0.38) (1.70) (0.69) Total 36.33 14.8 20.9 21.22 6.75 (1.33) (1.02) (1.23) (1.23) (0.77) *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Number in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining when having dif�culty performing activi- ties becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. T A B L E A 3 . Regression of Asset Index Explanatory variable Coef�cient Standard error p.t Log of per capita expenditure 1.9768 0.0512 0.0000 DISLOW – 0.2281 0.0508 0.0000 Ratio of children (before age 15) – 0.3940 0.1096 0.0000 Ratio of elderly (after age 60) – 0.3389 0.0898 0.0000 Household size 0.2366 0.0163 0.0000 constant – 9.8166 0.4585 0.0000 R-squared 0.345 Number of observations 9,189 Note: DISLOW is the lower of two thresholds for determining when having dif�culty per- forming activities becomes a disability (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. T A B L E A 4 . Regressions of Natural Logarithm of Per Capita Consumption Expenditure (Log of Thousands of Dong) 344 Model 1: ordinary Model 2: ordinary Model 3: district Model 1: ordinary Model 2: ordinary Model 3: district Explanatory variable least squares least squares �xed effects least squares least squares �xed effects DISLOW Before age 18 – 0.2468*** – 0.1327*** – 0.1285*** (0.0352) (0.0272) (0.0242) 18 and older – 0.0427** – 0.0268* – 0.0353*** (0.0189) (0.0158) (0.0132) Two or more family members, 0.0472 – 0.0213 one before age 18 and one (0.0710) (0.0519) 18 or older DISHIGH Before age 18 – 0.2358*** – 0.1222*** – 0.1314*** (0.0407) (0.0319) (0.0267) 18 and older – 0.0530** – 0.0082 – 0.0124 THE WORLD BANK ECONOMIC REVIEW (0.0246) (0.0208) (0.0165) Two or more family members, – 0.0593 – 0.0559 one before age 18 and 18 or (0.0857) (0.0783) older Urban (yes ¼ 1) 0.3941*** 0.1946*** 0.3943*** 0.1955*** (0.0158) (0.0150) (0.0158) (0.0150) Household in Red River Delta Omitted Household in North East – 0.1444*** – 0.1464*** (0.0193) (0.0193) Household in North West – 0.2886*** – 0.2894*** (0.0329) (0.0329) Household in North Central – 0.2272*** – 0.2276*** Coast (0.0204) (0.0204) Household in South Central 0.0151 0.016 Coast Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (0.0209) (0.0208) Household in Central – 0.03 – 0.0323 Highlands (0.0324) (0.0325) Household in South East 0.3243*** 0.3230*** (0.0224) (0.0224) Household in Mekong River 0.1677*** 0.1678*** Delta (0.0188) (0.0188) Head age 0.0105*** 0.0067*** 0.0107*** 0.0071*** (0.0032) (0.0025) (0.0032) (0.0025) Head age squared * 1,000 – 0.0707** – 0.0439* – 0.0741** – 0.0489** (0.0304) (0.0237) (0.0306) (0.0237) Head without education Omitted degree Head with primary school 0.1740*** 0.1637*** 0.1738*** 0.1638*** degree (0.0161) (0.0131) (0.0161) (0.0131) Head with lower secondary 0.2953*** 0.2897*** 0.2961*** 0.2910*** school (0.0174) (0.0143) (0.0174) (0.0143) Head with upper secondary 0.4799*** 0.4235*** 0.4810*** 0.4246*** school (0.0249) (0.0202) (0.0249) (0.0203) Head with technical degree 0.5657*** 0.5206*** 0.5667*** 0.5223*** (0.0218) (0.0182) (0.0218) (0.0182) Head with post – secondary 0.8632*** 0.7713*** 0.8652*** 0.7739*** school (0.0295) (0.0247) (0.0294) (0.0247) Ratio of children (before age – 0.5144*** – 0.5352*** – 0.5124*** – 0.5311*** Daniel Mont and Nguyen Viet Cuong 15) (Continued ) 345 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 TABLE A4. Continued 346 Model 1: ordinary Model 2: ordinary Model 3: district Model 1: ordinary Model 2: ordinary Model 3: district Explanatory variable least squares least squares �xed effects least squares least squares �xed effects (0.0329) (0.0282) (0.0331) (0.0282) Ratio of elderly (after age 60) – 0.2150*** – 0.2082*** – 0.2241*** – 0.2202*** (0.0331) (0.0267) (0.0325) (0.0265) Household size – 0.0610*** – 0.0578*** – 0.0616*** – 0.0589*** (0.0040) (0.0032) (0.0040) (0.0032) Constant 8.5406*** 8.1967*** 8.3724*** 8.5339*** 8.1894*** 8.3625*** (0.0082) (0.0845) (0.0665) (0.0076) (0.0849) (0.0667) R-squared 0.01 0.48 0.30 0.01 0.48 0.30 Number of observations 9,189 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining THE WORLD BANK ECONOMIC REVIEW when having dif�culty performing activities becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 T A B L E A 5 . Regressions of Natural Logarithm of Per Capita Consumption Expenditure (Log of Thousand Dong): Disabilities with more Exogenous Causes Model 2: Model 1: Model 2: Model 1: ordinary ordinary least Model 3: district ordinary least ordinary least Model 3: district Explanatory variable least squares squares �xed effects squares squares �xed effects DISLOW Before age 18 –0.3004*** – 0.1596*** – 0.1422*** (0.0420) (0.0350) (0.0313) Age 18 and older –0.0291 – 0.0215 – 0.0316 (0.0345) (0.0297) (0.0253) Two or more family members, 0.0207 – 0.0449 one before age 18 and one (0.0958) (0.0678) 18 or older) DISHIGH Before age 18 – 0.3060*** – 0.1537*** – 0.1452*** (0.0479) (0.0410) (0.0332) Age 18 or older – 0.0289 – 0.0004 – 0.0022 (0.0439) (0.0382) (0.0309) Two or more family members, – 0.111 – 0.1745 one before age 18 and one (0.1156) (0.1080) 18 or older Urban (yes ¼ 1) 0.3943*** 0.1962*** 0.3947*** 0.1967*** (0.0158) (0.0150) (0.0158) (0.0150) Household in Red River Delta Omitted Household in North East – 0.1452*** – 0.1459*** (0.0193) (0.0193) Household in North West – 0.2900*** – 0.2898*** (0.0329) (0.0330) Daniel Mont and Nguyen Viet Cuong (Continued ) 347 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 TABLE A5. Continued 348 Model 2: Model 1: Model 2: Model 1: ordinary ordinary least Model 3: district ordinary least ordinary least Model 3: district Explanatory variable least squares squares �xed effects squares squares �xed effects Household in North Central – 0.2267*** – 0.2273*** Coast (0.0204) (0.0204) Household in South Central 0.0153 0.0156 Coast (0.0209) (0.0209) Household in Central – 0.0322 – 0.0338 Highlands (0.0323) (0.0325) Household in South East 0.3221*** 0.3219*** (0.0224) (0.0224) THE WORLD BANK ECONOMIC REVIEW Household in Mekong River 0.1673*** 0.1675*** Delta (0.0188) (0.0188) Head age 0.0109*** 0.0074*** 0.0108*** 0.0072*** (0.0032) (0.0025) (0.0032) (0.0025) Head age squared * 1000 – 0.0760** – 0.0513** – 0.0746** – 0.0502** (0.0302) (0.0236) (0.0303) (0.0236) Head without education degree Omitted Head with primary school 0.1731*** 0.1636*** 0.1728*** 0.1633*** degree (0.0161) (0.0131) (0.0162) (0.0131) Head with lower secondary 0.2951*** 0.2904*** 0.2956*** 0.2910*** school (0.0174) (0.0143) (0.0174) (0.0143) Head with upper secondary 0.4790*** 0.4226*** 0.4799*** 0.4236*** school Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (0.0249) (0.0203) (0.0249) (0.0203) Head with technical degree 0.5654*** 0.5210*** 0.5656*** 0.5216*** (0.0218) (0.0182) (0.0218) (0.0182) Head with postsecondary 0.8642*** 0.7728*** 0.8644*** 0.7737*** school (0.0295) (0.0247) (0.0295) (0.0247) Ratio of children (before – 0.5114*** – 0.5313*** – 0.5120*** – 0.5305*** age 15) (0.0329) (0.0281) (0.0329) (0.0282) Ratio of elderly (after age 60) – 0.2242*** – 0.2196*** – 0.2263*** – 0.2229*** (0.0322) (0.0263) (0.0322) (0.0263) Household size – 0.0620*** – 0.0591*** – 0.0620*** – 0.0594*** (0.0040) (0.0032) (0.0040) (0.0032) Constant 8.5302*** 8.1885*** 8.3583*** 8.5279*** 8.1908*** 8.3603*** (0.0072) (0.0843) (0.0665) (0.0072) (0.0843) (0.0665) R-squared 0.01 0.48 0.30 0.01 0.48 0.30 Number of observation 9,189 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining when having dif�culty performing activities becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 349 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 T A B L E A 6 . Logit Regressions of Employment by Degree of Disability on People ages 16 –60 (Odds Ratios) 350 Worked for a wage or Worked for the Had done any Worked for a wage or Worked for the Had done any Explanatory variable salary household work salary household work DISLOW 0.3506*** 0.2985*** 0.0670*** (0.0815) (0.0497) (0.0149) DISLOW at age 18 or 1.2843 1.9983*** 2.7735*** older (0.3460) (0.3647) (0.7000) DISHIGH 0.2939*** 0.2554*** 0.0514*** (0.0834) (0.0512) (0.0134) DISHIGH at age 18 or 0.8359 1.1853 1.1520 older (0.3156) (0.2863) (0.3559) Age 1.2170*** 1.2483*** 1.9014*** 1.2216*** 1.2507*** 1.9271*** (0.0120) (0.0107) (0.0259) (0.0121) (0.0108) (0.0260) Age squared 0.9971*** 0.9977*** 0.9921*** 0.9970*** 0.9977*** 0.9918*** THE WORLD BANK ECONOMIC REVIEW (0.0001) (0.0001) (0.0002) (0.0001) (0.0001) (0.0002) No education diploma Omitted Completion of primary 0.8400*** 1.2230*** 1.1129 0.8429*** 1.2215*** 1.1087 school (0.0483) (0.0620) (0.0937) (0.0482) (0.0619) (0.0936) Completion of secondary 2.1238*** 0.3404*** 0.4674*** 2.1400*** 0.3408*** 0.4701*** school (0.1366) (0.0203) (0.0412) (0.1372) (0.0203) (0.0414) Sex (male ¼ 1; female ¼ 0) 1.9721*** 0.6736*** 1.5603*** 1.9840*** 0.6779*** 1.6090*** (0.0601) (0.0205) (0.0646) (0.0607) (0.0207) (0.0671) Urban (yes ¼ 1) 1.6867*** 0.3918*** 0.4540*** 1.6861*** 0.3909*** 0.4480*** (0.0779) (0.0194) (0.0247) (0.0779) (0.0193) (0.0243) Household size 0.9714** 1.0558*** 1.0285* 0.9720** 1.0571*** 1.0330** (0.0133) (0.0131) (0.0159) (0.0133) (0.0132) (0.0158) Household in Red River Omitted Delta Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Household in North East 0.5020*** 2.3173*** 1.4348*** 0.4975*** 2.3023*** 1.4083*** (0.0345) (0.1617) (0.1125) (0.0342) (0.1607) (0.1108) Household in North West 0.3307*** 3.5289*** 1.8697*** 0.3288*** 3.5075*** 1.8368*** (0.0355) (0.3822) (0.2593) (0.0351) (0.3802) (0.2577) Household in North 0.5257 1.5660*** 0.8046** 0.5255*** 1.5653*** 0.8103** Central Coast (0.0398) (0.1121) (0.0690) (0.0398) (0.1121) (0.0690) Household in South 0.9735 0.9370 0.8653 0.9739 0.9386 0.8791 Central Coast (0.0741) (0.0763) (0.0802) (0.0742) (0.0764) (0.0817) Household in Central 0.5375*** 2.0206*** 1.2969** 0.5331*** 1.9983*** 1.2589** Highlands (0.0549) (0.2081) (0.1454) (0.0544) (0.2054) (0.1396) Household in South East 1.2887*** 0.6487*** 0.7173*** 1.2825*** 0.6442*** 0.7016*** (0.0902) (0.0479) (0.0580) (0.0894) (0.0476) (0.0563) Household in Mekong 1.0313 0.9636 0.9023 1.0285 0.9606 0.8946 River Delta (0.0651) (0.0589) (0.0660) (0.0647) (0.0589) (0.0650) Number of observations 24,710 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining when having dif�culty performing activities becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 351 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 352 T A B L E A 7 . District Fixed Effects Logit Regression of Employment by Degree of Disability on People Ages 16–60 (Odds Ratios) Worked for a wage or Worked for the Had done any Worked for a wage Worked for the Had done any Explanatory variable salary household work or salary household work THE WORLD BANK ECONOMIC REVIEW DISLOW 0.3155*** 0.2618*** 0.0498*** (0.0648) (0.0399) (0.0083) DISLOW at age 18 1.3364 2.1272*** 3.0159*** or older (0.3234) (0.3740) (0.5769) DISHIGH 0.2574*** 0.2258*** 0.0390*** (0.0624) (0.0396) (0.0073) DISHIGH at age 18 0.9081 1.3102 1.4810* or older (0.3020) (0.2909) (0.3427) Age 1.2273*** 1.2652*** 1.9576*** 1.2320*** 1.2673*** 1.9792*** (0.0109) (0.0100) (0.0243) (0.0108) (0.0100) (0.0247) Age squared 0.9970*** 0.9976*** 0.9917*** 0.9969*** 0.9975*** 0.9915*** (0.0001) (0.0001) (0.0002) (0.0001) (0.0001) (0.0002) No education Omitted diploma Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Completion of 0.8129*** 1.2898*** 1.1420* 0.8155*** 1.2892*** 1.1425* primary school (0.0407) (0.0598) (0.0854) (0.0408) (0.0599) (0.0860) Completion of 1.9437*** 0.3825*** 0.4851*** 1.9599*** 0.3839*** 0.4920*** secondary school (0.1088) (0.0207) (0.0393) (0.1096) (0.0207) (0.0400) Sex (male ¼ 1; 2.0499*** 0.6483*** 1.5674*** 2.0620*** 0.6517*** 1.6032*** female ¼ 0) (0.0654) (0.0200) (0.0676) (0.0658) (0.0201) (0.0694) Urban (yes ¼ 1) 1.4263*** 0.5270*** 0.5532*** 1.4262*** 0.5270*** 0.5526*** (0.0732) (0.0268) (0.0388) (0.0732) (0.0268) (0.0388) Household size 0.9691*** 1.0447*** 1.0070 0.9703*** 1.0464*** 1.0137 (0.0103) (0.0104) (0.0139) (0.0103) (0.0105) (0.0140) Number of 23,938 24,383 24,037 23,938 24,383 24,037 observations Number of districts 587 607 585 587 607 585 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining when having dif�culty performing activities becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 353 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 T A B L E A 8 . Enrollment Logit Results, Children Ages 6–17 (Odds Ratios) 354 Model 3: district �xed effect Model 3: district �xed effect Explanatory variable Model 1: logit Model 2: logit logit Model 1: logit Model 2: logit logit DISLOW 0.1571*** 0.0842*** 0.0775*** (0.0292) (0.0206) (0.0189) DISHIGH 0.1179*** 0.0682*** 0.0586*** (0.0270) (0.0200) (0.0173) Age 0.6811*** 0.6907*** 0.6825*** 0.6928*** (0.0129) (0.0097) (0.0130) (0.0097) Sex (male ¼ 1; female ¼ 0) 0.7819*** 0.8106*** 0.7811*** 0.8057*** (0.0516) (0.0559) (0.0516) (0.0556) Urban (yes ¼ 1) 1.3730*** 1.3703*** 1.3648*** 1.3716*** (0.1634) (0.1822) (0.1610) (0.1824) Per capita income (million 1.0736*** 1.0800*** 1.0725*** 1.0790*** dong) (0.0161) (0.0130) (0.0161) (0.0129) THE WORLD BANK ECONOMIC REVIEW Household size 0.9371*** 0.9250*** 0.9352*** 0.9240*** (0.0234) (0.0213) (0.0234) (0.0213) Household in Red River Delta Omitted Household in North East 1.1286 1.1331 (0.1625) (0.1620) Household in North West 0.7711 0.7819 (0.1534) (0.1556) Household in North Central 0.8336 0.8344 Coast (0.1242) (0.1252) Household in South Central 1.1343 1.1366 Coast (0.1872) (0.1853) Household in Central 0.7460 0.7619 Highlands (0.1358) (0.1387) Household in South East 0.5678*** 0.5781*** Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 (0.0903) (0.0919) Household in Mekong River 0.5385*** 0.5466*** Delta (0.0716) (0.0727) Head without education Omitted degree Head with primary school 1.6258*** 1.6487*** 1.6177*** 1.6356*** degree (0.1463) (0.1500) (0.1456) (0.1488) Head with lower –secondary 2.8462*** 2.7594*** 2.8434*** 2.7927*** school (0.3102) (0.3118) (0.3099) (0.3156) Head with upper secondary 4.5997*** 4.0878*** 4.5951*** 4.0552*** school (0.8923) (0.8216) (0.8915) (0.8151) Head with technical degree 8.1011*** 7.6141*** 8.2482*** 7.8381*** (2.2764) (1.7817) (2.3425) (1.8420) Head with post–secondary 7.9486*** 8.8110*** 7.8538*** 8.8640*** school (3.5292) (4.0266) (3.5421) (4.0509) Number of observations 9,880 9,880 8,352 9,880 9,880 8,352 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. DISLOW is the lower of two thresholds for determining when having dif�culty performing activities becomes a disability; DISHIGH, the higher threshold, excludes people with lesser dif�culties. DISLOW and DISHIGH (see box 1). Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 355 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 T A B L E A 9 . Primary School Completion Logits by Disability Status at age 10, Adults Ages 18 –62 (Odds Ratios) 356 Model 3: district �xed Model 3: district �xed Explanatory variable Model 1: logit Model 2: logit effect logit Model 1: logit Model 2: logit effect logit DISLOW before age 10 0.0631*** 0.0269*** 0.0138*** (0.0117) (0.0075) (0.0033) DISHIGH before age 10 0.0560*** 0.0223*** 0.0103*** (0.0116) (0.0070) (0.0028) Age 0.9512*** 0.9427*** 0.9512*** 0.9427*** (0.0019) (0.0019) (0.0019) (0.0019) Sex (male ¼ 1; female ¼ 0) 1.6871*** 1.8349*** 1.6837*** 1.8331*** (0.0624) (0.0752) (0.0623) (0.0752) Urban (yes ¼ 1) 2.2910*** 2.7020*** 2.2887*** 2.6885*** (0.1833) (0.1999) (0.1831) (0.1963) Per capita income (million 1.1377*** 1.1480*** 1.1388*** 1.1491*** dong) THE WORLD BANK ECONOMIC REVIEW (0.0137) (0.0069) (0.0137) (0.0069) Household size 0.9194*** 0.9734** 0.9213*** 0.9753*** (0.0156) (0.0117) (0.0157) (0.0117) Household in Red River Omitted Delta Household in North East 0.2549*** 0.2541*** (0.0285) (0.0285) Household in North West 0.0883*** 0.0887*** (0.0123) (0.0123) Household in North Central 0.5262*** 0.5278*** Coast (0.0700) (0.0702) Household in South Central 0.2078*** 0.2095*** Coast (0.0260) (0.0264) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Household in Central 0.1116*** 0.1109*** Highlands (0.0152) (0.0151) Household in South East 0.1143*** 0.1147*** (0.0133) (0.0134) Household in Mekong River 0.0669*** 0.0671*** Delta (0.0064) (0.0065) Number of observations 23,012 23,012 21,368 23,012 23,012 21,368 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 357 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 T A B L E A 1 0 . Secondary School Completion Logits by Disability Status, Adults Ages 18–62 (Odds Ratios) 358 Model 3: district �xed Model 3: district �xed Explanatory variable Model 1: logit Model 2: logit effect logit Model 1: logit Model 2: logit effect logit DISLOW before age 17 0.1520*** 0.1185*** 0.1161*** (0.0426) (0.0389) (0.0332) DISHIGH before age 17 0.1812*** 0.1488*** 0.1372*** (0.0536) (0.0500) (0.0424) Age 0.9550*** 0.9522*** 0.9550*** 0.9531*** (0.0019) (0.0019) (0.0019) (0.0019) Sex (male ¼ 1; 1.4492*** 1.5235*** 1.4477*** 1.5204*** female ¼ 0) (0.0449) (0.0518) (0.0449) (0.0517) Urban (yes ¼ 1) 3.2871*** 3.1740*** 3.2838*** 3.1772*** (0.1972) (0.1746) (0.1970) (0.1747) Per capita income 1.1052*** 1.1041*** 1.1052*** 1.1041*** THE WORLD BANK ECONOMIC REVIEW (million dong) (0.0077) (0.0033) (0.0077) (0.0033) Household size 0.9380*** 0.9550*** 0.9389*** 0.9550*** (0.0141) (0.0115) (0.0141) (0.0115) Household in Red River Omitted Delta Household in North 0.7672*** 0.7664*** East (0.0598) (0.0598) Household in North 0.4848*** 0.4853*** West (0.0664) (0.0665) Household in North 0.8033** 0.8033** Central Coast (0.0707) (0.0707) Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Household in South 0.5764*** 0.5781*** Central Coast (0.0530) (0.0532) Household in Central 0.4185*** 0.4173*** Highlands (0.0477) (0.0476) Household in South 0.3910*** 0.3914*** East (0.0352) (0.0352) Household in Mekong 0.2161*** 0.2165*** River Delta (0.0177) (0.0178) Number of observations 23,012 23,012 22,247 23,012 23,012 22,247 *** Signi�cant at p , .01; ** signi�cant at p , .05; signi�cant at p , .1. Note: Numbers in parentheses are robust standard errors clustered at the commune level. Source: Authors’ analysis based on the 2006 Vietnam Household Living Standards Survey. Daniel Mont and Nguyen Viet Cuong 359 Downloaded from wber.oxfordjournals.org at International Monetary Fund on August 12, 2011 Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW • What Constrains Africa’s Exports? Caroline Freund and Nadia Rocha • Does the Internet Reduce Corruption? Evidence from U.S. States and across Countries Thomas Barnebeck Andersen, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya • Do Labor Statistics Depend on How and to Whom the Questions Are Asked? Results from a Survey Experiment in Tanzania Elena Bardasi, Kathleen Beegle, Andrew Dillon, and Pieter Serneels SYMPOSIUM ON ENTREPRENEURSHIP AND DEVELOPMENT • Entrepreneurship and Development: the Role of Information Asymmetries Leora Klapper and Inessa Love • Getting Credit to High Return Microentrepreneurs: The Results of an Information Intervention Suresh de Mel, David McKenzie, and Christopher Woodruff • The Impact of the Business Environment on Young Firm Financing Larry W. Chavis, Leora F. Klapper, and Inessa Love • Does a Picture Paint a Thousand Words? Evidence from a Microcredit Marketing Experiment Xavier Giné, Ghazala Mansuri, and Mario Picón • Entrepreneurship and the Extensive Margin in Export Growth: A Microeconomic Accounting of Costa Rica’s Export Growth during 1997-2007 Daniel Lederman, Andrés Rodríguez-Clare, and Daniel Yi Xu THE WORLD BANK 1818 H Street, NW Washington, DC 20433, USA World Wide Web: http://www.worldbank.org/ E-mail: wber@worldbank.org