77480 On the Unequal Inequality of Poor Communities ¨ zler, and Ken Simler Chris Elbers, Peter F. Lanjouw, Johan A. Mistiaen, Berk O Communities differ in important ways in their needs, capacities, and circumstances. Because central governments are not able to discern these differences fully, they seek to achieve their policy objectives by relying on decentralized mechanisms that use local information. Household and individual characteristics within communities can also vary substantially. A growing body of theoretical literature suggests that inequality within communities can influence policy outcomes in ways that are either harmful or helpful, depending on the circumstances. Until recently, empirical investigations into the impact of inequality have been held back by a lack of systematic evidence on community-level inequality. This study uses household survey and population census data to estimate per capita consumption inequality within communities in three devel- oping economies. It finds that communities vary markedly in their degree of inequality. It also shows that there should be no presumption that inequality is less severe in poor communities. The kind of community-level inequality estimates generated here can be used in designing and evaluating decentralized antipoverty programs. Governments commonly implement decentralized antipoverty programs that are designed to distribute assets or cash to individuals or households. Usually, the central government distributes antipoverty funds to communities, which then decide how to allocate the funds. One example is social fund projects, a type of community-based development initiative in which poor communities identify projects, apply for funding, and design, implement, and manage their projects (Mansuri and Rao 2004).1 These initiatives intend to improve poverty Chris Elbers is a professor at the Vrije (Free) University Amsterdam; his e-mail address is celbers@ feweb.vu.nl. Peter F. Lanjouw is a lead economist in the Development Research Group at the World Bank; his e-mail address is planjouw@worldbank.org. Johan A. Mistiaen is an economist/statistician in the Development Data Group at the World Bank; his e-mail address is jmistiaen@worldbank.org. Berk O¨ zler is an economist in the Development Research Group at the World Bank; his e-mail address is bozler@worldbank.org. Ken Simler is a research fellow at the International Food Policy Research Institute; his e-mail address is k.simler@cgiar.org. The authors are grateful to Francois Bourguignon, Francisco Ferreira, Emanuela Galasso, Ravi Kanbur, Jenny Lanjouw, Vijayendra Rao, and Martin Ravallion for comments and helpful discussions. They would also like to thank the journal editor and three anonymous referees for guidance. 1. Mansuri and Rao (2004) distinguish community-based development from community-driven development, popularized by the World Bank, which refers to projects in which communities have direct control over key decisions as well as management of investment funds. Community-based development can be thought of as a broader term that accommodates but is not restricted to the World Bank’s community-driven development concept. THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3, Ó The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi:10.1093/wber/lhh046 18:401–421 401 402 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 targeting and project implementation by using local information and inviting local participation. In practice, however, these potential benefits of local involve- ment may be outweighed by the possibility of resources being captured by local elites.2 In a review of the community-based development approach, Mansuri and Rao (2003) argue that although potential gains are large, there are also important risks inherent in the basic precepts of the approach. Uncertainty about the ultimate impact of such programs implies that a blanket application of a given approach in all communities may not be approp- riate. Again, Mansuri and Rao (2004) caution against the wholesale scaling up of best practices identified in a few pilot settings, because the success of such pilot projects might depend crucially on local conditions that are not found elsewhere. Still, large projects such as a countrywide cash transfer or social fund program cannot take into account the full range of local characteristics that could possibly affect project performance. Hence, policymakers must confront the challenge of designing schemes that take critical local information into account but are not prohibitively costly to implement. Governments have traditionally dealt with this problem by categorizing communities by easily observable characteristics and adapting schemes for each group. Lacking local-level data on poverty, government programs may draw on proxy indicators—believed to be correlated with local poverty condi- tions—to determine the eligibility of communities for various projects. But despite emerging theoretical analysis and empirical evidence that local inequal- ity may also affect local development outcomes, such information has rarely made its way into program design. One reason is that estimates of local inequal- ity have not been widely available until recently.3 Another is that inequality may not be considered of primary importance when the target of an intervention is a small, poor community in a developing economy. The natural assumption is that where livelihoods are at the subsistence level there is little likelihood that well-being would vary much across households and individuals. This article addresses both these issues. Applying a newly developed metho- dology, it estimates local-level welfare outcomes using the detailed information available from household surveys and the large-scale representation of the population census for Ecuador, Madagascar, and Mozambique. These techni- ques can be used to derive meaningful estimates of income or expenditure inequality for small areas for many countries, using readily available data. The article examines the importance of local-level inequality by decomposing national inequality in each country into a within-community and between- community component. This decomposition exercise produces a summary sta- 2. A vivid illustration of elite capture problems in practice and a theoretical treatment of this issue are provided in Platteau and Gaspart (2003). 3. McKenzie (2003) provides a recent attempt to proxy local inequality on the basis of easily observed correlates of household income. Elbers and others 403 tistic that masks significant heterogeneity in inequality across communities. The article provides additional evidence that this heterogeneity in inequality is evident even among poor rural communities. It demonstrates that information on local inequality can help program implementers further categorize commu- nities after conditioning on local poverty and type of area. I. HOW CAN LOCAL INEQUALITY AFFECT WELFARE OUTCOMES? Mansuri and Rao (2004) present a comprehensive overview of the theoretical and empirical literature on the relationship between local inequality and devel- opment outcomes. Two critical issues emerge. How does inequality within a community influence the targeting impact of a particular project? How does local inequality affect collective action within communities? Recent theoretical analysis suggests that inequality may affect targeting out- comes of social fund projects or antipoverty transfer schemes by reducing the relative power of the intended beneficiaries (Galasso and Ravallion forthcom- ing; Bardhan and Mookherjee 1999). In such cases, the advantage of such decentralized approaches to make use of better community-level information about priorities and the characteristics of residents could be offset by the possibility that the local governing body is controlled by elites, who may have different objectives than the poor within their communities. Although the predictions from this theoretical work are ambiguous, limited empirical evidence shows that both the pros and the cons of decentralized decisionmaking are at work in various countries. Alderman (2002) finds that communities in Albania were able to improve targeting by using information unavailable to the central government. By contrast, Galasso and Ravallion (forthcoming) find that high levels of local inequality (as measured by land- holding) were associated with worse targeting performance under the Food for Education program in villages in Bangladesh. A detailed case study of the small north Indian village of Palanpur from the late 1950s through the early 1990s shows how local elites appropriated public resources and opportunities that were to be made available to the whole com- munity (Dre ` ze and others 1998). The study documents the introduction of 18 types of government-provided programs into the village, including a public works road-building program, free schooling, free basic health care, old-age pensions, a fair-price shop, and a farmer cooperative. The sobering diagnosis is that most of these programs were nonfunctional, particularly programs that had a redistributive component. Dre ` ze and others argue that a key explanation for this dispiriting record is that village institutions were dominated by privileged groups and that only programs that enjoyed their backing were allowed to succeed. Dre ` ze and others (1998, p. 211) see ‘‘little prospect of major improve- ment in the orientation and achievements of government intervention without a significant change in the balance of political power, both at the state and at the 404 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 local level.’’4 There is also a rich body of literature on the relationship between inequality and collective action, with implications for the provision of public goods, management of common pool resources, and group participation (Olson 1973; Balland and Platteau 1999, 2001, 2003; Dayton-Johnson and Bardhan 2002). This literature points to the possibility that some inequality may be necessary to mobilize the collective action needed for group provision of a public good. If a community is large and homogeneous, no single individual could make any significant difference in the provision of the public good, and so all would want to free-ride, resulting in no provision of the good. Again, the theoretical relationship among inequality, participation, and col- lective action is complex. Most of the empirical evidence, however, seems to point to a negative or U-shaped relationship, with increased inequality lead- ing, at least initially, to a decline in collective action (Dayton-Johnson 2000; Dayton-Johnson and Bardhan 2002; Khwaja 2002; Alesina and La Ferrara 2000; La Ferrara 2002). The growing literature on the relationship between local inequality and development outcomes thus suggests several ways that local inequality could influence development efforts of the kind described in this article. The empirical literature, though still far from complete, suggests that on balance inequality is likely to hamper local development efforts. Incorporating information on inequality into the design of development efforts might therefore be necessary. II. DATA AND METHODOLOGY This study drew on data from both household surveys and population censuses in Ecuador, Madagascar, and Mozambique (see appendix table A.1 for details on data sources and coverage). The construction of comprehensive ‘‘geographic profiles’’ of inequality across localities has been constrained by the limitations of conventional distributional data. Detailed household surveys, which include reasonable measures of income or consumption, are samples and thus are rarely representative or of sufficient size at low levels of disaggregation to yield statistically reliable estimates. Census data (or large sample surveys) of sufficient size to allow disaggregation either have no information about income or consumption or measure these variables poorly (see Alderman and others 2003). This article uses a recently developed statistical procedure to combine data sources in a way that takes advantage of the detailed information available in 4. The review by Dre ` ze and colleagues (1998) does not cover any specific community-based devel- opment projects in Palanpur. It is possible that performance of such projects might have been different. The review does indicate, however, that any notion that the villagers in Palanpur all have the same objectives, interests, and influence would be sorely mistaken. That villagers differed in economic well- being was clearly discernible in the study: income inequality within Palanpur was on the same order of magnitude as measures of inequality for India as a whole (Lanjouw and Stern 1998). Elbers and others 405 household sample surveys and the comprehensive coverage of censuses to estimate inequality at a level of disaggregation previously unattainable. The methodology is developed in detail by Elbers and others (2002, 2003a), and applications are described elsewhere (see Demombynes and others 2004; Elbers and others forthcoming; Mistiaen and others 2002), so only a brief description is provided here.5 A model of log per capita household expenditures, y, is estimated using the sample survey data, restricting the explanatory variables to those common to both the survey and the census or to those in a tertiary data set that can be linked to both of these data sets.6 Then, the expected level of an indicator of poverty or inequality, W, is estimated given the census-based observable char- acteristics of the population of interest using parameter estimates from the first stage model of y. The same approach could be used with other household measures of well-being, such as assets, income, or employment. The first-stage estimation is carried out using the household sample survey. The first concern is to develop an accurate empirical model of household consumption. Consider the following model: ð1Þ ln ych ¼ E½ln ych jxch T Š þ uch % xch T a ˇ þ Zc þ ech where household h is located in sample cluster c, and Z and e are uncorrelated with each other and with observables. This specification allows for an intraclus- ter correlation in the disturbances. For any given disturbance variance, sch2, the greater the fraction due to the common component Zc, the less the benefit from aggregating over more households. Welfare estimates become less precise. Furthermore, failing to account for spatial correlation in the disturbances could bias the inequality estimates. A Hausman test (described in Deaton 1997) is used to determine whether to estimate with household weights. The R2 is generally high, ranging from 0.45 and 0.77 in Ecuador to 0.29 to 0.63 in Madagascar and 0.27 to 0.55 in Mozambique. (For details see Elbers and others 2002, 2003a; Mistiaen and others 2002; Simler and Nhate 2002.) Next, the variance of the idiosyncratic part of the disturbance, se,ch2, is modeled. To model heteroscedasticity in the household-specific part of the residual, 5–20 variables, zch, are chosen that best explain variation in ech2 of all potential explanatory variables, their squares, and interactions.7 5. Elbers and others (2003a) is an earlier version of the present study and does not explore the potential relevance of local inequality to policy or the association between poverty and degree of inequality in communities. 6. As described in Elbers and others (2002, 2003a), a separate model is estimated for each stratum, rather than forcing the models and the parameter estimates to be the same for the whole country. 7. The number of explanatory variables is limited to avoid overfitting, and a bounded logistic functional form is used. 406 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 " c and Finally, the distribution of h and e is determined using the cluster residuals Z standardized household residuals. Normal or t distributions are used with varying degrees of freedom, or the actual standardized residual distribution is used when a semiparametric approach is taken. The estimated variance-covariance matrix is used to obtainfinalgeneralizedleast squares estimatesof the first-stage consumption model. Welfare estimates are obtained from 100 simulations in each of the three countries. For each simulation a vector of simulated parameters is drawn from a multivariate normal distribution with variance-covariance matrix estimated in the survey-based consumption and heteroscedasticity regressions. In addition, distur- bance terms at the cluster and household level are drawn from their standardized parametric or semiparametric distributions. These parameter and disturbance draws are then applied to the census-level regressors to predict per capita con- sumption. For the next simulation, a new set of parameters and disturbances is drawn and a new value of per capita consumption is calculated for each household. For a given locality a separate measure of inequality is calculated for each vector of simulated per capita consumption. The average across the 100 simulations yields the estimate of inequality for the locality, and the standard deviation of the inequality measures yields the estimate of the standard error. III. ESTIMATES OF LOCAL INEQUALITY IN THREE COUNTRIES This section presents the census-based estimates of inequality produced using the methodology described and compares them with estimates from household surveys at the level at which those surveys are representative.8 If the methodology is applied with proper attention to data comparability, first-stage regression models, and the error structures used in simulating the inequality measures, then stratum-level esti- mates,virtuallyby construction, should correspondcloselywith those inthe household survey. As such, these comparisons cannot be considered a test of the methodology. Clearly, however, if the census-based estimates are wildly different from the corre- sponding survey comparators, there would be little basis for proceeding further. Estimated per capita consumption for each country from the household survey was compared with the census data at the stratum level (for which the household survey is representative). The results show that in nearly every case the hypothesis that estimates of average per capita consumption are the same across the two data sources (at the 95 percent confidence level) cannot be rejected (table 1). With few exceptions, point estimates match closely. Note that the standard errors of the per capita consumption estimates are often smaller for estimates based on census data than for those based on household survey data. Although the census estimates are predicted with error due mainly to the imprecision of the first-stage regressions, they are free of sampling error, making them more precise than estimates from the household survey. 8. For a similar analysis, focusing specifically on poverty, see Demombynes and others (2004). T A B L E 1 . Comparison of Survey- and Census-Based Average per Capita Consumption Estimates at the Stratum Level in Ecuador, Madagascar, and Mozambique Ecuador (Sucres per Capita) Madagascar (Francs per Capita) Mozambique (Meticais per Capita) Stratum Survey Census Stratum Survey Census Stratum Survey Census Quito 126,098 125,702 Antananarivo 513,818 576,470 Niassa 4,660 5,512 (11,344) (8,026) Urban (48,455) (23,944) (355) (484) Sierra 121,797 122,415 Fianarantsoa 360,635 372,438 Cabo Delgado 6,392 6,586 Urban (8,425) (4,642) Urban (42,613) (21,878) (416) (433) Sierra 66,531 63,666 Toamasina 445,514 417,823 Nampula 5,315 5,547 Rural (4,067) (2,213) Urban (73,099) (15,406) (287) (279) Guayaquil 89,601 77,432 Mahajanga 613,867 580,775 Zambezia 5,090 5,316 (5,597) (2,508) Urban (74,092) (31,025) (208) (274) Costa 86,956 90,209 Toliara 343,111 321,602 Tete 3,848 4,404 Urban (3,603) (2,391) Urban (76,621) (32,193) (267) (176) Costa 57,619 61,618 Antsiranana 504,841 693,161 Manica 6,299 6,334 407 Rural (4,477) (2,894) Urban (46,148) (93,437) (741) (527) Oriente 110,064 174,529 Antananarivo 312,553 324,814 Sofala 3,218 4,497 Urban (9,078) (56,115) Rural (23,174) (14,378) (191) (379) Oriente 4,7072 59,549 Fianarantsoa 319,870 251,312 Inhambane 4,215 4,177 Rural (4,420) (3,051) Rural (45,215) (18,091) (359) (134) Toamasina 275,943 279,239 Gaza 6,024 6,521 Rural (22,832) (15,838) (356) (355) Mahajanga 325,872 321,398 Maputo Province 5,844 8,559 Rural (30,209) (19,385) (613) (745) Toliara 233,801 259,537 Maputo City 8,321 11,442 Rural (22,174) (16,222) (701) (4,956) Antsiranana 486,781 442,431 Rural (91,181) (54,869) Note: Numbers in parentheses are standard errors. All household survey estimates are computed using weights that are the product of household survey weights and household size. The census-based estimates are weighted by household size. Source: Authors’ calculations based on household survey and census data. 408 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 Comparing stratum-level estimates of inequality is less straightforward. Inequality measures tend to be sensitive to the tails in the distribution of expendi- ture. Since far-off portions of the tails are typically not observed in the survey (because of its small sample size), survey estimates of inequality will often be below the true level of inequality. Perhaps more important, nonresponse may be an issue in household surveys, and to the extent that nonresponse is more prevalent among rich households, the resulting selection bias will lead to further downward bias of survey-based estimates (see Mistiaen and Ravallion 2003). To the extent that a census suffers less from such problems of observation, and assuming that the expenditure model is correct, the expenditure of rich households will be better represented in the census-based estimates of inequality. These considerations lead to expectations of higher inequality estimates from census-based imputation. Reflecting the complex sample design of the household survey for the survey- based estimates and the imputation procedure for the census-based estimates, standard errors are presented for all estimates of Gini coefficients (table 2). For Ecuador and Mozambique, the census-based estimates of consumption inequal- ity tend to be higher than the survey-based estimates, although not generally to such an extent that one can reject the hypothesis that they are the same.9 For some provinces in Mozambique, such as Sofala, Maputo Province, and Maputo City, the estimates from the census are not only higher than those in the survey but are also imprecisely estimated.10 In Madagascar the standard errors on the survey estimates of inequality are quite high. This serves as a reminder that although stratum-level estimates of welfare in household surveys are often referred to as representative, the sample size in these strata can be small, so the accompanying welfare estimates are not always very precise. Nonetheless, it is encouraging that the point estimates of the Gini coefficient from the survey and the census data in Madagascar are often quite close. Elbers and others (2002, 2003a) demonstrate that standard errors on census- based estimates are inversely correlated with the size of the target population. Thus, although estimates of inequality may look good at the stratum level, they could become quite imprecise for smaller localities. 9. These issues are subjects of current research. If anything, the true difference between census-based and survey-based inequality estimates is expected to be even larger, because extreme draws of the error terms were ignored in the simulations underlying the poverty maps. Again, this might lead to under- representation of high-expenditure cases. To the extent that extreme draws of the error terms were not culled severely enough, census-based average consumption estimates would also be expected to exceed their survey-based counterparts. Mean per capita consumption, unlike the median, is directly affected by tails of the consumption distribution. A quick scrutiny of the consumption and inequality estimates for Mozambique suggests that trimming was possibly too light and that as a consequence both mean and inequality estimates are higher in the census than in the survey. 10. There is no evidence that the census-based estimates become even noisier at lower levels of aggregation in Mozambique. T A B L E 2 . Comparison of Survey- and Census-Based Inequality Estimates (Gini Coefficients) at the Stratum Level in Ecaudor, Madagascar, and Mozambique Ecuador Madagascar Mozambique Stratum Survey Census Stratum Survey Census Stratum Survey Census Quito 0.490 (0.023) 0.465 (0.012) Antananarivo Urban 0.492 (0.027) 0.469 (0.012) Niassa 0.355 (0.020) 0.402 (0.025) Sierra Urban 0.436 (0.020) 0.434 (0.011) Fianarantsoa Urban 0.430 (0.038) 0.426 (0.015) Cabo Delgado 0.370 (0.025) 0.413 (0.021) Sierra Rural 0.393 (0.034) 0.457 (0.013) Toamasina Urban 0.434 (0.042) 0.402 (0.015) Nampula 0.391 (0.026) 0.400 (0.020) Guayaquil 0.378 (0.014) 0.416 (0.011) Mahajanga Urban 0.371 (0.027) 0.392 (0.016) Zambezia 0.324 (0.017) 0.366 (0.012) Costa Urban 0.359 (0.015) 0.382 (0.011) Toliara Urban 0.514 (0.052) 0.504 (0.030) Tete 0.346 (0.019) 0.394 (0.018) 409 Costa Rural 0.346 (0.036) 0.400 (0.015) Antsiranana Urban 0.362 (0.025) 0.433 (0.039) Manica 0.413 (0.036) 0.449 (0.020) Oriente Urban 0.398 (0.035) 0.563 (0.104) Antananarivo Rural 0.376 (0.023) 0.404 (0.015) Sofala 0.405 (0.031) 0.529 (0.032) Oriente Rural 0.431 (0.034) 0.478 (0.014) Fianarantsoa Rural 0.470 (0.050) 0.437 (0.018) Inhambane 0.382 (0.037) 0.398 (0.012) Toamasina Rural 0.352 (0.036) 0.362 (0.017) Gaza 0.380 (0.024) 0.421 (0.023) Mahajanga Rural 0.320 (0.026) 0.306 (0.015) Maputo Province 0.424 (0.029) 0.518 (0.029) Toliara Rural 0.383 (0.029) 0.377 (0.017) Maputo City 0.444 (0.033) 0.560 (0.108) Antsiranana Rural 0.518 (0.110) 0.453 (0.048) Note: Numbers in parentheses are standard errors. All household survey estimates are computed using weights that are the product of household survey weights and household size. The census-based estimates are weighted by household size. Source: Authors’ calculations based on household survey and census data. 410 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 This study produced estimates of inequality at the third administrative level (the firaisana in Madagascar, the parrqouia in rural Ecuador, the administrative post in Mozambique). Does this imply that at such fine levels of disaggregation, these inequality estimates are too noisy to be useful? Elbers and others (forthcoming) document that standard errors correspond to about 5–15 percent of point esti- mates of inequality for these localities (see also later discussion)—the same range that is generally judged to be acceptable at the stratum level in household surveys. Elbers and others also show that the explanatory power of simple, descriptive ordinary least square regressions of inequality at the smallest administrative level on a set of simple community characteristics is quite high in these three countries (with an R2 ranging between 0.57 and 0.78 in urban areas and between 0.38 and 0.57 in rural areas). They find that community-level inequality is typically higher in communities with large shares of the elderly, whereas in rural (but not urban) areas it is generally also positively correlated with total population.11 If the inequality estimates produced with this methodology were just noise, such correla- tions and explanatory power would not be expected.12 From the evidence presented here, it can be concluded that the applied estimation technique can yield meaningful estimates of inequality for small areas. The next section turns to inequality decompositions by administrative units and the heterogeneity of inequality across communities. IV. DECOMPOSING INEQUALITY BY GEOGRAPHIC SUBGROUPS Decomposition of inequality by geographic subgroups in both developed and developing economies has a long tradition. The policy implications may be quite different when national inequality is attributable largely to differences in mean incomes across localities (between-group inequality) than when national inequality is basically an expression of heterogeneity that already exists at the local level (within-group inequality). To decompose inequality using the general entropy class of inequality measures, a class of measures that is particularly well suited for this exercise:13 11. A conditional correlation with estimated per capita consumption is also observed in rural areas, although evidence of a Kuznet’s inverted U-curve is not particularly strong. Elbers and others (forth- coming) suggest that although the inequality measures included in these regressions have been estimated, this does not invalidate their use for these purposes (although they do advocate correcting standard errors for model error when estimated variables are included as regressors). 12. Demombynes and O ¨ zler (forthcoming) find evidence of a strong association between local inequality estimates produced on the basis of this methodology in South Africa and official crime statistics collected at the local level. 13. Following Bourguignon (1979), Shorrocks (1980), and Cowell (1980). Cowell (2000) provides a useful recent survey of methods of inequality measurement, including a discussion of the various approaches to subgroup decomposition. Sen and Foster (1997) and Kanbur (2000) discuss some of the difficulties in interpreting results from such decompositions. Elbers and others 411 GEc ¼ ½1=cðc À 1ފÆi fi ½ðyi =mÞc À 1Š for c 6¼ 0; 1 ð2Þ ıÞ for c ¼ 0 ¼ ÀÆi fi logðyi =` ıÞ logðyi =` ¼ Æi fi ðyi =` ıÞ for c ¼ 1 where fi is the population share of household i, yi is per capita consumption of household i, m is average per capita consumption, and c is a parameter to be selected by the user.14 This class of inequality measures can be decomposed into between- and within-group components along the following lines: GEc ¼ ½1=cðc À 1ފ½Æj gj ðmj =mÞc À 1Š þ Æj GEj gj ðmj =mÞc for c 6¼ 0; 1 ð3Þ GEc ¼ ½Æj gj ðm=mj Þc À 1Š þ Æj GEj gj for c ¼ 0 GEc ¼ ½Æj gj ðmj =mÞ logðmj =mފ þ Æj GEj gj ðmj =mÞ for c ¼ 1 where j refers to subgroups, gj refers to the population share of group j, and GEj refers to inequality in group j. The between-group component of inequality is captured by the first term on the right. It can be interpreted as measuring the level of inequality in the population if everyone within the group had the same (the group-average) consumption level, mj. The second term on the right reflects within-group inequality, or the overall inequality level if there were no differ- ences in mean consumption across groups but each group had its actual within- group inequality, GEj. Ratios of the respective components with the overall inequality level provide a measure of the percentage contribution of between- group and within-group inequality to total inequality. At one extreme, when inequality is measured at the national level, all inequal- ity is by definition within groups. At the other extreme, when each individual household is taken as a separate group, the within-group contribution to overall inequality is zero and all inequality is between groups. But where does the between-group component start to outweigh the within-group component? Is it reasonable to suppose that at a sufficiently low level of disaggregation, such as the village or community, inequality within groups is small, and most of overall inequality is due to differences between groups? The highest between-group inequality at the community level (measured as the mean log deviation, or GE[0])15 is observed in Ecuador, at about 41 percent (table 3). The share of inequality that can be attributed to mean expenditure differences between communities is much smaller in the other two countries— 14. Lower values of c are associated with greater sensitivity to inequality among the poor, and higher values of c place more weight to inequality among the rich. A c value of 1 yields the well known Theil entropy measure, a value of 0 provides the Theil L or mean log deviation, and a value of 2 is ordinally equivalent to the squared coefficient of variation. 15. Results remain virtually identical for other values of c. 412 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 T A B L E 3 . Decomposition of Inequality between and within Communities in Ecuador, Madagascar, and Mozambique Level of Decomposition Number of Within-Group Between-Group Subgroups Inequality (%) Inequality (%) Ecuador All communities 1,579 58.8 41.2 Urban 664 76.7 23.3 Rural 915 85.9 14.1 Madagascar All communities 1,248 74.6 25.4 Urban 131 76.7 23.2 Rural 1,117 81.9 18.1 Mozambique All communities 424 78.0 22.0 Note: Communities are defined at the third administrative level (1,000–10,000 households): the zona (urban) and parroquia (rural) in Ecuador, the firiasana (commune) in Madagascar, and administrative posts in Mozambique. Source: Authors’ computations based on household survey and census data. 25 percent in Madagascar and 22 percent in Mozambique.16 There is also evidence, particularly for Ecuador, that the observed between-group inequality is due mainly to differences between urban and rural communities. When attention is focused solely on rural communities in Ecuador, the between- group component of inequality falls to less than 15 percent of total inequality. In Madagascar the share of between-group inequality in rural areas is signifi- cantly lower (18 percent) than the combined share for rural and urban areas. In all three countries overall inequality is attributable mostly to inequality within communities, even when the community is defined as the lowest level of central government administrative unit.17 16. Elbers and others (2003a) show that the between-group share of inequality at higher levels of aggregation (first and second administrative levels) is below 10 percent in all countries other than urban Madagascar. 17. Inequality estimates produced on the basis of the methodology described in the article are averages calculated over a number of simulations (100 in our case). It is possible that a decomposition of inequality carried out after this averaging procedure has occurred overstates the within-group compo- nent of inequality because differences in inequality across communities have been smoothed out. To check this, the decomposition was carried out for each of the 100 simulations and then averaged across the decomposition results. The between-group component of inequality increased by at most 1–2 percent, and the qualitative results were unchanged. There is no other reason to suspect that the methodology for estimating local-level inequality is associated with any built-in tendency to overstate within-group inequality. One way to test this is to carry out the imputation exercise described here for a data set that also contains directly collected information on welfare and then to compare decomposition results on the basis of imputed welfare with those on the basis of observed welfare. Elbers and others (2003b) show for Brazil that a decomposition of inequality based on imputed consumption reaches virtually identical conclusions as a decomposition based on observed income. Elbers and others 413 Interpretations of decompositions like these are not completely straightfor- ward, however. For example, the decomposition results reported (documenting a large within-group component of inequality) do not imply that local inequality levels are uniformly high or even that the majority of communities exhibit high levels of inequality. Rather, the decomposition provides a summary statistic, suggesting that on average within-group inequality is not particularly low at the third administrative level. In other words, it is possible that a country has both highly equal and highly unequal communities. A simple example can illustrate this. Consider a population of eight indivi- duals for whom the vector of consumption values is (1, 1, 2, 2, 4, 4, 5, 5). This population could be divided into two communities of (1, 2, 4, 5) and (1, 2, 4, 5) or two of (1, 1, 5, 5) and (2, 2, 4, 4), both cases having the same average consumption. The between-group inequality component from a decomposition exercise such as that carried out is always zero (and the within-group share is thus 100 percent in both cases). However, in the first case inequality in the two communities is exactly equal to national inequality, whereas in the second case one community has a higher and the other a lower level of inequality than the national level. Thus, finding a high within-group share of inequality in a decomposition exercise of a large number of communities is consistent with great heterogeneity in inequality levels across those same communities. The obvious question to ask, then, is whether communities vary widely in their degree of inequality. This can be answered by plotting community-level inequality estimates and comparing them with overall inequality. Communities in rural Ecuador are ranked from most equal to most unequal, and 95 percent confidence intervals on each community-level estimate are included as scatter- plots (figure 1). Although the within-group share from the decomposition is as high as 86 percent, the summary statistic masks considerable variation in inequality levels across parroquia. A large majority of parroquia-level point estimates are well below the national level. Even allowing for the imprecision around the parroquia-level estimates (which are typically 5–15 percent of the point estimate), equality is unambiguously greater in a large proportion of parroquias than at the national level. Another sizable proportion is not obviously less or more unequal than the country as a whole, and a smaller proportion is considerably more unequal.18 In urban Ecuador the proportion of zonas that have lower inequality than the national inequality rate is even higher than in rural areas (figure 2). The precision of point estimates is somewhat higher in urban areas of Ecuador than in rural areas; accordingly, more zonas lie 18. Note that there are more communities with inequality below the national level than above the national level because between-group inequality, although relatively small, is not absent. Differences in average per capita consumption ensure that at least some of total inequality is attributable to differences between groups. If there were no within-group inequality, or if all communities had the same level of within-group inequality, overall inequality would be greater than or equal to inequality in each of the individual communities. 414 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 F I G U R E 1 . Distribution of Parroquia-Level GE(0) Inequality across Parroquias in Rural Ecuador Note: The 915 parroquias are ranked from most equal to most unequal, and the average number of households per parroquia is 1,050. The scatterplot indicates the 95 percent confidence interval. Source: Authors’ calculations based on the 1994 National Survey of Living Conditions (ECV) and the 1990 population census. F I G U R E 2 . Distribution of Zona-Level GE(0) Inequality across Zonas in Urban Ecuador Note: The 664 zonas are ranked from most equal to most unequal. The average number of households per zona is 1,325. The scatterplot indicates the 95 percent confidence interval. Source: Authors’ calculations based on the 1994 ECV and the 1990 population census. Elbers and others 415 unambiguously below the national inequality level. The pattern is similar in rural and urban Madagascar and in Mozambique (not shown). In all three countries there is a large group of communities with lower inequality than inequality in the country as a whole, another large group for which inequality is not significantly different from inequality in the country as a whole, and a small third group of communities with inequality higher than the national level. V . A R E P O O R C O M M U N I T I E S ‘‘ M O R E E Q U A L ’’ T H A N O T H E R S ? Although most of the inequality in Ecuador, Madagascar, and Mozambique is attributable to inequality within communities, there is considerable heterogene- ity in inequality across communities within each country. This section looks at whether inequality is less marked if the focus is on poor communities. Commu- nity-based development programs are often targeted primarily to poor commu- nities. If the communities have low levels of inequality, it may be less important that policymakers incorporate information on inequality into the design and implementation of community-based development projects. That turns out not to be the case for the countries examined in this study. The Gini index at the community level by quintiles of the imputed headcount index (see Demombynes and others 2004)19 ranges from 0.299 to 0.501 in Ecuador (figure 3), 0.231 to 0.466 in Madagascar (figure 4), and 0.261 to 0.534 in Mozambique (figure 5).20 In all three countries median inequality in the poorest quintile is no lower than in any of the richer quintiles. Furthermore, the range of inequality levels across communities is among the widest in the poorest quintile—even when only rural communities are considered. Thus, inequality in a typical poor community in any of these three countries—even in rural areas—is at least as great as in other communities, and the range of inequality among poor communities is no narrower than for the country as a whole. At the beginning of this article, reference was made to the literature suggest- ing that inequality at the local level has a bearing on the political economy of communities and in this way can affect the performance of community-based development initiatives. Another way of thinking about the implications of the finding is to consider how local inequality would undermine the effectiveness of policies to alleviate poverty through fine geographic targeting of transfers. In a parallel study Elbers and others (2004) illustrate that with relatively high 19. It is possible that high levels of inequality would be observed in high-poverty areas simply because these two measures of welfare are highly correlated. However, the results presented in this section are the same when communities are ranked by their mean consumption levels instead of the headcount index. 20. These reported ranges exclude the top and bottom 1 percent of communities (in terms of the Gini index) in each country. 416 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 F I G U R E 3 . Unequal Inequality of Communities in Ecuador Note: The quintiles are based on the headcount poverty index of all communities (rural and urban), and the box-whisker percentiles shown are the median, 25/75 and 1/99. Source: Authors’ calculations based on the 1994 ECV and the 1990 population census. F I G U R E 4 . Unequal Inequality of Communities in Madagascar Note: The quintiles are based on the headcount poverty index of all communities (rural and urban), and the box-whisker percentiles shown are the median, 25/75 and 1/99. Source: Authors’ calculations based on the 1993/94 National Household Survey and the 1993 population census. Elbers and others 417 F I G U R E 5 . Unequal Inequality of Communities in Mozambique Note: The quintiles are based on the headcount poverty index of all communities (rural and urban), and the box-whisker percentiles shown are the median, 25/75 and 1/99. Source: Authors’ calculations based on the 1996/97 National Household Survey on Living Conditions (IAF96) and the 1997 population census. inequality in poor communities, even fine geographic targeting of communities would result in a considerable amount of mistargeting. An attempt to quantify such mistargeting reveals that in Ecuador targeting at the community level (and implicitly treating everyone within the commune as equally poor) would achieve at best only half the poverty reduction that would be attainable if the consump- tion level of each individual household had been observable. VI. POLICY IMPLICATIONS There has been a massive increase in resources devoted to community-based development programs in the past 10 years. The review by Mansuri and Rao (2004) suggests that funding for such projects rose from around $325 million in 1996 to $2 billion in 2003. Although the main goal is to achieve better out- comes by involving local communities in the decisionmaking process and man- agement of projects, governments also need some basic indicators for targeting communities and tailoring basic features of these projects to different types of communities. So far, governments have commonly used type of area (urban or rural) and proxy information on poverty at the community level for such purposes. 418 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 This article proposes measures of inequality at the community level as possible additional indicators to inform the design of decentralized antipoverty programs and community-based development projects. Recent theory and lim- ited empirical evidence suggest that inequality may be related to outcomes at the community level. Inequality at the community level may lead to the capture of the intended benefits by the local elite, or inequality may be highly correlated with another (not easily observed) factor that leads to capture by the elite. Collective action and the provision of public goods may also be correlated with the level of inequality within communities. A recently developed small area estimation technique can provide estimates of inequality at the local level. In the three countries examined here, although most of the consumption inequality on average is attributable to inequality within communities, local inequality varies widely across communities. Further- more, inequality is highly heterogeneous even in the poorest communities. Not only is inequality as high in a typical poor community as it is in other commu- nities, but the range of inequality levels is at least as wide in poor communities as it is in richer communities. This finding remains true even when attention is restricted to rural areas. These findings suggest that even after controlling for the type of area and the poverty levels of communities, local inequality measures can provide additional information that can enhance desired outcomes. For example, for transfer programs that expect local communities to identify poor beneficiaries, eligible communities could be categorized broadly as low, middle, and high inequality. Random audits and means-tested targeting by the central government (as, for example, in Mexico’s Progresa program) could then be considered to improve propoor targeting in the middle- and high-inequality communities. Clearly, a first priority is to undertake more systematic research into the relationship between local inequality and various development outcomes. Cri- tical questions are the manner and extent to which current development pro- cesses and practices interact with local inequality.21 Better estimates of consumption inequality at the local level using the techniques described here and related approaches promise important new insights. Microlevel estimation of welfare based on the methodology has been completed or is under way in some 25 developing economies. These estimates can be combined with detailed information on the operation of antipoverty programs and community-based development projects in these countries, with an eye toward uncovering sys- tematic relationships, positive or negative. 21. An even more basic question concerns the relationship between community-level consumption or income inequality and the welfare inequality that is of ultimate interest. Dasgupta and Kanbur (2001) show how the presence of community-specific public goods could imply that the distribution of nominal income provides a very misleading picture of real inequalities and tensions in society, both between and within communities. Elbers and others 419 APPENDIX T A B L E A . 1 . Data Summary Instrument Ecuador Madagascar Mozambique Household survey Year 1994 1993/94 1996/97 Source National Survey National Household National Household of Living Conditions Survey Survey on Living Conditions Sample size 4,500 households 4,508 households 8,250 households References Hentschel and Lanjouw Mistiaen and Simler and Nhate (1996); Hentschel and others (2002) (2002) others (2000) Population census Year 1990 1993 1997 Coverage About 10 million About 11.9 million About 16 million individuals in individuals in individuals in 2 million 2.4 million 3.6 million households households households REFERENCES Alderman, Harold. 2002. ‘‘Do Local Officials Know Something We Don’t? Decentralization of Targeted Transfers in Albania.’’ Journal of Public Economics 83(3):375–404. Alderman, Harold, Miriam Babita, Gabriel Demombynes, Nthabiseng Makhatha, and Berk O ¨ zler. 2003. ‘‘How Low Can You Go? Combining Census and Survey Data for Mapping Poverty in South Africa.’’ Journal of African Economies 11(2):169–200. Alesina, Alberto, and Eliana La Ferrara. 2000. ‘‘Participation in Heterogeneous Communities.’’ Quarterly Journal of Economics 115(3):847–904. Baland, Jean-Marie, and Jean-Philippe Platteau. 1999. ‘‘The Ambiguous Impact of Inequality on Local Resource Management.’’ World Development 27(5):773–88. ———. 2001. ‘‘Collective Action and the Commons: The Role of Inequality.’’ In Jean-Marie Baland, Pranab Bardhan, and Samuel Bowles, eds., Inequality, Cooperation and Environmental Sustainabil- ity. Princeton: Princeton University Press. ———. 2003. ‘‘Institutions and the Efficient Management of Environmental Resources.’’ In Karl-Goran Mahler and Jeffrey R. Vincent, eds., Handbook of Environmental Economics, vol. 1. Amsterdam: Elsevier Science, North Holland. Bardhan, Pranab, and Dilip Mookherjee. 1999. ‘‘Relative Capture of Local and Central Governments: An Essay in the Political Economy of Decentralization.’’ CIDER Working Paper. University of California, Center for International and Development Economics Research, Berkeley. Bourguignon, Francois. 1979. ‘‘Decomposable Income Inequality Measures.’’ Econometrica 47(4): 901–20. Cowell, Frank. 1980. ‘‘On the Stucture of Additive Inequality Measures.’’ Review of Economic Studies 47(3):521–31. ———. 2000. ‘‘Measurement of Inequality.’’ In Anthony B. Atkinson and Francois Bourguignon, eds., Handbook of Income Distribution, vol. 1. Amsterdam: Elsevier Science, North Holland. Dasgupta, Indraneel, and Ravi Kanbur. 2001. ‘‘Class, Community, Inequality.’’ Cornell University, Department of Applied Economics and Management, Ithaca, NY. 420 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 3 Dayton-Johnson, Jeff. 2000. ‘‘Determinants of Collective Action on the Local Commons: A Model with Evidence from Mexico.’’ Journal of Development Economics 62(1):181–208. Dayton-Johnson, Jeff, and Pardhan Bardhan. 2002. ‘‘Inequality and Conservation on the Local Com- mons: A Theoretical Exercise.’’ Economic Journal 112(481):577–602. Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to Develop- ment Policy. Baltimore, MD: Johns Hopkins University Press: Demombynes, Gabriel, and Berk O ¨ zler. Forthcoming. ‘‘Crime and Local Inequality in South Africa.’’ Journal of Development Economics. Demombynes, Gabriel, Chris Elbers, Jean O. Lanjouw, Peter Lanjouw, Johan A. Mistiaen, and Berk O¨ zler. 2004. ‘‘Producing an Improved Geographic Profile of Poverty: Methodology and Evidence from Three Developing Countries.’’ In Anthony Shorrocks and Rolf van der Hoeven, eds., Growth, Inequality and Poverty: Prospects for Pro-Poor Economic Development. Oxford: Oxford University Press. ` ze, Jean, Peter Lanjouw, and Naresh Sharma. 1998. ‘‘Economic Development in Palanapur, Dre 1957–1993.’’ In Peter Lanjouw and Nicholas Stern, eds., Economic Development in Palanpur over Five Decades. Oxford: Oxford University Press. Elbers, Chris, Jean O. Lanjouw, and Peter Lanjouw. 2002. ‘‘Micro-Level Estimation of Welfare.’’ Policy Research Working Paper 2911. Development Research Group, World Bank, Washington, DC. ———. 2003a. ‘‘Micro-Level Estimation of Poverty and Inequality.’’ Econometrica 71(1):355–64. ———. Forthcoming. ‘‘Imputed Welfare Estimates in Regression Analysis.’’ Journal of Economic Geo- graphy. ¨ zler, and Kenneth Simler. Forthcoming. Elbers, Chris, Peter Lanjouw, Johan A. Mistiaen, Berk O ‘‘Are Neighbors Equal? Estimating Local Inequality in Three Developing Countries.’’ In Ravi Kanbur and Anthony J. Venables, eds., Spatial Inequality and Development. Oxford: Oxford University Press. Elbers, Chris, Jean O. Lanjouw, Peter Lanjouw, and Phillippe G. Leite. 2003b. ‘‘Poverty and Inequality in Brazil: New Estimates from Combined PPV-PNAD Data.’’ Development Research Group, World Bank, Washington, DC. Elbers, Chris, Tomoki Fujii, Peter Lanjouw, Berk O ¨ zler, and Wesley Yin. 2004. ‘‘Poverty Alleviation through Geographic Targeting: Does Disaggregation Help?’’ Development Research Group, World Bank, Washington, DC. Galasso, Emanuela, and Martin Ravallion. Forthcoming. ‘‘Decentralized Targeting of an Anti-Poverty Program.’’ Journal of Public Economics. Hentschel, Jesko, and Peter Lanjouw. 1996. ‘‘Constructing an Indicator of Consumption for the Analysis of Poverty: Principles and Illustrations with Reference to Ecuador.’’ LSMS Working Paper 124. Devel- opment Research Group, World Bank, Washington, D.C. Hentschel, Jesko, Jean O. Lanjouw, Peter Lanjouw, and Javier Poggi. 2000. ‘‘Combining Census and Survey Data to Trace the Spatial Dimensions of Poverty: A Case Study of Ecuador.’’ World Bank Economic Review 14(1):147–65. Kanbur, Ravi. 2000. ‘‘Income Distribution and Development.’’ In Anthony B. Atkinson and Francois Bourguignon, eds., Handbook of Income Distribution, vol. 1. Amsterdam: Elsevier Science, North Holland. Khwaja, A. 2002. ‘‘Can Good Projects Succeed in Bad Communities? Collective Action in the Himalayas.’’ Harvard University, Department of Economics, Cambridge, MA. La Ferrara, Eliana. 2002. ‘‘Inequality and Participation: Theory and Evidence from Rural Tanzania.’’ Journal of Public Economics 85(2):235–73. Lanjouw, Peter, and Nicholas Stern. 1998. ‘‘Inequality.’’ In Peter Lanjouw and Nicholas Stern, eds., Economic Development in Palanpur over Five Decades. Oxford: Oxford University Press. Mansuri, Ghazala, and Vijayendra Rao. 2004. ‘‘Community Based (and Driven) Development: A Review.’’ World Bank Research Observer 19(1):1–39. Elbers and others 421 McKenzie, David J. 2003. ‘‘Measuring Inequality with Asset Indicators.’’ BREAD Working Paper 042. Bureau for Research in Economic Analysis of Development. Available online at www.cid.harvard. edu/bread. Mistiaen, Johan, and Martin Ravallion. 2003. ‘‘Survey Compliance and the Distribution of Income.’’ Policy Research Working Paper 2956. World Bank, Development Research Group, Washington, DC. ¨ zler, Tiaray Razafimanantena, and Jean Razafindravonona. 2002. ‘‘Putting Welfare Mistiaen, Johan, Berk O on the Map in Madagascar.’’ Africa Region Working Paper 34. World Bank, Washington, DC. Olson, Mancur. 1973. The Logic of Collective Action: Public Goods and the Theory of Groups. Cambridge, MA: Harvard University Press. Platteau, Jean-Philippe, and Frederic Gaspart. 2003. ‘‘The ‘Elite Capture’ Problem in Participatory Development.’’ Working Paper 253/14. FUNDP, University of Namur, Belgium. Sen, Amartya K., and James Foster. 1997. On Economic Inequality (Expanded Edition with Substantial Annexe). Oxford: Oxford University Press. Shorrocks, Anthony. 1980. ‘‘The Class of Additively Decomposable Inequality Measures.’’ Econometrica 48(3):613–25. Simler, Kenneth, and Virgulino Nhate. 2002. ‘‘Poverty, Inequality and Geographic Targeting: Evidence from Small-Area Estimates in Mozambique.’’ International Food Policy Research Institute, Washington, DC.