economic growth di 48795 v2 h economic grow h ec no growth disparitiesdisp economic growth Determinants of Regional Welfare Disparities Within LatinAmerican Countries VOLUME 2: COUNTRY CASE STUDIES BRAZIL, MEXICO, ECUADOR, PERU Determinants of Regional Welfare Disparities within LatinAmerican Countries VOLUME 2:COUNTRY CASE STUDIES (BRAZIL,MEXICO,ECUADOR,PERU) May,2009 Emmanuel Skou as Gladys Lopez-Acevedo Determinants of Regional Welfare Disparities within LatinAmerican Countries Copyright © 2009 by The International Bank for Reconstruction and Development / The World Bank. 1818 H Street, N.W. Washington, D.C. 20433, U.S.A. Internet: www.worldbank.org.mx All Rights Reserved Printing and Manufactured in Mexico / 2009 First Printing: May, 2009 The ndings, interpretations, and conclusions expressed in this book are entirely those of the authors and should not be attributed in any manner to the World Bank, to its af liated organizations, or to members of its Board of Executive Directors or the countries they represent. The World Bank does not guarantee the accuracy of the data included in this publication and accepts no responsibility for any consequence of their use. The boundaries, colors, denominations, and other information shown on any map in this volume do not imply on the part of the World Bank Group any judgment on the legal status of any territory or the endorsement or acceptance of such boundaries. Rights and Permissions The material in this publication is copyrighted. The World Bank encourages dissemination of its work and normally will promptly grant permission to reproduce portions of the work. For permission to reprint individual articles or chapters, please fax a request with complete information to Fernanda Zavaleta, Communications Of cer ­ The World Bank Colombia and Mexico Fax (55) 5480-4222.All other queries on rights and licenses should be addressed to the Communications Of cer, The World Bank Mexico at the above faxed to (55) 5480-4222. The World Bank - Mexico Edition: Christopher S. Humphrey Cover and Design:Alejandro Espinosa World Bank Determinants of Regional Welfare Disparities within LatinAmerican Countries. ­ Mexico : The World Bank, 2009. 2 vol. : il. 1. Urban Poverty - LatinAmerica - Regional Disparities. ­ 2. Rural Poverty - LatinAmerica. 3. LatinAmerica ­ Urban Poverty ­ Regional Disparities. ­ 4. LatinAmerica ­ Rural Poverty. 339.4098/W67 TABLE OF CONTENTS THE SOURCES OFWELFARE DISPARITIESACROSS AND WITHIN REGIONS OF BRAZIL: EVIDENCE FROM THE 2002-03 POF HOUSEHOLD SURVEY 1 Emmanuel Skou as and Roy Katayama SOURCES OFWELFARE DISPARITIES ACROSS REGIONS IN MEXICO 35 Hector Valdes Conroy POVERTY IN LATINAMERICA: SHOULD POLICIES FOCUS ON POOR REGIONSAND/OR POOR PEOPLE? THE CASE OF ECUADOR 97 Monica Tinajero and Gladys López-Acevedo SPATIALDISPARITIES IN LIVING CONDITIONS IN PERU: THE ROLE OF GEOGRAPHIC DIFFERENCES IN RETURNS VS. DIFFERENCES IN MOBILE HOUSEHOLD ASSET ENDOWMENT 125 Carmen Ponce and Javier Escobal REFERENCES 160 Vice President: Pamela Cox Chief Economist: Augusto de la Torre PREM Director: Marcelo Giugale, Ana L. Revenga Sector Manager: Jaime Saavedra, Louise Cord Task Managers: Emmanuel Skou as and Gladys LopezAcevedo DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES ACKNOWLEDGEMENTS T he report is the product of a research program led by Emmanuel Skou as and Gladys Lopez Acevedo with analysis and contributions from Roy Katayama, Hector Valdes Conroy, Francisco Haimovich, Pe- nelope Brown, Erika Strand, Monica Tinajero, Guillermo Rivas, Javier Escobal, Jorge Mario Soto and Christopher Humphrey. The team is grateful for funding and support from the Regional Studies Program of the Of ce of the Chief Econ- omist, with support from the Poverty Reduction and Economic Management (PREM) Poverty Unit in the anchor and the PREM Department of the LatinAmerica and Caribbean Regional Of ce of the World Bank. We would like to thank our peer reviewers, Martin Ravallion, Paul Dorosh, John Nash, Somik Lall, Gabriel De- mombynes, and Alain De Janvry, for their comments and suggestions on this research program. The team also gratefully acknowledges the following people for the support and comments on various drafts: Marcelo Giugale, Tito Cordella, Augusto de la Torre, Ana Revenga, Louise Cord, Jaime Saavedra, Pierella Paci, Farhad Shilpi, Kenneth Simlerm, Pablo Fajnzylber and Ambar Narayan. We are also grateful to the individuals who provided comments and suggestions at the annual meeting of NIP and LACEAin Rio de Janeiro, Brazil, 2008. The ndings, interpretations and conclusions expressed in this paper are entirely those of the authors, and do not necessarily represent the opinions of the World Bank, its Board of Directors or the countries it represents. THE SOURCES OF WELFARE DISPARITIES ACROSS AND WITHIN REGIONS OF BRAZIL: EVIDENCE FROM THE 2002-03 POF HOUSEHOLD SURVEY Emmanuel Skoufias Roy Katayama Abstract B razil's inequalities in welfare and poverty across and within regions can be accounted for by differences in household characteristics and returns to those characteristics. Using Oaxaca- Blinder decompositions at the mean as well as at different quantiles of welfare distributions on regionally representative household survey data, this paper finds that household characteristics account for most of the welfare differences between urban and rural areas within regions. However, comparing the lagging Northeast region with the leading Southeast region, differences in returns to characteristics account for a large part of welfare disparities, in particular in metropolitan areas, pointing to the presence of agglomeration effects in booming areas. JEL classification: Keywords: Brazil, Leading and Lagging Regions, Welfare, Poverty Corresponding author: Emmanuel Skoufias, The World Bank, 1818 H Street NW, Washington DC 20433-USA. Tel: (202) 458-7539. Fax: (202) 522-3134. E-mail: eskoufias@worldbank.org. This paper is part of a Latin America and Caribbean regional World Bank study financed by the LAC Chief Economist office and the PREM Poverty anchor unit. The findings, interpretations, and conclusions in this paper are entirely those of the authors and do not necessarily reflect the view of the World Bank. 1 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 1 1. Introduction Although Brazil is now one of the world's ten largest economies, the country's high levels of poverty and inequality continue to pose major challenges. Like most countries, Brazil's development is spatially uneven. People in metropolitan areas tend to be better off than people in urban non-metropolitan areas, and people in urban areas tend to be better off than people in rural areas. Of Brazil's five major regions, the North and Northeast lag behind the South and Southeast, where economic centers like São Paulo are booming. In 2002, the Northeast was home to 28 percent of the country's population but accounted for 50 percent of the country's poor, and its poverty rate of 38 percent was the highest in the country. On the other hand, the Southeast was home to 43 percent of the population but accounted for 25 percent of the poor and had a poverty rate of 13 percent. Although recent decreases in inequality have been encouraging, the Gini index has nonetheless remained high--above 55--over the last 25 years. These persistently high levels of inequality in Brazil have raised many economic and social concerns. First, high inequality may compromise economic efficiency and growth. For instance, credit and insurance market failures may prevent poorer households from investing in and contributing to the economy at an optimal level, thereby undermining efficiency and growth. Also, inequality in political influence may lead to an inefficient allocation of resources for the public services necessary for greater output. Furthermore, lower social cohesion and greater crime caused by inequality may increase the cost of doing business (World Bank, 2003). Second, high initial levels of inequality may undermine the poverty reduction potential of growth. Based on empirical studies, the growth elasticity of poverty tends to be low in countries with high levels of initial inequality (World Bank 2003; Ravallion 2004). Third, addressing the inequality of opportunities is fundamental to the pursuit of social justice, and doing so constitutes a development objective in itself (Roemer 1998, World Bank 2003 and 2005). Several complementary factors are important in determining the observed spatial distribution of welfare across Brazil. One factor is the concentration and availability of skilled and unskilled labor in a particular location. Since migration is not restricted, the sorting of certain attributes may occur over time to produce variation in the concentration of household attributes in a region or area. Other factors such as the quality of infrastructure, the distance to and size of markets, and the ability of local government to finance public investments and to create the right incentives for private sector development can influence the scope of available opportunities and the rate of return to attributes such as occupational skills or education. In exploring the factors that may be driving this spatial inequality of welfare within Brazil, a fundamental question is whether the observed differences are due primarily to the spatial concentration of individuals with characteristics that tend to leave them in poverty, or due to the geographical differences in the returns to these characteristics.1 In other words, would individuals who live in different regions of the country, but are otherwise identical, have comparable standards of living, indicating that the returns to identical characteristics are similar? Or are their returns and standard of living quite different across regions? Several studies have investigated this question using income as a measure of household welfare, typically obtained from the annual National Household Sample Survey (Pesquisa Nacional por Amostra de Domicílios--PNAD) surveys and employing different methods. Duarte et al. (2004) utilized a semi- parametric model (following DiNardo, Fortin & Lemieux, 1996) to investigate the educational disparities between the Northeast and Southeast regions as a partial determinant of income differences. They concluded that "more than 50 percent of the income difference is explained by the difference in 1This question was posed by Ravallion and Wodon (1999) for Bangladesh. 2 2 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES schooling." However, factors other than education were not considered in their study. Bourguignon et al. (2002) find that "most of Brazil's excess income inequality is due to underlying inequalities in the distribution of two key endowments: access to education and to sources of non-labor income, mainly pensions." This conclusion is based on a comparison of the relative roles of three components--the distribution of population characteristics, the returns to these characteristics, and the occupational structure of the population--in accounting for the income distributions between Brazil and the United States. Guimarães et al. (2006) account for differences in labor income between the metropolitan areas of the Southeast and Northeast regions using the quantile regression decomposition method in Machado and Mata (2004). The paper finds that the difference in returns to education accounts for a larger share of the income gap than the difference in the distribution of education, suggesting that policy interventions focused on education alone are not likely to be sufficient in decreasing regional inequality. It is possible that the higher returns to education that Guimarães et al. (2006) find in the Southeast may be due to agglomeration effects in booming metropolitan areas. As described in the New Economic Geography literature, agglomeration economies are characterized by increasing economies of scale. With well developed infrastructure, a high degree of market specialization, greater competition, information exchange, and more efficient matching in the labor market, the environment is conducive to lowering costs and producing higher returns (Venables 2005, Krugman 1998). Thus, one could expect metropolitan areas of leading regions to have both high returns--from increasing economies of scale-- and a higher concentration of individuals with valuable human capital assets (both observable education and unobservable ability and motivation), as talented workers are attracted to the higher rates of return and wider range of employment opportunities. This study deepens our understanding of poverty and the spatial welfare disparities in Brazil through an investigation of the role of demographics, human capital, occupation, and structural geographical differences in returns. This is done by employing a recent survey in Brazil, the 2002-2003 Household Budget Survey (Pequisa de Orcamentos Familiares, or POF) that allows us to use consumption, rather than income, as a measure of household welfare. While regional disparities in income and welfare have clearly existed for some time in Brazil, there is no consensus on whether this is mainly due to returns to characteristics or the distribution of characteristics. Our paper explores these issues in more depth. Aside from using consumption instead of income as a measure of welfare, the approach here is different from previous studies in that it (i) disaggregates urban 7areas into metropolitan and non-metropolitan areas within regions to obtain a more refined picture and (ii) analyzes both differences in mean welfare and differences between distributions, using Oaxaca- Blinder and quantile regression decomposition techniques. This paper is organized as follows. Section 2 presents the methodology for the decomposition of mean welfare differentials and the quantile regression decomposition along with a brief summary of the 2002- 03 POF survey and the variables used in the analysis. Section 3 presents the results of these decompositions, and Section 4 concludes and discusses some of the implications for policy. 3 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 3 2. Methodology This section outlines the methodology for investigating the factors behind Brazil's spatial disparities in the standard of living. We begin with a brief discussion of the measure used for the standard of living of households and their members, and then summarize the Oaxaca-Blinder methodology used to decompose differences in mean welfare within regions and between regions. The Oaxaca-Blinder decomposition allows the estimation of the relative contributions of differences in household characteristics and in returns in accounting for differences in living standards. Next, we describe the quantile regression decomposition methods used to determine the relative importance of covariate and returns effects at different quantiles of the distributions. Lastly, we describe the data used in our analysis. Comparing Living Standards Within and Across Regions The standard of living measure used in the analysis is the welfare ratio,2 defined as nominal consumption expenditure per capita deflated by the appropriate region-specific poverty line. The region-specific poverty lines are assumed to incorporate all the cost of living differences faced by the poor in different regions and areas.3 As a result, this permits comparisons both within and across regions. Consumption expenditures available from household surveys are preferable for the purpose of poverty and inequality analysis to an indicator such as household income for both conceptual and pragmatic reasons. Consumption expenditures reflect not only what a household is able to command based on its current income, but also whether a household can access credit markets or household savings at times when current incomes are low or even negative (due perhaps to seasonal variation or a harvest failure) to smooth consumption. Hence, a consumption measure is thought to provide a better picture of a household's longer run standard of living than a measure of current income. Furthermore, consumption expenditures for the poor are often better captured than household incomes. While poor households are probably purchasing and consuming only a relatively narrow range of goods and services, their total income may derive from multiple different activities with strong seasonal variation and with associated costs that are not always easily assigned. Lastly, respondents may perceive incentives to under-report income in household surveys. Oaxaca-Blinder Type Decomposition Oaxaca-Blinder decompositions--first proposed by Oaxaca (1973) and Blinder (1973)--decompose the differential in mean log welfare ratios between different regions and areas of a country into two components: one consisting of the differences in average characteristics, and another due to the differences in the returns to characteristics. Various determinants of the welfare ratio are classified into two broad groups: a set of "covariates" that summarize the portable or non-geographic attributes of the household, denoted by the vector , and a set of structural parameters denoted by the vector that summarize the marginal effects or "returns" to these household attributes. The variables in vector include the number and age of household members (excluding housekeepers and renters), education levels, marital status, ethnicity, gender of the household head, and occupations. Education is categorized into five groups: no education, incomplete elementary I (1-3 years), incomplete elementary II (4-7 years), incomplete secondary (8-10 years), and at least 2The welfare ratio and its theoretical properties are discussed by Blackorby and Donaldson (1987). More practical applications of the welfare ratio in the measurement of poverty can be found in Ravallion (1998) and Deaton and Zaidi (2002). 3Since the poverty lines are constructed based on the cost of basic needs (CBN) approach, which in essence is a Laspeyres price index with fixed weights, the welfare ratio is also analogous to "real expenditures". 4 4 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES secondary completed (11 years or more). The various occupations are included in the following categories: (1) professional or military, (2) technician, (3) administrative services, (4) service workers and vendors (reference case), (5) agriculture, (6) manufacturing and industrial services, and (7) other occupation or missing. We would have liked to include land ownership variables, but they were not available in the POF. A variable for religion (Roman Catholic) was included initially but was found to be statistically insignificant and subsequently dropped. Given any two sub-populations, A and B, defined by the region (and area) of residence, the logarithm of the welfare ratio of each sub-population, denoted by is assumed to be summarized by the linear regression: , and (1) , (2) where is a random disturbance term with the usual properties, for summarizing the influence of all other factors on the standard of living. In this specification, the "returns" to characteristics summarize the influence of a variety of factors on the standard of living for the sub-population living in a particular region and area. Basic infrastructure, ease of access to markets and other basic services are some of the most important of these factors. In addition, returns to characteristics are also affected by the role of institutions, social customs and other cultural factors that are typically too difficult to quantify. Based on the specifications above (1 and 2), and given that estimated regression lines always cross through the mean values of the sample, the mean difference in the standard of living between groups A and B can then be expressed as: (3) where the bar over the relevant variables denotes the sample mean values of the respective variables. We have assumed that for . After adding and subtracting the term to the above differences in (3), the difference above can be expressed as: , or (4) (4a) Alternatively, if one were to add and subtract the term , the difference in (3) could be expressed as: , or (5) (5a) 5 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 5 Both expressions (4) and (5) imply that the differential in the mean log welfare ratios between regions A and B can be decomposed into two components: one consisting of the differences in average characteristics summarized by the term , and another due to the differences in the coefficients or returns to characteristics in different regions (and areas) of a country summarized by the term, . The decompositions given by expressions (4) and (5) are equally valid. The only difference between them lies in how the differences in the characteristics and the differences in coefficients are weighted. In expression (4), the differences in the characteristics are weighted by the returns to characteristics in region B, whereas the differences in the returns are weighted by the average characteristics of households in region A. In contrast, in expression (5) the differences in the characteristics are weighted by the returns to characteristics in region A, whereas the differences in the returns are weighted by the average characteristics of households in region B. Since the original decompositions by Oaxaca and Blinder, numerous papers have extended the method by proposing alternative weights for the differences in the characteristics and the differences in returns , (e.g., Reimers, 1983; Cotton, 1988; and Neumark, 1988).4 The general expression allowing for these alternative weights is: , (7) where I is the identity matrix and D is a matrix of weights. The traditional Oaxaca-Blinder decompositions can be considered to be special cases, in which D=0 yields (4) and D=1 yields (5). In addition to using D=0 and D=1, we have followed Reimers (1983) and used as weights the average of the coefficients and the average of the characteristics, that is, the diagonal of the D matrix = 0.5. The use and interpretation of the decomposition method discussed above involves a number of caveats. First of all, these decompositions are simple descriptive tools that provide a useful way of summarizing the role of endowments and returns in explaining existing welfare differentials. For this reason, we refrain from attributing causality to either endowments or returns for the welfare differences between or within regions. Secondly, the variables used in the vector X are composed only of portable non-geographic household characteristics. Our specification intentionally excludes infrastructure and access to basic services. The influence of infrastructure, as well as other omitted variables, is captured by default by the estimated coefficients of the portable characteristics of the household. As the formula for omitted variable bias suggests, the estimated coefficients of household characteristics can be considered as including the direct effect of the omitted variables (such as infrastructure, local institutions and other household variables possibly correlated with the location of the household) on welfare and their correlation with the included household characteristics. Thirdly, the decomposition results may exhibit selection bias. To the extent there is free internal migration within and between different regions, the current place of residence may not be exogenous. The role of selection bias in the decomposition results was explored, as in Ravallion and Wodon (1999), 4For a comprehensive summary, see O'Donnell, et al. (2008) or Ben Jann (2008). In our study, the decompositions employed are done using the Stata command "oaxaca" written by Ben Jann using the "weight ()" option. 6 6 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES and the decomposition results did not change significantly after correcting for selection bias (see Appendix A). Lastly, the decomposition formula in equation (7) holds only at the mean of the two regions being compared, and the decompositions are performed at only one point in time. It is worth considering the extent to which the results of these decompositions change substantially over time.. The potential variation at different points of distributions is examined through quantile decompositions. 7 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 7 Quantile Regression Decomposition Since the role of household characteristics and returns can vary across distributions, we look beyond averages and explore differences across the entire spectrum of welfare distributions by applying the quantile regression decomposition methodology used in recent studies (Machado and Mata, 2005; Nguyen et al., 2007; Shilpi, 2008) that extend the Oaxaca-Blinder decomposition to any quantile of the distribution of living standards. We focus on the leading Southeast and lagging Northeast regions of Brazil, and estimate the relative importance of the returns and covariate effects in accounting for welfare differences across distributions. While the basic idea is the same as traditional Oaxaca-Blinder decompositions, the quantile regression decomposition technique requires the construction of a counterfactual distribution to separate covariate and returns effect. Following Nguyen et al. (2007), we construct a counterfactual distribution by running quantile regressions at each percentile ( ) for sub-population A to estimate coefficient vectors ( ). Each of these coefficient vectors is then used to generate fitted values ( ) of the natural logarithm of welfare ratios using the covariates of sub-population B ( ) for each household i. From each set of fitted values, a sample of 100 values is randomly selected with replacement, and these are combined to yield a counterfactual welfare distribution of households that possess Group B's characteristics but receive Group A's returns; the distribution can be denoted by: . The counterfactual distribution is then used to decompose the difference between two welfare distributions, for example, between metro areas of leading and lagging regions, or between metro and urban distributions of the same region. For any given quantile (q) of the distributions, we can estimate the covariate effect that accounts for differences in household characteristics and the returns effect that accounts for differences in returns and the constant terms. The decomposition can be expressed as: (a) Covariate effect + Returns effect where the first set of brackets represents the covariate effect and the second represents the returns effect. Since an alternative counterfactual distribution, denoted by , could be constructed such that households possess Group A's characteristics and receive Group B's returns, an alternative specification of the decomposition would be: (b) Returns effect + Covariate effect where the first set of brackets represents the returns effect and the second represents the covariate effect. We use both specifications to see whether results are sensitive to the choice of the counterfactual distribution used. 8 8 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Data The 2002-2003 POF survey in Brazil is a household budget survey designed to measure consumption, expenditures, and income. Unlike other Brazilian surveys, the POF is representative at both the national and regional levels for metropolitan, urban (i.e. non-metropolitan urban), and rural areas. The 48,568 households (181,747 individuals) in the POF represented 48,534,638 households (175,331,798 individuals).5 The regional breakdown of households in the sample is shown in Table 2.1. Table 2.1 Sample Households by Region Metro Urban Urban Rural Total North 2,472 2,452 1,957 6,881 Northeast 5,524 8,921 4,218 18,663 Southeast 2,578 , 4,254 , 1,835 , 8,667 , South 1,423 3,666 1,023 6,112 Center West 1,851 4,779 1,615 8,245 Total 13,848 24,072 10,648 48,568 The measure of standard of living used is the log of the "welfare ratio", defined as the nominal per capita consumption deflated by the region-specific poverty line that summarizes the cost of meeting minimum livelihood needs. The 2007 World Bank study on measuring poverty in Brazil estimates separate poverty lines for 21 different metropolitan, urban, and rural areas of each of Brazil's five major regions using the cost of basic needs method (Table 2.2).6 5When the sample size is a small fraction of the population, the finite population correction is close enough to unity and can be ignored (Deaton, p. 43). Thus, for the Brazil POF, where one out of every 1,000 households were sampled, we chose to ignore the finite population correction for simplicity with the understanding that the calculated standard errors may be slightly larger than if the finite population correction were accounted for. 6 The CBN poverty lines used are based on the lower estimate of the adjustment to the food poverty line for basic nonfood expenses. For a detailed discussion of the construction of the region-specific poverty lines in Brazil, see World Bank, 2007. Note that a welfare ratio equal to one, or equivalently a log welfare ratio equal to 0, represents a household per capita consumption equal to the poverty line. 9 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 9 Table 2.2: Regional Poverty Lines and Mean Expenditures Mean Per Capita Lower Lower Expenditures Expenditures Poverty Line Mean Welfare Mean elfare Region (R$/month) (R$/m (R$/month) (R$/m Ratio 1 Metro Belém 299.0 105 2.8 2 North N th Urban 238.2 102 2.3 3 Rural 135.0 93 1.5 4 Metro Fortaleza 309.4 99 3.1 5 Metro Recife 331.3 104 3.2 6 Northeast h Metro Salvador 386.8 108 3.6 7 Urban 207.6 100 2.1 8 Rural 111.9 92 1.2 9 Metro Rio De Janeiro 547.7 107 5.1 10 Metro São Paulo 525.3 115 4.6 11 Southeast Metro Belo Horizonte 429.1 103 4.2 12 Urban 381.3 109 3.5 13 Rural 207.0 97 2.1 14 Metro Curitiba 522.8 105 5.0 15 Metro Porto Alegre g 485.0 111 4.4 South 16 Urban 368.3 99 3.7 17 Rural 236.9 90 2.6 18 Brasília 596.2 109 5.5 19 Goiânia municipality 425.9 103 4.1 Center West est 20 Urban 268.5 105 2.6 21 Rural 217.7 100 2.2 Total 335.9 103 Source: Table 12 in World Bank (2007) and authors' estimates using the 2002-03 POF. While mean welfare ratios and poverty rates vary considerably both within and across regions, clear patterns emerge. Within regions, mean welfare ratios are the lowest in rural areas, as one might expect, and highest in metropolitan areas. Likewise, poverty rates are the highest in the rural areas and lowest in metropolitan areas (Figure 2.1). Across regions, the Northeast and North exhibit the lowest welfare ratios, while the Southeast, South, and Center West regions have the higher welfare ratios. The Northeast consistently has the highest poverty rates in metropolitan, urban and rural areas while the South has the lowest (Figure 2.2). The next section will investigate the factors behind these trends. 10 10 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 2.1: Poverty Within Regions Figure 2.2: Poverty Across Regions 11 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 11 3. Results This section begins with a discussion of regression results for metropolitan, urban, and rural areas, followed by results of the Oaxaca-Blinder decompositions of differences in mean consumption expenditures within and across regions. Lastly, the quantile regression decompositions within and between the Northeast and Southeast regions are discussed. Regression Results The regression results for metropolitan, urban, and rural areas are presented in Table 3.1 and summarized below. When interpreting the regression results, note that the reference household in the analysis is comprised of a single person of mixed race ancestry ("parda") who has no children, no schooling, is a vendor/service worker, and lives in the Southeast region. The Southeast is the reference region included in the constant term, and four binary variables identifying the North, Northeast, South and Center-West regions were included in the regression. The Oaxaca-Blinder decompositions discussed subsequently are not based on the coefficients presented in Table 3.1 since the groups used in the decompositions are divided not only by regions but by metro, urban, and rural areas as well. Also, the null hypothesis of equality of coefficients between metropolitan and urban areas, and between urban and rural areas, can be rejected for all categories of variables (see Appendix A). Region of residence: The metro, urban, and rural areas of the Northeast region appear to have structural regional disadvantages relative to the Southeast reference region. The rural area of the South has the largest regional advantage. The metro and urban areas of the North and Center West have slight disadvantages. All other areas exhibit little difference relative to corresponding areas of the Southeast. Household demographics: Welfare tends to decrease as family size increases, as indicated by the negative coefficients on variables representing the number of household members. In general, additional household members will decrease per capita welfare ratios, but given the functional form (log welfare ratio = (# in age group) + (# in age group squared)), the incremental change for each additional household member is not as great. Other demographics: Controlling for other factors, households with a head who is white or Asian tend to have higher welfare than households with heads of other ethnicities. The ethnicity of the head is categorized into three groups: parda (reference case), white or Asian, or other (black, indigenous, or not identified). Controlling for other factors, having a spouse tends to decrease per capita welfare ratios by about 9-13 percent. The coefficients for the age of head and age of head squared are both positive, indicating that welfare tends to increase with the age of household. The regression estimates for age have a slight convex shape to the curve, but the relationship is nearly linear. This may be due to a generous pension system. 12 12 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table 3.1 Regressions for Log Welfare Ratios for Metro, Urban and Rural areas of Brazil Dependent var: De endent var: ln(welfare ratio welfare ratio) Metro etro Urban Urban Rural Rural Constant 0.773 *** 0.012 0.469 *** Geographic Regions North -0.137 *** -0.040 0.017 Northeast -0.242 *** -0.242 *** -0.248 *** South 0.029 0.018 0.134 ** Center West -0.081 *** -0.053 ** -0.007 Demographics hics # age 0-2 0 2 -0.378 *** -0.358 *** -0.281 *** 0.378 0.358 0.281 # age 0-2 squared uared 0.050 *** 0.050 0.060 *** 0.060 0.034 ** 0.034 # age 3-11 -0.339 *** -0.292 *** -0.229 *** # age 3-11 squared uared 0.035 *** 0.035 0.030 *** 0.030 0.013 *** 0.013 # age 12-17 -0.245 *** -0.208 *** -0.196 *** # age 12-17 squared uared 0.031 *** 0.021 *** 0.022 *** # age 18-59 -0.171 *** -0.183 *** -0.212 *** # age 18-59 squared uared 0.012 *** 0.012 0.018 *** 0.018 0.020 *** 0.020 # age 60+ -0.199 *** -0.155 *** 0.004 # age 60+ squared uared 0.016 0.035 ** -0.019 Female head -0.060 ** -0.054 ** -0.113 *** Spouse -0.163 *** -0.166 *** -0.144 *** White or Asian 0.171 *** 0.158 *** 0.108 *** Black indigenous, or not identified -0.020 Black, indi enous 0.007 -0.029 Age of head 0.018 *** 0.032 *** 0.019 *** Age of head squared uared / 100 -0.010 ** -0.023 *** -0.015 *** Education of head 1-3 ears 1-3 years -0.051 0.130 *** 0.112 *** 4-7 years 0.139 *** 0.299 *** 0.241 *** 8-10 years 8-10 0.276 *** 0.490 *** 0.444 *** 11+ years 0.629 *** 0.806 *** 0.605 *** Education of spouse 1-3 years 0.010 0.080 *** 0.096 *** 4-7 years 4-7 0.095 ** 0.181 *** 0.198 *** 8-10 years 0.189 *** 0.302 *** 0.351 *** 11+ years ears 0.438 *** 0.438 0.518 *** 0.518 0.483 *** 0.483 Education differential 1-3 years 1-3 ears 0.048 0.005 0.031 4-6 years 0.125 *** 0.115 *** 0.098 *** 7-9 years ears 0.140 *** 0.140 0.168 *** 0.168 0.212 *** 0.212 10+ years 0.531 *** 0.545 *** 0.385 *** Occupation ation Professional 0.542 *** 0.443 *** 0.177 *** Technician Technician 0.234 0.234 *** 0.213 *** 0.213 0.028 0.028 Administrative 0.088 ** 0.083 ** -0.006 Agriculture riculture -0.088 -0.099 *** -0.170 *** Manufacturing / Industry 0.001 0.027 -0.033 Missing Missin / not defined 0.060 ** -0.047 * -0.222 *** N 13 848 13,848 24 24,072 10,648 10 648 R-squared uared 0.583 0.583 0.534 0.534 0.438 0.438 *p<0.10, <0.10 ** p<0.05, ***p<.01 <0.05 *** .01 13 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 13 Education: Education plays an important role in household welfare. All else equal, higher educational attainment is positively correlated to higher welfare. Attainment of at least 11 years of education by the household head tends to increase the welfare ratio by 63 percent in metro areas, 81 percent in urban areas, and 60 percent in rural areas. Having a spouse with education also tends to increase household welfare (by 50 percent on average with 11 or more years of education), although to a lesser extent than the head. When household members (e.g. children) have more education than either the head or spouse, household welfare also tends to be higher, in particular when the difference is large. For instance, with a difference of 10 years or more, per capita welfare ratio increases by over 50 percent in metropolitan and urban areas and 38 percent in rural areas.7 Occupation: There is a large disparity in returns for professional and technical occupations between metropolitan/urban and rural sectors (0.548 and 0.477 versus 0.188). A household head whose primary occupation is agriculture tends to have a welfare ratio about 9 percent lower in urban areas and 16 percent lower in rural areas, relative to service workers (the reference case). Mean welfare exceeded the poverty line by 177 percent in metro areas, 97 percent for urban areas, and 22 percent for rural areas (Table 3.2). The largest negative factor is household size and composition, and the largest positive factors include education and other demographics. Also, the constant term accounts for a substantial part of the expected mean welfare for metro and rural areas. Table 3.2: Contributing Factors to Average Levels of Living and Metro-Urban and Urban-Rural Disparities Metro Urban Urban Rural Mean ean log welfare ratio 1.02 0.68 0.20 Constant Constant 0.77 0.01 0.47 Geographic dummies g p -0.06 -0.07 -0.09 HH Composition -0.77 -0.72 -0.81 HH Demographics g p 0.52 0.85 0.45 Education 0.47 0.57 0.31 Occupation 0.10 0.04 -0.13 Oaxaca-Blinder Decomposition--Within Regions Both characteristics and returns play a major role in explaining metropolitan-urban welfare differentials. Oaxaca-Blinder decompositions8 indicate that returns account for about 60 to 70 percent of the log welfare difference in the South and Southeast regions, whereas characteristics account for about 60 percent of the log welfare difference in the North, Northeast, and Center-West regions (Figure 3.1). Household characteristics seem to be the dominant explanation for urban-rural welfare differences within regions. Oaxaca-Blinder decomposition results indicate that characteristics account for nearly all of the urban-rural differences (Figure 3.1).9 The results from poverty simulations, following Ravallion and Wodon (1997), are consistent with these findings (see Appendix B). 7 Elementary education is compulsory and lasts eight years, divided into two four-year cycles. Secondary education is another three years. 8The decomposition was done as in Reimers (1983) using the average of the two groups' coefficients for the so-called "nondiscriminatory" coefficient vector. 9The results shown in Figures 3.1 and 3.2 are for decompositions using a weighting matrix of D=0.5, as in Reimers (1983). Results for other weighting specifications (i.e. D=0 and D=1) are included in Appendix C. While the choice of the weighting specification can shift these results 14 14 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 3.1: Oaxaca-Blinder Decompositions of Welfare Differential Within Regions South Southeast Between Metro No rtheast and Urban N o rth C enter West -10 0 10 20 30 40 50 60 70 80 90 100 110 South C enter West Between Urban Northeast and Rural Southeast N orth -10 0 10 20 30 40 50 60 70 80 90 100 110 % of difference in log welfare ratios Endow ments Coefficients Oaxaca-Blinder Decomposition--Across Regions Although both household characteristics and returns play important roles in accounting for welfare differences between regions, characteristics tend to be a slightly greater factor in most cases, in particular when comparing metropolitan areas. The Oaxaca-Blinder decomposition results indicate that 45 to 75 percent of the difference in log welfare ratios between the lagging Northeast and North regions and the leading South and Southwest regions is explained by differences in household characteristics (Figure 3.2). However, in comparisons involving the Northeast region, both returns and characteristics are about equally important, with returns playing a slightly more dominant role in Northeast-Southeast and Northeast-South comparisons of metro areas. Figure 3.2: Oaxaca-Blinder Decompositions of Welfare Differentials Across Regions S-Northeast SE-Northeast CW-Northeast Across Metro S-North SE-North CW-North -10 0 10 20 30 40 50 60 70 80 90 100 110 slightly, the relative importance of either the characteristics or returns components in accounting for welfare differences is fairly robust to the choice of the weighting matrix D. 15 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 15 CW-Northeast S-Northeast SE-Northeast Across Urban S-North SE-North CW-North -10 0 10 20 30 40 50 60 70 80 90 100 110 CW-Northeast SE-Northeast S-Northeast Across S-North Rural SE-North CW-North -10 0 10 20 30 40 50 60 70 80 90 100 110 % of difference in log welfare ratios Endow ments Coefficients 16 16 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Differences in Welfare Distributions The set of figures below (Figure 3.3) summarizes the welfare differences along the per capita consumption distribution between metropolitan and urban areas and between urban and rural areas within the Southeast region and within the Northeast region. Figure 3.3: Welfare Differences Within the Southeast and Northeast Regions (a) (b) (c) (d) The difference in welfare ratios between metropolitan and urban areas within the Southeast region has a U shape, meaning that welfare differences are higher at the bottom and at the top of the distribution of welfare in the leading region. In contrast, the differences in welfare between urban and rural areas in the Southeast increase almost monotonically with the level of welfare. Within the Northeast, the difference in welfare between metropolitan areas and urban areas is also higher at higher level of welfare, suggesting that the wealthier households in the metropolitan areas of the Northeast are much better off than the wealthier households in the urban areas of the Northeast. Lastly, the differences in welfare between households in the urban and rural areas of the Northeast suggest that the differences are larger primarily among relatively wealthier households (above the 50th percentile of the per capita consumption distribution). See Appendix D for welfare differences for other regions. Figure 3.3 suggest that decompositions of the welfare differences at the mean may yield a misleading picture about the relative role of characteristics and their returns in explaining these differences. 17 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 17 Quantile Decompositions Within Regions (Southeast and Northeast) Decompositions of the metro-urban and urban-rural differences within the Southeast region and within the Northeast region along the consumption distribution reveal that both returns and covariate effects play a role in explaining the large differential in living standards between metro and urban areas (Figure 3.4). See Appendix D for decompositions using alternative counterfactual distributions. Figure 3.4: Quantile Decomposition of Welfare Differences Within Southeast and Northeast Regions (a) Between metro and urban areas of SE (b) Between urban and rural areas of SE (c) Between metro and urban areas of NE (d) Between urban and rural areas of NE In the Southeast region (Figure 3.4a), the returns effect is dominant throughout most of the distribution, which is consistent with the result in the previous section, and further supports the idea of agglomeration effects in the leading metropolitan area.10 In the Northeast region, the metro-urban comparison (Figure 3.4c) shows that both the returns and covariate effects play a role in the observed difference, with the covariate effect at least as large as the returns effect. The returns effect increases for households from the middle to the top of the welfare ratio distribution, indicating that those in the metro area obtain higher returns, in particular for those better off. However, the returns effect is smaller for poorer individuals relative to those better off. 10We also estimated bootstrapped standard errors around the estimated returns effect and covariate effects. The standard error bands do not overlap for most of the welfare distribution, which implies that there are significant differences in the estimated returns and characteristics effects. For more details, see the Brazil case study report. 18 18 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Decomposing welfare differences between urban and rural areas (Figures 3.4b and 3.4d), most of the welfare disparity is accounted for by the covariate effect, with little difference in returns, throughout the distribution. Thus, households possess more favorable attributes in urban areas vis-à-vis rural areas, while the differences in returns are relatively small. These results are consistent with the results from the mean decomposition. One possible explanation for little to no returns effect between these areas is that migration may be equalizing returns across areas. Migration between urban and rural areas within the Southeast and within the Northeast may involve lower costs/risks (e.g. more likely to have social networks to assist with the transition, less distance from home, temporary employment) than between regions. Relatively cost-free migration is expected to facilitate migration flows and thereby equalize the returns of portable characteristics between the origin and destination areas, as long as there are no agglomeration effects in the destination region. Also, through migration, sorting based perhaps on education or some unobserved ability may occur such that we arrive at the observed concentration of poor people and covariate effects. In conclusion, the results of within region decompositions indicate: (i) a large returns effect in accounting for higher welfare in the metro Southeast relative to the urban Southeast; (ii) a combination of covariate and returns effect in accounting for higher welfare in the metro Northeast relative to the urban Northeast; and (iii) a dominant covariate effect and little to no returns effect in accounting for higher welfare in urban over rural areas. These results are consistent with the findings from the analysis of the means presented earlier in this section. Quantile Decompositions Across Regions (Southeast and Northeast) The next set of decompositions explores contribution of the returns and covariate effects in accounting for differences in the distributions of welfare between similar areas of the leading Southeast and the lagging Northeast regions. In comparing the metro areas of the Southeast and Northeast regions (Figure 3.5a), the returns effect is greater than the covariate effect in accounting for the differences between welfare distributions. Considering that metro Southeast (i.e. São Paulo and Rio de Janeiro) has a high density of economic activity, better infrastructure, and serves as a hub for trade, a large returns effect suggests the presence of agglomeration effects. Between urban areas (Figure 3.5b) and between rural areas (Figure 3.5c) of these two regions, both the covariate and returns effects contribute to the difference. The upward slope of the returns effects indicates that the returns effect is smaller for the poor relative to those better off. 19 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 19 Figure 3.5: Quantile Decomposition of Welfare Differences Between Southeast and Northeast Regions (a) Between SE and NE metro areas (b) Between SE and NE urban areas (c) Between SE and NE rural areas In conclusion, the results indicate: (i) a large returns effect in accounting for higher welfare in the metro Southeast relative to the metro Northeast; and (ii) a combination of covariate and returns effects in accounting for higher welfare in the urban and rural Southeast relative to urban and rural Northeast, respectively. Again, these results are broadly consistent with the findings from the analysis of the means presented earlier. 20 20 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 4. Conclusions And Policy Implications Welfare disparities between metropolitan, urban, and rural areas within and between the five regions of Brazil remain large, in spite of many federal and state government programs devoted to alleviating such disparities. As discussed in the introduction, this could be explained either because individuals with personal attributes that inhibit growth in their living standards concentrate in certain areas, or that in areas not well endowed with local public goods, such as better infrastructure and other basic services, geographic externalities lead to lower returns to personal attributes, and hence to lower overall welfare. This paper classifies factors associated with spatial differences in the standard of living into two broad groups: a set of "covariates" that summarize the portable or non-geographic attributes of the household, such as age, level of education, type of occupation etc., and a set of parameters that summarize the marginal effects or "returns" to these characteristics (either at the mean or at different points of the welfare distribution). Based on this framework, we then address the question of whether the spatial disparities in welfare and poverty are better explained by the sorting of people with low portable characteristics in some areas, or by persistent spatial differences in the returns to portable characteristics such as human capital. The goal is to provide more guidance for the design and prioritization of policies aimed to reduce poverty in different areas. The decomposition of means indicates that the welfare disparities across regions are associated more with the concentration of people with similar observable household attributes than the differences in returns to those attributes. Moreover, in the Northeast, both household characteristics and returns are much lower than in other regions of Brazil. Regarding differences in the standard of living within regions, our findings reveal that the covariate effect is the primary explanation for the differences between metropolitan and urban areas or between urban and rural areas in different regions of Brazil, with the exception of the Southeast. While returns effects also accounted for a non-trivial part of the difference in living standards, covariate effects were generally the dominant explanation. In the Southeast, the returns effect turns out to be the dominant explanation for differences in the standard of living between metropolitan and urban areas. This result is supportive of the existence of agglomeration effects in the metropolitan areas of the Southeast (i.e., São Paulo and Rio de Janeiro) where a high density of economic activity and better infrastructure exist. However, even in the Southeast, the covariate effect is the dominant explanation for living standards differences between urban and rural areas. As well, the findings indicate that differences in returns between urban and rural areas play a very small role in explaining living standards differentials within the Southeast and Northeast regions. Thus, most of the welfare disparity between urban and rural areas in both the Southeast and the Northeast is explained almost exclusively by the covariate effect. That is, households in urban areas possess more favorable attributes than households in rural areas. Comparisons between metropolitan areas in the Southeast and the Northeast regions provide further validation of the presence of substantial agglomeration effects in the metropolitan areas of the Southeast. These agglomeration effects lead to substantially higher returns to portable assets, such as education, in the metropolitan areas of the Southeast in comparison to the returns in metropolitan areas of the Northeast. Further comparisons between urban areas and rural areas of these two regions suggest that both the covariate and returns affect contribute to the difference, with the covariate effect dominating especially for the poorer households. 21 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 21 The results also shed some light on the role of labor migration on welfare disparities across and within regions of Brazil. On the one hand, the absence of any significant role for the returns effect as an explanation for welfare disparities between urban and rural areas in the Southeast or the Northeast regions suggests that migration of labor between urban and rural areas within regions is able to equalize returns to individual attributes within regions. Therefore, welfare differences between urban and rural regions seem to be primarily due to the sorting or concentration of people with higher attributes in the urban areas of these regions. On the other hand, the dominant role of the returns effect in explaining varying living standards between metropolitan areas in the Southeast and metropolitan areas in the Northeast suggests that the persistent and large migration of workers from the lagging Northeast to the metropolitan areas of the Southeast is not able to diminish the differences in the returns across regions. In fact, migration from the Northeast to the Southeast may also be a leading cause of these welfare inequalities. To the extent that the migration of workers from the Northeast to the Southeast enhances overall productivity and economic growth due to the positive externalities associated with clustering human capital in the metropolitan areas of the Southeast, then it should be encouraged and facilitated in spite of the magnitude and apparent persistence of inequality in the living standards in the Northeast region. Overall, the study's findings validate the recent change in strategy towards poverty alleviation in Brazil encapsulated by the Bolsa Familia program. Since the 1950s, government policies in Brazil focused on diminishing regional inequality in Brazil through direct government investments in infrastructure, public credits subsidizing private initiative, and related territorial development programs in the Northeast. By contrast, the key characteristic of Bolsa Familia is that it considers the lack of sufficient human capital rather than geography as the primary cause of extreme poverty, and it employs monetary and in-kind benefits as instruments for encouraging poor families to invest in the education, health and nutrition of their children. In Bolsa Familia, as in other conditional cash transfer programs, geographic targeting is only a means of finding the areas where poor households are likely to be located. This new public policy emphasis represents a major step in the right direction not only in the fight against poverty but also towards reducing spatial welfare disparities in the long run. Although quite tentative, the inferences regarding the role of migration in Brazil also suggest a set of policies complementary to Bolsa Familia. As long as some people migrate out of the Northeast because of push rather than pull factors, such as limited access to or low quality of basic social services like health and education, then programs focusing on the Northeast region should concentrate on increasing access to and quality of these basic services. More empirical evidence on the determinants of migration between the lagging and leading regions of Brazil, along the lines of Lall, Timmins and Yu (2008), can be particularly helpful in guiding the design of government interventions that can enhance both equality of opportunity in the lagging regions as well as aggregate productivity. 22 22 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Appendix A Appendix A contains: a map of Brazil's major regions, GDP per capita and inequality trends, various descriptive statistics on population and poverty distributions, and a brief discussion of selection bias. Figure A1: Major Regions and States of Brazil Source: http://gosouthamerica.about.com Figure A2: GDP Per Capita and Gini Index: 1981-2005 Source: WDI 2008 23 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 23 Table A1: Descriptive Statistics of Regional Distribution of Population and Poverty Total Population Number Number Distribution Headcount Headcount Poverty Poverty Region Region Population Share Share of Poor of the Poor Poverty Gap Gap Severity Severity 1 Metro Belém 1,845,708 1.05% 307,931 0.8% 16.7% 4.2% 1.7% 2 North Urban 8,229,439 4.69% 2,252,145 6.0% 27.4% 8.9% 4.0% 3 Rural 3,533,713 2.02% 1,615,266 4.3% 45.7% 16.3% 7.7% Metro 4 Fortaleza 2,985,823 1.70% 620,452 620,452 1.6% 20.8% 6.4% 2.7% 5 Metro Recife 3,331,278 1.90% 588,597 588,597 1.6% 17.7% 5.4% 2.5% Northeast 6 Metro Salvador 3,088,893 3,088,893 1.76% 444,994 1.2% 14.4% 4.4% 2.1% 7 Urban 25,579,176 25,579,176 14.59% 14.59% 9,420,720 25.0% 36.8% 13.3% 6.4% 8 Rural 13,940,461 13,940,461 7.95% 7.95% 7,664,897 20.3% 20.3% 55.0% 55.0% 22.3% 22.3% 12.0% 12.0% Metro Rio de 9 Janeiro 11,052,249 6.30% 970,581 2.6% 8.8% 2.3% 0.9% Metro São 10 Paulo 17,696,179 10.09% 1,278,930 3.4% 7.2% 1.7% 0.7% Southeast Metro Belo 11 Horizonte 4,437,346 2.53% 316,681 0.8% 7.1% 1.4% 0.5% 12 Urban 35,016,773 19.97% 5,045,726 13.4% 14.4% 4.7% 2.2% 13 Rural 6,586,851 3.76% 1,859,310 4.9% 28.2% 9.4% 4.4% 14 14 Metro Curitiba 2,641,166 2,641,166 1.51% 125,209 125,209 0.3% 0.3% 4.7% 4.7% 1.2% 1.2% 0.4% 0.4% Metro Porto 15 15 South South Alegre 3,663,574 3,663,574 2.09% 263,083 263,083 0.7% 7.2% 1.6% 0.5% 16 16 Urban 15,083,301 15,083,301 8.60% 1,825,518 4.8% 4.8% 12.1% 12.1% 3.6% 3.6% 1.6% 1.6% 17 17 Rural 4,438,516 4,438,516 2.53% 704,361 704,361 1.9% 1.9% 15.9% 15.9% 4.7% 4.7% 2.0% 2.0% 18 Brasília 2,151,035 1.23% 173,966 0.5% 8.1% 1.5% 0.4% Goiânia 19 Center municipality 1,121,683 0.64% 57,457 0.2% 5.1% 1.4% 0.6% West 20 Urban 7,526,053 4.29% 1,722,950 4.6% 22.9% 7.2% 3.3% 21 Rural 1,382,581 0.79% 427,710 1.1% 30.9% 10.9% 5.7% Total 175,331,798 100.00% 37,686,485 37,686,485 100.0% 21.5% Source: Authors' estimates using the 2002-03 POF; World Bank (2007). Table A2 : Poverty by Geographic Regions Poverty Headcount Distribution of the Distribution Distribution of Rate Rate Poor Population 2002 2002 2002 Urban 17.5 67.4 82.9 Rural Rural 41.0 41.0 32.6 32.6 17.1 North 30.7 11.1 7.8 Northeast 38.3 49.7 27.9 Southeast Southeast 12.7 25.1 42.7 42.7 South 11.3 7.7 14.7 Center Center 19.6 6.3 6.3 6.9 West West Total Total 21.5 21.5 100.0 100.0 100.0 100.0 24 24 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure A3: Distributions of Welfare Ratios by Region Frequency weighted Epanechnikov, halfwidth of kernel = 0.15 ----- Rural ----- Urban (non-metro) ----- Metropolitan Note: Since a welfare ratio of one, or equivalently a log welfare ratio equal to zero, means that an individual is living at the poverty line, the area to the left of the dashed vertical line in Figure A.3 represents the population living in poverty. Table A3: Tests of Equality of Coefficients between Metro, Urban, and Rural Regressions Metro=Urban tro=Urban Urban=Rural Urban=Rural F-test F-test F-test F-test Year 2002-2003 Restrictions F-value (1 percent level) F-value (1 percent level) Nongeographic Variables g g p HH Structure (hhmem) (hhm 10 14.46 14.46 Rejected Rejected 24.03 Rejected Rejected HH Demographics (hhdem) g p ( 5 15.39 Rejected j 24.83 Rejected j Education 12 25.84 Rejected 20.89 Rejected Occupation p 6 5.83 Rejected j 12.13 Rejected j Geographic dummies Geographic 4 9.75 Rejected 3.91 Rejected 25 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 25 Table A4: Selection Bias Corrected and Uncorrected Coefficients Urban Northeast Rural Northeast Dependent: log welfare ratio movestay ols movestay ols # age 0-2 -0.322*** -0.313*** -0.285*** -0.296*** # age 0-2 squared 0.050** 0.046** 0.032 0.037** # age 3-11 -0.292*** -0.289*** -0.203*** -0.204*** # age 3-11 squared 0.033*** 0.033*** 0.009* 0.009* # age 12-17 -0.181*** -0.180*** -0.206*** -0.204*** # age 12-17 squared age squared 0.018** 0.016* 0.023** 0.024** # age 18-59 -0.168*** -0.169*** -0.206*** -0.205*** # age 18-59 squared age squared 0.017*** 0.017*** 0.018*** 0.018*** # age 60+ g -0.122** -0.097* 0.083 0.059 # age 60+ squared age squared 0.023 0.019 -0.014 -0.010 Female head -0.042 -0.028 -0.064 -0.074 Spouse Spouse -0.146*** -0.158*** -0.169*** -0.159*** White or Asian 0.083*** 0.075*** 0.039 0.050 Black, Indigenous, or Missing Black, Indigenous, ssing 0.019 0.012 -0.076 -0.066 Age of head 0.023*** 0.025*** 0.023** 0.021*** Age of head squared / 100 Age squared -0.016*** -0.018*** -0.018* -0.016*** Education of head 1-3 years years 0.133*** 0.147*** 0.112*** 0.105*** 4-7 yyears 0.224*** 0.269*** 0.260* 0.224*** 8-10 years years 0.373*** 0.468*** 0.497 0.380*** 11+ yyears 0.655*** 0.750*** 0.725 0.588*** Education of spouse spouse 1-3 yyears 0.125*** 0.126*** 0.134*** 0.134*** 4-7 years years 0.231*** 0.254*** 0.247*** 0.228*** 8-10 yyears 0.334*** 0.392*** 0.513** 0.450*** 11+ years years 0.592*** 0.655*** 0.715** 0.630*** Education differential 1-3 years years 0.036 0.032 0.046 0.048 4-6 yyears 0.095*** 0.116*** 0.138* 0.120*** 7-9 years 7-9 0.191*** 0.260*** 0.327 0.261*** 10+ years 10 0.351*** 0.436*** 0.383 0.298*** Occupation Professional 0.468*** 0.461*** 0.113 0.114 Technician 0.103* 0.107** -0.036 -0.051 Administrative 0.115** 0.124** 0.027 0.010 Agriculture 0.046 -0.152*** -0.347 -0.188*** Manufacturing / Industry g y -0.027 -0.023 0.029 0.027 Missing / not defined -0.088*** -0.096*** -0.263*** -0.253*** _cons 0.092 -0.116 0.326 0.128 rho 1 (urban) -0.506** rho 2 (rural) 0.382 note: *** p<0.01, ** p<0.05, * p<0.1; Urban excludes metro areas p<0.1; 26 26 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Since households may be making the decision to migrate or not based on expected welfare gains or losses given household characteristics, the location of residence should not be assumed to be exogenous. Using a full information maximum-likelihood estimation of an endogenous switching regression model, we explore the direction of the bias for urban and rural areas of the Northeast region. However, as in most cases, the correct specification of the switching equation always poses a challenge. Without good instrumental variables at hand, the model was simply identified by non-linearities. A comparison of the selection bias-corrected and uncorrected (OLS) coefficients are presented in Table A4 (below). In the urban areas, the uncorrected education coefficients are higher than corresponding selection bias corrected coefficients. In rural areas, the converse is true. Since the urban coefficients are lower and rural coefficients higher, the difference in the coefficients, that is, urban minus rural, would be smaller, and since the returns effect is essentially the difference in coefficients, the returns effect would be smaller as well. Assuming the direction of the bias is correct, these results do not undermine the finding that the urban-rural differences in the Northeast are accounted for by the covariate effect and not the returns effect. 27 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 27 Appendix B Simulated Poverty Profiles We have followed Ravallion and Wodon (1988) and simulated the poverty profiles that would prevail if the mean values for individual portable characteristics for different regions and areas were fixed at national means and thereby did not vary between and within regions (geographic or returns poverty profile) and poverty profiles that would prevail if the returns to characteristics did not differ between and within regions (concentration or endowment poverty profiles). In many cases this alternative approach allows one to determine visually whether it is differences in the returns to household characteristics or differences in the characteristics themselves that can better explain the differences in the standard of living between and within regions. First, in each region k a regression equation is estimated separately for metro, urban, and rural areas, as in: . . (1) . The simulated geographic poverty profile is constructed using the constants and coefficients from equations (1) using the expressions: (2) where are the standard deviation of errors for each region/area and is the national mean value of the individual endowments (urban and rural areas pooled). The simulated concentration poverty profiles are constructed as: (3) where and . 28 28 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure B1 below presents the simulated geographic and concentration poverty profiles together with the actual poverty profile in the metropolitan, urban and rural areas of Brazil. In rural areas, the similarity of the concentration poverty profile across regions to the actual or unconditional poverty profile suggests that household endowments are primarily responsible for the level and the variation of poverty across regions. Figure B1: Simulated Poverty Profiles Metro Poverty Profiles (B) 60% 50% ytre 40% ovp t 30% uno dca 20% he 10% 0% Geographic Concentration Unconditional Northeast (Metro) North (Metro) Center West (Metro) Southeast (Metro) South (Metro) Urban Poverty Profiles (B) 60% 50% ytre 40% pov 30% untocdaeh 20% 10% 0% Geographic Concentration Unconditional Northeast (Urban) North (Urban) Center West (Urban) Southeast (Urban) South (Urban) Rural Poverty Profiles (B) 60% 50% ytrev 40% po 30% untocda 20% he 10% 0% Geographic Concentration Unconditional Northeast (Rural) North (Rural) Center West (Rural) Southeast (Rural) South (Rural) 29 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 29 Appendix C Table C1: Oaxaca-Blinder Decompositions Within Regions (a) Difference in log welfare ratios explained by differences in endowments and coefficients Metro-Urban Metro-Urban Urban-Rural Urban-Rural threefold threefold twofold twofold threefold threefold twofold twofold North D=0 D=0.5 D=1 D=0 D=0.5 D=1 endowments 0.181 0.175 0.178 0.181 0.311 0.356 0.333 0.311 coefficients 0.121 0.121 0.118 0.115 -0.021 -0.021 0.002 0.024 interaction -0.006 - - - 0.045 - - - Northeast Northeast endowments 0.273 0.284 0.278 0.273 0.273 0.391 0.464 0.427 0.391 coefficients 0.191 0.191 0.197 0.197 0.203 0.203 -0.024 -0.024 0.012 0.012 0.049 0.049 interaction 0.011 - - - 0.073 - - - Southeast endowments 0.105 0.125 0.115 0.105 0.185 0.248 0.216 0.185 coefficients 0.176 0.176 0.186 0.195 -0.028 -0.028 0.004 0.035 interaction 0.019 - - - 0.063 - - - South South endowments 0.098 0.093 0.096 0.096 0.098 0.227 0.273 0.250 0.227 0.227 coefficients 0.203 0.203 0.201 0.198 0.016 0.016 0.040 0.063 0.063 interaction -0.005 - - - 0.046 - - - Center West endowments 0.280 0.299 0.290 0.280 0.168 0.284 0.226 0.168 coefficients 0.170 0.170 0.179 0.189 -0.043 -0.043 0.015 0.073 interaction 0.019 - - - 0.116 - - - (b) Percentage of difference in log welfare ratio explained by endowments and coefficients Metro-Urban Metro-Urban Urban-Rural Urban-Rural threefold threefold twofold twofold threefold twofold twofold North D=0 D=0.5 D=1 D=0 D=0.5 D=1 endowments 61% 59% 60% 61% 93% 106% 99% 93% coefficients 41% 41% 40% 39% -6% -6% 1% 7% interaction -2% - - - 13% - - - Northeast Northeast endowments 57% 60% 59% 59% 57% 57% 89% 106% 97% 97% 89% 89% coefficients 40% 40% 41% 41% 43% 43% -6% -6% 3% 11% interaction 2% - - - 17% - - - Southeast endowments 35% 41% 38% 35% 84% 113% 98% 84% coefficients 59% 59% 62% 65% -13% -13% 2% 16% interaction 6% - - - 29% - - - South South endowments 33% 31% 32% 33% 33% 78% 94% 86% 78% coefficients 69% 69% 68% 68% 67% 67% 6% 6% 14% 22% interaction -2% - - - 16% - - - Center West endowments 60% 64% 62% 60% 70% 118% 94% 70% coefficients 36% 36% 38% 40% -18% -18% 6% 30% interaction 4% - - - 48% - - - 30 30 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table C2: Oaxaca-Blinder Decompositions Between Northeast and Southeast (a) Difference in log welfare ratios explained by differences in endowments and coefficients Southeast - Northeast threefold twofold Metro D=0 D=0.5 D=1 endowments 0.185 0.188 0.187 0.185 coefficients 0.233 0.233 0.234 0.235 interaction 0.003 - - - Urban Urban endowments 0.307 0.379 0.343 0.343 0.307 coefficients 0.216 0.216 0.252 0.252 0.288 0.288 interaction interaction 0.072 - - - Rural endowments 0.239 0.288 0.263 0.239 coefficients 0.202 0.202 0.227 0.251 interaction 0.049 - - - (b) Percentage of difference in log welfare ratio explained byy endowments and coefficients p Southeast - Northeast threefold twofold Metro D=0 D=0.5 D=1 endowments 44% 45% 44% 44% coefficients 55% 55% 56% 56% interaction 1% - - - Urban Urban endowments 52% 64% 58% 58% 52% 52% coefficients 36% 36% 42% 42% 48% 48% interaction interaction 12% 12% - - - Rural endowments 49% 59% 54% 49% coefficients 41% 41% 46% 51% interaction 10% - - - 31 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 31 Appendix D Figure D1: Metro-Urban and Urban-Rural Differences in Log Welfare Ratios by Region 32 32 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure D2: Differences in Log Welfare Ratios Between Southeast and Northeast Regions As mentioned in the methodology section, alternative counterfactual distributions can be constructed for the quantile decompositions. The results using these alternative counterfactual distributions are presented in Figure D3 below. Figure D3: Quantile Decompositions Using Alternative Counterfactual Distributions Within Regions (a) Between metro and urban areas of SE (b) Between urban and rural areas of SE (c) Between metro and urban areas of NE (d) Between urban and rural areas of NE 33 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 33 Between Regions (e) Between SE and NE metro areas (f) Between SE and NE urban areas (g) Between SE and NE rural areas 34 34 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES SOURCES OF WELFARE DISPARITIES ACROSS REGIONS IN MEXICO Hector Valdes Conroy Abstract W elfare differences between two areas are ultimately due to two factors: the characteristics of the people living in those areas and the returns they can obtain from those characteristics in the places where they live, which in turn are shaped by the specific characteristics of those places. Following a methodology similar to the Oaxaca-Blinder decomposition, this paper analyzes welfare differences across different geographical areas in Mexico. The characteristics of the population explain the great majority of the difference in livelihoods between urban and rural areas. When comparing two regions such as the North and South, then returns account for most of the difference. These results hold throughout the welfare distribution, which is analyzed using a quantile decomposition methodology. 35 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 35 1. Introduction Mexico is one of Latin America's most developed countries. In 2006, its PPP-adjusted per capita GDP was USD $11,531, the third highest in the region after Argentina and Chile.11 However, a large fraction of the population still lives in poverty. Mexico's National Social Policy Evaluation Council (Consejo Nacional de Evaluación de la Política de Desarrollo Social--CONEVAL) estimates that, in 2005, 47 percent of the population lived in poverty and 18.2 percent in extreme poverty. Inequality is also very high, with an expenditure-based Gini index of 46.1.12 Over the last ten years, the country has made significant efforts at reducing poverty--mostly through direct conditional transfers to the poor--and as a result poverty rates have gone down substantially even though the economy has grown only at an average 2.4 percent per annum.13 In spite of these improvements, welfare is not evenly spread across the whole territory. Rural areas are significantly poorer than urban ones, and the Southeast of the country has substantially lower standards of living than the North and Mexico City. This may be due to the fact that Mexico's North can benefit from its proximity to the U.S. whereas the South and Southeast are not as well-linked to the international economy. This problem of market accessibility in the South and Southeastern regions of the country may be further accentuated by insufficient basic infrastructure, in turn caused partially by institutions (e.g. land tenure problems limiting economic growth in the South) and geography (the South is highly mountainous whereas the North is much flatter). Then again, spatial differences in welfare could be driven by the characteristics of the inhabitants of different regions. People with higher endowments of human capital typically enjoy higher levels of welfare and so if they concentrate in specific areas--such as cities or cities of a specific region--they will produce spatial disparities that would not otherwise occur. Which of these two factors is mostly responsible for the observed differences in welfare across the Mexican territory? Is human capital evenly spread across the territory, but people receive different returns to their skills due to the different characteristics of the places in which they live? Or do people with similar levels of human capital tend to concentrate in the same areas? These questions have enormous implications for the design of policies aimed at eliminating poverty. The Mexican government has invested heavily in the human capital of the poor through the OPORTUNIDADES program, which is widely recognized to have helped bring down poverty rates. However, it is not yet clear that this investment will indeed reduce poverty in the long run, once cash transfers to poor families stop. If welfare is mostly determined by the characteristics of the area--which would go against the premise on which OPORTUNIDADES is founded--then investments in the characteristics of the locations through programs such as MICRORREGIONES would be more effective to eliminate poverty. The objective of this paper is to provide a quantitative assessment of the relative contribution of the two factors mentioned above to the differences in welfare observed in Mexico across urban and rural areas as well as across five sub-national regions. To do this, the paper will closely follow the methodology employed by Ravallion and Wodon (1999), which uses the Oaxaca-Blinder method to decompose urban- rural differences in welfare in two factors: the part due to the endowment of characteristics explaining welfare, and the part due to the returns to those characteristics. We do this analysis first at the national level and then at the regional level. Regions in Mexico are so different from each other in terms of welfare and other attributes that welfare disparities between urban and rural areas are likely due to different factors in different regions. 11World Development Indicators. 12World Development Indicators. The calculation is based on percentiles of the expenditure distribution. A calculation based on income is likely to be higher because expenditure is only a fraction of income at the upper end of the distribution. 13Real GDP per capita growth between 1995 and 2004. 36 36 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES The relative importance of the factors driving welfare differences across territories is also likely to change along the welfare distribution, a situation that is overlooked if one concentrates on the means only. Hence we will also look at the relative importance of characteristics and returns to characteristics in explaining urban-rural differences in welfare at various quantiles of the expenditure distribution, using the methodology proposed by Nguyen, et al. (2007). The following section presents some background on the economy of Mexico as a whole and of the different regions that make the country, as well as a brief explanation of the potential causes of the current spatial disparities. Section 3 of the paper presents the two methodologies used and discusses the general characteristics of the data used in the analysis. Section 4 presents the results in detail so as to provide a rich portrait of the factors affecting current welfare differences across various regions of Mexico and along several points of the expenditure distribution. Section 5 closes the paper with a summary of the findings. 2. Background Mexico is one of Latin America's most developed countries. With an estimated GDP per capita of almost $12,00014 in 2007, it ranks third in the region, behind Chile and Argentina. However, the country's level of development is far from being evenly distributed. Latin America is the world's most unequal region, and although Mexico ranks relatively low among Latin American countries, still its Gini index for 2004 was estimated at a 46.05. Just as interpersonal inequality is high, spatial inequality is also high. While people living in some neighborhoods of Mexico's largest cities enjoy extremely high levels of welfare, people living in small towns in the southern sierras are subject to abject poverty. These stark differences in welfare can be observed at the regional level as well: in 2005, Mexico City had a poverty rate of 31.8 percent while the corresponding figure for the three southern states of Guerrero, Oaxaca, and Chiapas was well over twice that level, at 71.6 percent. Geography is one potential explanation for such marked welfare difference across Mexico's territory. The most impoverished parts of the country are usually located in the mountains, in remote areas that are hard to access and therefore far from markets and centers of economic activity. A second and perhaps stronger reason has to do with institutions and historical developments. Mexico has a large number of different indigenous peoples, most of them concentrated in the center and southern parts of the territory. During colonial times, the Spanish dominated these ethnic groups and established an economic system based on slavery and a society with profound class divisions. This highly unequal society carried over to the country's first century of independence in the nineteenth century, as the political class was dominated by a small, wealthy elite. Just as during the Colony, the economy was centered around haciendas, which concentrated huge portions of land among a small minority of landholders. After the 1910-1920 revolution, a land reform was implemented to distribute land among peasants. The land, however, was not given as private property but as a communal asset called ejido. These consisted of small plots assigned to each household head in the community, and they could not be sold, rented or left idle for longer than the time necessary for fallowing. This land-tenure system created strong disincentives for investment, severely limited economies of scale in agricultural production, and often tied people to their lands. 14PPP-adjusted 2005 US dollars. Source: World Bank, World Development Indicators. 37 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 37 Ejidos were an especially important institution in southern Mexico, where large numbers of landless peasants concentrated during the war of revolution with "land and liberty" as their motto. Furthermore, the pre-Hispanic system of communal ownership of land adopted by the ejidos had an especially strong cultural appeal in the ethnically-diverse south. In 1991 a reform was enacted to allow ejidos to become private property; however a large number of them still operate in the south of the country. The above reasons may help explain the marked differences in welfare between the North and South of Mexico. However, welfare differences can be found throughout the territory. To explore this further, this paper adopts the National Population Council's (Consejo Nacional de Población--CONAPO) division of the territory in five regions (Figure 1). Figure 1. CONAPO's Regional Division of Mexico Reviewing basic regional statistics, it is evident that the regions are not balanced in terms of population (Table 1). Region 2 (Center) contains almost half of the population while regions 3, 4, and 5 contain roughly 10 percent of the population each. Region 4 is much more rural than urban, followed to a lesser degree by region 5, whereas Mexico City is purely urban. Having a region which represents a very small proportion of the total national population can result in little statistical power; however, our sample is sufficiently large as to circumvent this problem. Nonetheless, it is clear that we cannot include Mexico City in the analysis of rural areas. People in Mexico City have an income per capita more than five times that of residents in the Southern Pacific states but a poverty rate that is slightly less than half of the poverty rate in those same states, suggesting that income distribution is much more unequal in Mexico City (Table 1). It is interesting to note that even though region 2 is much more urbanized than region 5 and that rural regions are typically poorer than urban ones, both regions have similar levels of income per capita. 38 38 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table 1. Summary Statistics of Mexico's Regions, 2005 Sources: Regional division: CONAPO (taken from World Bank, 2004, p. 98); Population and GDP: Informe de Gobierno, 2007; Poverty: CONEVAL (www.coneval.gob.mx). 1. Figures are expressed in 1993 pesos and correspond to the year 2004. Population was linearly interpolated between 2000 and 2005. 2. Annual rate of growth for the period 2000-2004. 3. Based on income. There are also marked differences in terms of economic activity (Table 3). The South Pacific region--the poorest--has the highest proportion of its GDP coming from primary activities. The three states conforming this region are located along the Sierra Madre del Sur mountain range and contain very few areas where large-scale agriculture can be carried out. Moreover, as explained above, most agricultural production in these states takes place in ejidos, with the consequent negative effect on investment. Production is therefore highly inefficient and often destined for self-consumption. Mexico City's economic activity is heavily concentrated on the service sector and secondarily on manufacturing. Table 3. Mexico's GDP by Type of Economic Activity, 2005 Source: World Bank staff calculations with data from INEGI. 39 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 39 3. Methodology The goal of this paper is to make a quantitative evaluation of the different factors determining welfare levels in Mexico. We do this in two parts: First, we follow Ravallion and Wodon (1999) and look at the differences in mean welfare. Second, we look at differences along the whole distribution of welfare using the approach suggested by Nguyen, et al. (2007). 3.1 Differences in mean welfare To examine differences in mean welfare, we first estimate log welfare ratios15 ( ) as a linear function of household mobile characteristics--i.e. characteristics such as demographic composition and education of its members, that are not specific to the location of the household: .16 Urban and rural areas are estimated separately: (1) (2) The expected values of equations (1) and (2) are called "unconditional profiles" because they are based on an unconditional expectation. We look at the difference in unconditional profiles between urban and rural areas throughout the whole country. Such differential allows us to analyze the magnitude and sources ( ) of mean welfare differences between urban and rural areas.17 The difference is given by: (3) where and are the mean characteristics among urban and rural areas. We estimate this differential using sample means and OLS estimates of the parameters. Two opposing factors could explain welfare differences: characteristics of the people and the returns they obtain from them. To look at the importance of each of these two factors, we analyze them separately. First, we look at the difference in standards of living between urban and rural areas within each region caused by the difference in returns. Unlike in equation (3) where differentials were considered as a whole, the intention here is to isolate the effect of returns on welfare. To do this, we set household characteristics in both rural and urban areas at a common reference point--in this case, the national mean. These are called "geographic profiles": (4) for urban areas, and (5) for rural areas. 15Log welfare ratios are the natural logarithm of a household's per capita consumption normalized by the applicable poverty rate. 16The methodology differs slightly from Wodon and Ravallion's (1999) mainly in that they also include dummy variables for each province. In this paper, the analysis will be replicated for each region separately and so regional comparisons will be made below in more detail than an intercept shifter allows. 17Throughout the paper, we will use the word "area" to mean urban or rural zones. 40 40 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES The difference between equations (4) and (5), given by equation (6) below, is a useful analytical construct that isolates the differences in welfare caused by location-specific factors. This is because the characteristics contained in in equation (1) are all mobile characteristics and so the returns associated to them ( ) capture all the area-specific factors affecting welfare. (6) As before, this difference is estimated through sample analogs and OLS estimates of the parameters. Another interesting analysis consists of assessing the effect that different endowments of household characteristics have on welfare differences across regions and urban/rural environments. Even if returns to characteristics were the same across all regions of Mexico, welfare differentials could be caused by people with high (or low) human capital concentrating in certain regions. In order to look at how strong this effect is, we secondly simulate "concentration profiles" which, as opposed to the geographic profiles, allow household characteristics to vary by location but fix returns at a national level: (7) (8) As before, the comparison of concentration profiles reveals how much of the welfare differentials between urban and rural areas are due to the endowment of characteristics. Formally: (9) An additional step after looking at mean welfare ratios is to analyze how poverty rates would be affected under the same scenarios. We first calculate the "unconditional empirical" poverty rates for each area that can be obtained directly from the data, based on the observed distribution of log welfare ratios. It is also possible to approximate poverty rates from mean levels of consumption if we are willing to assume a specific distributional form. Here we assume that log consumption is normally distributed, which allows us to estimate poverty as the probability that the log welfare ratio is less than zero: (10) for urban areas, and (11) for rural areas; where is the average log welfare ratio in urban (rural) areas, and is the standard deviation. This produces "unconditional normal" poverty rates based on a normal distribution. Following the same procedure, we approximate the poverty rates that would be observed under the geographic and concentration scenarios described above. 3.2 Differences throughout the welfare distribution The methodology described above analyzes differences in welfare that occur only at the mean of the distribution. However, the determinants of welfare differentials between urban and rural areas as well as between regions are likely to vary depending on the point of the overall per capita expenditure distribution at which the comparison is made. To explore this possibility, we follow Nguyen, et al. 41 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 41 (2007) and use quantile regressions to estimate log-welfare levels at various points along the per capita expenditure distribution of urban and rural areas. To determine how much of the difference between the two distributions is due to returns and how much to characteristics, we construct a counterfactual distribution based on urban characteristics and rural returns. Formally, for every qth quantile (Qq) of the per capita expenditure distribution we estimate quantile regressions18 of the following type: (12) for urban areas, and (13) for rural areas. The estimated parameters and are used to replicate the urban and rural distributions, and respectively. A third distribution is constructed using rural characteristics and the estimated urban returns . This `intermediate' distribution is then used as counterfactual to decompose the difference between urban and rural welfare levels into the part that is due to returns-- "returns effect"--and the part that is due to characteristics--"covariate effect". The decomposition is explained by the following equation: (14) . The first term on the right-hand side of equation (14) is the covariates effect because the only difference between the urban and the counterfactual distributions is the covariates. Similarly, the second term is the returns effect because the only difference between the counterfactual and rural distributions is given by the bs. 3.3 Characteristics of the data The data used come from the 2006 National Household Income and Expenditure Survey (Encuesta Nacional de Ingreso y Gasto de los Hogares--ENIGH). This survey is conducted biannually by Mexico's National Statistics, Geography and Informatics Institute (Instituto Nacional de Estadística, Geografía e Informática--INEGI) with the objective of providing "information regarding the distribution, level and structure of households' income and consumption."19 With a sample of over 20,800 households, the 2006 wave is representative at the national level as well as of urban and rural areas separately. The 1996 wave has the same representativeness but with a smaller sample of about 13,000 households. The urban-rural classification is based on population. In 1996, an area was categorized as rural if its population was below 2,500 people; by 2006, this cutoff had changed to 15,000. Log-welfare ratios were calculated using household per capita gross total consumption aggregates and dividing them by the moderate poverty lines of the corresponding years. 18For details on quantile regression see Koenker and Bassett (1978, 1982). 19Translated from INEGI's website: http://www.inegi.gob.mx/est/contenidos/espanol/proyectos/metadatos/encuestas/enigh_211.asp?s=est&c=10748 42 42 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 4. Results 4.1 Urban ­ rural differentials in Mexico as a whole To formally analyze the determinants of welfare, we first construct welfare ratios--a household's per capita consumption divided by the poverty line--and regress them on a series of household characteristics.20 The first group of characteristics summarizes the demographic composition of the household. It contains the number of babies (up to two years of age), number of children (three to 11), number of teenagers (12 to 17), number of adults (18 to 59), number of seniors (age 60+), and the square of each of these variables. This group also contains the number of males in the household. A second group of variables contains demographic characteristics of the head of household: a dummy indicating whether the head is male, three dummies indicating his/her civil and cohabitation status,21 the age and age squared of the head, and the number of males in the household. A third set of regressors are educational variables. These are dummy variables indicating the years of schooling completed by the head of household and his/her spouse. Category 1 comprises one to three years of education; category 2, four to six years; category 3, seven to 10 years; and category 4, 11 or more years (the omitted category is 0 years of education). Another set of dummy variables categorizes the educational differential within the household, defined as the years of schooling of the most educated person in the household minus the years of schooling of the head or the spouse (whoever is more educated). The categorization is done in the same way as with the education of the head. The education differential captures the possibility that one member of the household--other than the head or spouse-- has a high level of education. The fourth group of variables is a set of dummies indicating the head of household's type of main economic activity. The first category includes arts and sciences professionals, public sector officials, directors, supervisors, and members of the armed forces; the second category includes technicians only; the third comprises coordinators and supervisors in administrative activities; the fourth incorporates all workers in the services sector; the fifth category includes all primary activity jobs; the sixth represents all production workers, including craftsmen; and the seventh category includes people who are retired and receive a pension or people whose main source of income is from financial interests or rents. The final group of variables would ideally include land ownership in order to control for both wealth and production possibilities in rural settings. However, the ENIGH does not present this information. A rather imperfect indicator is included which indicates whether some household member owns the house in which they live. This list of variables is intentionally long so that it incorporates as many of a household's mobile assets as possible. To be complete, the list would include other important aspects like ethnicity (which Ravallion and Wodon (1999) include) and unobservable characteristics such as preferences and skills. Unfortunately, the information contained in the ENIGH does not allow us to include any more variables than those listed above. Table 4 presents the mean values of all the variables included in the analysis, for urban and rural areas separately. The last column on the right presents the p-value of a difference-in-means test. It shows that rural and urban areas are significantly different in all but one of the variables considered. The first thing 20The specification of the regression follows Ravallion and Wodon (1999) as closely as the data allowed. 21The three dummies correspond to the following situations: Head of household is married but spouse is not present; head of household is single; head of household is divorced, separated or widowed. The omitted category is the case where the spouse is present 43 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 43 to notice is that welfare ratios are, on average, substantially lower in rural areas. Explaining this differential is precisely the objective of this paper. Rural households tend to be larger than urban ones but have fewer adults, which implies a higher dependency ratio among the former. Also, heads of household and their spouses are less educated in rural areas, but the high frequency on the fourth category of the education differential suggests that rural households have other members--presumably young adults--with high levels of education. These two observations coincide with the observations made by the migration literature which has found that young adults with high levels of education tend to migrate from rural to urban areas, where the returns to their skills are higher.22 We explore this hypothesis below. Another readily observable difference between rural and urban households which could account for their welfare differences is the economic activity to which they are mainly dedicated. Heads of urban households work mostly in the service sector (group 4) or as production workers (group 6), and a sizeable proportion (17 percent) belong to a typically high-paying sector (professionals, directors, public officials, supervisors). In rural areas, by contrast, an overwhelming majority of heads of household are dedicated to primary activities (group 5), and then either as production workers or in the service sector. 22This situation need not be negative because migrants often contribute to their original household's welfare through remittances. In fact, investing in the education of future migrants may be a long-term, income-maximizing strategy for rural households. 44 44 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES These differences, however, do not provide a satisfactory explanation of the observed welfare differences between urban and rural households in Mexico because we do not know the returns to those characteristics. Table 5 presents the results of estimating equations (1) and (2) using the variables described above as regressors. In these regressions, the reference household has a female head with her spouse living in the household; she is not working and does not have any source of income; she and her spouse have zero years of education and they do not own the house in which they live. In line with the observations made from Table 4, the first thing to notice in Table 5 is that the presence of any additional person in the household reduces welfare.23 The returns to all demographic composition variables look very similar across urban and rural households. However a test for joint difference (Table 6) shows that, taken together, they are significantly different between urban and rural areas. Heads of household in urban areas receive higher returns to their education than they would in rural areas. Together with the observation made above that heads of household in urban areas are more educated than in rural ones, the difference in returns to education helps explain why urban areas have higher levels of welfare. 23If we take into account the quadratic term and contrast it to the support of the corresponding distribution we observe that, for every variable, at least 95 percent of the distribution resides under the part where the welfare returns to the number of people in the household are negative. 45 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 45 46 46 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 47 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 47 Nonetheless, education of the spouse and education differentials are better rewarded in rural areas, which makes the total effect of education ambiguous. Economic occupations have relatively similar returns, although they are jointly significantly different across rural and urban areas (Table 6). Nevertheless, while heads of urban households are mostly dedicated to high-paying activities (group 1, especially), a great majority of rural heads are dedicated to primary activities, which is associated with lower levels of welfare than those of the reference household where the head does not receive any income. To get a more accurate account of whether these factors are indeed responsible for the observed welfare differentials, average welfare ratios are broken down by each of the variable groups (Table 7).24 Education is indeed an important factor determining the difference in observed welfare levels, accounting for two thirds of the net differential. Household composition factors and main occupation of the head account each for approximately a third of the difference in mean welfare ratios between urban and rural households. What is most striking, however, is the large counteracting effect of the variables describing the head of household. Indeed, all factors analyzed play against rural households' welfare except for the characteristics of the head of household. This last result is persistent in subsequent variations of the analysis. Looking within the group of characteristics of the head of household, the age of the head is the specific variable driving this result. It is not clear why this variable would have such higher returns in rural areas. One possibility is that the estimated coefficient is capturing the effect of another variable that is correlated both with expenditures and with the age of the head of household, and that this variable is especially important in rural areas. One such variable could be remittances. Households with older heads could have migrant members who send remittances and hence raise the household's welfare. In rural areas, such remittances could imply 24The effect of each group of variables is formed by the sample average of those variables within the corresponding area (urban or rural) multiplied by their estimated returns. Since regressions included a constant, the estimated average welfare ratios coincide with the sample averages of actual welfares. 48 48 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES larger increases in expenditures in rural than in urban areas. To explore this possibility, a separate analysis (results not shown) also controlled for whether the household currently receives transfers, but the results do not change, which rules out the possibility that transfers are driving the results found for the age of the head of household. Another variable we do not control for and that could have such effect is land ownership. Ravallion and Wodon (1999) control for this important variable; unfortunately, the ENIGH does not include this information. In order to determine how much of the urban-rural welfare differential is due to household characteristics and how much is due to returns that households receive for those characteristics, we simulate two alternative welfare profiles for urban and rural areas. In the first simulation ("geographic profile"), urban and rural mean log-welfare ratios are calculated using their specific returns but using national mean endowments (as in equations 5 and 6 above). In the second simulation ("concentration profile"), urban and rural mean log-welfare ratios are calculated using their specific endowments but using national returns (equations 8 and 9 above). The concentration profile is much closer to the actual mean log-welfare ratio than the geographic profile (Table 9). This means that a big part of the actual log-welfare ratios is explained by endowments, not returns. This is because when we move from the unconditional to the concentration profile we are changing the returns and leaving endowments untouched, which results in a small change. In contrast, the change is large when we move from the unconditional to the geographic profile, which leaves returns unaltered but changes endowments. A second important result is that rural areas would be wealthier than urban areas if they had the same characteristics as urban areas. Under this scenario, urban and rural households would on average have the same characteristics, returns being the only difference between one area and the other. This implies that returns in rural areas are larger than in urban ones. At first, this might sound contradictory as one would expect market returns to human capital to be larger in urban areas. However, this depends on what returns we are talking about. If it is the returns to education or other characteristics that are important for labor markets, then indeed one would expect such returns to be higher in urban areas. Just as we did for the unconditional profile in Table 7 above, By far the largest difference in returns in the geographic simulations occurs with the head of household characteristics, specifically the age (Table 10). 49 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 49 A central motivation of this paper is exploring the factors behind not just mean levels of welfare, but also poverty. As with log-welfare ratios, one would like to know how poverty rates change if we modify returns and characteristic endowments. This analysis is complicated because calculating poverty rates requires more knowledge of the expenditure distribution than just the mean; however, we can simulate poverty rates under the geographic and concentration profiles if we assume a specific distributional form. We use the normal distribution to produce the poverty rates (Table 11). The results are qualitatively very similar to those presented in Table 9 for log-welfare ratios. However, in this case the simulations depart from the actual poverty figures in two ways: first in the changes in either returns or characteristic endowments, and second in the distributional assumption. For this reason, the second line of the table recalculates poverty rates for the unconditional profile assuming a normal distribution (instead of the empirical distribution). As we can see by comparing the first two lines, this functional choice yields relatively good approximations for rural areas but rather bad ones for urban areas. But once we have accounted for the distributional effect, we can compare the poverty rates in the geographic and concentration profiles with our unconditional (normal) poverty measures. 4.2 Urban ­ rural differences within Mexico's regions The results presented thus far reveal that characteristics, more than returns to those characteristics, explain the difference in mean welfare levels between urban and rural areas. The leap from the unconditional to the geographic profile is not only big, but also makes rural areas better off than urban ones, indicating that returns are actually in favor of the poorer areas. However, these results do not give a very complete picture. Mexico is a highly diverse country with substantial interregional variation. The results thus far correspond to means for the whole country and so are likely to miss significant regional variation. To overcome this deficiency, we repeat the analysis for each of the country's five regions. The analytical gain of doing this, however, comes with a cost: the ENIGH is representative only of urban and rural areas at the national level. Therefore, the following results must be interpreted with care. Table 50 50 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 12 shows the sample sizes for each region. Region 3, Mexico City, is overwhelmingly urban and so we cannot draw reliable statistical inferences for its rural population. Appendix Tables A1 through A5 replicate the analysis presented in Table 5 for each of the five regions separately. As in Table 6, Table A6 presents tests of joint difference between the different groups of independent variables across rural and urban settings in each region. Although the findings are generally in line with the results from the national regression, a few results show important regional departures and are hence worth highlighting. In region 1, the richest after Mexico City, higher education differentials have a larger positive impact on urban welfare levels than in the national average. However, the positive impact observed among rural areas in the national regression is not present in this region. This suggests that labor markets in urban areas of the North assign a high value to education but that the same is not the case in rural areas. Another important difference between this region and the national average is that rural households whose head is mainly dedicated to primary activities are not significantly worse off than other rural households. Although more evidence is necessary to make a formal claim, this finding is suggestive of a more developed rural economy in which primary activities are well remunerated. In Central Mexico, a striking difference with respect to the rest of the country is that education differentials in urban areas are not associated to higher levels of welfare. In contrast with the North, labor markets in Central cities do not seem to attach a high value to the education of workers. This is consistent with the additional finding that the positive association observed in rural areas at the national level between households whose heads are mainly dedicated to administrative activities or services and welfare is not present in this region. Results for Mexico City have to be interpreted very carefully as the sample size is the smallest relative to the actual population. Not surprisingly, most of the differences with respect to the national regression are only in terms of the statistical significance. One important difference, however, is that in Mexico City, households whose head is mainly dedicated to primary activities have much lower levels of welfare than the rest of the households; in Mexico as a whole, these urban households are not significantly poorer than the rest of urban households. The two southern regions show perhaps the most notable differences relative to the country as a whole. In the South Pacific states of Guerrero, Oaxaca and Chiapas, the education of the heads of household is associated to higher increases in welfare relative to households where the head is uneducated, both in urban and rural settings. However, the education of other household members is not associated with higher levels of welfare. Together, these findings speak of an economy in which the heads of household have a much more prominent role in the economy of the household than in other parts of the country. However, this assertion needs to be moderated because these households could be benefitting from remittances sent by family members who reside elsewhere. Indeed, migration from this region to other 51 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 51 parts of the country and to the U.S. is highly prevalent, and so it could be that migrants' economic contribution is significant even though they are not part of the household for the purposes of the survey. A final difference between the South Pacific region and the country as a whole is that rural households whose head is mostly dedicated to primary activities are substantially worse off than similar households with a head who is unemployed or out of the labor force and without financial income (the omitted category). The difference in welfare levels is substantially larger than for the country as a whole. In the Southern Gulf and Caribbean region, the education of the heads of households and their spouses is associated to higher levels of welfare (relative to households whose head and spouse are not educated) than in the country as a whole. This phenomenon is present both in urban and rural areas, and seems to reflect a more developed labor market in both urban and rural settings. However, this would be a surprising conclusion to draw since this region has the second highest poverty rate and the second lowest mean per capita income. In all five regions, the main positive contributor to mean log-welfare ratios in urban areas is the group of educational variables (Table A7). In rural areas, the same group of variables is the main contributor in the North and Southern Gulf and Caribbean regions. However in the Center and South Pacific regions, the characteristics of the head of the household are most important. In the four regions where the comparison can be made, urban areas are substantially better off than rural ones. The decomposition of the differences in all four regions shows that urban households have an advantage over rural ones in every group of variables, except the characteristics of the head of household (only the Southern Gulf and Caribbean region is the exception). Despite the regularities just mentioned, the decompositions show substantial heterogeneity across regions. While in the North all groups of variables have a similar contribution to the differential (except, once again, for the characteristics of the head of household which counteract the contribution), in the South Pacific education and occupation variables explain most of the differential. In the Center, education and the demographic composition of the households matter most, while in the Southern Gulf and Caribbean it is mostly the demographic composition. To analyze whether the differences between urban and rural areas are mostly due to differences in returns or endowed characteristics, we simulate mean log-welfare ratios under the geographic and concentration profiles for each region and urban-rural setting separately (Figures 2a and 2b). Mexico City is excluded because, since it is purely urban, the decomposition cannot be carried out. For reference, its (urban) mean log-welfare ratio is 0.55. 52 52 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES As can be seen, concentration profiles are much closer to the unconditional profiles than the geographic ones, both in urban and rural areas. This suggests that most of the urban-rural differentials in welfare in all four regions are due to differences in the characteristics of their households. The result can be seen with the naked eye: For any given region, the difference between the urban and rural geographic profiles reflects differences in returns. This difference is relatively small in each of the four regions. On the other hand, the difference between the urban and rural concentration profiles reflects differences in household characteristics. In the four regions analyzed, these differences are very large. When we compared urban and rural areas for the whole country, under the geographic profile urban areas would be poorer than rural areas, whereas in reality the opposite is true. Conversely, the urban-rural differential increased under the concentration profile. This finding is also present in all four regions for which we carry out the decomposition and it implies that returns in rural areas are higher than in urban areas. But are all returns higher in rural areas, and if not, which ones are driving this differential? Itemizing the geographic profiles and their urban-rural difference for all regions shows that in the Center and South Pacific, household head characteristics drive the bulk of the difference in favor of rural areas, implying that the returns to this group of variables are much higher in rural than in urban areas (Appendix Table A8). In the North, the returns to all groups of variables (except house ownership) are higher in rural areas; in the Southern Gulf and Caribbean, returns to education and to occupational groups are higher in rural areas. Repeating the exercise for concentration profiles, two results are notable (Appendix Table A9). First, in every category of characteristics except house ownership, urban households are better endowed than rural ones. Second, in all four regions for which we can draw a comparison as well as in Mexico as a whole, education characteristics are responsible for the majority of the urban-rural differential. In the country as a whole, and in the Southern Gulf and Caribbean region, education characteristics account for more than 50 percent of the differential; in the North and Center regions, they account for approximately two thirds. After education, the occupation of the household head and household demographic composition are important in explaining the differential. These findings, coupled with the observation that the concentration profiles are very close to the unconditional profiles, suggest that education--followed by the occupation of the head of household and the demographic composition--is the major characteristic driving welfare differentials between urban and rural areas. 53 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 53 As before, we can also analyze what part of the observed differences in poverty rates across urban and rural areas is due to differences in returns and what part is due to differences in endowments. Taking the normal distribution as a basis, we generate the geographic and concentration profile poverty rates (Figures 3a and 3b). As before, we present the actual poverty rates as well as the poverty rates we would obtain by forcing a normal distribution to the data so as to draw a cleaner comparison with the geographic and concentration profiles. Mexico City is excluded because the geographic and concentration profiles coincide with the unconditional poverty rate at 25.9 percent. The analysis based on poverty rates shows qualitatively the same results as the one based on mean log- welfare ratios: Most of the urban-rural differentials are driven by differences in the characteristics of the households. Aggregate returns to all characteristics are higher in rural than in urban areas. When one looks at a more disaggregated picture, o7nly some groups of variables have higher returns in rural areas, and these groups vary by region. 54 54 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 4.3 Differences between leading and lagging regions The analysis presented thus far has revealed very interesting results that could speak of a labor mobility story. The fact that returns between urban and rural areas are similar and that welfare differences between areas are mostly due to the characteristics of the people is an indication that people could be moving to areas where their skills are best rewarded. Such labor mobility is more likely to occur among areas that are relatively close, which is consistent with our observations within regions. Across distant regions, however, the situation may be different. Considering the stark differences in livelihoods between the North and the South, it is well-worth exploring the factors explaining welfare differences between similar areas of different regions within the country, especially between leading and lagging areas. To do so, we concentrate on the comparison between urban and rural areas in the North (region 1) and the South (region 4). For urban areas only, we also compare Mexico City (region 3) with the South. The analysis is carried out in the exact same fashion as before, except that this time geographic profiles are estimated using mean characteristics of urban or rural areas from the two regions being compared, while concentration profiles use the returns obtained from a regression pooling the urban or rural areas of the two regions being compared.25 The results tell a different story from the ones analyzed previously (Table 13).26 First, welfare differences between the South and the other regions are extremely large. Second, they are due to different factors in each case. 25See the methodology section above. The equations used in this part of the analysis are the same if instead of using the sub-indices "U" and "R", one uses "R1" and "R4" (meaning regions 1 and 4). The equations presented in that section apply for the whole population of Mexico; in this case they would apply only to the urban or rural population of the two regions involved. 26As before, the ENIGH is representative of urban and rural areas at the national, not regional level. Therefore, the results of this part of the analysis have to be interpreted with care. 55 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 55 Note: The first three columns compare the North and South regions while the last three compare Mexico City with the South region. The top panel makes the comparison for urban areas and the bottom panel for rural ones. The comparison between urban areas in the North and South regions reveals that welfare differences are mostly due to returns, rather than characteristics, although both factors play a role. In other words, the North has better endowments than the South (the concentration profile is higher in the former region), but this difference is approximately a third of the difference between the geographic profiles. This result suggests that labor mobility between urban citizens of the South and North is low. The comparison between Mexico City and urban areas in the South tells a similar story, although this time both factors are equally strong. The magnitude of both differentials is comparable to the geographic differential between the North and South. That is, Mexico City has returns that are similar to those of the North, but the characteristics of its inhabitants result in higher welfare levels than those in urban areas of the North, and much higher than those in the urban South. The contrast of rural areas in the North and South reveals that, again, returns play a larger role than characteristics. However, this time characteristics explain welfare differentials that are about two-thirds of the welfare differentials explained by returns (0.189 versus 0.273). Notice, however, that this relatively larger weight is due only to a larger differential in characteristics and not because returns are closer together between the two regions. Welfare differences between rural areas in the North and South are higher than between urban areas only because rural areas in the South have even larger deficiencies in endowments than urban areas, relative to their northern counterparts. 56 56 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES The same analysis based on poverty rates instead of log welfare ratios draws qualitatively the same conclusions, but comparisons drawn against unconditional rates have to be based on the normally- distributed unconditional poverty rates (Table 13). 4.4 Quantile analysis: Welfare differences throughout the expenditure distribution We now analyze the relative contribution of returns and characteristics to the urban-rural differential in welfare levels along the consumption distribution. The analysis presented thus far concentrates solely on the mean of the distributions. However, what happens at the mean of the consumption distribution is not necessarily the same as what happens at the extremes. As an illustration of this, consider the distribution of the urban-rural differential in log-welfare ratios (Figure 4). The U pattern reveals that welfare differentials are larger towards the tails of the distribution and always in favor of urban areas. This suggests that the findings made at the mean do not apply throughout the distribution. Several points are noteworthy in the results of quantile decomposition for the whole country (Figure 5). First, the returns effect in rural areas is higher than in urban areas throughout most of the consumption distribution, which is consistent with our previous findings. However, this is not the case at the bottom decile of the distribution, where urban returns are higher than rural ones.27 Also, the covariates (characteristics) effect is larger in magnitude than the returns effect (measured by the vertical distance of either line from the 0 line). Once again, this confirms our previous finding that urban-rural differentials are mostly due to differences in household characteristics. Lastly, the complete distributions show that both effects increase in magnitude moving up the per capita expenditure distribution, and at comparable rates. 27This observation must be taken carefully because income and consumption measures are usually noisier at the ends of the distribution. 57 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 57 Considering that education characteristics drive most of the urban-rural differential in the concentration profile (from Table A9), the upward slope of the covariates effect curve in Figure 5 suggests that educational differentials increase with per capita expenditure. That is, the difference in years of schooling between urban and rural households is larger among the wealthier households from both settings than among the poorer households. This is not a surprising finding, since in developing countries such as Mexico high levels of education are usually not well rewarded in rural areas. As a result, people with high levels of education tend to concentrate in the cities and they usually belong to the upper deciles of the income distribution. In contrast, at the bottom deciles of the distribution--in both cities and rural areas--the levels of education are not so dissimilar. Although the shape of the covariates effect curve can be easily explained, the shape of the returns effect curve is somewhat puzzling. As one moves to the upper end of the income distribution, returns in rural areas become increasingly larger than in urban areas. We now repeat the previous quantile analysis for each region separately to discern important regional differences (Figures 6a through 6d), and find substantial heterogeneity across regions. Region 5-- Southern Gulf and Caribbean--resembles the patterns observed at the national level. However the South Pacific (region 4) shows a very different pattern. First, we do not observe the increasing trend in either effect; both curves remain relatively flat throughout the per capita expenditure distribution. Second, the magnitude of the covariates effect is substantially lower than the national aggregate--and the same is true of the returns effect over the upper half of the expenditure distribution. The North (region 1) and Center (region 2) also have flatter returns and covariates effect curves than at the national level, however their magnitudes are closer to the national-level curves. 58 58 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES This quantile analysis can also be used to decompose cross-regional welfare differences. Table 13 above showed that welfare differences between Mexico's North and South are mostly due to returns, rather than characteristics. This is especially true among urban areas but holds also among rural areas. Figure 7a presents the log-welfare ratio differentials between urban areas in the North and South of Mexico along the expenditure distribution, and Figure 7b shows the quantile decomposition of those differences. Ignoring the tails of the distribution, where expenditures are usually measured with a large degree of error, the gap in welfare remains, on average, stable. Consistent with the results presented in Table 13, both the covariates and return effects are positive, implying that the North has better endowments and higher returns. Also, the returns effect is generally larger than the covariates effect. Notice that the latter effect seems to decline steadily between the second and sixth deciles before leveling off; the returns effect, on the other hand, starts roughly at the same level as the covariate effect and rises until the sixth decile where it then levels off. In other words, characteristics explain relatively little of the welfare differential among the upper segments of the expenditure distribution. 59 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 59 Figures 8a and 8b repeat this analysis for rural areas in the North and South. Just as in the case of urban areas, welfare differentials between rural areas in the North and the South remain relatively constant along the expenditure distribution. However, the decomposition unveils a very interesting pattern: a constant tendency for the covariates effect to decrease and for the returns effect to increase throughout the distribution. As with urban areas, characteristics are ever less important in driving rural welfare differentials as one moves along the expenditure distribution. However, an important distinction is that in rural areas, characteristics are more important than returns at the lower end of the distribution and are equally important as returns at the middle of the distribution (recall that in urban areas both effects were equally important at the lower end). Figures 9a and 9b present the same analysis for Mexico City and urban areas in the South region. In this case, the most salient result is that welfare differences among the upper quarter of the distribution become extremely high due to both factors. Mexico City is the center of economic activity in the country and so concentrates most of the country's human capital and employing firms. However, it is also a very unequal city, where a large part of the population has much lower levels of welfare than the upper deciles. 60 60 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 4.5 Evolution of welfare differentials between 1996 and 2006 To explore how welfare differentials in Mexico have evolved over time, we repeat the analysis done thus far using data for ten years earlier. In 1996, Mexico was in the process of recovering from the severe financial crisis that started in December 1994. Poverty rates were hence substantially higher for that period (Table 14).28 Annex Tables A10 to A15 present the results of the basic regressions (similar to Tables 5 and A1 through A5), and Table A16 presents the tests of joint equality of coefficients (comparable to Table A6). Annex Tables A17 to A19 present the contributing factors to the three welfare profiles that compare urban and rural areas within regions (comparable to Tables A7 through A9). Finally, Table A20 (compare with Table 13) presents the comparison between the North and the South regions as well as between Mexico City and the South. The comparison of unconditional profiles for urban and rural areas across the whole country reveals that urban-rural differentials were larger in 1996 than in 2006 (0.41 versus 0.33 points in log-welfare ratios). This differential was driven mostly by educational variables and by the demographic composition of households (see last panel of Table A17). In 2006, these factors also played an important role (demographic composition less so than in 1996), but they were completely outweighed by the effect of the age of the head of household, which in 1996 was much smaller. The last panels on Tables A18 and A19 show that most of these differentials can be attributed to better endowments of education and demographic composition of households in urban areas. 28All poverty lines are recalculated every year the ENIGH is conducted, based on different consumption bundles. Official poverty rates are calculated on the basis of income. 61 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 61 On a region-by-region basis, characteristics also seem to drive most of the welfare differentials. Concentration profiles are closer to the unconditional profiles, implying that homogenizing returns across areas does not have a major impact on welfare (Figures 10a and 10b). In contrast, making both urban and rural areas have the same endowments of characteristics would make the former substantially worse off and the latter much better off. To see exactly which characteristics are mostly responsible for the observed welfare differentials, we need to look at the different factors contributing to the concentration profiles (Table A19). As in 2006, education differentials make up the greatest part of the difference, followed by demographic composition factors. In 1996, however, the latter factors are relatively more important than ten years later. These similarities notwithstanding, an important difference emerges between both years. For 2006, the characteristics of the head of household--more specifically, his or her age--were the main factors driving the difference between geographic profiles in urban and rural areas. That is, although differences in returns across urban and rural areas were small (relative to differences in endowments), these differences were mostly due to different returns to characteristics of the head of household, and favored rural areas. In 1996 this result is not observed. First, although the magnitude of this difference is large in most regions, it is largest in the Center and South Pacific regions. Second, it does not always favor rural areas. In the North and South Pacific regions, the returns differential plays in favor of urban areas, while in the Center region it favors rural ones. Identifying the cause of these observations would require a separate study beyond the scope of this report. 62 62 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table A20 presents the results of the decomposition of welfare differences between the North and South regions as well as between Mexico City and the South in 1996. Contrary to our results for 2006 (Table 13), welfare differences in urban areas--North versus South and Mexico City versus South--are mostly due to characteristics. This suggests that differences in endowments between the North and South narrowed over that ten-year period. The same is true for the difference in welfare between the rural North and the rural South. In 1996, the welfare gap was more than 150 percent larger than the gap in 2006 and it was almost equally due to returns as to characteristics. By 2006, differences in returns had been reduced by 30 percent while differences in characteristics were almost halved. As before, we decompose welfare differences along the expenditure distribution through quantile analysis (Figures 11 and 12, analogous to Figures 4 and 5 for 2006). Once more, we observe a U shape in log welfare differentials between urban and rural areas throughout the country, but the curve is shifted upwards with respect to 2006. Hence, in terms of the aggregate urban-rural welfare differential, this ten- year period brought about a reduction that is roughly the same across all income groups. This uniformity is lost when one looks at the decomposition of the differential. In 1996, returns effects were equally large throughout the distribution; by 2006, they were significantly reduced for all the distribution except the highest decile in such a way that poorer households saw greater reductions in returns differences across urban and rural areas. 63 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 63 A quantile urban-rural decomposition analysis for all regions in 1996 (Mexico City excluded for being exclusively urban) shows some heterogeneity, with three salient patterns (Figures 12a through 12d). First, covariate effects are much larger than returns effects. Furthermore, the former are positive and the latter are negative, showing that in every region returns are higher in rural areas but that the better endowments of urban areas make them wealthier than rural areas. Second, covariate effects tend to follow a U shape, although in the North the right-hand tail does not climb up again and the Southern Gulf and Caribbean is rather flat. Finally, returns effects have a slightly increasing tendency (decreasing in absolute value, since they are negative), except in the Center. We conclude our analysis presenting a similar quantile analysis for the cross-regional comparison, in 1996: urban North versus urban South, Mexico City versus urban South, and rural North versus rural South (Figures 13a through 13c). The analysis reveals substantial differences between the urban comparisons (North-South and Mexico City-South) and the rural comparison (North-South), but also with respect to 2006. In 1996, welfare differences between leading urban areas and the urban South were almost equally made up of returns and endowments. Indeed, although the return effect curves are highly unstable along the expenditure distribution, on average they remain fairly close to the covariates effect curves. Furthermore, both effects are positive, confirming that leading urban areas have both better endowments and returns than lagging areas. Among rural areas, in contrast, returns were responsible for the larger part of welfare differentials between the North and South except among the poorer households, where characteristics were just as important. 64 64 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 65 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 65 5. Conclusions This paper has analyzed welfare differences between urban and rural areas of Mexico, within and across regions, and both at the mean and throughout the distribution of per capita expenditure. The comparison was made first for the country as a whole, and then by region. We have compared log-welfare levels at the mean and every percentile of the distribution. The analysis reveals that, within regions and for the country as a whole, urban-rural differentials are mostly due to different endowments of household characteristics, as opposed to returns on those characteristics. These characteristics include an extensive list of households' mobile assets, including human capital, demographic composition, and main occupation of the head of household. Among all these characteristics, education stands out as driving most of the urban-rural differential in living standards. This result holds both at the national level and in each of the four regions for which an urban- rural comparison could be made. Another result of the analysis is that, within regions, returns in rural areas are higher than in urban ones. This, however, is an aggregate result. When one looks into the different returns, it becomes evident that the returns to the age of the head of household drive most of this result. That is, returns associated to most characteristics are higher in urban areas, but the returns to the age of the head of household are higher in rural areas and their magnitude is such that the aggregation of all these returns turns out to be in favor of rural areas. Results are different when one compares leading and lagging regions. In this case, both returns and characteristics explain welfare differentials, although a greater weight can be assigned to returns. Together, these results suggest that labor mobility may be higher within as opposed to across regions. Standard economic models of migration predict that people will move to the areas that offer the highest returns to their characteristics and skills, after discounting the cost of migrating. Our observation of returns being roughly equal across the different areas of a region suggests that enough migration has taken place as to equalize these returns. However, across regions we do observe large differences in returns, indicating that potential migrants could expect significant gains from moving to those areas. The comparison between 1996 and 2006 reveals two main findings. First, welfare differences have reduced over the ten-year period. This is true for all regions and throughout the distribution of mean per capita expenditure. Second, returns were more important in explaining differences in livelihoods between urban areas of the leading region and urban areas of the lagging region in 2006 than in 1996. However, the importance of returns is relative to the importance of characteristics, so this change is due to an equalization in characteristics rather than to an opening in the returns gap. In sum, rural areas in Mexico are poorer than urban areas of the same region because their inhabitants have worse endowments: they are less educated, the demographic composition of their households is not optimal--i.e., they have higher dependency ratios--and their heads of household are mainly dedicated to low-paying economic activities. However, differences between leading and lagging regions--i.e. between the North and South regions--are due to differences in endowments, but even more importantly to differences in returns. In 1996, urban-rural differentials within regions of the country could be explained in the same way as in 2006. Welfare differentials across regions in 1996, however, were equally explained by characteristics and returns because differences in characteristics were even higher then. 66 66 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Appendix 67 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 67 68 68 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 69 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 69 70 70 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 71 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 71 72 72 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 73 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 73 74 74 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 75 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 75 76 76 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 77 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 77 78 78 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 79 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 79 80 80 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 81 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 81 82 82 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 83 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 83 84 84 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 85 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 85 86 86 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 87 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 87 88 88 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 89 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 89 90 90 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 91 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 91 92 92 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 93 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 93 94 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 95 96 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES POVERTY IN LATIN AMERICA: SHOULD POLICIES FOCUS ON POOR REGIONS AND/OR POOR PEOPLE? THE CASE OF ECUADOR Monica Tinajero Consultant The World Bank Tinajerobm@gmail.com Gladys López-Acevedo Senior Economist, LCSPP The World Bank gacevedo@worldbank.org Abstract T his paper contributes to the analysis of spatial poverty in Ecuador by deepening the understanding of the constraints faced by the poor in Ecuador through an investigation of the role of portable characteristics (e.g. human capital) and geography in explaining welfare. At a national level, the results indicate that characteristics explain 72 percent of the differences in welfare level between urban and rural areas, while returns to characteristics account for 28 percent of the difference. Comparing a leading and a lagging region such as the coast versus the Amazon, characteristics explain about 90 percent of the welfare differential in urban areas, while returns explain about 30 percent of the welfare differential in rural areas. Among the characteristics analyzed, education is the most important variable for explaining differences in living conditions between urban and rural areas in Ecuador. Background paper for the regional study on poverty in Latin America at the World Bank. These are views of the authors and do not necessarily reflect those of the World Bank, its executive directors, or the countries they represent. 97 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 97 Introduction Countries in Latin America continue to struggle to share the benefits of economic growth with the poor segments of their population. The progress towards achieving the first Millennium Development Goal of eradicating extreme poverty by 2015 is still far from being achieved in most Latin American countries. Another feature common in many countries in the region is geographic disparities in living standards that persist over time in spite of overall economic growth. A better understanding of which factors influence disparities in living standards is crucial to provide better guidance to poverty reduction strategies. Two opposing views present divergent explanations on which factors are likely to influence the spatial distribution of poverty. The "concentration view" holds that poor areas arise from the persistent concentration in these areas of individuals with personal attributes that inhibit growth in their living standards. According to this view, otherwise identical individuals will have the same growth prospects independently of where they live. Thus, geography does not play a causal role in explaining the level of and growth in living standards. The other view holds that "geography" itself is the cause of the high level of poverty and weak growth of living standards over time. In areas better endowed with local public goods, such as better access to infrastructure and other basic services (electricity, water and sanitation), geographic externalities may facilitate the exit of poor households from poverty. This paper provides an applied framework to analyze which of these two views better explains spatial disparities in poverty in Ecuador. Ecuador is located in northwestern South America, with an ethnically diverse population comprised of mestizos (62 percent), Amerindians (25 percent), immigrants from European countries as well as unmixed descendants of early Spanish colonist (10 percent) and Afro-Ecuadorians (3 percent). The population of over 13,750,000 people (in 2008) is growing at a rate of 1.1 percent per year,29 and one of every four Ecuadorians is between 15 and 29 years old. Nearly 60 percent of its population lives in urban areas and the rest in the rural sector. Ecuador's economy depends heavily on oil exports, which accounts for an average of 35 percent of GDP (1997-2001). In the past three decades, GDP per capita growth has been low, at an average of 1.7 percent annually during 1970-2002. GDP volatility has been high, mainly as a consequence of external shocks such as downturns in commodity prices, as well as unstable fiscal policy. At the end of 1999 Ecuador faced a severe economic crisis due to several factors: external economic shocks, "El Niño" weather phenomenon (1997) and poor economic management. The economy recovered after the 1999 crisis, with GDP per capita rising to 2.1, 6.5 and 4.5 percent in 2003, 2004 and 2005, respectively. Average inflation fell from an annual rate of 52.2 percent in 1999 to 7.9 percent in 2003, 2.7 percent in 2004 and 2.1 percent in 2005.30 Poverty and inequality remain a major concern for Ecuador. According to the World Bank's Poverty Assessment for Ecuador (2004), poverty increased from 40 percent in 1990 to 45 percent in 2001 (mainly as a consequence of the 1999 crisis), and fell to around 36 percent between 2001 and 2004 (mainly as a consequence of macroeconomic stabilization). While poverty declined in both urban and rural areas, in 2004 the rural poverty rate was still more than twice the urban poverty rate. With these differences in mind, the goal of this paper is to evaluate how much of the difference in standards for living across areas and regions in Ecuador can be attributed to disparities in the mobile non- geographic variables, and how much to location differences in the returns to those characteristics. The paper follows closely the methodological framework of Ravallion and Wodon (1999), in which they 29 World Development Indicators database, April 2008. 30Idem. 98 98 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES propose a methodology for decomposing the differentials in living standards between geographical areas into: a) differences in non-geographic characteristics and b) differences in returns to those characteristics. The paper is structured as follows. Section 1 presents a brief description of the data, and Section 2 presents the model. Section 3 discusses the national-level results from the model using regression analysis to test the relationship between different household characteristics and welfare. Section 4 carries out the decompositions at the national level to analyze the relative importance of characteristics and returns to characteristics on welfare. Section 5 discusses the extension of the model to allow for comparisons within and across regions. Section 6 concludes. 1. Data Source This analysis uses data from the Living Conditions Survey (Encuesta de Condiciones de Vida--ECV) for 2005-2006. The ECV is a micro-level dataset collected by Ecuador's National Statistics and Census Institute (Instituto Nacional de Estadística y Censos--INEC). The ECV collects information about individual characteristics like gender, age, education, main activity, income, household's attributes, and detailed information on the expenses of the household. The survey is representative at the national level and by urban and rural areas. Other representative areas include Quito, Guayaquil, Cuenca and Machala. The survey design was stratified, multi-staged and clustered. At the first and second stages the units were selected with probability proportional to the number of households on them, while at the third stage households were selected with equal probability. Weights are then necessary to get suitable estimations. The final sample size was 13,581 households, 8,065 in urban area and 5,516 in rural areas (Table 1). Table 1. Households by Region Region Urban Rural Total Sierra 3,815 3,451 7,266 Costa Costa 3,770 1,537 5,307 Oriente 480 528 1,008 Total 8,065 5,516 13,581 Source: Own calculations based on ECV. Nation-wide, the ECV indicates that urban households have an average consumption of one and a half times the poverty line, while the mean consumption of rural households is almost equal to the poverty line. 2. Model A linear model is used to estimate the contribution of location differences in living standards after controlling for non-geographic attributes. The dependent variable is the logarithm of the welfare ratio, which is a proxy for living standards, and the independent variables are household attributes. These are two different regression equations, one for each area (urban and rural). The model is as follows: , (1) , (2) where denotes the welfare ratio, defined as the per capita consumption deflated or divided by a national poverty line at the household , is a vector of non-geographic variables for the household , 99 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 99 is a vector of geographical dummy variables (provinces), and and denote the urban and rural areas. The error terms, and , are assumed to be independent and identically distributed with zero mean; , , , , and are the parameters to be estimated.31 Based on the above specification, differences in living standards between different geographic areas may be attributed to two main factors: 1. differences in the mobile non-geographic characteristics between urban and rural areas, holding all else constant, i.e. differences between and ; or 2. differences in the returns to characteristics, holding all else constant, i.e. differences between and . The vectors and of explanatory variables include the following: · Demographics. The number of babies, children, teenagers and adults; age of the household head and their squared values; gender of the household head; household structure (household head married with spouse, single without spouse, separated/divorced/widowed without spouse, and married without spouse present); and ethnicity of household head (indigenous, mixed raced, white, black, other). · Education. Level of education according to the household head and his/her spouse years of schooling (none, one to four, five to seven, eight to 10, 11 to 13 and 14 or more years). · House ownership. Four categories: own, rent, own and paying, other. · Occupation status. Whether the household head is an employee, employer, self-employed, employee without pay, farm laborer, owner-farmer, self-employed farm laborer or not working. · Geography. In addition to urban and rural areas, there are dummies for each of the 21 provinces32 (the ECV excludes the Galápagos islands): Azuay, Bolívar, Cañar, Carchi, Cotopaxi, Chimborazo, El Oro, Esmeraldas, Guayas, Imbabura, Loja, Los Ríos, Manabí, Morona Santiago, Napo, Orellana, Pastaza, Pichincha, Tungurahua, Sucumbíos and Zamora. Several of the variables mentioned above are categorical, therefore it is necessary to leave one category as a reference group. Those categories are: Pichincha province (the location of Quito, Ecuador's capital), male household head, married with spouse, no education of household head, no education of spouse, owner of house, and employee. Analogous equations to (1) and (2) are estimated by region in order to allow for variation of coefficients across regions, as follows: , (1') , (2') where and denote the urban and rural areas of region , respectively; denotes Sierra ("mountain"), Costa ("coast") and Oriente ("east" or "jungle"). These regions are conformed by the following provinces: 31 The omitted variables are assumed to be uncorrelated with place of residence, otherwise the estimates of geographic effects will be biased. 32Note: The provinces of Santa Elena and Santo Domingo de los Tsachilas were created in 2007, and were not used in the survey data employed by this study. 100 100 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES · Sierra: Azuay, Bolívar, Cañar, Carchi, Cotopaxi, Chimborazo, Imbabura, Loja, Tungurahua and Pichincha provinces, which account for 47 percent of households in Ecuador. · Costa: El Oro, Esmeraldas, Guayas, Los Ríos and Manabí provinces, which account for 49 percent of households in Ecuador. · Oriente: Morona Santiago, Napo, Orellana, Pastaza, Sucumbíos and Zamora provinces, which account for 4 percent of the households in the country. The urban-rural composition in the first two regions is similar, with 65 percent of the households in the urban area in the case of Sierra and 75 percent in the case of Costa. In the Oriente region, approximately 61 percent of households are in rural areas. 3. Estimates from the Model at the National Level This section estimates the returns to each one of the household characteristics at the national level, by urban and rural areas, using linear regression. This is the necessary first step before subsequently moving on to the decomposition analysis. The explanatory variables from equations (1) and (2) included in the model explain approximately 59 percent and 52 percent of the variability in the welfare ratio for urban and rural areas, respectively (Table 2).33 The estimated effects differ by areas. However, these effects are significant and with the expected sign by area. The analysis indicates the following: · An increase in household size and having a female household head is correlated with lower consumption; the welfare ratio for a household head with spouse is less than the welfare ratio of a household head without spouse; an increase in the household head's age increases the welfare ratio; and the welfare ratio is lower for indigenous household heads. · The welfare ratio of a household head with some education is greater than the welfare ratio of a household head with no education, and the same result applies to the education of the spouse. The returns to education of the household head's spouse are significant in both urban and rural areas, but higher in urban areas. · Consumption is lower in households that do not own their house (rent, paying, other) compared to households with their own house. · Welfare ratios are higher for employers or owner-farmers than for employees, and lower for the self-employed, employees without pay, farm laborers, self-employed farm laborers or those not working. · On average, consumption in several provinces is lower than in Pichincha province, which includes the national capital. Estimated coefficients for rural and urban areas are different for some provinces. 33 In order to estimate the parameters and their standard errors, the survey design was taken into account. 101 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 101 Table 2. Regressions for Log Welfare Ratio Urban Urban Rural Standard Standard Standard Standard Explanatory variables Coefficient Error Coefficient oefficient Error Constant -0.336 * 0.101 -0.279 * 0.107 Province Azuay 0.142 * 0.040 -0.001 0.085 Bolívar -0.265 * 0.077 -0.238 * 0.090 Cañar -0.027 0.049 0.071 0.091 Carchi Carchi -0.333 * 0.047 -0.438 * 0.095 Cotopaxi -0.169 * 0.055 -0.063 0.085 Chimborazo Chim -0.198 * 0.056 -0.202 * 0.089 El Oro -0.186 * 0.039 -0.145 0.094 Esmeraldas -0.191 * 0.044 -0.174 * 0.084 Guayas -0.247 * 0.036 -0.057 0.089 Imbabura Im -0.200 * 0.042 -0.294 * 0.100 Loja -0.077 ** 0.043 -0.272 * 0.094 Los Ríos -0.382 * 0.039 -0.112 0.086 Manabí -0.358 * 0.047 -0.189 * 0.082 Morona Santiago 0.020 0.083 -0.601 * 0.166 Napo -0.221 0.139 -0.351 * 0.141 Pastaza -0.123 * 0.052 -0.374 * 0.158 Tungurahua -0.035 0.039 -0.089 0.077 Zamora -0.197 * 0.084 -0.298 * 0.092 Sucumbíos -0.002 0.044 -0.209 * 0.085 Orellana Orellana -0.039 0.055 -0.335 ** 0.175 Demographics g p Number of babies -0.296 * 0.036 -0.376 * 0.032 Number of babies squared 0.035 0.021 0.076 * 0.018 Number of children -0.272 * 0.014 -0.264 * 0.015 Number of children squared 0.024 * 0.004 0.024 * 0.004 Number of teenagers -0.209 * -0.209 0.022 0.022 -0.141 -0.141 * 0.023 0.023 Number of teenagers squared g q 0.015 0.010 0.012 ** 0.007 Number of adults -0.109 * 0.016 -0.076 * 0.017 Number of adults squared q 0.005 * 0.002 0.007 * 0.003 Sex of the head -0.022 0.027 -0.091 * 0.036 No spouse, single 0.253 * 0.052 0.092 ** 0.048 No spouse, separated/divorced/widowed 0.227 * 0.048 0.206 * 0.037 No spouse, married 0.422 * 0.065 0.459 * 0.070 Age of household head 0.028 * 0.003 0.020 * 0.003 Age of household head squared g q -0.0002 * 0.000 -0.0002 * 0.000 Mixed race (mestizo) household head (mestizo) 0.071 ** 0.040 0.214 * 0.038 White household head 0.167 * 0.049 0.261 * 0.054 Black household head Black 0.026 0.053 0.274 * 0.060 Mixed race ((mulato)) household head 0.041 0.050 0.262 * 0.068 Other ethnicity household head 0.093 0.350 ---- 102 102 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Education of Household Head 1 - 4 years 0.255 * 0.047 0.156 * 0.028 5 - 7 yyears 0.452 * 0.046 0.276 * 0.029 8 - 10 years 0.572 * 0.047 0.397 * 0.037 11 - 13 years y 0.765 * 0.051 0.520 * 0.050 14 - + years 1.135 * 0.056 0.913 * 0.101 Education of Spouse 1 - 4 years 0.020 0.043 0.036 0.030 5 - 7 years 0.124 * 0.044 0.086 * 0.029 8 - 10 years 0.156 * 0.046 0.144 * 0.040 11 - 13 years 0.267 * 0.045 0.216 * 0.052 14 - + years 0.472 * 0.048 0.512 * 0.075 House Ownership Rent Rent -0.130 * -0.130 0.018 0.018 0.042 0.042 0.060 0.060 Own and paying -0.004 0.036 0.269 * 0.127 Other Other -0.146 * 0.020 -0.057 * 0.027 Position in Occupation p Employer 0.313 * 0.027 0.297 * 0.052 Self-employed p y -0.051 * 0.017 -0.062 * 0.031 Employee no pay 0.057 0.048 -0.154 * 0.054 Farm labourer -0.131 * 0.030 -0.169 * 0.026 Owner farmer Owner 0.238 * 0.055 0.217 * 0.039 Self-employed farm labourer -0.232 * 0.057 -0.197 * 0.029 Not working -0.046 ** 0.027 -0.302 * 0.041 Source: Own calculations based on ECV. Note: Number of observations: 7950 (urban) and 5481 (rural). R2=0.59 (urban) and 0.52 (rural), * indicates that the coefficient is significant at the 5 percent level, and ** at the 10 percent level. The base categories are: Pichincha area (Quito belongs to that area), male household head, married with spouse, indigenous household head, no education of household head, no education of spouse, own house and employee. According to the F tests, the null hypothesis , where is the coefficient for the category of variable and is the number of categories of this variable--i.e., the hypothesis that the effects of all categories of variable are the same in urban and rural areas--is rejected for almost all variables (Table 3). In other words, all the explanatory variables but education of the spouse have different effects in urban than in rural sectors. Table 3. Test of Equality of Coefficients Between Urban and Rural Regressions Number of of T value Variable Variable restrictions restrictions /F value P value value Test (5% level) (5% level) Constant Constant 1 -2.61 0.01 Rejected Province 20 5.38 0.00 Rejected j Demographics ographics 18 5.87 0.00 Rejected Rejected Education variables 10 2.04 0.03 Rejected Education of Household Head 5 2.95 0.01 Rejected Rejected Education of Spouse p 5 0.63 0.68 Not rejected j House Ownershipnership 3 5.16 5.16 0.00 0.00 Rejected Rejected Position in occupation 7 5.58 0.00 Rejected Source: Own calculations based on ECV Survey. 103 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 103 Annex 1 presents the results for the three regions chosen in the analysis using equations (1') and (2'). The Oriente region (mainly the rural part) is very different to the rest of the regions in characteristics and on returns. By contrast the Sierra (including Quito) and Costa (including Guayaquil) appear to be more similar on characteristics and returns. 4. Decompositions at the National Level In light of the salient differences in per capita income between rural-urban areas and across provinces, the following analysis considers the contribution of household characteristics to per capita income differences. These differences may come from differences in characteristics (for example, a lower level of education in rural areas) or from differences in the returns to characteristics (for example, a lower impact of education on earnings and thereby a lower per capita income in rural areas). In some cases, the differences in characteristics and in the returns to characteristics reinforce each other, but in other cases they might not. This is analyzed using the Oaxaca-Blinder decomposition of the income gap between assets and returns to those assets (Blinder 1973 and Oaxaca 1973). The estimates from equations (1) and (2) are used to analyze the comparisons. Three types of geographic comparisons are examined: i) the difference in mean welfare ratios between urban and rural areas, ii) the difference within the urban and rural sectors across provinces, and iii) the difference between urban and rural areas within a given province. The first questions concern the overall differential in living standards between urban and rural areas at the national level. This entails a comparison of urban-rural differentials in mean welfare ratios, given by (3): (3) where and are the sample means for urban and rural areas respectively, and and are the proportions of province 's population in each area. Table 4 shows the result obtained for equation (3) using the coefficients showed in Table 2 and the means reported in Annex Table A1. The difference in the intercepts, -0.057, gives the difference between the fitted log welfare ratio for a married couple, both illiterate, with a male indigenous household head who owns a house and is an employee living in the urban Pichincha area, and a household with the same characteristics located in the rural Pichincha area. Among the non-geographic variables that could cause the differential impact in urban and rural areas, house ownership is a minor factor. The next variable in order of significance is the occupation of household head, while most of the differential is due to demographic variables and education. In fact, demographics accounts for around 32 percent of the differential in log welfare ratio and education for 68.5 percent approximately, which is explained by both higher levels and higher returns to education in urban sector. The difference due to the geographic variable is small, which indicates that, controlling for other characteristics, the gap between Pichincha province and all other provinces in urban areas is almost of the same order in magnitude as the gap in rural areas. 104 104 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table 4. Contribution of Variables to Average Level of Log Welfare Ratio Urban- Urban- Urban Urban Rural Rural Rural Difference Difference Mean log welfare ratio 0.689 -0.024 0.713 Decomposition Constant -0.336 -0.279 -0.057 Geographic variables Geographic -0.158 -0.141 -0.017 Demographic variables 0.418 0.185 0.233 Education variables 0.817 0.328 0.489 House ownership variable -0.054 -0.007 -0.046 Occupation variable 0.001 -0.110 0.111 Source: Own calculations based on ECV Survey. The comparison between urban areas or rural areas of two provinces is based simply on the comparison of the coefficients of province dummy variables in Table 2. For the urban areas, only in Azuay is the coefficient estimate positive, while in the rural areas all of the significant province coefficient estimates are negative. Households living in Pichincha appear to be better off or equal to their urban and rural counterparts from other provinces, after controlling for the non-geographic characteristics. The second component of the right side in equation (3) reflects differentials in urban and rural consumptions due to both differences in returns and in characteristics. It does not quantify structural differences, which is accomplished by comparing the expected gain in consumption from living in urban areas of a given province over rural areas, given the national means of all non-geographic characteristics, . For province this is given by: (4) where is a vector with zeros in all its entries except for entry if the household belongs to province . The first term in equation (4) is the same as in equation (3), and the second term gives the effect of urban-rural differences in the returns to household characteristics. The sum of these two terms is 0.24, which can be obtained from coefficients in Table 2 and means at national level in Table A1, and accounts for the difference between the conditional log consumption of households living in urban and rural areas of the Pichincha province when conditioning on national means. The third term is close to zero for seven of the provinces, meaning that in these provinces the differences in expected log consumptions between urban and rural areas are similar to Pichincha province (Table 5). For most of the remaining provinces, the differences are moderate to large; this suggests that residence in a given province is related to differences in expected consumption, once we control for household characteristics. The province of Morona Santiago has the largest difference in consumption between urban and rural area. 105 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 105 Table 5. Differences and Expected Gain by Province Difference Expected Province Urban - Rural gain Azuayy 0.142 0.380 Bolívar -0.027 0.211 Cañar -0.098 0.140 Carchi 0.105 0.343 Cotopaxi -0.106 0.132 Chimborazo 0.004 0.242 El Oro -0.041 0.197 Esmeraldas -0.017 0.221 Guayas -0.189 0.049 Imbabura 0.094 0.332 Loja 0.195 0.433 Los Ríos Ríos -0.270 -0.032 Manabí -0.169 0.069 Morona Santiago Santiago 0.621 0.859 Napo 0.130 0.368 Pastaza Pastaza 0.250 0.250 0.488 0.488 Tungurahua 0.054 0.292 Zamora 0.101 0.339 Sucumbíos 0.207 0.445 Orellana Orellana 0.296 0.534 Source: Own calculations based on ECV. To estimate the importance of characteristics and returns obtained from these characteristics in explaining welfare differences between urban and rural areas, we analyze each of these factors separately. i) Geographic profile of living standards. This reflects the differentials in returns between urban and rural areas, isolating the structural component, i.e., controlling for all non-geographical attributes. The geographic log welfare ratio is defined as: (5) . (6) Any differences between the log of welfare ratio in (5) and (6), for a given province, are due to differences in the returns to the common characteristics . ii) Concentration profile of living standards. This reflects the spatial concentration of non-geographic characteristics, by isolating the effects of non-geographic attributes controlling for geographic variables. The concentration log welfare ratio is defined as: , 106 106 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES where and represent the mean characteristics of the households living in the province for the urban and rural areas, respectively; and are the weighted average parameters: , . For most of the provinces, as was expected, both geographic and concentration urban welfare ratios are lower than the unconditional ones (Table 6). However, the reasons are not the same. In the geographic profile, this is because urban households tend to have better characteristics than households at national level. In the concentration profile, this is because the returns to better characteristics tend to be larger in urban areas than nationally. By analogous reasoning, average consumption for geographic and concentration simulations in rural areas is larger compared to an unconditional profile. Table 6. Welfare Ratios by Province and by Urban/Rural Area (logarithms) Geographic Concentration oncentration Unconditional Unconditional Province Profile Profile Profile Urban Urban Rural Urban Urban Rural Urban Urban Rural Pichincha 0.69 0.45 0.68 0.26 0.91 0.25 Azuay Azuay 0.83 0.45 0.70 0.70 0.17 1.06 0.16 Bolívar 0.42 0.21 0.70 0.04 0.66 -0.23 Cañar Cañar 0.66 0.52 0.43 0.12 0.63 0.16 Carchi 0.36 0.01 0.59 0.18 0.49 -0.26 Cotopaxi 0.52 0.39 0.68 0.68 0.09 0.74 0.00 Chimborazo 0.49 0.25 0.77 -0.03 0.81 -0.26 El Oro Oro 0.50 0.31 0.59 0.59 0.34 0.63 0.16 Esmeraldas 0.50 0.28 0.51 0.05 0.53 -0.07 Guayas 0.44 0.39 0.59 0.59 0.11 0.58 0.06 Imbabura 0.49 0.16 0.55 0.03 0.59 -0.26 Loja 0.61 0.18 0.75 0.15 0.90 -0.15 Los Ríos 0.31 0.34 0.49 0.22 0.33 0.11 Manabí 0.33 0.26 0.55 0.09 0.42 -0.08 Morona Santiago 0.71 -0.15 0.60 -0.12 0.84 -0.74 Napo 0.47 0.10 0.84 -0.13 0.86 -0.50 Pastaza 0.57 0.08 0.63 -0.24 0.72 -0.61 Tungurahua 0.65 0.36 0.68 0.68 0.18 0.88 0.07 Zamora 0.49 0.15 0.57 0.05 0.59 -0.25 Sucumbíos Sucum 0.69 0.24 0.64 0.18 0.86 -0.02 Orellana 0.65 0.12 0.26 -0.11 0.42 -0.43 Source: Own calculations based on ECV Survey. There are important and positive correlations between geographic and unconditional profiles (0.73 for urban, 0.91 for rural and 0.91 for national), and between concentration and unconditional profiles (0.72 for urban, 0.87 for rural and 0.96 for national), meaning that both structural differentials (returns) and disparities in characteristics (endowments) are important to explain urban-rural differences in welfare ratios within provinces. Mean consumption--conditional on non-geographic variables set at national level--is higher in urban than in rural areas, meaning that the returns are larger for the urban sector than for rural areas. When the 107 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 107 welfare ratio is simulated setting all parameters at the national level and explanatory variables at the mean of the province, urban measures are also higher compared to the rural ones. This is due to household characteristics that tend to increase living standards, such as education, which is higher in urban than in rural sectors. In order to calculate profiles at the national level, equations analogous to (1) and (2) were estimated without province dummies (Table 7). The results indicate that characteristics are more important in explaining differences between urban and rural sector at the national level. Characteristics explain around 72 percent of the differential in welfare ratio, while returns explain 28 percent. Table 7. Welfare Ratios at National Level by Urban/Rural Area (logarithms) Profile Urban Rural Ru Geographic 0.527 0.326 Concentration tion 0.625 0.112 Unconditional 0.689 -0.024 Source: Own calculations based on ECV Survey. Poverty Measures This section presents the same analysis as above, but based on poverty rates instead of log welfare ratios. Assuming normally distributed errors, and conditioning on national sample means, the conditional probabilities of being poor for a household living in province j of urban and rural areas are represented as: Pª¬logWi <0|iU,Gi =Gj,Xi = Xº¼ = ¬ª- U + ( U X + Uj U º ) ¼ Pª¬logWi <0|iR,Gi = Gj,Xi = Xº¼ = ¬ª- R + ( R X + Rj R ¼º ) where U and R are the standard deviations of the errors in the urban and rural regressions and is the cumulative density of the standard normal distribution. Similarly, to calculate the simulated poverty rates based on the concentration profile, the weighted average of urban and rural parameters are used. The sources of the differences between simulations and the unconditional (actual) poverty rates are characteristics or changes in returns to those characteristics, and the assumption that logarithm of welfare ratio follows a normal distribution. Thus, the poverty rates for the unconditional profile were calculated assuming a normal distribution. For most of the provinces, the concentration, geographic, and unconditional measures of poverty are all lower for urban than for rural areas (Table 8). The correlations between the geographic and unconditional (normal) profiles are 0.72 for urban and 0.91 for rural, and are very close to the correlations between concentration and unconditional (actual) profiles, which are 0.69 and 0.88 for urban and rural areas, respectively. The conclusions are similar to those discussed above, meaning that both returns and characteristics explain the differences in welfare ratios within each province, while at the national level characteristics are more important. 110 108 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table 8. Poverty Rates by Province and by Urban/Rural Area Geographic Concentration oncentration Unconditional Unconditional Province Profile Profile Profile (normal) (normal) Profile Profile Urban Urban Rural Rural Urban Rural Urban Urban Rural Urban Rural Rural Urban Urban Rural Rural Pichincha 8.2 18.4 8.5 30.1 11.5 36.3 12.7 42.5 Azuay Azua 4.7 4.7 18.4 18.4 8.0 8.0 36.8 36.8 7.9 7.9 41.2 41.2 5.0 5.0 43.0 43.0 Bolívar 19.6 33.5 8.0 47.0 19.1 62.3 11.6 65.9 Cañar Cañar 9.1 9.1 14.8 19.4 40.8 20.1 41.5 41.5 13.8 41.1 41.1 Carchi 23.6 48.9 11.7 35.9 25.9 64.2 28.0 64.3 Coto axi Cotopaxi 14.7 14.7 21.9 8.5 8.5 43.0 43.0 16.5 16.5 49.9 49.9 14.6 14.6 48.5 48.5 Chimborazo 16.1 30.9 6.0 52.0 14.3 64.2 16.0 68.3 El Oro ro 15.5 15.5 27.1 11.9 24.9 24.9 20.2 20.2 41.0 41.0 15.3 15.3 41.6 41.6 Esmeraldas 15.8 29.0 15.0 46.0 24.1 53.6 25.9 53.3 Gua as Guayas 18.6 21.6 11.6 41.5 41.5 22.2 46.9 46.9 22.4 22.4 46.2 46.2 Imbabura 16.2 37.7 13.2 47.8 21.7 64.2 21.0 68.1 Lo Loja 10.8 10.8 36.0 36.0 6.5 6.5 38.6 38.6 11.7 11.7 58.4 58.4 8.4 8.4 63.4 63.4 Los Ríos 26.7 24.9 16.2 33.1 33.0 44.0 33.7 45.5 Manabí Manabí 25.2 25.2 30.0 30.0 13.4 13.4 42.6 42.6 29.0 29.0 54.5 54.5 31.8 31.8 61.2 61.2 Morona Santiago 7.6 61.7 11.4 59.3 13.4 84.8 17.0 77.2 Napo 17.3 42.0 42.0 4.6 4.6 60.1 60.1 12.9 75.4 75.4 16.9 71.3 Pastaza 12.7 43.8 10.3 68.4 16.9 79.9 11.9 76.0 Tungurahua Tun urahua 9.3 9.3 23.5 23.5 8.4 8.4 36.4 36.4 12.1 12.1 46.2 46.2 11.7 11.7 46.9 46.9 Zamora 16.1 38.0 12.5 46.4 21.9 63.5 21.4 68.2 Sucumbíos Sucumbíos 8.3 8.3 31.4 31.4 9.7 9.7 36.0 36.0 12.7 12.7 50.8 50.8 11.4 11.4 50.8 50.8 Orellana 9.5 40.8 30.2 58.8 28.9 72.6 23.9 66.5 Source: Own calculations based on ECV Survey. 5. Decompositions at the Regional Level This section expands the analysis from national to regional level. The analysis clearly reveals that for both the Sierra and the Costa, demographics and education account for the largest share in total log welfare ratio (Table 9). In the Costa, demographic factors account for about 28 percent of the differential in log welfare ratio, and education for more than 79 percent. In the Sierra, demographic factors account for about 32 percent of the differential in log welfare ratio, and education for more than 74 percent. In Oriente, the difference in the constant term accounts for the largest share in living standards, which means that a married couple, both illiterate, with male indigenous household head who owns a house and is an employee living in the urban Oriente has a mean log welfare ratio 0.72 higher than a household with the same characteristics located in the rural Oriente. Table 9. Contribution of Variables to Average Level of Log Welfare Ratio by Region Urban-Rural Difference rban-R Sierra Sierra Costa Costa Oriente Oriente Mean log welfare ratio 0.868 0.516 1.142 Decomposition Decom Constant term -0.090 -0.075 0.720 Demographic variables 0.242 0.166 -0.272 Education variables 0.682 0.383 0.494 House ownership ownership variable -0.089 -0.089 -0.024 -0.042 Position in occupation p variable variable 0.122 0.122 0.065 0.065 0.242 0.242 Source: Own calculations based on ECV Survey. 109 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 109 Comparison between urban and rural areas within each of the regions is analogous to that of national level: the concentration profile uses weighted parameters from specifications (1') and (2') and the means for urban and rural areas of the respective region, while the geographic profile was obtained by setting the mean at the regional level and the parameters estimated using (1') and (2'). To compare across the regions of the same area (urban/rural), the concentration measure uses the means of each region only for the area of interest (urban/rural) and weighted returns for the two regions in the comparison from the regressions that were estimated separately for each region. The geographic effect uses the weighted means for the two regions in the comparison and the estimated coefficients for each region and area of interest. The results are presented below. In the Sierra and Costa, the disparities in the concentration profile between urban and rural areas are larger than the differences in the geographic profile, which means that characteristics are the major explanation behind welfare differentials in urban and rural areas. In the Oriente, which is mainly rural (approximately 61 percent of households), both returns and characteristics are important in order to explain the differences in welfare ratio between urban and rural sector (Figure 1). The same conclusions emerged using the poverty rates shown in Table 10. Figure 1. Decomposition of Welfare Ratios (logarithms) Within Regions Source: Own calculations based on ECV Survey. 110 110 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table 10. Poverty Rates by Region and Within Region Geographic Concentration oncentration Unconditional Unconditional Unconditional Unconditional Province Province Profile Profile Profile Profile Profile (normal) (normal) Profile Profile Urban Rural Urban Urban Urban Rural Urban Urban Rural Urban Rural Urban Sierra 9.4 22.3 5.9 36.5 12.5 49.8 12.8 52.6 Costa Costa 19.2 32.4 16.6 43. 43.6 23.0 23.0 48.6 48.6 24.2 24.2 51.0 51.0 Oriente 23.6 59.4 12.0 43.6 16.5 68.7 16.3 67.6 Source: Own calculations based on ECV Survey. For the Oriente, characteristics and returns each explain around 50 percent of the difference in log- welfare ratios between urban and rural sector (Figure 2). In the Sierra and Costa, characteristics explain approximately 30 percent and 34 percent of the differentials, respectively. Figure 2. Decomposition of Welfare Differences (logaritms) Within Regions Source: Own calculations based on ECV Survey. The urban-rural composition of the Sierra and Costa are similar, with 65 percent of the households in urban areas in the case of Sierra and 75 percent in the case of Costa, while Oriente is mainly rural, at 61 percent approximately. In terms of poverty, Costa is the lagging region in the urban sector while Oriente is behind in the rural area. The comparison between Oriente and Costa on the one hand and Sierra and Costa on the other in urban areas reveals that urban welfare differences are mostly due to returns, rather than characteristics. In the former case, concentration plays a role also, but in spite of the fact that Oriente has better characteristics than Costa, which yields a difference in the concentration profile, it is smaller than the difference in the returns. In the second comparison, both regions have similar characteristics, which lead to a differential only in returns (Figure 3). In the rural sector, both geographic and concentration effects explain the disparities in poverty rates between Sierra and Oriente. Characteristics explain the main differential between Costa and Oriente, although there is a difference in returns as well (Figure 4). 111 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 111 Figure 3. Decomposition of Poverty Rates. Urban Areas Figure 4. Decomposition of Poverty Rates. Rural Areas Source: Own calculations based on ECV Survey. In sum, returns explain about 90 percent of the differences in log welfare ratio between the urban areas of Oriente and Costa and about 70 percent between the urban areas of Sierra and Costa (Figure 5). The differential between the rural areas of Sierra and Oriente is explained by returns and characteristics, while the differences between Costa and Oriente are mainly due to characteristics. 112 112 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 5. Decomposition of Welfare Differences Between Regions 6. Conclusions Overall, the results of the study indicate that at the national level, urban-rural differences in welfare ratios are mainly due to characteristics, which account for almost three-quarters of the difference. Comparisons in living standards within a given region across urban and rural areas reveal that in the Sierra and Costa regions, characteristics are also the main source of differences in welfare. In the Oriente, both characteristics and returns are important to explain the welfare level differences between urban and rural areas. The empirical results indicate that among the characteristics analyzed, education was the variable with the largest contribution to explain the differential in mean log welfare ratio between urban and rural areas, as was expected. Other non-geographic characteristics also contribute to the difference, but to a lesser extent. According to the results obtained from both urban and rural areas in the simulated welfare ratios and the simulated poverty rates within each province, differences in living standards between urban and rural areas can be attributed to disparities in non-geographic variables and to differences in the returns to those characteristics. This indicates that poor areas are poor not only because households with observable non- geographic characteristics that favor poverty are settled there, but because there are differences in the returns to those characteristics by location. Comparisons in living standards between a leading and a lagging region reveal that for urban areas, the differences between Oriente and Costa and between Sierra and Costa are due to returns. In the case of rural areas, the differentials between Costa and Oriente are explained by characteristics, while the differences between Sierra and Oriente are explained by both returns and characteristics. 113 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 113 Annex Table A1. Means/Percentages and Standard Errors Urban Rural National Std. Std. Std. Std. Std. Std. Mean / Err. of of Mean / Err. of Mean / Err. of Percentage Mean / Percentage Mean / Percentage Mean / Variable Variable Percent Percent Percent Percent Percent Percent Province Azuayy 4.0% 0.2% 7.0% 1.2% 4.9% 0.4% Bolívar Bolívar 0.5% 0.2% 3.0% 3.0% 0.6% 0.6% 1.3% 1.3% 0.2% 0.2% Cañar 0.9% 0.2% 3.1% 0.6% 1.6% 0.3% Carchi Carchi 1.0% 0.2% 1.9% 1.9% 0.5% 0.5% 1.3% 1.3% 0.2% 0.2% Cotopaxi p 1.1% 0.3% 6.3% 1.0% 2.7% 0.4% Chimborazo Chim 2.0% 0.4% 6.2% 6.2% 1.1% 1.1% 3.3% 3.3% 0.5% 0.5% El Oro 5.3% 0.7% 3.7% 0.9% 4.8% 0.5% Esmeraldas 2.5% 0.4% 3.9% 3.9% 0.8% 0.8% 3.0% 3.0% 0.4% 0.4% Guayas y 33.5% 1.3% 11.7% 2.4% 26.7% 1.2% Imbabura 2.8% 0.5% 3.1% 3.1% 0.8% 0.8% 2.9% 2.9% 0.4% 0.4% Loja j 2.1% 0.4% 5.6% 1.1% 3.2% 0.4% Los Ríos Ríos 4.6% 4.6% 0.7% 0.7% 7.5% 7.5% 1.4% 1.4% 5.5% 5.5% 0.7% 0.7% Manabí 7.7% 1.1% 12.1% 2.0% 9.1% 1.0% Morona Santiago 0.4% 0.1% 2.1% 2.1% 0.7% 0.7% 0.9% 0.9% 0.2% 0.2% Napo 0.3% 0.1% 0.9% 0.4% 0.5% 0.2% Pastaza 0.5% 0.5% 0.2% 0.2% 1.1% 1.1% 0.7% 0.7% 0.7% 0.7% 0.2% 0.2% Tungurahua 2.7% 0.5% 6.2% 1.0% 3.8% 0.5% Zam Zamora 0.2% 0.2% 0.1% 0.1% 1.1% 1.1% 0.4% 0.4% 0.5% 0.5% 0.2% 0.2% Sucumbíos 0.5% 0.2% 1.9% 0.6% 1.0% 0.2% Orellana 0.4% 0.2% 1.0% 1.0% 0.4% 0.4% 0.6% 0.6% 0.2% 0.2% Pichincha 26.8% 1.3% 10.4% 2.2% 21.7% 1.1% Demographics Number of babies 0.2 0.0 0.3 0.0 0.3 0.0 Number of babies squared 0.3 0.0 0.4 0.4 0.0 0.0 0.3 0.3 0.0 0.0 Number of children 0.8 0.0 1.1 0.0 0.9 0.0 Num Number of children children squared squared 1.6 0.0 2.7 0.1 1.9 0.0 Number of teenagers 0.5 0.0 0.6 0.0 0.5 0.0 Number of teenagers squared squared 0.8 0.0 1.1 1.1 0.0 0.0 0.9 0.0 Number of adults 2.4 0.0 2.4 0.0 2.4 0.0 Number of adults squared 5.8 0.1 5.5 5.5 0.1 0.1 5.7 0.1 Female household head 23.3% 0.6% 15.7% 0.7% 20.9% 0.5% Male household head 76.7% 0.6% 84.3% 84.3% 0.7% 0.7% 79.1% 79.1% 0.5% 0.5% No spouse, single 7.2% 0.3% 5.8% 0.4% 6.8% 0.3% 114 114 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES No spouse, separate/divorced/widow 22.8% 0.6% 17.1% 17.1% 0.6% 0.6% 21.0% 21.0% 0.5% 0.5% No spouse, married p 1.6% 0.2% 2.1% 0.2% 1.7% 0.1% Spouse, married 68.4% 0.7% 75.0% 75.0% 0.8% 0.8% 70.5% 70.5% 0.5% 0.5% Age of household head 45.9 0.2 48.1 48.1 0.3 0.3 46.6 46.6 0.2 0.2 Age of household head g squared squared 2346.6 24.2 2585.3 35.7 2421.6 2421.6 20.2 20.2 Mixed race (mestizo) household head 82.0% 0.7% 73.7% 1.6% 79.4% 0.7% White hite household head 8.7% 0.5% 5.6% 5.6% 0.5% 0.5% 7.7% 7.7% 0.4% 0.4% Black household head 2.9% 0.3% 3.1% 0.5% 3.0% 0.3% Mixed race (mulato) household household head 2.9% 0.3% 1.5% 1.5% 0.3% 0.3% 2.5% 2.5% 0.2% 0.2% Other ethnicity household head 0.1% 0.0% 0.0% 0.0% 0.0% 0.0% 0.05% 0.05% 0.02% 0.02% Indigenous 3.4% 0.4% 16.2% 1.5% 7.4% 0.6% Education of Household Head 1 - 4 years 8.6% 0.4% 22.1% 22.1% 0.8% 0.8% 12.8% 12.8% 0.4% 0.4% 5 - 7 years 28.9% 0.8% 43.2% 1.0% 33.4% 0.6% 8 - 10 years 14.2% 0.5% 7.3% 7.3% 0.4% 0.4% 12.0% 12.0% 0.4% 0.4% 11 - 13 years y 21.2% 0.6% 7.4% 0.5% 16.9% 0.4% 14 - + years 23.5% 1.1% 4.3% 4.3% 0.6% 0.6% 17.5% 17.5% 0.8% 0.8% None 3.6% 0.3% 15.7% 0.7% 7.4% 0.3% Education of Spouse 1 - 4 years 5.1% 0.3% 13.9% 0.6% 7.8% 0.3% 5 - 7 years 18.0% 0.6% 32.4% 32.4% 0.9% 0.9% 22.5% 22.5% 0.6% 0.6% 8 - 10 years 10.9% 0.5% 7.6% 0.5% 9.9% 0.4% 11 - 13 years 16.7% 0.5% 5.6% 5.6% 0.5% 0.5% 13.2% 13.2% 0.4% 0.4% 14 - + years y 14.8% 0.7% 2.4% 0.3% 10.9% 0.5% None / No spouse spouse 34.6% 0.7% 38.2% 38.2% 1.0% 1.0% 35.7% 35.7% 0.6% 0.6% House Ownership ouse nership Rent 25.0% 0.8% 3.2% 0.5% 18.1% 0.6% Own and paying 3.6% 0.4% 0.8% 0.8% 0.1% 0.1% 2.7% 2.7% 0.3% 0.3% Other 14.5% 0.5% 19.6% 1.0% 16.1% 0.5% Own and paid paid 56.9% 0.9% 76.4% 76.4% 1.1% 1.1% 63.0% 63.0% 0.7% 0.7% Position in Occupation ccupation Employer p y 8.5% 0.4% 2.1% 0.2% 6.5% 0.3% Self-employed 27.8% 0.7% 10.3% 10.3% 0.8% 0.8% 22.3% 22.3% 0.6% 0.6% Employee without pay 1.5% 0.1% 1.6% 0.2% 1.5% 0.1% Farm labourer 3.3% 0.4% 21.7% 21.7% 1.2% 1.2% 9.1% 9.1% 0.5% 0.5% Owner farmer 1.0% 0.1% 6.2% 0.6% 2.7% 0.2% Self-employed farm 1.6% 0.2% 32.4% 1.3% 11.3% 0.5% 115 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 115 labourer Not working 13.2% 0.5% 6.6% 6.6% 0.6% 0.6% 11.2% 11.2% 0.4% 0.4% Employee 42.9% 0.8% 19.0% 1.0% 35.4% 0.7% Table A2. Means/Percentages and Standard Errors, Sierra Region Urban Urban Rural Sierra Region Region Std. Std. Std. Err. Std. Std. Err. Std. Mean / Mean / Mean / Err. of of of Mean of ean of Mean Percentage Percentage Percentage Mean / / Percent / Percent Variable Variable Percent Percent Demographics Number of babies 0.2 0.0 0.3 0.0 0.2 0.0 Number of babies squared 0.3 0.0 0.4 0.0 0.3 0.0 Number of children 0.7 0.0 1.0 0.0 0.8 0.0 Number of children squared 1.4 0.1 2.5 0.1 1.8 0.1 Number of teenagers 0.5 0.0 0.6 0.0 0.5 0.0 Number of teenagers squared 0.8 0.0 1.1 0.0 0.9 0.0 Number of adults ber adults 2.4 0.0 2.4 0.0 2.4 0.0 Number of adults squared 5.6 0.1 5.1 0.1 5.4 0.1 Female household head 22.6% 1.0% 20.1% 20.1% 0.9% 0.9% 21.7% 21.7% 0.7% 0.7% Male household head 77.4% 1.0% 79.9% 0.9% 78.3% 0.7% No spouse, single 9.2% 0.6% 7.3% 7.3% 0.5% 0.5% 8.5% 8.5% 0.4% 0.4% No spouse, separate/divorced/widow 18.9% 0.8% 17.3% 0.8% 18.3% 0.6% No spouse, married 2.0% 0.3% 3.1% 3.1% 0.4% 0.4% 2.4% 2.4% 0.2% 0.2% Spouse, married 69.9% 1.0% 72.3% 1.0% 70.7% 0.7% Age of household head 46.0 0.4 49.2 0.5 47.1 0.3 Age of household head squared 2358.9 35.9 2700.1 47.5 2480.2 29.3 Mixed race (mestizo) household household head 83.7% 1.0% 68.9% 68.9% 2.2% 2.2% 78.5% 78.5% 1.0% 1.0% White household head 7.3% 0.7% 4.9% 0.5% 6.4% 0.5% Black Black household head 1.7% 1.7% 0.3% 1.7% 1.7% 0.6% 0.6% 1.7% 1.7% 0.3% 0.3% Mixed race (mulato) household head household 1.6% 0.3% 1.0% 1.0% 0.3% 0.3% 1.4% 1.4% 0.2% 0.2% Other ethnicity household head head 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.00% 0.00% 0.00% 0.00% Indigenous 5.7% 0.8% 23.5% 2.2% 12.0% 1.0% Education of Household 116 116 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Head 1 - 4 years 7.0% 0.6% 21.1% 21.1% 1.0% 1.0% 12.0% 12.0% 0.6% 0.6% 5 - 7 years y 28.6% 1.3% 44.1% 1.3% 34.1% 1.0% 8 - 10 years 12.9% 0.7% 6.2% 6.2% 0.6% 0.6% 10.5% 10.5% 0.5% 0.5% 11 - 13 years 21.0% 0.9% 6.8% 0.6% 16.0% 0.6% 14 - + years 27.5% 1.8% 4.9% 4.9% 1.0% 1.0% 19.5% 19.5% 1.3% 1.3% None 3.0% 0.4% 16.9% 1.0% 7.9% 0.5% Education of Spouse 1 - 4 years 5.0% 0.5% 14.1% 14.1% 0.7% 0.7% 8.3% 8.3% 0.4% 0.4% 5 - 7 years 17.0% 0.9% 29.7% 1.1% 21.5% 0.7% 8 - 10 years 11.3% 0.7% 5.8% 5.8% 0.6% 0.6% 9.3% 9.3% 0.5% 0.5% 11 - 13 years 16.6% 0.8% 4.6% 0.6% 12.3% 0.6% 14 - + years 17.0% 1.1% 3.0% 3.0% 0.6% 0.6% 12.0% 12.0% 0.8% 0.8% None / No spouse 33.1% 1.0% 42.8% 1.4% 36.5% 0.8% House Ownership Rent 31.6% 31.6% 1.1% 1.1% 4.2% 4.2% 0.8% 0.8% 21.8% 21.8% 0.8% 0.8% Own and paying p y g 4.3% 0.7% 1.1% 0.2% 3.2% 0.5% Other 15.9% 0.8% 20.1% 20.1% 1.2% 1.2% 17.4% 17.4% 0.7% 0.7% Own and paid 48.2% 1.2% 74.6% 1.4% 57.6% 0.9% Position in Occupation Employer 8.9% 0.6% 2.6% 2.6% 0.3% 0.3% 6.6% 6.6% 0.4% 0.4% Self-employed 24.4% 0.8% 11.5% 1.0% 19.8% 0.7% Employee without pay 1.3% 0.2% 1.8% 1.8% 0.2% 0.2% 1.5% 1.5% 0.1% 0.1% Farm labourer 1.6% 0.3% 14.1% 1.2% 6.0% 0.5% Owner farmer farm 0.9% 0.2% 4.2% 4.2% 0.6% 0.6% 2.1% 2.1% 0.2% 0.2% Self-employed farm labourer 2.2% 0.3% 35.3% 1.8% 14.0% 0.8% Not working 13.3% 0.8% 6.6% 6.6% 0.9% 0.9% 10.9% 10.9% 0.6% 0.6% Employee 47.4% 1.1% 24.0% 1.5% 39.1% 0.9% Source: Own calculations based on ECV Survey. 117 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 117 Table A3. Means/Percentanges and Standard Errors, Costa Region Urban Urban Rural Costa Region Costa Std. Std. Std. Std. Std. Std. Mean / Err. of of Mean / Err. of Mean / Err. of of Percentage Mean / Percentage Mean / Percentage Mean / Variable Variable Percent Percent Percent Percent Percent Demographics Number of babies Number of babies 0.3 0.3 0.0 0.0 0.3 0.3 0.0 0.0 0.3 0.3 0.0 0.0 Number of babies squared 0.3 0.0 0.4 0.0 0.3 0.0 Number of children 0.9 0.0 1.0 1.0 0.0 0.0 0.9 0.0 Number of children squared 1.8 0.1 2.4 0.1 2.0 0.1 Number of teenagers 0.5 0.0 0.5 0.5 0.0 0.0 0.5 0.5 0.0 0.0 Number of teenagers squared 0.9 0.0 1.1 0.1 0.9 0.0 Number ber of adults adults 2.5 0.0 2.5 2.5 0.0 0.0 2.5 2.5 0.0 0.0 Number of adults squared 6.0 0.1 5.8 0.2 6.0 0.1 Female household head 24.1% 0.8% 10.3% 10.3% 0.9% 0.9% 20.7% 20.7% 0.7% 0.7% Male household head 75.9% 0.8% 89.7% 0.9% 79.3% 0.7% No spouse, single 5.5% 0.4% 4.1% 4.1% 0.6% 0.6% 5.2% 5.2% 0.4% 0.4% No spouse, separate/divorced/widow 26.3% 0.9% 17.7% 1.1% 24.1% 0.7% No spouse, married 1.2% 0.2% 0.8% 0.8% 0.3% 0.3% 1.1% 1.1% 0.2% 0.2% Spouse, married 67.0% 0.9% 77.4% 1.3% 69.6% 0.8% Age of household head Age 46.1 0.3 47.4 0.6 46.4 46.4 0.3 0.3 Age of household head squared 2355.0 33.7 2518.6 59.1 2395.9 29.4 Mixed race (mestizo) household household head 80.7% 0.9% 85.0% 85.0% 1.6% 1.6% 81.8% 81.8% 0.8% 0.8% White household head hi h h ld h d 9.9% 0.7% 6.2% 0.9% 9.0% 0.5% Black Black household head 4.0% 0.5% 5.6% 5.6% 1.0% 1.0% 4.4% 4.4% 0.4% 0.4% Mixed race (mulato) household household head 3.9% 0.5% 2.5% 2.5% 0.5% 0.5% 3.6% 3.6% 0.4% 0.4% Other ethnicity household head head 0.1% 0.1% 0.0% 0.0% 0.0% 0.0% 0.10% 0.10% 0.05% 0.05% Indigenous 1.3% 0.3% 0.8% 0.3% 1.1% 0.2% Education of Household Head 1 - 4 years 10.0% 0.6% 25.0% 25.0% 1.4% 1.4% 13.8% 13.8% 0.6% 0.6% 5 - 7 years 29.1% 1.1% 40.9% 1.5% 32.1% 0.9% 8 - 10 years 15.1% 0.7% 7.7% 7.7% 0.7% 0.7% 13.3% 13.3% 0.6% 0.6% 11 - 13 years 21.3% 0.8% 7.7% 0.8% 17.9% 0.7% 118 118 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 14 - + years 20.2% 1.3% 3.4% 0.6% 16.0% 1.0% None 4.3% 4.3% 0.4% 0.4% 15.4% 15.4% 1.2% 1.2% 7.0% 7.0% 0.5% 0.5% Education of Spouse 1 - 4 years 5.1% 0.4% 13.2% 13.2% 1.1% 1.1% 7.2% 7.2% 0.4% 0.4% 5 - 7 years 18.6% 0.9% 35.3% 1.7% 22.8% 0.9% 8 - 10 years 10.5% 0.6% 9.7% 9.7% 1.1% 1.1% 10.3% 10.3% 0.5% 0.5% 11 - 13 years 16.6% 0.8% 6.4% 0.8% 14.1% 0.6% 14 - + years 13.0% 0.9% 1.3% 1.3% 0.3% 0.3% 10.1% 10.1% 0.7% 0.7% None / No spouse 36.1% 1.0% 34.2% 1.4% 35.6% 0.8% House Ownership p Rent 19.2% 19.2% 1.1% 1.1% 1.4% 1.4% 0.4% 0.4% 14.7% 14.7% 0.9% 0.9% Own and paying 3.1% 0.5% 0.4% 0.2% 2.5% 0.4% Other 13.4% 0.7% 20.0% 20.0% 1.9% 1.9% 15.1% 15.1% 0.7% 0.7% Own and paid 64.3% 1.2% 78.1% 2.0% 67.8% 1.1% Position in Occupation p Employer 8.2% 0.6% 1.5% 1.5% 0.4% 0.4% 6.5% 6.5% 0.5% 0.5% Self-employed 31.1% 1.0% 9.4% 1.4% 25.7% 0.9% Employee without pay 1.6% 0.2% 1.3% 1.3% 0.3% 0.3% 1.6% 1.6% 0.2% 0.2% Farm labourer 4.7% 0.6% 34.4% 2.1% 12.1% 0.9% Owner farmer farm 1.2% 0.2% 9.8% 9.8% 1.2% 1.2% 3.3% 3.3% 0.4% 0.4% Self-employed farm labourer 0.9% 0.2% 24.6% 1.8% 6.9% 0.7% Not working 13.5% 0.7% 7.2% 7.2% 0.9% 0.9% 11.9% 11.9% 0.6% 0.6% Employee 38.7% 1.1% 11.7% 1.3% 32.0% 1.0% 119 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 119 Table A4. Means/Percentanges and Standard Errors, Oriente Region Urban Urban Rural Oriente Region riente Std. Std. Std. Std. Mean / Mean / Mean / Std. Err. Std. Err. of of Err. of of Percentag Percentag Percentag of Mean / Mean / Mean / e e e Percent Percent Variable Variable Percent Percent Percent Demographics Number of babies ber babies 0.3 0.0 0.5 0.0 0.4 0.0 Number of babies squared 0.3 0.0 0.7 0.1 0.5 0.1 Number of children 0.9 0.1 1.5 0.1 1.3 0.1 Number of children squared 1.8 0.2 4.7 0.5 3.6 0.3 Number of teenagers 0.6 0.1 0.8 0.1 0.7 0.1 Number of teenagers squared 1.0 0.1 1.8 0.2 1.5 0.2 Number of adults 2.1 0.1 2.4 0.1 2.3 0.1 Number of adults squared 4.7 0.3 5.9 0.4 5.4 0.3 Female household head 18.7% 2.1% 12.6% 2.0% 15.0% 1.5% Male household head 81.3% 2.1% 87.4% 87.4% 2.0% 2.0% 85.0% 85.0% 1.5% 1.5% No spouse, single 9.9% 1.7% 3.7% 0.9% 6.2% 0.9% No spouse, separate/divorced/widow /di d/ id 16.5% 2.0% 13.1% 1.8% 14.4% 1.4% No spouse, married 1.4% 0.5% 1.7% 0.7% 1.5% 0.5% Spouse, married 72.2% 2.5% 81.5% 81.5% 2.3% 2.3% 77.9% 77.9% 1.8% 1.8% Age of household head g 41.6 1.0 44.0 0.9 43.1 0.7 Age of household head squared 1933.6 99.3 2157.1 91.6 2069.5 69.1 Mixed race (mestizo) household head 77.0% 2.1% 50.3% 6.8% 60.8% 4.6% White household head hite 8.7% 1.8% 6.6% 6.6% 1.5% 1.5% 7.4% 7.4% 1.2% 1.2% Black household head 2.2% 0.7% 0.2% 0.2% 1.0% 0.3% Mixed race (mulato) household head household 2.6% 2.6% 1.6% 1.6% 0.8% 0.8% 0.5% 0.5% 1.5% 1.5% 0.7% 0.7% Other ethnicity household head head 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.00% 0.00% 0.00% 0.00% Indigenous 9.5% 1.7% 42.0% 7.4% 29.3% 5.2% Education of Household Head 1 - 4 years 6.4% 1.2% 14.2% 14.2% 2.1% 2.1% 11.1% 11.1% 1.4% 1.4% 5 - 7 years 30.3% 2.4% 49.0% 3.4% 41.7% 2.6% 8 - 10 years 15.4% 1.4% 12.1% 12.1% 1.7% 1.7% 13.4% 13.4% 1.2% 1.2% 11 - 13 yyears 23.6% 1.7% 9.5% 1.7% 15.0% 1.5% 14 - + years 23.2% 3.0% 5.4% 5.4% 1.3% 1.3% 12.4% 12.4% 1.7% 1.7% None 1.0% 0.4% 9.9% 1.6% 6.4% 1.1% 120 120 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Education of Spouse 1 - 4 years 4.6% 0.9% 15.1% 15.1% 2.0% 2.0% 11.0% 11.0% 1.5% 1.5% 5 - 7 yyears 20.7% 1.9% 35.9% 2.6% 29.9% 2.0% 8 - 10 years 12.9% 2.0% 10.1% 10.1% 1.7% 1.7% 11.2% 11.2% 1.3% 1.3% 11 - 13 years 20.9% 1.9% 8.2% 1.5% 13.2% 1.4% 14 - + years 12.1% 1.8% 3.5% 3.5% 1.0% 1.0% 6.9% 6.9% 1.0% 1.0% None / No spouse 28.9% 2.6% 27.1% 2.2% 27.8% 1.7% House Ownership p Rent 33.3% 33.3% 3.4% 3.4% 5.3% 5.3% 1.8% 1.8% 16.2% 16.2% 2.3% 2.3% Own and paying 2.4% 0.9% 0.8% 0.4% 1.4% 0.4% Other 14.8% 1.9% 14.0% 14.0% 2.3% 2.3% 14.3% 14.3% 1.6% 1.6% Own and paid 49.5% 3.3% 80.0% 3.1% 68.0% 2.9% Position in Occupation Employer 9.4% 1.8% 2.3% 2.3% 0.9% 0.9% 5.1% 5.1% 0.9% 0.9% Self-employed 18.0% 1.5% 6.9% 1.6% 11.2% 1.3% Employee without pay 1.2% 0.7% 1.9% 1.9% 0.6% 0.6% 1.7% 1.7% 0.5% 0.5% Farm labourer 4.0% 0.9% 10.8% 2.1% 8.2% 1.3% Owner farmer farmer 0.7% 0.4% 2.0% 2.0% 0.6% 0.6% 1.5% 1.5% 0.4% 0.4% Self-employed farm labourer 5.3% 1.3% 50.8% 5.4% 33.0% 4.5% Not working 6.8% 1.3% 3.8% 3.8% 1.0% 1.0% 5.0% 5.0% 0.8% 0.8% Employee 54.5% 2.7% 21.4% 3.4% 34.4% 3.2% 121 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 121 Table A5. Regressions for Log Welfare Ratio, Sierra Region Urban Sierra Urban Sierra Rural Sierra Sierra Standard Standard Standard Standard Explanatory variables Coefficient Error Coefficient Coefficient Error Constant Constant -0.543 * 0.150 -0.453 * 0.130 Demographics Number of babies -0.290 * 0.072 -0.369 * 0.047 Number of babies squared 0.013 0.050 0.060 * 0.025 Number of children -0.286 * 0.024 -0.267 * 0.024 Number of children squared q 0.033 * 0.007 0.026 * 0.006 Number of teenagers teenagers -0.223 -0.223 * 0.041 -0.126 * 0.033 Number of teenagers squared 0.019 0.020 0.008 0.012 Number of adults Number adults -0.138 * 0.026 -0.058 * 0.026 Number of adults squared 0.006 0.004 0.008 * 0.004 Sex of the head Sex -0.065 0.044 -0.068 0.051 No spouse, single p g 0.144 0.091 0.092 0.064 No spouse, separate/divorced/widow spouse, separate/divorced/widow 0.211 * 0.087 0.251 * 0.057 No spouse, married 0.333 * 0.098 0.479 * 0.082 Age of household head 0.030 * 0.030 0.005 0.020 * 0.004 Age of household head squared 0.000 * 0.000 0.000 * 0.000 Mixed race (mestizo) household head mestizo 0.103 * 0.049 0.132 * 0.042 White household head 0.212 * 0.069 0.212 * 0.071 Black household head Black 0.102 0.103 0.026 0.141 Mixed race (mulato) household head 0.183 * 0.078 0.276 ** 0.143 Other ethnicity household head Education of Household Head 1 - 4 years 0.423 * 0.090 0.180 * 0.035 5 - 7 yyears 0.705 * 0.086 0.352 * 0.041 8 - 10 years 0.843 * 0.084 0.499 * 0.056 11 - 13 years y 1.064 * 0.094 0.701 * 0.074 14 - + years 1.479 * 0.100 1.233 * 0.141 Education of Spouse p 1 - 4 years -0.014 0.081 0.029 0.038 5 - 7 years 0.060 0.082 0.096 * 0.043 8 - 10 years 0.078 0.085 0.124 * 0.059 11 - 13 years y 0.167 * 0.084 0.264 * 0.064 14 - + years 0.364 * 0.089 0.408 * 0.088 House Ownership p Rent -0.203 * 0.023 0.050 0.078 Own and paying p y g -0.006 0.051 0.395 * 0.125 Other Other -0.214 * 0.030 -0.081 * 0.033 Position in Occupation Employer 0.259 * 0.039 0.330 * 0.063 Self-employed -0.010 0.026 -0.011 0.041 Employee without pay 0.075 0.063 -0.207 * 0.071 Farm labourer -0.148 ** 0.083 -0.258 * 0.043 Owner farmer 0.242 * 0.090 0.281 * 0.052 Self-employed farm labourer p y -0.227 * 0.081 -0.203 * 0.039 Not working -0.032 0.043 -0.255 * 0.058 Source: Own calculations based on ECV Survey. Note: Number of observations: 3780 (urban) and 3436 (rural). R2=0.59 (urban) and 0.48 (rural), * indicates that the coefficient is significant at 5 percent level, and ** at 10 percent level. The base categories are: male household head, married with spouse, indigenous household head, no education of household head, no education of spouse or no spouse, own house and employee. 122 122 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table A6. Regressions for Log Welfare Ratio, Costa Region Urban Costa Urban Costa Rural Costa Costa Standard Standard Standard Standard Explanatory variables Coefficient Error Coefficient Coefficient Error Error Constant Constant -0.308 * 0.133 -0.233 0.151 Demographics Number of babies -0.275 * 0.042 -0.348 * 0.049 Number of babies squared 0.028 0.023 0.069 * 0.025 Number of children -0.271 * 0.017 -0.278 * 0.020 Number of children squared q 0.022 * 0.004 0.030 * 0.005 Number of teenagers teenagers -0.198 * 0.025 -0.133 * 0.034 Number of teenagers squared 0.010 0.010 0.011 0.010 Number of adults -0.089 * 0.019 -0.079 * 0.028 Number of adults squared 0.004 0.003 0.002 0.004 Sex of the head Sex 0.024 0.036 -0.022 0.062 No spouse, single p g 0.297 * 0.063 0.150 ** 0.083 No spouse, separate/divorced/widow spouse, separate/divorced/widow 0.201 * 0.055 0.171 * 0.054 No spouse, married 0.550 * 0.091 0.647 * 0.183 Age of household head 0.027 * 0.027 0.004 0.020 * 0.004 Age of household head squared 0.000 * 0.000 0.000 * 0.000 Mixed race (mestizo) household head mestizo -0.153 * 0.074 0.033 0.134 White household head -0.077 0.083 -0.027 0.142 Black household head household head -0.188 * 0.080 0.021 0.143 Mixed race (mulato) household head -0.199 * 0.084 0.014 0.146 Other ethnicity household head -0.212 0.377 Education of Household Head 1 - 4 years 0.180 * 0.049 0.162 * 0.047 5 - 7 years y 0.330 * 0.047 0.218 * 0.043 8 - 10 years 0.447 * 0.052 0.316 * 0.055 11 - 13 years y 0.629 * 0.055 0.354 * 0.065 14 - + years 0.939 * 0.061 0.650 * 0.101 Education of Spouse p 1 - 4 years 0.002 0.046 0.063 0.053 5 - 7 years 0.140 * 0.047 0.100 * 0.046 8 - 10 years 0.188 * 0.049 0.190 * 0.054 11 - 13 years y 0.309 * 0.049 0.188 * 0.082 14 - + years 0.504 * 0.055 0.596 * 0.144 House Ownership p Rent -0.044 0.028 0.025 0.102 Own and paying p y g -0.031 0.055 -0.251 0.281 Other Other -0.097 * 0.027 0.011 0.039 Position in Occupation Employer 0.345 * 0.038 0.250 * 0.113 Self-employed -0.081 * 0.024 -0.123 * 0.059 Employee without pay 0.020 0.071 -0.078 0.113 Farm labourer -0.178 * 0.032 -0.077 0.049 Owner farmer Owner 0.178 * 0.074 0.180 * 0.062 Self-employed farm labourer p y -0.206 * 0.097 -0.156 * 0.051 Not working -0.071 * 0.036 -0.326 * 0.070 Source: Own calculations based on ECV Survey. Note: Number of observations: 3697 (urban) and 1520 (rural). R2=0.55 (urban) and 0.50 (rural), * indicates that the coefficient is significant at 5 percent level, and ** at 10 percent level. The base cathegories are: male household head, married with spouse, indigenous household head, no education of household head, no education of spouse or no spouse, own house and employee. 123 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 123 Table A7. Regressions for Log Welfare Ratio, Oriente Region Urban Oriente Urban riente Rural Oriente Standard Standard Standard Standard Explanatory variables Coefficient Error Coefficient Coefficient Error Constant Constant -0.390 0.388 -1.111 * 0.311 Demographics Number of babies -0.540 * 0.159 -0.413 * 0.091 Number of babies squared 0.221 ** 0.128 0.123 * 0.048 Number of children -0.377 * 0.058 -0.289 * 0.042 Number of children squared 0.038 * 0.019 0.020 * 0.008 Number of teenagers -0.225 -0.225 * 0.081 -0.193 * 0.066 Number of teenagers squared 0.044 0.034 0.016 0.019 Number Number of adults adults -0.038 0.050 -0.118 * 0.048 Number of adults squared -0.005 0.007 0.014 * 0.006 Sex of the head Sex -0.124 0.089 -0.089 0.096 No spouse, single p g 0.931 * 0.307 0.164 0.172 No spouse, separate/divorced/widow 0.820 * 0.293 0.205 0.151 No spouse, married 1.161 * 0.393 0.741 * 0.227 Age of household head 0.020 **** 0.011 0.046 * 0.011 Age of household head squared 0.000 0.000 0.000 * 0.000 Mixed race (mestizo) household head mestizo 0.151 ** 0.084 0.569 * 0.072 White household head 0.127 0.113 0.548 * 0.121 Black Black household head 0.053 0.185 0.646 * 0.146 Mixed race (mulato) household head -0.091 0.127 0.571 * 0.140 Other ethnicity household head Education of Household Head 1 - 4 years 0.037 0.269 0.088 0.109 5 - 7 years y 0.108 0.257 0.215 * 0.098 8 - 10 years 0.263 0.258 0.318 * 0.097 11 - 13 years y 0.273 0.265 0.509 * 0.124 14 - + years 0.555 * 0.258 0.307 * 0.144 Education of Spousep 1 - 4 years 0.427 ** 0.249 0.074 0.148 5 - 7 years y 0.651 * 0.305 0.049 0.123 8 - 10 years 0.701 * 0.279 0.268 * 0.126 11 - 13 years y 0.859 * 0.288 0.244 ** 0.132 14 - + years 1.005 * 0.310 0.660 * 0.181 House Ownership p Rent -0.138 * 0.054 -0.069 0.094 Own and paying p y g 0.169 0.180 0.000 0.111 Other Other -0.079 0.082 -0.053 0.083 Position in Occupation p Employer 0.245 * 0.069 0.147 0.191 Self-employed p y -0.126 ** 0.069 -0.024 0.081 Employee without pay -0.011 0.243 -0.464 * 0.170 Farm labourer -0.367 * 0.087 -0.372 * 0.091 Owner farmer 0.412 0.424 0.190 ** 0.098 Self-employed farm labourer p y -0.421 * 0.134 -0.435 * 0.063 Not working -0.059 0.104 -0.412 * 0.117 Source: Own calculations based on ECV Survey. Note: Number of observations: 473 (urban) and 525 (rural). R2=0.58 (urban) and 0.70 (rural), * indicates that the coefficient is significant at 5 percent level, and ** at 10 percent level. The base cathegories are: male household head, married with spouse, indigenous household head, no education of household head, no education of spouse or no spouse, own house and employee. 124 124 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES SPATIAL DISPARITIES IN LIVING CONDITIONS IN PERU: THE ROLE OF GEOGRAPHIC DIFFERENCES IN RETURNS VS. DIFFERENCES IN MOBILE HOUSEHOLD ASSET ENDOWMENT Carmen Ponce Consultant The World Bank cponce@grade.org.pe Javier Escobal Consultant The World Bank jescobal@grade.org.pe Abstract W elfare levels in Peru vary enormously across the country's territory. While the urban coastal areas enjoy relatively high levels of welfare, rural areas in the mountains and the jungle areas have extremely high poverty rates. Using a decomposition methodology similar to that proposed by Oaxaca and Blinder, this paper analyzes which of two factors--endowments of characteristics and their returns--is mostly responsible for the observed spatial differences in welfare levels. The paper finds that different endowments of productive characteristics are the main factor behind urban-rural welfare differences. However, short-term changes in poverty rates appear to be mainly due to short-term changes in the returns to those characteristics. Using quantile analysis, the paper also shows that the welfare gap between urban and rural areas increases as one moves along the income distribution and that this is chiefly due to different endowments of productive characteristics. Background paper for the regional study on poverty in Latin America at the World Bank. These are views of the authors and do not necessarily reflect those of the World Bank, its executive directors, or the countries they represent. 125 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 125 1. Introduction Socio-economic dynamics in Peru have historically shown differentiated spatial patterns between urban and rural areas and between the coastal (Costa), highland (Sierra) and Amazon (Selva) regions, with the rural sections of the Sierra and Selva regions significantly worse off. In 2006, rural poverty in these regions was 77 percent and 62 percent, respectively, contrasting with urban Costa's poverty rates of less than 30 percent. Extreme poverty continues to be widely spread in rural Sierra, affecting almost half of its population, whereas less than 5 percent of urban Costa dwellers live in such precarious condition. During the last decade the Peruvian economy grew at an annual rate of almost 4 percent, accelerating to 5.7 percent per annum in from 2002 to 2007. Urban poverty decreased by six percentage points between 2004 and 2006, consistent with this economic growth. However, rural poverty rates continue to be high and very unresponsive to growth. The question arises as to whether structural differences in living conditions, as opposed to differences in mobile individual characteristics, explain a significant amount of these spatial disparities. In other words, are people in poor regions poor because they lack important characteristics that would enable them to improve their lot in life (such as a good education, for example), or are they poor because the returns to those characteristics are lower where they live (better education does not translate into a better wage, for example). This paper explores regional patterns of poverty in Peru and attempts to identify the role that geographic differences in returns to household characteristics play in such differentiated spatial patterns, taking into account differences in households´ endowment of mobile characteristics. The paper is structured as follows. After first reviewing the data, a methodological framework similar to that proposed by Ravallion and Wodon (1999) is used to decompose the contribution of geography to welfare, controlling for mobile individual characteristics. Next, access to infrastructure is added to the equation in order to evaluate changes in the marginal contribution of geography when controlling for these non-mobile characteristics. Finally, we use quantile regressions, following the methodology outlined by Nguyen et al. (2007), to evaluate the robustness of the decomposition when allowing for heterogeneity in the rate of return to endowments for individuals with different income levels. 2. A summary of differences in household characteristics The data used for this paper come from the 2002 and 2006 Peruvian National Household Survey (Encuesta Nacional de Hogares--ENAHO), which was designed to achieve representativeness for Costa, Sierra and Selva regions, within both urban and rural strata. Differences in wealth and consumption between Peruvian urban and rural areas are striking (Table 1). Perhaps the most eloquent figure is the extreme poverty rate, which captures the capability of a household to afford a basket of food items that cover basic nutritional needs. While rural extreme poverty rises to 37 percent, the urban figure is only 5 percent. These spatial differences exist in a culturally and geographically diverse country. While 70 percent of the Peruvian urban population lives in the Costa, the most linguistically homogeneous region, over 60 percent of the rural population lives in the Sierra, the most linguistically heterogeneous region.34 Ample evidence shows that urban household's education and physical assets are greater than those of rural households (Table 2). Among the assets analyzed here, education clearly shows the greatest 34Six out of every 10 households in rural Sierra have a native ethnic background, defined as households where both head and spouse have a native mother tongue. In rural Selva this proportion is almost 40 percent. 126 126 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES differences. Although education coverage has increased in the last decades and primary school attendance by children between six and 11 years was 93 percent by 2006, differences in education among the heads of households in Peru is still significant between rural and urban areas. Almost half of the rural population lives in households headed by a person who did not finish primary school, compared to only 18 percent of urban households. At the same time, 64 percent of rural population belongs to households with the head's spouse having either no or incomplete primary schooling, compared to 24 percent in urban areas. Household size is not significantly different across areas, but rural households tend to be younger. This finding is striking since typically census and living standards survey data reported for Peru a decade or more ago showed that rural families were, on average, larger than urban families. A small but statistically significant reduction in birth rates that occurred in the past decade can explain part of the result. However, it is more likely that migration patterns, where a member of the family goes to another region (or another country) and send back remittances to their family, is behind this structural change. It is also noteworthy that one out of five household heads in urban areas are not working, whereas this rate decreases to only 3 percent in rural areas.35 35A person is considered not to be working if he or she does not work or works less than 15 hours per week. 127 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 127 128 128 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 129 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 129 3. The basic decomposition exercise: controlling for geographic and mobile individual characteristics The methodology here follows the framework developed by Ravallion and Wodon (1999), which uses the Oaxaca-Blinder method to decompose urban-rural differences in welfare in two factors: the part due to the endowment of characteristics explaining welfare, and the part due to the returns to those characteristics. Modeling living standards in Peru This section discusses the national-level results from the model using regression analysis to test the relationship between different household characteristics and welfare. The welfare measure used for the analysis is the logarithm of the welfare ratio--household per capita expenditure spatially deflated by the corresponding poverty line (see Annex Table A.1). This captures how far each household consumption level is from the level required to be considered non-poor. The model assumes a linear relation between the welfare ratio, log Ci, and mobile household characteristics, X, with region-specific parameters, Di (taking different values for Costa, Sierra and Selva regions).36 log Ci = U + U ´ Xi + ´ Di + U Ui (i U) (1) log Ci = R + R ´ Xi + ´ Di + R Ri (i R) (2) As far as U and R capture the returns to private assets, the role of U and R is to capture fixed regional characteristics as well as the differential access to public infrastructure. Table 3 shows the point estimates and their statistical significance. The standard errors were adjusted by the survey sample design in order to allow for heteroscedasticity and adjust for the additional uncertainty derived from a multistage stratified sampling (as opposed to random sampling).37 When specifying the dummy variables, the reference household was set as one headed by a married woman with a native mother tongue and no formal education, with no higher education from other household members, who works in an elementary (unskilled) job, with no motorized transportation assets and living in the Sierra region in a house that is not owned by the family. 36We assume that omitted household characteristics are not correlated with place of residence (so they do not bias the estimates of geographic effects). 37Regression estimates are available upon request. 130 130 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 131 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 131 Even after controlling for household characteristics, living in the Sierra is negatively associated with household welfare among rural households. In urban areas, living in the Costa is positively associated with welfare, and no differentiation was seen between Sierra and Selva. The coefficients on household characteristics are quite intuitive. Larger households tend to be poorer. Gender composition is important in rural households but unimportant in urban ones. Household head's gender is unimportant, controlling for other covariates, in either urban or rural households. As expected, more education is associated with higher welfare outcomes. Coefficients in household head's main occupation are positive, showing a negative effect of elementary (unskilled) jobs on household welfare. This is true in general, with one exception in rural areas: skilled agricultural and fishery. Even when at least one of the family income sources is agriculture, self-employment for most rural households is not the only source of labor income. This may arise because of two different situations. One possibility is that for those better endowed with complementary assets, other sources such as crafts and related trade jobs may be more profitable. Alternatively, this may come about just because land restrictions make it difficult to employ all family members in farming, in which case employment diversification arises as a way of coping with such vulnerability. House ownership is unimportant within rural areas, whereas it is significant for urban welfare outcomes. As for other private assets, motorized transportation assets are strongly associated with welfare ratios in both areas. The size of the correlation between household characteristics and welfare outcomes is different between rural and urban regressions for all groups of variables, at least at 1 percent of confidence level (Table 4). Only at the 5 percent confidence level can we not reject the null hypothesis of parameter equality across urban and rural areas for just one asset group (variables related to house ownership). Thus it is safe to say that rate of returns on assets between urban and rural areas are indeed significantly different. 132 132 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Evaluating geographic disparities in living standards in Peru In order to identify the set of variables explaining most of the difference between urban and rural per capita expenditure, differences are decomposed as follows: E [ logCi| i U, Xi = XU ] - E [ logCi| i R, Xi = XR ] = ( - U R ) + ( ´ XU - U R´ XR ) + k( sUk Uk­ sRk Rk ) (3) Table 5a shows the decomposition derived from the regression in Table 3. First, note the significant difference between urban and rural average welfare ratio. The characteristics that contribute the most to this urban/rural welfare difference are education, home ownership and household head's main occupation. Before proceeding to disentangle the geographic and non-geographic components of these differences in the following section, it is worth showing that the results obtained in Table 5a are robust even if the specification is modified. For example, the ethnicity proxy variable was not included here. However, given the importance of ethnic diversity in both urban and rural Peru, it is worth mentioning the result if this variable is included. The regression estimates including ethnicity show that this variable is significant in both urban and rural areas. Households with both head and spouse having a native mother tongue tend to have lower welfare outcomes than households with the head or the spouse having Spanish as his or her mother tongue. Including ethnicity does not alter the decomposition results showed in Table 5a. The sets of variables most influencing the urban-rural difference in living standards are again education, occupation and ownership (Table 5b). This of course does not mean that ethnicity does not matter when explaining living standards differences among Peruvian households. Indeed, the relative importance of demographic variables within both urban and rural areas increases when ethnicity is included (compare Tables 5a and 5b). However, this increased contribution takes place in the same proportion in both areas, so the urban- rural difference is not affected. 133 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 133 The regression in Table 3 does not include age of the head of household squared. The collinearity problem induced by this variable does have consequences in the decomposition exercise, inflating the contribution of the set "other demographic variables" and so the constant term as well (Table 5c). In consequence, the variable was excluded from the regression and following decompositions. Simulated welfare ratios and poverty measures Although informative, the decomposition in the previous section does not allow disentangling the structural component of the urban-rural difference in living standards. Again, following Ravallion and Wodon (1999) we perform a Oaxaca-Blinder decomposition which reveals the structural component of the difference. 134 134 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES First, we isolate the structural component by controlling for mobile household characteristics: log GEOiU = U + U ´ X* + ´ Dj U (4a) log GEOiR = R+ R ´ X* + ´ Dj R (4b) where X* represents the national mean characteristics and Dj is a geographic vector with zeroes in all rows except row j (for the jth region). Second, we isolate the effect of the non-geographic household characteristics by controlling for geographic differences: log CONjU = N+ N´ XjU (5a) log CONjR = N+ N´ XjR (5b) where XjU and XjR represent the sample mean characteristics of households living in urban and rural areas of region j, respectively, and the parameters are computed as population weighted means: N = sU ( U + k Uk Uk s ) + sR (R + k sRk Rk) (6a) N = sU U + sR R (6b) in which sU and sR denote the urban and rural population shares and sUk and sRk are region k´s share of the urban and rural populations, respectively. Table 6 shows the simulation results in terms of welfare ratios and Table 7 shows the results in terms of poverty rates. Notice that the geographic (returns) component is unimportant for explaining differences in living standards in rural Selva, where only the concentration effect matters. In the other regions, differences in both returns and endowments matter. As expected, rural geographic profiles are higher than the corresponding unconditional profiles. The opposite is true for urban areas of the three regions. This is consistent with the significant difference in household endowments between urban and rural households. While setting the national levels for urban households will reduce the welfare outcomes (given that they typically have better endowment than the national average), setting such levels for rural households simulation will show larger welfare outcomes than the actual ones. For the concentration profile, results are mixed. First, urban Costa, rural Costa and rural Selva show lower concentration profiles compared to unconditional profiles, indicating that returns to household characteristics in those areas are higher than national average returns. This result is fully explained by the significant location effect for these three areas showed in Table 3. On the other hand, both the urban and rural Sierra and the urban Selva show lower returns than the national average. 135 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 135 Another way of looking at the decomposition is to calculate what the poverty rates would be if private asset distribution were more equal or if the return on assets was the same across regions. Results of this exercise show clearly that the Sierra, where poverty rates are the highest, would benefit the most if the rate of returns to assets were raised to the national averages (Table 7). Similar results are shown in the rural Selva. That is to say, there may be important bottlenecks (inequality in infrastructure provision being a critical one) at the root of spatial inequality differences within Peru. Replicating the decomposition at two points in time In order to assess the robustness of the decomposition, we have reproduced the same exercise for 2002. To construct the welfare ratios, per capita expenditures of 2002 were deflated by 2002 poverty lines. If one compares the unconditional profile to the geographic and concentration profile, most of the differences in poverty can be attributed to differences in asset endowments (Table 8). As such, the decomposition exercise is robust.38 Comparing Table 8 to Table 7, although different asset endowments are the main driver of differences in poverty rates in both years, short-term changes in poverty rates may be driven by short-term differences in rate of returns (see Annex Figures A.2 and A.4). This result is crucial because it highlights the importance of a different policy mix when we consider different time-frames. When looking at structural differences across space, the importance of endowments in shaping wellbeing is obvious. At the same time we must recognize that in the short run, policies will tend to affect more the rate of return to assets than the access to them. 38Tables A.2, A3 and A.4 in the annex show the test of equality of coefficients, the contributing factors to urban rural disparities and the simulated welfare ratios for 2002. 136 136 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 4. The effect of controlling for access to infrastructure services So far the decomposition exercise has considered household-specific variables (size and other household demographics, human capital, occupation and private asset variables) as non-geographic. The role of geography may be biased upward due to not controlling for non-mobile characteristics like access to infrastructure services. To evaluate the role of infrastructure, the log welfare ratio equations are estimated controlling for access to three infrastructure services: water, electricity and roads. Drainage and sanitation facilities are omitted since they appear highly correlated in the sample. The time spent traveling from the district capital to the closest city of at least 75,000 inhabitants and the additional time required to get to a city of at least 100,000 inhabitants is used as a proxy for improved roads. The idea around this variable is to capture the degree of connection to regional markets. As with the previous estimation that excludes infrastructure-related variables, the coefficients on household characteristics are quite intuitive. Larger households tend to be poorer and education increases the likelihood of been non-poor. Again, gender composition is important to household welfare in rural households but unimportant in urban ones. All infrastructure service variables are highly significant in the urban areas, whereas access to water and access to improved roads are significant factors in rural areas. Electricity does not seem to be correlated to improved wellbeing in rural areas. However, these are mere profiles and as such they may be capturing, in the best case scenario of no endogeneity, reduced form effects. Because of this, care should be taken interpreting the sign and significance of the estimated parameters in these equations. Lastly, the regression results indicate that even after controlling for household characteristics and access to infrastructure services, living in the Sierra is significantly negatively associated with household welfare among rural households. In urban areas, living in the Costa is positively correlated, while the association of living in the Sierra and Selva cannot be differentiated. Table 10 evaluates the significance of the coefficients presented in Table 9 grouped by type of endowment. In all cases the parameters are significant at the 1 percent confidence level. Only house ownership variables look marginally significant at the 1 percent confidence level. Table 11 uses the estimated parameters to decompose the estimated log welfare ratio between geographic and non-geographic explanatory variables. The introduction of access to infrastructure induced dramatic changes in the constant term. Since the constant is a proxy for mean of omitted variables, this was an expected result. The contribution of the regional dummies to the urban-rural log welfare ratio gap diminishes once access to infrastructure services is controlled. Lastly, although the impact of non- geographic household related variables are similar when controlling for infrastructure services, there seems to be a slight reduction in the contribution of education in explaining the urban-rural wellbeing gap. 137 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 137 138 138 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES With the decomposed mean log welfare ratio it is again possible to follow Ravallion and Wodon (1999) to construct counterfactuals of interest, such as geographic and concentration profiles (Table 12). The geographic profile indicates what the log welfare ratio would be if all regions in both urban and rural areas shared the same endowments (now including public infrastructure endowments). The concentration profile shows what the log welfare ratio would be if all regions in both urban and rural areas shared the same rate of returns to endowments observed at the national level. As expected, the rural geographic profiles show higher levels of welfare than the corresponding unconditional profiles, and the opposite is true for urban sections of Costa and Sierra, reflecting the significant difference in household endowments between urban and rural households. In spite of these similarities, urban Selva results differ from those observed in Table 6. This difference highlights the lower access to infrastructure that urban Selva has compared to other urban areas of the country. This result is also highlighted in Table 13, where poverty rates in the geographic profile of urban Selva show lower figures than the unconditional profile (similar to the results in rural areas), which means that if urban Selva was given national-average endowments it would be better off. It is also worth noting that compared to Table 6, the geographic profile for urban Costa shows much higher poverty (lower welfare ratio) than its unconditional profile. 139 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 139 These results confirm the intuition that including infrastructure in the analysis improves the understanding of the regional differences derived from differences in endowments, widening the gap between urban Costa and the rest of the country, and placing urban Selva in a less advantaged condition than if those key assets are not considered. Compared to the previous exercise (see Table 6), including infrastructure yields larger differences between the concentration and unconditional profiles for rural Costa and rural Selva, whereas differences in urban Selva and rural Sierra almost disappear. This means that including infrastructure tends to equalize returns in urban Selva and rural Sierra with the national average, while it widens the differences in favor of rural Selva and rural Costa. Another way of evaluating the effect of equalizing endowments or rate of returns across regions is to simulate what the poverty rates would be if this were achieved. Table 13 shows these results and Figure 1 portrays the same results compared to Table 7, which does not control for access to infrastructure services. The comparison shows clearly that the omission of the variables associated to infrastructure services downplays somewhat the role of equalizing endowments in rural areas. For all regions, the simulated poverty rates will be 2 percentage points lower if we equalize endowments, including access to infrastructure services. Differences in the concentration profiles between the two estimations show that omitting access to infrastructure services underestimates the importance of differences in rates of return, as poverty rates will be even larger than originally estimated if we control for these services. The difference in urban and rural poverty rates for the Costa region using the geographic profile is larger if we control for access to infrastructure services. This indicates that rates of return to endowments are larger in urban than in rural Costa once we control for access to infrastructure services. 140 140 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 1 In conclusion, although the results controlling for infrastructure services are in general consistent with the previous estimations, they reflect the role of these services in enhancing the rate of returns to assets. Moreover, including access to key infrastructure assets gives a more precise picture of the wellbeing gaps across regions. 141 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 141 5. Quantile decomposition So far the analysis has allowed rates of returns to endowments to differ between urban and rural regions, by introducing separate models for each area. However the rates of return are expected to vary not only between urban and rural areas but also within areas. In particular, poorer households are likely to have different rates of return from more wealthy households. To allow for this fact, the decomposition exercise was performed allowing for parameter heterogeneity. The procedure is an adapted version of the methodology proposed in Nguyen et al (2007): Qr ( Yi / X, u) = r+ X r (7) where Qr ( Yi / X, u) represents the th conditional quantile of the log welfare ratio, Yi. This equation was estimated only for rural areas, in order to account for differences in parameters across the corresponding income distribution. X represents all conditioning variables (both geographic and non-geographic). Estimating (7) allows for the rate of return to assets to change across quantiles in both the urban and rural sectors. Based on the estimation of (7), we simulate the counterfactual distribution, which estimates what the log welfare ratio would be in rural areas across quantiles if rural households had their urban counterparts´ asset endowments. Based on those estimations, we decompose the urban-rural log welfare ratio gap, as follows: Yu ( ) ­ Yr ( ) = { Yu ( ) ­ Y* ( ) } + { Y* ( ) ­ Y* ( ) } (8) where Yi ( ) stands for the observed log welfare ratio corresponding to the th quantile of the urban distribution, i= urban, rural; and Y* ( ) stands for the log welfare ratio corresponding to the th quantile of the counterfactual distribution. The first term in the right-hand side of (8) is the returns effect, and it measures the contribution to the urban-rural gap at th quantile of the differences in returns between urban and rural areas. The second term is the covariate effect, and measures the contribution to the urban-rural gap at th quantile of differences in endowments between these two areas. Results excluding access to infrastructure The decomposition allowing for quantile variation was done using ENAHO 2002 and then with ENAHO 2006. Both cases start with the base case scenario which ignores infrastructure services and controls only for non-geographic mobile endowments. As expected, the distribution of the log welfare ratios for 2002 in rural areas is shifted to the left, indicating a wellbeing gap in favor of urban areas (Figure 2). The urban-rural wellbeing gap for all quantiles of the income distribution shows that, again as expected, the urban­rural gap is increasingly larger, which means that urban richer households are better off than their rural counterparts to a greater extent than the urban poor when compared to the rural poor (Figure 3). 142 142 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 2 Figure 3 Using the national urban and rural quantile estimates we can construct a counterfactual distribution indicating the rural log welfare ratio if rural dwellers had the same endowment base as their urban counterparts. The comparison of this counterfactual distribution with the observed urban and rural log welfare ratios shows clearly that the urban-rural gap can be eliminated completely if asset endowments were equal across the urban-rural divide for the diferent segments of the quantile distribution (Figure 4). Figure 4 143 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 143 Figure 5 uses these estimates to plot the decomposition of the urban-rural log welfare ratio gap between the endowment effect (or covariate effect, as it is called in the figure) and the return effect referred in (8). Here the return effect is indeed small and may only be different from zero around the 70th to 90th quantile. By contrast, the endowment effect is driving the increasing urban-rural log welfare ratio gap. Figure 5 To evaluate how this pattern may differ across regions, the urban/rural log welfare ratio gap is decomposed for the Costa, Sierra and Selva regions (Figure 6). The pattern observed for the Costa is similar to the national average, with the counterfactual distribution slightly above the urban log welfare ratio (Panel a). The Sierra shows a similar pattern, where endowments differences are by far the most important factor explaining the increasing urban/rural log welfare ratio gap throughout the quantile distribution (Panel b). In this case, however, the counterfactual distribution lays a bit above the urban log welfare ratio curve, indicating that rural returns are slightly higher than urban returns in the Sierra for most quantiles, except for those of the very poor and the very rich. In the Selva, the counterfactual distribution lies well above the log welfare ratio curve of the urban area, indicating clearly that returns to endowments are larger for rural Selva as compared to urban Selva (Panel c). This is especially clear for the most wealthy residents of the Selva. 144 144 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 6: Regional Decomposition of the Urban/Rural Gap for 2002 (a) Costa (b) Sierra 145 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 145 (C) Selva 146 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Results for 2006 The results for 2006 indicate that the urban/rural log welfare ratio gap has declined marginally since 2002 (Figure 7). Figure 7: Urban/Rural Log Welfare Ratio Gap 2002- 2006 Figure 8 presents the log welfare ratio curves across quintiles for both urban and rural areas for 2006, together with the counterfactual distribution, in which the rural sector is endowed with the same assets as their urban counterpart sharing the same quantile position. The results look very similar to the 2002 results, with the counterfactual distribution just above the urban log welfare ratio curve. Figure 8 Comparing the 2002 and 2006 decompositions indicates that the endowment or covariate effect has reduced its contribution to the urban/rural gap in the upper tail of the distribution, while the return effect slightly increased its contribution in the second half of the distribution (Figure 9). 147 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 147 Figure 9: Urban/Rural Gap Decomposition 2002-2006 (a) 2002 (b) 2006 The same decomposition exercise for 2006 for the three regions of Costa, Sierra and Selva gives similar results to those for 2002 (Figure 10). 148 148 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure 10: Regional Decomposition of the Urban/Rural Gap for 2006 (a) Costa (b) Sierra 149 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 149 (c) Selva 150 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Controlling for access to infastructure in the quantile decomposition exercise To evaluate the robustness of these results controlling for non-mobile endowments, in particular access to infrastructure, the quantile decomposition exercise was reworked to control for access to water, electricity and roads. The urban and rural log welfare ratio curves as well as the counterfactual distribution for the rural area have the same endowment base than their urban counterparts in the same income quantile (Figure 11). The results are remarkably similar to the results without access to infrastructure, indicating the robustness of the results. Figure 11 The results are very similar when comparing the relative contribution of the returns and endowment (covariate) effects to the urban-rural gap estimated with infrastructure versus without infrastructure (Figure 12). Figure 12: Urban/Rural Gap Decomposition 2006 (a) Without Infrastructure (b) With Infrastructure 151 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 151 The results are also remarkably similar when performing the decomposition exercise by region, further confirming the robustness of the results (Figure 13). Figure 13: Regional Decomposition of the Urban/Rural Gap for 2006 When Controlling for Access to Infrastructure (a) Costa: Without Infrastructure With Infrastructure (b) Sierra: Without Infrastructure With Infrastructure (c) Selva: Without Infrastructure With Infrastructure 152 152 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 6. Final Remarks This paper decomposed log welfare rations between urban and rural Peru for 2002 and 2006, a period of high but unequal growth for the Peruvian economy. The decomposition exercises looked at the role of differences in asset endowments and differences in returns across regions in urban and rural Peru. The first decomposition methodology allowed us to compare rural and urban sections of the three regions and find, for example, the important role that infrastructure has in urban Selva when analyzing differences with the national broad picture. The second methodology showed the role of endowment and returns in the urban-rural divide across the income distribution. Following the Ravallion and Wodon (1999) decomposition methodology, we have shown that endowment differences matter the most when explaining the urban-rural wellbeing gap in Peru. Comparing the 2002 and 2006 decompositions shows that, although differences in asset endowments are the main driver of differences in poverty rates in both years, short-term changes in poverty rates may be driven by short-term differences in rates of returns. This result highlights the critical importance of a different policy mix when considering different time-frames. When looking at structural differences across space, the importance of endowments in shaping wellbeing is obvious. At the same time, in the short run policies will tend to affect the rate of return to assets more than the access to them. It is thus clear that policies to improve households' endowments and those to improve returns to those endowments are both needed. Although increasing access to endowments of characteristics that improve income and employment opportunities is critical in the long run, market reforms that enhance the rates of return to the endowments of the poor is also a move in the right direction. Allowing returns to change across the expenditure distribution results in a similar albeit slightly more complex picture. The quantile decomposition shows that the urban/rural wellbeing gap increases throughout quantiles of the income distribution, which means richer urban households are better off than their rural counterparts to a greater extent than the urban poor when compared to the rural poor. This urban-rural wellbeing gap pattern is explained mostly by the differences in endowments, and very little can be explained by differences in returns once we control for the relative position in the distribution. This of course does not mean that rate of return does not matter, but rather that it differs across quantiles, and is higher for those that have a better relative position in the wellbeing distribution. 153 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 153 Annex Table A.1. Poverty lines Used to Construct the Welfare Ratios Table A.2 154 154 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Table A.3 Table A.4 155 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 155 Figure A.1. Decomposition for 2002 75% 60% 45% 30% 15% 0% Geographic Profile Concentration Profile Unconditional Profile Costa-Urban Sierra-Urban Selva-Urban Costa-Rural Sierra-Rural Selva-Rural Figure A.2. Decomposition for 2006 75% 60% 45% 30% 15% 0% Geographic Profile Concentration Profile Unconditional Profile Costa-Urban Sierra-Urban Selva-Urban Costa-Rural Sierra-Rural Selva-Rural 158 156 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure A.3. Differences between Profiles 2002 Figure A.4. Differences between Profiles 2006 157 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 157 Figure A.5. Differences between Profiles 2002-2006 158 158 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Figure A.6. Distribution of Regional log Welfare Ratios between Urban and Rural Peru (a) 2002 (b) 2006 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 159 REFERENCES Bourguignon, Francois, Francisco Ferreira, and Phillippe Leite. 2002. "Beyond Oaxaca-Blinder: Accounting for Differences in Household Income Distributions across Countries." World Bank Policy Research Working Paper 2828. Washington DC: World Bank. Cotton, J. 1988. "On the Decomposition of Wage Differentials." The Review of Economics and Statistics 70: 236-243. Deaton, Angus. 1997. The Analysis of Household Surveys. Washington DC: World Bank. Duarte, Angelo Jose Mont'Alverne, Pedro Cavalcanti Ferreira, and Marcio Antonio Salvato. 2003. "Regional or Educational Disparities? A Counterfactual Exercise." Ensaios Economicos da EPGE No 532. Efron, Bradley, and Robert J. Tibshirani. 1993. Introduction to the Bootstrap. Boca Raton: Chapman and Hall/CRC. Guimarães, Juliana Ferraz, Tiago Cavalcanti, and Raul Silveira Neto. (2006). "Accounting for Labor Income Differences in Brazil: the Role of Human Capital." Anais do XXXIV Encontro Nacional de Economia [Proceedings of the 34th Brazilian Economics Meeting] 35, Associação Nacional dos Centros de Pósgraduação em Economia [Brazilian Association of Graduate Programs in Economics]. Hendricks, Wallace, and Roger Koenker. 1992. "Hierarchical Spline Models for Conditional Quantiles and the Demand for Electricity." Journal of the American Statistical Association 87(417): 58-68. Jalan, Jyotsna and Martin Ravallion. 2002. "Geographic Poverty Traps? A Micro Model of Consumption Growth in Rural China." Journal of Applied Econometrics 17: 329-346. Jann, B. 2008. "A Stata Implementation of the Blinder-Oaxaca Decomposition." ETH Zurich Sociology Working Paper No. 5. Kanbur, Ravi. 2006. "The Policy Significance of Inequality Decompositions." Journal of Economic Inequality 4(3): 367-374. Koenker and Bassett. 1978. "Regresion Quantiles." Econometrica 46 (1): 33-50. Koenker and Hallock. 2001. "Quantile Regression." Journal of Economic Perspectives 15: 143-156. Krugman, P. 1998. "What's New About the New Economic Geography." Oxford Review of Economic Policy 14(2): 7-17. Lall, Somik, and Z. Shalizi. 2003. "Location and Growth in the Brazilian Northeast." Journal of Regional Science 43(4): 663-681. Lall, S., C. Timmins, S. Yu. 2008. "Moving to Opportunity? Successful Integration of Bright Lights." Presentation at World Bank, June 2008. 161 160 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES Machado, Jose A. F., and Jose Mata. 2005. "Counterfactual Decomposition of Changes in Wage Distributions Using Quantile Regression." Journal of Applied Econometrics 20: 445-465. Neumark, D. 1988. "Employers' Discriminatory Behavior and the Estimation of Wage Discrimination." The Journal of Human Resources 23: 279-295. Nguyen, Binh T., James W. Albrecht, Susan B. Vroman, and M. Daniel Westbrook. 2007. "A Quantile Regression Decomposition of Urban-Rural Inequality in Vietnam." Journal of Development Economics 83: 466 ­ 490. Oaxaca, Ronald, L. 1973. "Male-Female Wage Differentials in Urban Labor Markets." International Economic Review 14: 693-709. Oaxaca Ronald, and Michael Ransom. 1994. "On Discrimination and the Decomposition of Wage Differentials." Journal of Econometrics 61: 2-21. Ravallion, Martin. 2004. "Pro-Poor Growth: A Primer." World Bank Policy Research Working Paper 3242. Washington DC: World Bank. Ravallion, Martin, and Quentin Wodon. 1999. "Poor Areas, or Only Poor People?" Journal of Regional Science 39 (4): 689-711. Reimers, C. W. 1993. "Labor Market Discrimination Against Hispanic and Black Men." The Review of Economics and Statistics 65(4): 570-579. Venables, A. 2005. "Spatial Disparities in Developing Countries: Cities, Regions, and International Trade." Journal of Economic Geography 5(1): 3-21. World Bank. 2003. "Brazil: Inequality and Economic Development" Washington, DC: The World Bank (Report 24487-BR). Washington DC: World Bank. World Bank. 2005. World Development Report 2006: Equity and Development. New York: The World Bank and Oxford University Press. World Bank (2005). Ecuador Poverty Assessment. Washington DC: World Bank. World Bank. 2007. "Brazil: Measuring Poverty Using Household Consumption" World Bank Report 36358-BR. Washington DC: World Bank. 162 DETERMINANTS OF REGIONAL WELFARE DISPARITIES IN LATIN AMERICAN COUNTRIES 161 o i g h t d economic growth urban disparities poverty ty www.worldbank.org