Report No. 24487-BR Brazil Inequality and Economic Development in Brazil (In Two Volumes) Volume II: Background Papers October 2003 A Joint Report by Instituto de Pesquisa Econômica Aplicada and Brazil Country Management Unit Poverty Reduction and Economic Management Sector Unit Latin America and the Caribbean Region Document of the World Bank Table of Contents 1.POVERTYAND INEQUALITYINBRAZIL: NEWESTIMATESFROM COMBINEDPPV-PNADDATA C.Elbers, J.Lanjouw, P.Lanjouw and P.Leite ................................................................................................................... 1 Introduction............................................................................................................................................... 1 2- Methodology........................................................................................................................................ 1- Data ..................................................................................................................................................... 3 4 4- Poverty and Inequality at the Regional Level ..................................................................................... 3- Implementation.................................................................................................................................... 7 9 6- Inequality Decompositions ................................................................................................................ 5- Poverty and Inequality at Lower Levels of Disaggregation .............................................................. 11 13 7- Conclusions ....................................................................................................................................... 14 2.BEYONDOAXACA-BLINDER: ACCOUNTINGFORDIFFERENCESIN HOUSEHOLDINCOMEDISTRIBUTIONSACROSS COUNTRIES ..................................................................... 27 F Bourguignon,F Ferreira and P Leite . . . Introduction............................................................................................................................................. 27 2- A General Statement of Statistical DecompositionAnalysis............................................................ 1- IncomeDistribution inBrazil, Mexico andthe UnitedStates........................................................... 29 34 4- The Brazil-US Comparison............................................................................................................... 3- The Decompositions inPractice: A Specific Model.......................................................................... 35 5- The Brazil - Mexico Comparison ...................................................................................................... 39 6- Conclusions ....................................................................................................................................... 48 46 3.INEQUALITY OFOUTCOMES, INEQUALITY OFOPPORTUNITIESAND INTERGENERATIONALEDUCATIONMOBILITY INBRAZIL ........................................................................ 65 F Bourguignon. F Ferreira and M Menendez . . . Introduction............................................................................................................................................. 65 67 2- Opportunities and the distribution of individual wages..................................................................... 1- Theoretical background..................................................................................................................... 4- The effects of the inequality of opportunities on the distribution of household income...................82 3- Simulatingthe effects of the inequality of opportunities on earnings............................................... 69 5- Summary and conclusion .................................................................................................................. 88 92 4.INDIRECTTAXATION REFORM: SEARCHINGFOR DALTON-IMPROVEMENTSINBRAZIL ....................................................................................................... 95 C.E. Vdez, S. Vianna,F.G. Silveira and C.Magalhiies Introduction............................................................................................................................................. 95 2- What is the impact of taxation on the distribution of secondary income?........................................ 1- Taxation in Brazil: recent evolution, trends and issues..................................................................... 96 3- Dalton improving tax reforms: analytical framework ..................................................................... 99 102 4- Identifyingpotential candidates for raising and reducing taxes...................................................... 104 5- The pairs of tax changes that satisfy the Dalton improvement condition........................................ 107 6- Conclusions ..................................................................................................................................... 110 5. SCHOOLINGEXPANSIONINDEMOGRAPHICTRANSITION: A TRANSIENTOPPORTUNITY FOR INEQUALITYREDUCTIONINBRAZIL S. Soares. M.Madeiros and C.E. Vdez ................................................. 117 Introduction........................................................................................................................................... 117 2- Methodology and Data.................................................................................................................... 121 1- Demographic Background............................................................................................................... 119 4- The IncreasingStock-to-Cohort Time Lagof EducationalAttainment......................................... 3- The Evolution of Education Between Cohorts ............................................................................... 122 124 6- Conclusions ..................................................................................................................................... 130 5- Simulations...................................................................................................................................... 127 6.EX-ANTEEVALUATIONOFCONDITIONALCASHTRANSFERPROGRAMS: THE CASEOF BOLSAESCOLA ................................................................................................................. 139 F.Bourguignon. F .Ferreira and P .Leite. 1- Main features of the Bolsa Escolaprogram...................................................................................... Introduction........................................................................................................................................... 139 2- A simple framework for modeling and simulatingBolsa Escola..................................................... 141 141 4- An ex-ante evaluation of BolsaEscola and altemative programdesigns......................................... 3- Descriptive statistics and estimation results ..................................................................................... 147 149 5- Conclusion........................................................................................................................................ 152 7.THEDYNAMICS OFTHE SKILL-PREMIUMINBRAZIL: GROWINGDEMAND A.Blom and C.E. Ve'lez ANDINSUFFICIENT SUPPLY? .................................................................................................................. 163 Introduction........................................................................................................................................... 163 2- Relative Supply, Relative Demand and the Skill-premium............................................................. 1- Wage-inequality and Education....................................................................................................... 166 172 181 4- Summary ......................................................................................................................................... 3- What if? Alternative paths for supply and wage-inequality inthe past........................................... 184 8.DISTRIBUICAO DETERRA EAS POL~TICAS PUBLICASVOLTADAS AO ME10RURALBRASILEIRO ...................................................................................................................... J.Assunpio 195 IntrodugBo ............................................................................................................................................. 196 2- A Ineficigncia da DistribuigBo de Terras: aspectostedricos e empiricos........................................ 1- Cenhrio Brasileiro............................................................................................................................ 198 207 222 4- Ligdes e Propostas........................................................................................................................... 3- Politicas Wblicas voltadas ao Meio Rural Brasileiro...................................................................... 234 1.POVERTYANDINEQUALITY INBRAZIL: NEWESTIMATES FROMCOMBINEDPPV-PNADDATA' Chris Elbers (Vrije Universiteit, Amsterdam), Jean Olson Lanjouw (Yale University and Brookings Institution), Peter Lanjouw (World Bank), Phillippe George Leite ( Pontificia Universidade Cat6lica do Rio de Janeiro) Introduction Inequality and poverty occupy a prominent place in debates surrounding the recent development experience of Brazil, its future prospects and available policy options. There i s an extensive literature on the distribution of well-being in Brazil - describing levels and dynamics of poverty and inequality outcomes; scrutinizing regional and sectoral disparities; studying the links to labour markets, human capital outcomes, public spending patterns; and so on.2 An important stylized fact that emerges from this body of research i s that, compared to other countries, Brazil i s a clear outlier in terms of inequality and also accounts for a dominant share of the total number of poor inLatin America. Conclusions regarding measured poverty and inequality levels, and trends, depend crucially on the underlying empirical foundations that support such analysis. Almost all of what i s known about the distribution of economic welfare in Brazil, at the level of the country as a whole, comes from the well- known PNAD (Pesquisa Nacional por Amostra de Domicilios ) household surveys. These are large surveys, fielded on an annual basis since the late 1960s, covering virtually all of Brazil (except the sparsely populated north of the country). The PNAD survey permits the construction of a measure of household income, and this indicator of economic welfare underpins much of the subsequent analysis of well-being that has drawn on PNADdata. A recent study by Ferreira, Lanjouw and Neri (2000) suggests that there are at least some reasons for concem regarding the welfare indicatoravailable inthe PNAD surveys. Becausethe survey i s essentially an earnings survey, it i s oriented towards formal sector employment. As a result income data from households engagedin self-employment activities are only cursorily collected. These problems may result ininaccurate measures of income from two groups of particular importance in distributional analyses: self-employed informal sector and cultivating households. The question thus arises whether the limitations of the PNAD income figures are driving some of the conclusions about welfare outcomes inBrazil -with respect to both levels andpatternsacross population subgroups. In 1996 a pilot household survey was fielded in Brazil's northeast and southeast regions. This survey, known as the Pesquisa sobre Padr6es de Vida (PPV), i s modelled after the World Bank Living Standard Measurement Survey (LSMS). The PPV i s a multi-module integrated survey which collects, in addition to information on incomes, data on household consumption. Fairly detailed information on consumption expenditures are collected and it is also possible to impute values of consumption streams from items such as housingand home-produced food products. In addition to being more detailed and comprehensive in construction, a consumption measure is generally perceived to provide a more reliable indicator of economic well-being than income, even when there are no clear biases in the income measure. This i s particularly so when the purpose of the analysis i s to study poverty. Essentially it i s argued that it i s easier to collect reliable and reasonably complete ' We are grateful to Francois Bourguignon,Francisco Ferreira, Pedro Luis do Nascimento Silva, Ricardo Paes de Bamos, and Martin Ravallion for useful discussions. Financial support was gratefully received from to the Bank-Netherlands Partnership programand the World Bank PREM InequalityThematic Group. The views presentedin this paper are those of the authorsonly andshouldnotbetakento reflect the views of the World Bank or any affiliatedinstitution. The literature is very large. Useful recent contributionsinclude Camargo and Ferreira (1999), Ferreira and Litchfield (1996), Ferreiraand Paesde Barros(1999) andWorld Bank (2001a, 2001b). 1 information on consumption than on income. In addition, there are theoretical arguments in favor of using consumption as a measure of welfare because consumption is thought to better proxy long term living standardsthan current in~ome.~ Despite this surface appeal, the PPV also has clear limitations. Compared to the PNAD survey, the PPV sample i s tiny. In addition, the PPV is not designed to be representative of the country as a whole - its coverage extends only to the northeast and southeast of Brazil (covering roughly three quarters of the nationalpopulation). The purpose of this paper i s to report on the results of an attempt to exploit the best features of the two datasets described above, so as to produce consumption-based estimates of poverty and inequality, but in the large PNAD sample. We employ a recently developed methodology which permits the analyst to impute a welfare indicator from one survey, the PPV inour case, into another survey, the PNAD. Imputingconsumption into the much larger PNAD survey allows us then to estimate summary measures of poverty and inequality at levels of regional disaggregation significantly lower than what would have been possible in the PPV survey. An additional feature of the methodology i s that we are able to assess the statistical precision of the welfare measures we estimate. One of our concerns inthis paper will be to determine whether the imputation methodology comes at an unacceptably highprice interms of statistical significance of the estimates. Aside from demonstrating the feasibility of imputingconsumption from the PPV into the PNAD survey, we also aim in this paper to ask whether the picture of poverty and inequality that derives via this approach differs significantly from that which obtains from standard analysis based on the conventional PNAD income measure. In that sense we are interested to use our approach as a means to gauge the robustnessof the conventional picture of poverty and inequality inBrazil. Itis clear that we cannot directly assess the reliability of the PNADincome measureby simply comparing the conventional results against those we obtain following our approach: the welfare concepts of income and consumption are different and could not be expected to yield identical quantitative estimates of poverty and inequality. However it may perhaps be arguable that if the two welfare measures were sound, their qualitative implications, in terms of the profile of poverty they yield, would be broadly similar. To that end we compare the spatial profile of poverty and inequality across these two approaches on the basis of simple cross tabulations and decompositions. The structure of the remainder of this paper i s as follows. Inthe next section we describe in greater detail the two data sources we draw on in this analysis. Section I11turns to an overview of the methodology we employ. Section IV describes how we implement the methodology in this particular setting. Section V presents results at the level of the PPV's representative region. We ask whether our estimates of poverty and inequality accord with those of the PPV, we assess the statistical precision of our estimates, and we compare our consumption based estimates against the estimates that would obtain had we used the income measure that i s conventionally analyzed using the PNAD survey. Section VI produces further consumption-based estimates, at levels of disaggregation which the PPV survey could not support. Once again we compare findings against the PNAD income concept in order to develop a further sense of whether, and where, the two approaches part company. In Section VI1 we report some basic inequality decomposition results, again in turn based on the PNAD income measure and the PNAD imputed consumption measure. We ask whether qualitative conclusions regarding the relative contribution to overall inequality from certain population subgroups are robust to the welfare measure that i s employed. Section VI11summarizes our findings and discusses directions for future enquiry. See Ravallion (1994) for further discussion. 2 1- Data As described above, the analysis in this paper draws on two sources of household survey of data: the combined 1996 and the 1997 rounds of the Pesquisu Nucionul por Amostru de Domicilios (PNAD); and the Pesquisa sobre Padr6es de Vidu (PPV) of 1996. The PNAD, implemented by the national statistical organization IBGE (Znstituto Brasileiro de Geogrufiu e Estutisticu), has been the main staple of country-wide (as opposed to metropolitan) distributional analysis in Brazil since the mid-1970s. It i s an annual survey covering both urban and rural areas (except inthe Northernregion), and is representative at the level of the state andall metropolitan areas. Its sample size, currently around 105,000 dwellings per survey-round, i s generally viewed as ample to produce reliable estimates of poverty or inequality at the regional, state, or possibly even lower, level. However, for such a large survey, and one which i s fielded so often, some of the PNADquestionnaire shortcomings are remarkable. The questionnaire has evolved a great deal between the mid-1970s and 1996, generally much for the better. Nevertheless, there is one aspect, crucial for poverty and income distribution analysis, which has remained rather problematic: the income questions for any income source other than wage employment are insufficiently disaggregated and detailed? Inprinciple, the nonsampling errors likely to arise from the absence of these more detailed questions could bias income measurement ineither direction. Too few questions about in-kindbenefits or the values of different types of production for own consumption are likely to lead to an underestimate of welfare, through forgetfulness. On the other hand, the absence of detailed questions about expenditure on inputs i s likely to lead to an overestimate of net incomes from home production. In practice, the international evidence suggests that the first effect often predominates, and the absence of such detailed questions can lead to income under-reporting by categories of workers which, as it happens, are quite likely to be poor. Ferreira, Lanjouw and Neri (2000) examine these issues for the case of Brazil in some detail and suggest not only that under-reporting of income in the PNADmay well be significant, but that the degree i s likely to vary considerably significantly across population subgroups. As mentioned earlier, our second data source, the PPV, is a household survey modeled on the Living Standard Measurement Survey. It was fielded in 1996-97 by IBGE to assess the poverty targeting of Government social spending in Brazil. The aim of the PPV was to supplement the information already available through the PNAD, in order to improve the data available for poverty monitoring and policy analysis inBrazil. The PPV was designed to fill some of the data gaps left by the PNAD. It provides a much more detailed picture of household expenditures and consumption, as well as utilization of various publicly subsidized services, particularly education, health, and transportation. The questionnaire i s much longer, and requires multiple visits to each household. This richer information comes at a price. To keep survey expenses within reason, the sample size i s much smaller (just under 5000 households in total) and the survey only covers the two most populous of Brazil's five regions, the Northeast and Southeast. These two regions together account for 73% of Brazil's population. The PPV i s representative for ten spatial units (the metropolitan areas of Sgo Paulo, Rio de Janeiro, Belo Horizonte, Salvador, Recife, and Fortaleza; the non-metropolitan urban Northeast; the non-metropolitan urban Southeast; the rural Northeast; and the rural Southeast). However, as we shall see below, even at the representative region level, estimates of poverty and other welfare indicators may be rather imprecisely estimated. The purpose of this paper is to report on the application of a technique to combine the PPV and the PNADdatasets, seeking to complement their respective strengths and to compensatefor their weaknesses. The data issues addressed in this section are more thoroughlydiscussed in Ferreira, Lanjouw and Neri (2000) 3 Because a maintained hypothesis of the imputation is that the consumption models estimated on the PPV data apply to PNAD households, it is most tenable to implement our method with reference to the Northeast and Southeast of Brazil only (that part of the country which the PPV i s representative of). All results presented in this paper, including those based on the PNAD, thus pertain to these two regions A final word on the data concerns the comparability of the PPV and the PNAD surveys. It is imperative for the successful implementation of our methodology, that the two data sources we draw on be closely comparable. The methodologies underlying sampling, data collection methods, questionnaire design, etc., across these two datasets are quite different. Nonetheless, Soares de Freitas, et al (1997) find little evidence, in a comparison of the PPV with the 1995 PNAD survey, that these basic methodological differences introduce major discrepancies across the two data sources in terms of population characteristics.6 The PPV survey was fielded during a period of one year spanning 1996 and 1997. The annual PNAD surveys are fielded on or around a given date, usually in September, in their respective survey years. In a further attempt to ensure that the two data sources we work with are as comparable as possible, and also in order to maximize the sample size of the database into which consumption is imputed, we have merged the 1996 and 1997 rounds of the PNAD on the grounds that these two neatly bridge the periodcovered by the PPV survey. Although the geographic coverage of this combined dataset i s confined to the Northeast and Southeast of Brazil only, the size of the sample i s sufficiently large (around 111,000households) to permit considerable disaggregation. 2- Methodology The methodology we implement here has been described in detail in Elbers, Lanjouw and Lanjouw (2001). The basic idea is straightforward. We estimate poverty and inequality based on a household per-capita measure of consumption expenditure, Yh. A model of Y h i sestimated using the PPV survey data, restricting explanatory variables to those that can be linked to households in both sets of data? Then, letting W represent an indicator of poverty or inequality, we estimate the expected level of W given the PNAD-based observable characteristics of the area of interest using parameter estimates from the `first-stage' model of y. The same approach could be used with other household measures of well-being, such as per-capita expenditure adjusted by equivalence scales, or to estimate inequalities in the distribution of household characteristics other than expenditures, such as assets or income. Definitions The basis of the approach i s that per-capita household expenditure, Yh, i srelated to a set of observable characteristics, xh, that can be linkedto households inboththe PPV and PNAD sample surveys:* In yh = E[ln Y Ixh ] -k uh. h (1) Using a linear approximation to the conditional expectation, we model the observed log per-capita expenditure for household h as: We retumin the concludingsecton to a discussionof the feasibilityof extendingthis analysisto areas not coveredby the PPV, such as the SouthandCenter-Westof Brazil. '*Elberset BianchiniandAlbieri (1998)providefurther detailson the survey designs of all of Brazil's majorhouseholdsurveys. a1(2001) describethe case when we imputeexpenditurefrom a householdsurvey into the populationcensus. The explanatory variables are observed values and thus need to have the same degree of accuracy in addition to the same definitions across data sources. From the point of view of our methodology it does not matter whether these variables are exogeneous. 4 where j?i s a vector of k parameters and uh i s a disturbance term satisfying E[uhlxh]= 0. The vector of disturbances inthe population i s distributedu -3(0,a. The model in (2) i s estimated usingthe PPV data. We are interested in usingthese estimates to calculate Although the disaggregation may be along any dimension- not necessarily geographic -for convenience the welfare of an area or group for which we do not have any, or insufficient, expenditure information. we will refer to our target population as a `UF' (union federaggo). There are M, households in UF v, M," households inthe PNADsample from UFv, and household h has mhfamily members. While the unit of observation for expenditure inthese data i s the household, we are more often interested in poverty and inequality measures based on individuals. Thus we write W (m, X, fl u,), where mi s a vector of household sizes, X i s a matrix of observable characteristics and u i s a vector of disturbances. Because the disturbances for households in the target population are always unknown, we consider estimatingthe expected value of the indicator given the PNADhouseholds' observable characteristics and the model of expenditure in(2): We denote this expectation as bSE[W m;, X,", 51, = I (3) where 5 i s the vector of model parameters, including those which describe the distribution of the disturbances, and the superscript `s' indicates that the expectation i s conditional on the sample of PNAD households from UFv rather than a census of households. Inconstructingan estimator of p,,' we replace the`unknown vector 5with consistent estimators, p ,from the first-stage expenditure regression. This yields ,b," = E[W I m,", X,", PI. This expectation i s generally analytically intractable so we use simulation to obtain our estimator, p;. Properties The difference between p;, our estimator of the expected value of W for the UF, and the actual level of welfare for the UFmay be written (suppressing the index v): w-ps =(W-p))+(p-pS)+(pS -ps)+(ps-p". (4) Thus the prediction error has four components: the first due to the presence of a disturbance term in the first stage model which implies that households' actual expenditures deviate from their expected values (idiosyncratic error); the seconddue to the fact that we are imputing into a sample rather than a census of households (sampling error); the third due to variance in the first-stage estimates of the parameters of the expenditure model (model error); and the forth due to using an inexact method to compute ,,!is (computation error). Elbers, Lanjouw and Lanjouw (2001) provide a detailed description of the properties of the first and last two components of the prediction error. To summarize, the variance in our estimator due to idiosyncratic error falls approximately proportionately in M,, the size of the actual population of households in the UF. In other words, the smaller the target population, the greater is this component of the prediction error, and there is thus a practical limit to the degree of disaggregation possible. At what population size this error becomes unacceptably large IfthetargetpopulationincludesPPVhouseholdsthensomeinformationisknown. Asapracticalmatterwedonotusethesefew pieces of direct information on y. depends on the explanatory power of the x variables in the expenditure model and, correspondingly, the importance of the remaining idiosyncratic component of the expenditure. We calculate sampling errors on our poverty estimates taking into account the fact that the PNAD surveys are complex samples which involve stratification and multi-stage clustering (see Howes and Lanjouw, 1998, Deaton, 1997) . We employ the delta methodto calculate the variance due to model error: VM = VTV(@)V,where v =[a,? /a{] It and V( @)is the asymptotic variance covariance matrix of the first-stage parameter estimators. Because this component of the prediction error i s determined by the properties of the first-stage estimators, it does not increase or fall systematically as the size of the target population changes. Its magnitude depends, in general, only on the precision of the first-stage coefficients and the sensitivity of the indicator to deviations inhousehold expenditure. For a given UFits magnitude will also depend on the distance of the explanatory variables for households in that UF from the levels of those variables inthe sample data. The variance in our estimator due to computation error depends on the method of computation used. As our calculations of the idiosyncratic and models errors are based on simulations, we can make the computation error become as small as desired by choosing a large enough number of simulation draws (at the cost of computational resources andtime). We use Monte Carlo simulations to calculate: 4" the expected value of the poverty or inequality measure conditional on the first stage model of expenditure; VI, the variance in W due to the idiosyncratic component of household expenditures; and, for use in determining the model variance, the gradient vectorV = [d,Ls /a{] Let the vector fir be the rth draw from our estimated disturbance distribution - a random draw from an M,-variate standard normal or t distribution, pre-multiplied by a matrix T, defined such that TTT =gV, where e,, i s the estimated disturbance covariance matrix for the population of households in UF v. With each vector of simulated disturbances we construct a value for the indicator, Wr =W(m, X,p,fir ), where m and X represent numbers of households and observable characteristics of PNAD households, respectively, each repeated in accordance with its expansion factor so as to have rows equal to the census number of households, M,. The simulated expected value for the indicator i s the mean over R replications: ps = R- c w r . I R A (5) r = l Having estimated p using the population number of households, the variance of W around its expected value due to the idiosyncratic component of expenditures can be estimated in a straightforward manner usingthe same simulated values: G 1\r=l Simulatednumerical gradient estimators are constructed as follows: We make a positive perturbationto a parameter estimate, say bk,by adding 6 1 bkI, and then calculate Ps+. A negative perturbationof the same size i s used to obtain P `-. The simulated central distance estimator of the derivativeaps/ d,8 i s (,Es+-p")/(26 I bkI). Having thus derived an estimate of the gradient vector, we can calculate QM =VTV( g,v 3- Implementation The first-stage estimation i s carried out usingthe PPV survey. As described in section I1this survey i s stratified into ten regions and i s intended to be representative at that level. Within each region there are several levels of clustering. At the final level, 8 households are randomly selected from a census enumeration area. Such groups we call a `cluster' and denote with a subscript c. Expansion factors, lcht allow the calculation of regional totals. Our first concern i s to develop an accurate empirical model of household consumption. Consider the following model: where 77 and E are independent of each other and uncorrelated with observables, &h. This specification allows for an intra-cluster correlation inthe disturbances. One expects location to be relatedto household income and consumption, and it i s certainly plausible that some of the effect of location might remain unexplained even with a rich set of regressors. For any given disturbance variance, a:,, ,the greater the fraction due to the common component 77, the less one enjoys the benefits of aggregating over more households within a UF. Welfare estimates become less precise. Further, the greater the part of the disturbance which i s common, the lower will be inequality. Thus, failing to take account of spatial correlation in the disturbances would result in underestimated standard errors on welfare estimates, and upward biased estimates of inequality. Since unexplained location effects reduce the precision of poverty estimates, the first goal i s to explain the variation in consumption due to location as far as possible with the choice and construction of xchvariables. We try to tackle this in four ways. First, we estimate different models for each of the ten regions in the PPV. Second, we include in our specification household level indicators of connection to various networked infrastructure services, such as connection to electricity, piped water, telephone. To the extent that all or most households within a given neighborhood or community are likely to enjoy similar levels of access to such infrastructure, these variables might capture unobserved latent location effects. Third, we calculate in the PPV and PNAD dataset cluster-mean values of household level variables, such as the average level of education of household heads per cluster, and also consider these variables for inclusion in the first-stage regression specification. These cluster-level variables might also serve to proxy location-specific correlates of expenditure. Finally, we have merged both the PPV and the PNAD datasets with an independently compiled municipio-level database (BIM) of variables (such as 7 employment rates, school attendance rates, etc.) and also consider these variables as candidate variables for inclusioninour household expenditure models." We apply a selection criterion when deciding on our final specification requiring a significance level of 5% of all household-level regressors. To select location variables (cluster means and BIM variables), we estimate a regression of the total residuals, t i , on cluster fixed effects. We then regress the cluster fixed- effect parameter estimateson our location variables and select those five that best explain the variation in the cluster fixed-effects estimates. These five location variables are then added to our household level variables inthe first-stage regression model. We apply a Hausman test described in Deaton (1997) to determine whether each regression should be estimated with household weights. In seven out of ten regions we find that weighting has no significant effect on the coefficients, and these first-stage regressions are thus estimated without weights. R2'son our models are generally high, rangingbetween 0.45 and0.77." We next model the variance of the idiosyncratic part of the disturbance, O:,ch.Note that the total first- stage residual can be decomposedinto uncorrelated components as follows: ti =tic, (iCh = 9, i + - ti,,) ech (10) where a subscript `.' indicates an average over that index. To model heteroskedasticity in the household- specific part of the residual, we choose the twenty variables, Zch,that best explain variation in e,`h out of all potential explanatory variables, their squares, and interactions." We estimate a logistic model of the variance of &chconditional on Zch, bounding the prediction between zero and a maximum, A, set equal to (1.OS) * max {ech} : 2 Letting exp{zTh&}=B and using the delta method, the model implies a household specific variance estimator for &ch of Finally, we check whether and E are distributed normally, based on the cluster residuals 9, and standardized household residuals e:h =-%-[ 1 F C c h 7 ech3 , respectively where H i s the number of " E x h O & , c h households in the survey. The second term in e:his not needed when first stage regressions are not weighted. In many cases normality i s rejected, although the standard normal does occasionally appear to loA municipio represents a higher level of aggregation than the census EA, and as such BIM variables are intended to capture loactionaleffects at this higherlevel,ratherthan the morelocalcluster-level means. For reasons of space we do not reproducehere the parameterestimates and full set of diagnostics for all ten regressionmodels. These can be fumisheduponrequest. We limit the numberof explanatory variables to twenty to be cautiousabout overfitting. 8 be the better approximation even ifformally rejected. Elsewhere we use t distributions with varying degrees of freedom (usually 5), as the better approximation. Before proceeding to simulation, the estimated variance-covariance matrix, E, i s used to obtain GLS 1, estimatesof the first-stage parameters, BGLs.and , theirvariance,Var( BGLs. ). 4- Poverty and Inequality at the RegionalLevel We beginour examination of empirical results at the level of the representative region inthe PPV survey. Table 1reports estimates of the incidence of poverty for the ten representative regions of the PPV. We report poverty estimates based on the three possible combinations of welfare concept and data-source that are available. In the first column we present estimates of the incidence of poverty in the combined 1996-97 PNAD survey based on the PNAD income measure of welfare. Column two provides our calculationof the standarderror on this income-based poverty measure. This standarderror comprises the sampling error described earlier, and our calculations of this have taken into account the complex sample design of the PNAD survey. The second set of poverty estimates and standard errors (columns 3 and 4) are based on the PPV survey and the per capita consumption aggregate that can be constructed from that survey. The standard errors are once again sampling errors incorporating the complex sample design of the PPV. Finally, columns 5 and 6, provide estimates of poverty from the PNADusingthe consumption indicator of welfare that we have imputed into the PNAD. The standard errors on these poverty estimates comprise both a sampling error as well as the model error described in sections 111 and IV. The idiosyncratic error is vanishingly small at the levels of disaggregation that we are concerned with in this paper, and i s therefore not rep~rted.'~Computational error has been pushed close to zero by employing at least 100simulations in all our calculations. We use the same poverty line to measure the incidence of poverty across all three cases. We employ the poverty line of R$65.07 in 1996 Siio Paul0 reais which was derived inFerreira, Lanjouw and Neri (2000) as an extreme poverty line, sufficient to permit consumption of a minimum bundle of food items only. Both the income and consumption measures of welfare have been adjusted to capture spatial price variation (see Ferreiraet al, 2000). Considering first a comparison of poverty based on PNAD income versus PPV consumption we are immediately struck by the much higher levels of measured poverty in the PNAD. This point has already been discussed at length by Ferreira, et a1 (2000) and it i s perhaps useful to add only that the differences in measuredpoverty between PNADincome and PPV consumption are generally statistically significant. I t i s important to note that even though the PPV i s designed to be representative at the level of these ten regions, the standard errors on the poverty estimates at this level are generally higher than 10% of the point estimate -indicatingthat confidence bounds around these point estimates are quite wide.I4 Even so, measured poverty in the PNAD i s so much higher than in the PPV that one can generally rule out that they are statistically indistinct. While levels of poverty across the two surveys and welfare concepts are clearly different, qualitative conclusions between the PNAD-income poverty profile and the PPV-consumption profile, are much more similar. Both approaches find clear evidence that rural poverty in the northeast i s highest of all ten l3 Calculation of this component is very computationallyintensive as it requires using expansion factors to explode the PNAD sample up to a meta-census level, and then carrying out simulations to estimate the idiosyncratic error on the point estimate of poverty. Elberset al (2001) document that that idiosyncraticerror becomes negligiblewhen welfare estimatesare for populations of 10,000 householdsor more. In no case do we estimatewelfare measures for populationsbelowthis size. (Notethat the criterion i s population not sample size). l4 Forthis reasonone would be very reluctant to disaggregatethe PPV downbelowthis level. 9 regions, followed by poverty in the urban northeast and then the rural southeast. Poverty in the metropolitan areas of the southeasti s clearly lowest. These conclusions are echoed when we return to the PNAD data but base our poverty estimates on consumption imputedaccording to the method described in sections I11and N. The ranking of poverty across regions i s identical to that obtained with the PNAD-income approach, but point estimates of poverty are now virtually the same as those obtained in the PPV. The method we have employed seems to work in that it provides us with estimates of poverty in the large PNAD dataset that are in close accordance with the PPV survey. An indirect implication i s that the two data sources are reasonably good samples of, and are describing, the same underlying population. Without that strict comparability of the surveys we could not have expected to obtain such close agreement between the PPV and the PNAD consumption-based estimates. Although the standard errors for the imputed-consumption based PNAD profile incorporate several error components that do not affect the PPV estimates, the precision of the PNAD consumption estimates i s generally greater than that of the PPV estimates, even at this high level of aggregation. This is because the PNADconsumption estimates are calculated over the much larger PNADdataset and sampling errors are thus commensurately smaller. Although the model errors on the PNAD consumption estimates are not negligible (as reflected in the higher total standard errors for these estimates than for the PNAD income estimates), it appears that they are not so large as to invalidate the exerci~e.'~ In Table 2 we turn to a similar examination of the three alternatives, but consider measuredinequality. The inequality measure we employ for this purpose is the General Entropy class measure, with a parameter value of c=OS This class of inequality measures has the attractive feature of being sub-group decomposable, and the choice of c allows the analyst to weight changes in inequality differently depending on which segments of the income distribution are affected (see Bourguignon, 1979, Cowell, 1980, and Shorrocks, 1980). We employ here a value for c of 0.5 which corresponds to a fairly high weighting to changes in inequality amongst the lower tail of the distribution.16 The first conclusion i s that, as with poverty measurement, measured inequality i s much higher with the PNAD income concept than based on consumption. Consumption based estimates in both the PPV and PNADare much lower than those in the first column. We are unable to state at this point whether these differences are statistically significant because we have not yet been able to calculate sampling errors on inequality measures which properly take into account the complex design of both the PNAD and PPV surveys. This means also that we are only able to report model errors on the PNAD consumption-based inequality estimates. Future work will address this concern. There is considerable disagreement between all three data-source/welfare-concept combinations in terms of the relative ranking of inequality across regions. We do note that the range of values of the inequality estimates i s more compressedthan the values of poverty estimates in Table 1, and that it i s therefore quite possible that rankings would not be statistically significant if we had information on sampling errors. We will find below that inequality differences at lower levels of disaggregation are more pronounced. l5Note efforts are ongoing to reduce these model errors further by retuming to the first-stage specification of the consumption modelsin the PPV. The estimates, anderrors, reported here shouldthus be viewed as preliminary. Infuture work we will be replicating the analysis here for other values of c, as well as other measuresof inequality. 10 The PNADconsumption-based estimates are generally of the same order of magnitude as those estimated in the PPV, and are similarly lower than the PNAD income-based estimates. The only outlier in this regard i s inequality in the rural southeast which is estimated at 0.443 with the PNAD consumption criterion and 0.240 with the PPV. The model error on the PNAD consumption-based estimate i s very high however, 0.06, suggesting that perhaps the large difference in estimates is not statistically significant. Certainly, the large model error on the estimate with the PNAD consumption criterion suggests that it may be worth returning to the imputation exercise to see whether a better first stage model can be estimated." The broad conclusion on which all three alternatives agree i s that inequality inthe ruralnortheast tends to be particularly low compared to urban areas as a whole. A similar uniform finding i s that metropolitan areas inthe southeast tend to be more unequalthan other urban areas inthe southeast. 5- Poverty and Inequality at Lower Levels of Disaggregation The discussion in the preceding section indicated that our methodology appears to allow us to impute consumptioninto the PNAD survey and obtain estimates of poverty and inequality that are not out of line with what we would expect (based on analysis at the representative region in the PPV). The next step i s to produce PNADconsumption-based estimates at levels of disaggregation that are below what we would be able to produce with the PPV. This step representsthe actual goal of the whole exercise: to employ a concept of welfare we are more comfortable with in a dataset which offers much more scope for disaggregation than the PPV. InTable 3 we produce estimates of poverty at the level of the UnionFederaGiio, breaking these states up, inturn, into metropolitan, other urban, and rural areas. Once again, inorder to compare broadqualitative conclusions, we produce both PNAD income estimates as well as estimates in the PNAD based on our imputed consumption measure. A first point to note i s that there i s considerable heterogeneity of estimated poverty rates across states, intotal as well as within urban and rural areas. The two approaches both clearly identify the Maranhiio and Piaui as the two poorest states in total. The high poverty in these two states i s attributable to both highrates inrural areas as well as inurban areas. Our two approachesto measuring poverty also agree that CBara, Alagoas and Bahia are among the next poorest group of four states; although the precise rankingwithin this group i s not the same for the two approaches. Consideringrural areas only, the two approaches reach some common conclusions (on the highpoverty in Ceara and Piauf) but they also indicate some clear differences (rural Paraiba i s least poor according to the PNAD consumption criterion, but is the third poorest state according to the PNAD income criterion). Considering non-metropolitan urban areas only, Maranhao stands out as most poor according to both criteria. Metropolitan areas are clearly less poor than both other urban and rural areas in the northeast, but this is less markedly the case in the Southeast, according to both the income and consumption criteria. In the southeast other urban areas are particularly low in the UF of Sao Paulo. This finding i s also common to both criteria. It i s important to note that, at this level of disaggregation, the sampling error has become sufficiently large that standard errors on the PNAD income-based measures are generally only slightly smaller than the PNAD consumption-based measures. The model error affecting the consumption based measures does not vary systematically with level of disaggregation, but sampling errors clearly rise. Eventually, at "Indeed,thelargemodelerrorfortheRSEestimateisattributabletotheparameterestimatesontheheterskedasticitymodelandit may be possible to obtaina specificationof this model whichi s more successful. 11 even lower levels of disaggregation, these would come to dominate in the overall standard errors of the consumption-based estimatesas well, and the two approacheswould become similarly (im)precise.18 It is useful to ask, given the size on the standard errors observed at this level of disaggregation, whether one i s clearly doing better in using the PNAD consumption-based estimates than one would have been able to with the PPV. Ifone restricted oneself to working only with the PPV data, then one's best estimate of poverty or inequality at the UFlevel would be the region level PPV estimate within which the UF is located. The question thus arises whether the point estimates in our PNAD consumption-based approach are sufficiently precise at the UF level to infer that the UF's do not all have the same regional average poverty rate. Figures 1-4 indicate (in turn for the rural northeast, other urban northeast, rural southeast and other urban southeast) that the point estimates from our PNAD consumption-based approach are often significantly different from the PPV-estimate for the region as a whole. Two-fifths to one half of UF-level poverty rates are significantly above or below the PPV estimate for their respective region as a whole. This indicates that the approach proposed here does offer a perspective on the spatial distribution of poverty that would not be achievable with only the PPV." Table 4 reports UF-level point estimates of inequality from the PNADbased on the income and imputed consumption criteria. Within the Northeast, both approaches find that inequality i s generally lower in ruralareas than in metropolitan or other urban areas. This pattern is not found, with either approach, in the Southeast. A difference between the two approaches is that according to the PNAD consumption- based inequality estimates, inequality in metropolitan areas of the northeast i s clearly lower than in other urban areas. The reverse i s found in the PNAD income-based estimates. In the southeast both approaches find higher inequality in metropolitan areas. Once again, although levels of measured inequality are markedly different, broad qualitative conclusions across the two approaches tend to be broadly similar, with only a few subtle differences. We briefly report, in Table 5, a further attempt at disaggregation. Here we confine ourselves only to poverty rates in the northeast, basedon the consumption measure of welfare imputedinto the PNAD. In this table we break up urban and rural areas further. In urban areas we draw a distinction between actually urbanized settlements, and those which have been delineated as urban but which may are still rather sparsely populated (a proxy for peri-urban areas). In rural areas we draw a distinction between those areas designated as rural but which are in fact somewhat built up, with certain minimal facilities and infrastructure, and those areas which are rather more remote and dispersed.*' Rural poverty i s unambiguously highest in dispersed, remote areas. This i s also where the bulk of the rural population resides. In those states which have a sizable rural population residing in built-up areas, poverty rates are generally markedly lower in those areas than in the dispersed regions. In states where the built-uprural sector is small, poverty rates are not particularly low. There is considerable variation across states in the distribution of poverty across these locations, with the unambiguous results that metropolitan areas are always least poor, followed by urbanized urban areas. The definition of peri-urban areas we have used does not appear to work terribly well, as it generally represents only a very small fraction of the urban population. '*Of course, as mentioned earlier, at very much lower levels of disaggregationidiosyncratic errors would also kick in, and the consumptionbasedstandarderrors would explode. l9Note that althoughonly 40-50% of UF's are significantly differentfrom the PPV regional-estimate,dependingon representative *'This region,this is an understatement of the numberof pairwisecomparisonsacrossUFsthat are significant. exercise is essentially intended as a cross-check on results reported in Ferreira and Lanjouw (2001) which implementeda very basic versionof the methodologyemployedhere. 12 6- Inequality Decompositions A final exercise we carry out in this study is a decomposition of inequality across different population subgroups, based on the PNAD data and our two different welfare concepts. As mentioned earlier, the General Entropy class of inequality can be readily decomposed into a within-group and between group component. With our parameter c=OS our decomposition takes the following form: where N individuals are placed in one of J groups subscripted byj,and the proportion of the population in the jth group, denoted 6,has weighted mean per-capita expenditure (or income) y j and inequality wj. The first term in this expression i s the inequality between groups and the second i s within groups. One can think of the share of the between group inequality to total inequality as the amount of inequality that i s due simply to differences in average expenditures between the groups. That portion of inequality that would remain if all differences across individuals within each group were to be eliminated.*l Table 6 reports the results from our decomposition exercise. We first take the country as a whole and ask how much of overall inequality i s attributable to the between-group component in a series of settings. We observe that if one breaks Brazil down into an urban and rural sector, that only 10-12 percent of overall inequality can be attributed to the difference in average consumption or income between these two sectors. Most of inequality would remain if this difference in averages would be removed. The conclusion holds irrespective of the welfare concept that i s being used. If the country were broken down into Northeast and Southeast only, then the between component rises slightly to 14-15% (once again remarkably similar across the two welfare concepts). When the country i s divided into four - urban northeast, rural northeast, urban southeast and rural southeast - the between-group component continues to rise slowly. Again, the two approaches give essentially the same result. Turning to the question of whether inequality i s largely attributable to differences between metropolitan areas and the rest of the country, we find that only 6 percent of inequality i s due to the difference in average welfare across these two sectors. Adding a further subgroup, other urban, raises the between group component to 13-14%. At the national level, it i s evident that much of overall inequality remains within the groups that have been considered here. An important point to note, given the purpose of this paper, i s that the decomposition results at the national level are qualitatively the same whether we use the PNAD income measure or our imputedconsumptionmeasure. When we look at rural areas only, we see that the two approachesto appear to give rather different results. The PNAD income approach suggests that rural inequality would fall by around 12% if the difference in average income between the northeast and southeast were removed. The consumption based approach in the PNAD suggests that the reduction in inequality would be higher - around 17%. If differences in average income across all states were removed, the PNAD income approach suggests that inequality would fall by approximately 16%. The consumption approach suggests that the fall in inequality would be about twice as high: 31%. The two approaches depart here in a quite significant way, with the consumption based approach suggesting that a much larger source of overall rural inequality i s due to differences across states in average rural incomes. It seems possible that the consumption approach i s capturing better the enormous distances and varied geography of the country, and the different agroclimatic conditions that i s associated with this. And the PNAD income measure may be failing to capture state variation because it i s more focused on formal sector earnings which might tend to be 2 'One minus this proportion can then be attributedto the share of inequalitythat is due to heterogeneitywithin the groups. 13 relatively homogeneous across states. It would seem worth exploring to what extent this finding i s robust to alternative choices of the parameter c, and possibly other measures of inequality. The final decompositions, within urban areas only, find again that the between group component i s quite modest irrespective of the approach that is beingused. 7- Conclusions This paper had two objectives. The first objective has been to demonstrate a methodology to impute a measure of consumption, as defined in the PPV household survey, into the much large PNAD household survey. The purpose of this exercise has been to estimate measures of welfare, such as poverty and inequality, defined in terms of consumption, at levels of disaggregation that are permitted by PNAD dataset. Although the results are still to be finalized we have shown that the methodology works quite well. We are able to validate the exercise at the representative region level in the PPV, and find that at that level, point estimates are very similar across the PPV and the PNAD. We have also shown that standard errors on the consumption-based point estimates in the PNAD are quite reasonable - certainly compared to the standardof typical household surveys. Our second objective has been to shed some light on the question of whether the analysis of poverty and inequality based on the PNAD income indicator yields different conclusions than an analysis based on consumption. We referred to the concern inthe literature on PNAD-baseddistributional analysis that the income measureinthe PNADmight suffer from serious biases. We have found that poverty and inequality, estimated on the basis of consumption in the PNAD, tend to be much lower than estimates based on the income concept. This i s not necessarily an indictment of income based analysis, however, as the two concepts of welfare are different and should not be expected to yield the same quantitative estimates. We demonstrated however, that differences in estimates of poverty and inequality between the PNADand the PPV are not attributable to non-comparability of these two surveys. Our PNAD consumption-based estimates are very close to those which obtain with the PPV. We pursued the comparability of income and consumption-based results further by examining whether there are important qualitative differences in the geographic profile of welfare across the two approaches. We found that, in fact, the two reach broadly similar findings. Inonly a few cases do we note differences across the two approaches that may need to pursued further. First, according to the consumption criterion, there is a clear basis for viewing metropolitan areas in the northeast as less poor than other areas. This distinction i s less clear-cut according to the income criterion. Second, within rural areas in the northeast, rural Paraiba i s least poor state according to the PNADconsumption criterion, but i s found to be the third poorest state according to the PNAD income criterion. Third, the PNAD consumption criterion finds that metropolitan areas in the northeast are markedly more equal than other urban areas in this region. The PNAD income criterion finds the reverse. Fourth, the consumption-based approach reflects much more strongly than the income-based one the contribution of differences in average incomes across states to overall rural inequality. Looking for differences in qualitative conclusions regarding the spatial distribution of poverty and inequality, may not be the best way to examine whether the income-based PNAD measures introduce important biases into distributional analysis in Brazil. As described in section 11, the PNAD income measure i s thought to be inadequately capturing income levels of certain population subgroups, notably those who are engaged in informal sector self-employment activities. A more effective direction to take might thus be to compare consumption-based estimates of poverty and inequality amongst population 14 subgroups defined in terms of occupations and education levels, rather than along geographic lines?* This seems an important next step. Still within a geographic focus, there would seem to be two promising directions for further work. First, it is important to examine whether the conclusions of Ferreira et a1 (2000) regarding the distribution of urban poverty across city-size i s robust to the application of a consumption-based indicator of welfare. It i s possible to link urban households in the PNAD to the size of conurbation in which they reside. Ferreira et a1observe a much higher incidence of poverty in smaller towns relative to large cities and metropolitan areas. Second, the results presented here suggest that there may be considerable variation in poverty rates within rural areas. So far we have split rural areas up quite crudely into dispersed and built-upareas. An important additional direction to take would be to divide rural areas into agro-ecological and climatic zones, as well as areas demarcatedby differences inaccessto facilities and infrastructure. The analysis in this paper has concentrated on the northeast and southeast of Brazil. As a result, some 25% of the population have not been included in the analysis. In principle it would be possible to extend the analysis carried out here, to regions such as the south and the center west of the country. But to do so would require making some important, unverifiable, assumptions. Because there are no PPV data applicable to these regions one would have to select a set of parameter estimates from the PPV data and impose the assumption that they are applicable for these regions which lie outside the PPV sampling domain. One might, for example, assume that the appropriate model to apply to the rural center west region i s a first stage model based on the combined rural northeast and southeast sample of the PPV. Similarly one might impose the rural southeast parameter estimates on the rural south PNAD data. This exercise i s possible but still pending. In the medium run there is a potential to apply the methodology reported here to the 2001 population census for Brazil. There are initiatives underway to implement a large new consumption survey in Brazil, covering both rural and urban areas. While the mooted sample size of 50,000 i s very large in absolute terms, it i s clear that these data will not permit disaggregations of poverty and inequality significantly below the UF level. If this new survey were to serve a basis for estimating first stage consumption models with which to impute consumption into the population census, it would then be possible to measure inequality and poverty, based on a consumption measure of welfare, at the town or village (and possibly neighborhood) level across the entire country. Such initiatives are being actively implemented and/or explored in a number countries in the last few years. Some, such as Mexico, Indonesia and China have total populations that, like Brazil, are very large. 22 Note the spatial dimension was a natural one to pursue in this paper given our interest to also validate results against the representative region level estimates in the PPV. References Bianchini, 2.M., and Albieri, S. (1998) "A Review of Major Household Sample Survey Designs Usedin Brazil", proceedings of the Joint IASS/IAOS Conference, Statistics for Economic and Social Development, September. Bourguignon, F. (1979) `Decomposable IncomeInequality Measures' Econometrica, 47:901:920. Camargo, J. M., and Ferreira, F.H.G. (1999) "A Poverty ReductionStrategy of the Government of Brazil: A RapidAppraisal", mimeo, Dept. of Economics, Catholic University of Rio de Janeiro. Cowell, F. (1980) `On the Structure of Additive Inequality Measures' Review of Economic Studies 47521-531. Deaton, A. (1997) The Analysis of Household Surveys: A Microeconometric Approach to Development Policy (World Bank: Johns Hopkins University Press). Elbers, C., Lanjouw, J.O., and Lanjouw, P. (2001) "Welfare inVillages and Towns: Micro-Level Estimation of Poverty and Inequality", mimeo, Development Economics ResearchGroup, the World Bank. Ferrerira, F.H.G., Lanjouw, P., and Neri, M.(2000) "A New Poverty ProfileFor Brazil UsingPPV, PNADand Census Data", Departamento de EconomiaPUC-Rio, TD #418, March. Ferreira, F.H.G. and Lanjouw, P. (2001) "Rural NonfarmActivities and Poverty inthe Brazilian Northeast", World Development Vol29, No. 3, pg509-528. Ferreira, F.H.G. and Litchfield, J. (1996) "Growing Apart: Inequality and Poverty Trends in Brazilinthe 1980s", LSE-STICERD -DARP Discussion PaperNo. 23, London (August). Ferreira, F.H.G. and Paes de Barros, R. (1999) "The Slippery Slope: Explaining IncreasesinExtreme Poverty in UrbanBrazil, 1976-1996", Brazilian Review of Econometrics, 19(2), 1999. Howes, Stephen and J. 0.Lanjouw (1998) "Does Sample Design Matter for Poverty Comparisons?' Review of Income and Wealth. Series 44, no. 1. pp. 99-109. Ravallion, M.(1994) Poverty Comparison (Chur: Hanvood Press). Shorrocks, A. (1980) `The Class of Additive Decomposable Inequality Measures' Econometrica 48: 613- 625. Soares de Freitas, M.,Nogueira Duarte, R., Carneiro Pessoa, D., Albieri, S. and do Naschimento Silva, P. (1997) `Comparando DistribuiGaes Et6rias em Pesquisaspor Amostragem: PNAD 1995 e PPV 96/97", mimeo Fundaqgo Instituto Brasileiro de Geografia e Estatistica, DPE, DEMET. World Bank (2001a) "Attacking Brazil's Poverty: A Poverty Report with a Focus on UrbanPoverty ReductionPolicies", World Bank, Washington D.C.. World Bank (2001b) "Brazil: Rural Poverty Report", World Bank, WashingtonD.C. 16 Table 1: Poverty Measuresby Regionfor Different Data Sets Headcount Source: PNAD 1996 and 1997,and PPV 1995. Note: 1. PNAD per capita income, PPV per capita consumption and PNAD imputed per capita consumption have been adjusted for spatial price variation (see Ferreira, Lanjouw and Neri, 2000). 2. Sampling errors incorporate adjustments for complex survey design (see text and also Howes and Lanjouw, 1998). 3. Poverty Line of R$65.07 in 1996 Ssio Paulo reais (see Ferreira, Lanjouw, and Neri, 2000). Table 2: Inequality Measures by Regionfor DifferentDataSets GeneralEntropy Classc=OS Source: PNAD 1996 and 1997, and PPV 1996. Note: PNAD per capita income, PPV per capita consumption and PNAD imputed per capita consumption have been adjusted for spatial price variation (see Ferreira, Lanjouw and Neri, 2000). 17 Table 3: PovertyEstimatesBy UFinthe Northeast: Headcount 1. PNAD per capita income, PPV per capita consumption and PNAD imputedper capita consumption have been adjusted for spatial price variation (see Ferreira, Lanjouw and Neri, 2000). 2. Sampling errors incorporate adjustments for complex survey design (see text and also Howes and Lanjouw, 1998). 3. Poverty Line of R$65.07 in 1996 Slo Paulo reais (see Ferreira, Lanjouw, and Neri, 2000). 18 Rural 0.193 0.024 0.076 0.028 Total 0.074 0.004 0.035 0.010 Southeast Urban 0. 105 0.015 0.046 0.015 Rural 0.383 0.003 0.271 0.026 Total 0.134 0.004 0.068 0.017 19 Statelsector PNADIncome PNADImputedConsumption I I GE0.5 I GE0.5 I Maranhzo (21) Urban 0.644 0.453 Rural 0.706 0.401 Total 0.681 0.426 Paiui (22) Urban 0.573 0.553 Rural 0.562 0.193 Rural 0.476 0.241 TotalI1 0.665 0.449 20 I State/Sector Table 4 cont: Inequality Estimates by UFinthe Southeast: General Entropy 0.5 I PNADIncome 1 PNADImputedConsumption I Sa0 Paulo (35) Metropolitan Region 0.496 0.316 Other Urban 0.435 0.285 Rural 0.380 0.380 Total 0.475 0.309 Rural Southeast Urban 0.535 0.315 Rural 0.490 0.444 Total 0.538 0.334 1. PNAD per capita income, PPV per capita consumption and PNAD imputedper capita consumption have been adjusted for spatial price variation (see Ferreira, Lanjouw and Neri, 2000). 21 Table 5: Incidence of Poverty inNortheast Brazil: By State and Location Type PNADImputed Consumption Union Federado Location Type Headcount Total Population CearA Metropolitan area 0.1757 5,200,874 Urban area 'urbanizadas' 0.43 14 3,822,943 Other urban 0.6054 58,101 Rural 0.6286 4,277,386 Other Rural (extensao urbana+povoado+nucleo) 0.3752 304,375 Total 0.3952 13,663,679 Pernambuco Metropolitan area 0.1363 6,143,549 Urban area 'urbanizadas' 0.3968 5,370,770 Other urban 0.4569 241,801 Rural 0.5169 3,050,896 Other Rural (extensao urbana+povoado+nucleo) 0.5350 155,883 Total 0.3167 14,962,899 Bahia Metropolitan area 0.2120 5,488,158 Urban area 'urbanizadas' 0.4363 10,506,684 Other urban Rural 0.4981 8,032,862 Other Rural (extensao urbana+povoado+nucleo) 0.3521 1,412,093 Total 0.4027 25,439,797 MaranhZo Urban area 'urbanizadas' 0.4811 4,664,666 Other urban Rural 0.6426 3,427,483 Other Rural (extensao urbana+povoado+nucleo) 0.3826 2,468,407 Total 0.5105 10,560,556 Piaui Urbanarea 'urbanizadas' 0.3829 3,144,276 Other urban Rural 0.5744 1,838,160 Other Rural (extensao urbana+povoado+nucleo) 0.4464 434,462 Total 0.4529 5,416,898 22 Rio Grande do Norte Urban area 'urbanizadas' 0.3013 3,344,104 Other urban Rural 0.5552 1,205,068 Other Rural (extensao urbana+povoado+nucleo) 0.4092 649,996 Total 0.3736 5,199,168 Paraiba Urban area 'urbanizadas' 0.2975 4,356,333 Other urban Rural 0.4471 2,068,372 Other Rural (extensao urbana+povoado+nucleo) 0.2099 248,292 Total 0.3406 6,672,997 Sergipe Urban area 'urbanizadas' 0.3494 2,268,819 Other urban 0.5542 80,481 Rural 0.5370 697,821 Other Rural (extensao urbana+povoado+nucleo) 0.5 172 230,398 Total 0.4062 3,277,519 A1agoas Urbanarea 'urbanizadas' 0.3264 3,359,234 Other urban 0.5625 61,935 Rural 0.5469 1,374,245 Other Rural (extensao urbana+povoado+nucleo) 0.4368 559,560 Total 0.3972 5,354,974 Source: PNAD 1996 and 1997 1. PNAD imputedper capita consumption has been adjusted for spatial price variation (see Ferreira, Lanjouw and Neri, 2000). 2. Poverty Line of R$65.07 in 1996SBo Paulo reais (see Ferreira, Lanjouw, and Neri, 2000). 23 Table 6: DecomposingInequality:PNADIncomeVersusPNAD Consumption GeneralEntropy Class(0.5) 3. UrbanAreas Total Inequality 0.5991 0.3859 % Attributableto BETWEENGroup Component Northeastvs Southeast 10.0 8.7 By State 12.0 10.7 24 Figure 1 Headcount Index Plus/Minus 1.645 Standard Errors UF-level estimates based on PNPD consumption Rural Northeast Headcount Index 0.8 0 . 7 1 0 . 6 TI TI -I T1 I I 1 I 0 . 5 L 1 1 T 0.4 0 . 3 0.2 0 10 States ranked from lowest headcount t o highest Figure 2 Headcount Index Plus/Minus 1.645 Standard Errors UF-level estimates based on PNPD consumption Urban Northheast Headcount Index 0.5 0.4 0 . 3 L 0 . 2 0.1 0 10 States ranked from lowest headcount t o highest 25 Figure 3 Headcount Index Plus/Minus 1.645 Standard Errors UF-level estimates based on PNm consumption Rural Southeast Headcount Index 0.5 0.4- TI 0.3- l- 0 . 2 0.1 0 . 0 0 5 States ranked from lowest headcount to highest Figure 4 Headcount Index Plus/Minus 1.645 Standard Errors UF-level estimates based on PNPD consumption Urban Southeast Headcount Index 0.1 0 . 0 0 5 States ranked from lowest headcount t o highest 26 2. BEYONDOAXACA-BLINDER:ACCOUNTINGFORDIFFERENCES HOUSEHOLD IN INCOME DISTRIBUTIONS ACROSSCOUNTRIES FranCois Bourguignon, Francisco H. G.Ferreira and Phillippe G.Leite23 Abstract This paper develops a micro-econometric method to account for differences across distributions of household income. Going beyond the determination of eamings in labor markets, we also estimate statistical models for occupational choice and for the conditional distributions of education, fertility and non-labor incomes. We import combinations of estimated parameters from these models to simulate countefactual income distributions. This allows us to decompose differences betweenfunctionals of two income distributions (such as inequality or poverty measures) into shares due to differences in the structure of labor market retums (price effects); differences in the occupational structure; and differences in the underlying distribution of assets (endowment effects). We apply the method to the differences between the Brazilian income distribution and those of the United States and Mexico, and 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. JEL ClassificationCodes:C15, D31,131, J13,J22 Keywords:Inequality, Distribution, Micro-simulations Introduction The distribution of personal welfare varies enormously across countries. The Gini coefficient for the distribution of household per capita incomes, for instance, ranges from 0.20 in the Slovak Republic to 0.63 in Sierra Leone (World Bank, 2002) and similar (or greater) international variation can be found for any alternative measure of inequality. Given that inequality levels within countries are generally rather stable, one would think that there ought to be considerable interest in understanding why income distributions vary so much across countries. I s it because the underlying distributions of wealth differ greatly, perhaps due to historical reasons?Or i s it because returns to education are higher in one country than in the other? What is the role of differences in labor market institutions? Do different fertility rates and family structures play a role? And if, as is likely, differences in income distributions reflect all of these (and possibly other) factors, in what manner and to what extent does each one contribute? Yet, applied research on differences across income distribution has not been as abundant as one might expect.24Increasingly, this seems to have less to do with lack of data and more to do with inadequate 23Bourguignonis with DELTA, Paris,andthe World Bank. Ferreira andLeite are at the Departmentof Economicsof the Pontificia UniversidadeCat6licado Rio de Janeiro.We thank DavidLam, DeanJolliffe, Klara Sabirianovaandseminar participants at PUC- Rio, IBMEC-Rio, the University of Michigan, the World Bank and DELTA for helpful comments; and Nora Lustig and Cesar Bouillon at the IDB for making the Mexican data available to us, ready to use. The opinions expressed here are those of the authors anddo not necessarily reflect those of the World Bank,itsExecutiveDirectorsor the countries they represent. 27 methodological tools. Through initiatives like the Luxembourg Income Study, the WIDER International Income Distribution Dataset and others, the availability of high-quality household-level data is growing. Methodologically, however, those seeking an understanding of why distributions are so different - and reluctant to rely exclusively on cross-country regressionswith inequality measures as dependent variables - have often resorted to comparing Theil decompositions across countries.25We will argue below that, while these can be informative, their ability to shed light on determinants of differences across distributions is inherently limited. Meanwhile, substantial progress has been made in our ability to understand differences in wage (or earnings) distributions. Some of this work, such as Almeida dos Reis and Paes de Barros (1991), Juhn, Murphy and Pierce (1993),Blau and Khan (1996) and Machado and Mata (2001),draws on variants of a decomposition technique based on simulating counterfactual distributions by combining data on individual characteristics (X) from one distribution, with estimated parameters (p) from another, which is due originally to Oaxaca (1973) and Blinder (1973).26Another strand, which includes DiNardo, Fortin and Lemieux (1996) and Donald, Green and Paarsch (2000), i s based on alternative semi-parametric coefficients - to generate counterfactual density functions that combine population attributes (or labor approaches. DiNardo et.al. (1996) use weighted kernel density estimators - instead of regression market institutions) from one period, with the structure of returns from another. Donald et. al. (2000) adapt hazard-function estimators from the spell-duration literature to develop density-function estimators, and use these to construct counterfactual density and distribution functions (comparing the US and Canada).27 These approacheshave been very fruitful, but they have not yet been generalized from wage distributions to those of household incomes, largely because the latter involve some additional complexities. The distribution of wages is defined over those currently employed. Taking the characteristics of these workers as given, earnings determination can be reasonably well understood by estimating returns to those characteristics in the labor market, through a Mincerian earnings equation: yi = Xip+E~.Most of the aforementioned recent literature on differences in wage inequality i s based on simulating counterfactual distributions on the basis of equations such as this, and many further restrict their samples to include prime-age, full-time male workers only. In addition, some authors are quite clear that they are interested in wages primarily as indicators of the price of labor, rather than as measures of welfare. Naturally, the distribution of household incomes also depends on the returns and characteristics of its employed members, and will thus draw on earnings models too. But it also depends on their participation and occupational choices and on decisions concerning the size and composition of the family. In addition, changes in some personal characteristics, such as education, affect household incomes through more than one channel. Suppose we ask what the effect of "importing" the US distribution of education to Mexico i s 24Theoretical models of why income distributions might differ across countries have been more abundant. Banejee and Newman (1993) andBCnabou (2000) are two well-knownexamples. See Aghionet. al. (1999) for a survey. 25Theil decompositions are known more formally as decompositions of GeneralizedEntropy inequality measures by population subgroups.They were developedindependentlyby Bourguignon(1979), Cowell (1980) andShorrocks(1980). 26Some of these studies, like Juhn, Murphy and Pierce (1993) and Machado and Mata (2001) decompose changes in the wage distributionof asingle country, over time. Others, like Almeida dos Reis andPaes de Barros(for metropolitanareas within Brazil) and Blau and Khan (for ten industrializedcountries) decompose differences across wage distributions for different spatial units. For a less well knownbut also pioneering work, see Langoni(1973). "Thedistinctionbetween"parametric"and"semi-parametric"methodsisnotterriblysharp.DiNardoet. al.(1996)useaprobit model to estimate one of their conditional reweighing functions. Donald et. al. (2000) rely entirely on maximum likelihood estimates of parameters in a proportional-hazards model, and what is non-parametric about their method is a fine double- partitioningof the income space, allowing for considerableflexibility inboththe estimationof the baselinehazardfunction, andin the manner in which it is shifted by the proportional-hazards estimates. Conversely, in the current paper, which follows a predominantly parametric route, some non-parametric reweighingof joint distributionfunctions is also used (see below). These techniques are often morecomplementarythan substitutable. 28 on the Mexican distributions of earnings and incomes. Whereas for earnings it might very well suffice to replace the relevant vector of X with US values, the distribution of household incomes will also be affected through changes in participation and fertility behavior. This greater complexity of the determinants of household income distributions seems to have prevented counterfactual simulation techniques from being applied to them, thus depriving those interested in understanding cross-country differences in the distribution of welfare from the powerful insightsthey can deliver. Nevertheless, a more general version of the Oaxaca-Blinder idea - of simulating counterfactual distributions on the basis of combining models estimated for different real distributions - can fruitfully be applied to household incomes. What i s required i s an expansion of the set of models to be estimated, to include labor market participation, fertility behavior and educational choices. In this paper, we first propose a general statement of statistical decompositions applied to household income distributions; and then suggest a specific model of household income determination that enables us to implement the decomposition empirically. In particular, we investigate the comparative roles of three factors: the distribution of population characteristics (or endowments); the structure of returns to these endowments, and the occupational structure of the population. We apply the method to an understanding of the differences between the income distributions in Brazil, Mexico andthe US.28 The paper i s organized as follows. Section 2 summarizes what can be learned from conventional comparisons of income distributions across these three countries, and presents an empirical motivation. Section 3 contains a general statement of statistical decomposition analysis, which encompasses all variants currently in use as special cases. Section 4 proposes a specific model of household income determination and describes the estimation and simulation procedures needed for the decomposition. The results obtained inthe case of the Brazil-US comparison are discussed in some detail in Section 5. Section 6 discussesthe Brazil-Mexico comparison and Section 7 concludes. 1- IncomeDistributioninBrazil,Mexico andthe UnitedStates. This section compares the distributions of household income in the three most populous countries in the Western Hemi~phere.~'The comparisons are based on an analysis of the original household-level data sets: the Pesquisa Nacional por Amostra de Domicilios (PNAD) 1999 i s used for Brazil; the Encuesta Nacional de Ingresos y Gastos de Hogares (ENIGH) 1994 for Mexico; and the Annual Demographic Survey in the March Supplement to the Current Population Survey (CPS) 2000, for the United States. As always with the March Supplement of the CPS, total personal income data refers to the preceding calendar year: 1999. Sample sizes for each data set (actually used) are as follows: the CPS 2000 contained 50,982 households (133,649 individuals); the ENIGH 1994 contained 6,614 households (29,149 individuals); and the PNAD 1999 contained 80,972 households (294,244 individuals). We use income, rather than consumption, data because the decompositions described in the remainder of the paper rely in part on the determination of earning^.^' InBrazil and Mexico, the income variable used was monthly total household income per capita, available in the surveys as a constructed variable from the disaggregated income questionnaire. Inthe US, the variable used was the sum (across individuals in the household) of annual total personal income and other incomes, excluding disability benefits, 28This approach i s a cross-country extension of a methodologypreviouslydevelopedto analyze the dynamics of the distributionof incomewithin asinglecountry. See Bourguignon,FerreiraandLustig (1998). 29Our emphasis here is purely comparative. We make no attempt to present a detailed analysis of inequalityor poverty in each of these countries. There is a large literatureon these topics for each of our three countries, but see Henriques (2000) for a recent compilation of work on Brazil, and SzCkely (1998) on Mexico. For earlier studies comparing the Brazilian and US earnings distributions, see LamandLevison(1992) and SacconatoandMenezes-Filho(2001). 30And also because consumption data for Brazil i s either very old (ENDEF, 1975) or incomplete in geographicalcoverage (POF, 1996;PPV, 1996). 29 educational assistance and child support, divided by 12.31All three income definitions are before tax, but include transfers. While total annual incomes are not top-coded in the CPS, some of their components might be. The US Census Bureau warns that weekly earnings, in particular, are "subject to top-coding at U$1923", so as to censor the distribution of annual earnings from the mainjob at U$lOO,OOO. Inspection of our sample revealed, however, that 2.1% (2.5%) of observations had reported weekly (annual) earnings above those value. The maximum reported weekly value was U$2884. We therefore did not correct for top-coding inthe US. Incomes are not top-coded in Brazil or Mexicoeither. As usual, there are reasons to suspect that incomes may be measured with some error. In the case of Brazil, the problem i s particularly severe in rural areas, to the extent that the usefulness of any estimate based on rural income data is thrown into For this reason, we prefer to confine our attention to urban areas only, in Brazil and Mexico.33Care i s taken to ensure that the distributions used are as comparable as possible, and this requires that we work with data unadjusted for misreporting, imputed rents, or for regional price level differences within countries. 34 Table 1below reports some key summary statistics of the income distributions for our three countries. In addition to population, GDP per capita and mean income from the household survey, three inequality measures are computed: the Gini Coefficient, the Theil T andL indices - inwhat follows, the last two are sometimes labeled E(1) and E(O), respectively, as members of the class of generalized entropy inequality as for a distribution of equivalised incomes, where the Buhmann et. al. (e = 0.5) equivalence scale i s used. measures. Each of these statistics is presentedfor the distribution of household income per capita, as well 35 All households are weighted by the number of individuals they comprise. Table 1:Descriptive Statistics Country Population GDPper capita Mean equivalised Gini Theil-T Theil-L (millions, (monthly, USD) income Coefficient 1999) (monthly, USD) 0 = 1.0 (household income per capita) Brazil 168 526.42 290.34 0.587 0.693 0.646 Mexico 97 643.25 280.90 0.536 0.580 0.511 USA 273 2550.00 1691.64 0.445 0.349 0.391 0 = 0.5 Brazil 168 526.42 551.08 0.560 0.613 0.572 Mexico 97 643.25 587.91 0.493 0.478 0.423 USA 273 2550.00 2791.78 0.415 Notes: Pouulation and GDP uer caoita figures are from World Bank (20011. The other figures are from calculations bv the . , - 0.298 0.344 authors from the household surveys. GDP per capita and mean equivalisedincome (MEY) are monthly and measured in 1999 I US dollars at PPP exchange rates. Mexicansurvey data is for 1994; Brazilian survey data is for 1999, andUS survey datais for 2000. Values of 0 are for the economy of scale parameterin the Buhmann et.al. (1988) equivalencescale - 0 = 1corresponds to income per capita. 31These income sources were excluded from the analysis because non-retirement public transfers are proportionately much more important in the US than in Brazil or Mexico, and their allocationfollows rules which are not modelledin our approach. When they were included, the residual term of the decompositionwas slightly larger, but all of our conclusions remainedqualitatively valid. 32For evidence on the weaknesses of income data for rural Brazil, see Ferreira, Lanjouw and Neri (2000) and Elbers, Lanjouw, LanjouwandLeite (2001). 33For the US, since the CPS does not disaggregate non-metropolitanareas into urban and rural, and the former dominate, we includedbothmetropolitanand non-metropolitanareas. 34 All three datasets are well-known in their respective countries. For more detailed information about the CPS, go to www.census.gov.Information on the PNAD i s available from www.ibge,crov.br. Information on the ENIGH is available from httR://www.ineg.i.gob.mx/. 35According to that method, the equivalised income of a householdwith income y and size Nis taken to be y/Ne. This definition coincideswith incomeper capitawhen6=1. 30 Similarities between Brazil (in 1999) and Mexico (in 1994) are immediately apparent. Across those different years, the two countries had broadly similar levels of GDP per capita. Mexico's was 22% higher than Brazil's , which pales in comparison to the difference between the two countries and the US: 384% higher than Brazil's. Brazil's inequality i s ranked highest by all three measures reported, followed by Mexico and the UnitedStates. The difference between Brazil's and Mexico's Ginis, at approximately five points, is not too large, while there are a full fourteen points between Brazil and the US. It i s interestingto note that the effect of allowing for (a good deal) of scale economies in household consumption differs across both countries and measures. Focusing on the Gini coefficient, the reduction in inequality in Mexico from reducing 8 from 1.O to 0.5 is larger than either inthe US or Brazil. The considerable differences in both mean incomes and inequality across these three countries must translate into different poverty levels as well. Table 2 below presents the three standard FGT36poverty measures for each country, based on the distribution of per capita household incomes. The first panel shows poverty rates for the entire countries, whereas the second panel shows them for urban areas only, which i s the universe for the analysis carried out in the next sections of the paper. In both cases, we use two alternative poverty thresholds. The first block in each panel employs an absolute poverty line, originally calculated as a strict indigence line for Brazil by Ferreira, Lanjouw and Neri (2000). Translated to 1999 values, it was set at R$74.48, or US$83.69 at PPP exchange rates. Having the lowest mean and the highest inequality of the three countries, Brazil has the most poverty by all three measures, in urban areas and overall. The United States has, by this ungenerous developing country standards, only traces of poverty. As for Mexico, it i s strikinghow much of its poverty i s rural: poverty incidence falls from 23% nationally, to less than 7% in urban areas. While being mindfulthat urban-rural definitions vary across countries, it would seem that poverty has an even more predominantly rural profile in Mexico than in Brazil. But when one considers welfare across countries at such different levels of development and per capita income as these three countries, a strong argument can be made that a relative poverty concept might be more appropriate. For this reason we also present the same poverty measures, in the same distributions, calculated with respect to a line set at half the median income in each distribution, in the second block of each panel. By these more relative standards, poverty in the US reaches a full quarter of the population, which happens to be quite similar to Brazil's urban incidence. Mexico's P(0) also rises to 15% in urban areas. FGT(a)measuresfor Urban and Ruralareas FGT(a) measuresfor Urbanareas ~ P(0) P(1) P(2) Poverty line I P(0) P(1) P(2) Poverty line I Brazil 29,18 12,lO 6,74 83,69 Brazil 22,33 8,40 4,37 83,69 Mexico 23,29 8,02 3,84 83,69 Mexico 6,66 1,52 0,51 83,69 USA 1,41 0,75 0,54 83,69 Brazil 30,02 12,22 6,82 84,27 Brazil 26,74 10,42 5,55 95,51 Mexico 17,86 5,59 2,57 70,ll Mexico 14,98 3,73 1,39 110,46 USA 25,02 10,19 5,92 687,70 Figure 1, which contains the Lorenz curves for the urban household income distributions for Brazil, Mexico and the US, i s a useful complement to the indices presented so far. Brazil i s Lorenz dominated by both Mexico and the United States, whereas those two countries, at least with only urban Mexico being 36Foster, Greer and Thorbecke (1984). In what follows, we use the three common measures of that family of poverty indices :P(O), the headcount,P(l), the povertygap andP(2), the cumulated squared gap. 31 considered, can not be Lorenz ranked. The Atkinson Theorem (1970) - which establishes the link between normalized second-order stochastic dominance and unambiguous inequality ranking - makes Lorenz Curves very useful diagrammatic tools to compare income distributions. Nevertheless, because they are two levels of integrationabove a density function, we can do even better interms of picturing the distribution. Figure 2 below plots kernel estimates of the (mean normalized) density functions for the distribution of (the logarithm of) household per capita income in our three countries. The greater dispersion of the Brazilian distribution is noticeable with respect to the Mexican, as is the greater skewness of the Brazilian and Mexican distributions, vis-&vis that of the United States. Figure1: UrbanLomz Curve For Brazil, Meximand the US. Figure2: Income Distributionsfor Brazil, Mexicoand The UnitedStates 0.60 7 Brazil 0 2 4 6 8 10 12 Log income Sources: PNADilBGE 1999, CPSiADS 2000, ENIGH 1994 Note: GaussianKernel Estimates(with optimal window width) of the density functions for the distributions of the iogarithmsof household per capita incomes.The distributionwere scaied so as to have the Braziiian mean. Brazil and Mexico are urbanareas only. Incomes were convertedto US doliar at PPP exchange rates. Finally, Table 3 reports on standard decompositions of E(O), E(1) and E(2) by population subgroups37, computing the RBstatistic developed by Cowell and Jenkins (1995). This statistic i s an indicator of the relative importance of each attribute used to partition the population, in the process of "accounting for" 37See Bourguignon(1979), Cowell (1980) and Shorrocks (1980). 32 attribute - rather than within those groups - the more likely it i s that something about the distribution of or the inequality. The idea is that the larger the share of dispersion which i s between groups defined by some returns to that attribute are causally related to the observed inequality. The attributes to be used include education of the household head (or main earner for the distribution of household incomes); his or her age; his or her race or ethnic group; his or her gender; as well as the location of the household (both regional and ruravurban) and its size or type. The results are suggestive. In Brazil, education of the head i s clearly the most important partitioning characteristic, followed by race and family type. In the US, family type dominates, with education a surprisingly low second, and age of head third. In Mexico, education and urbadrural vie for first place, with family type third. It i s clear that education accounts for more inequality in Brazil (and Mexico) than inthe US, althoughthis technique can not tell us whether this is due predominantly to different returns or different endowments of education - i.e. a different distribution of the population across educational levels. The greater role of the urbadrural partition in Mexico i s in line with our findings regarding total and urbanpoverty rates there. Strikingly little of overall US inequality is between different regions of the country, reinforcing the widespread perception of a well-integrated economy. This i s in contrast to the two Latin American countries, where some 10% of the Theil-L i s accounted for by the regional ,partition.38Finally, it i s interesting to note that inequality between households headed by people of different races - which one would expect to be prominent inthe US - i s five to six times as large inBrazil. Table3: Theil Decompositionsof Inequality by PopulationCharacteristics Brasil USA Mexico RB(0) RB(1) RB(2) RB(0) RB(1) RB(2) RB(0) RB(1) RB(2) Region 0,092 0,076 0,031 0,003 0,004 0,003 0,113 0,103 0,050 HouseholdType 0,126 0,121 0,060 0,192 0,210 0,155 0,194 0,180 0,092 Urban/ Rural 0,101 0,073 0,026 0,253 0,194 0,079 Genderof the Head 0.000 0,000 0,000 0,002 0,002 0,002 0,000 0,000 0,000 Raceof the Head 0,137 0,119 0,051 0,024 0,024 0,016 EducationLevel 0,266 0,316 0,213 0,129 0,133 0,093 0,247 0,255 0,150 Age Group 0,051 0,047 0,021 0,082 0,091 0,066 0,042 0,037 0,017 Note: Entries reflect share of overall inequality which is between subgroups for each partition. See Cowell and Jenkins (1995). But although this is a useful preliminary exercise, there are at least three reasons why one would wish to go further. First, none of these decompositions control for any of the others: some of the inequality between regions in Mexico i s also between individuals with different races, and there is no way of telling how much. Second, the decompositions are of scalar measures, and therefore "waste" information on how the entire distributions differ (along their support). Although some information can be recovered from knowledge of the different sensitivities of each measure, this i s at best a hazardous and imprecise route. Finally, even to the extent that one i s prepared to treat inequality between subgroups defined by age or education, say, as being driven by those attributes -rather than by correlates -the share of total inequality 38The regional breakdowns used in this decomposition were standard for each country. Brazil was divided up into five regions: North, Northeast, Centre-West, Southeast and South. Mexico was divided up into nine regions: "Noroeste", "Noreste", "Norte", "Centro Occidente", "Centro", "Sur", "Sureste", "Suroeste" and "Distrito Federal". The US was broken down into four regions: Northeast, Midwest, South and West. For a much more detailed analysis o f the importance o f regional effects in Mexican inequality, see Legovini, Bouillon and Lustig (2000). 33 attributed to that partition tells us nothing of whether it is the distribution of the characteristic (or asset), or the structure of its returns that matters. In the next section, we propose an alternative approach, which suffers from none of these shortcomings. 2- A General Statement of Statistical Decomposition Analysis. In order to understand the differences between two distributions of household incomes, fA(y) and fB(y), it seems natural to depart from the joint distributions (pc(y, T), where T i s a vector of observed household characteristics, such as family size, the age, gender, race, education and occupation of each individual member of the household, etc.. The superscript C (= A, B) denotes the country. Because a number (but not all) of the characteristics in T clearly depend on others (e+ family size, via the number of children, will vary with the age and education of the parents), it will prove helpful to partition T = [V, W] where, possibly on some other elements of vh, but for any given household h in Cyeach element of Vh may be thought of as logically depending on wh, and Wh is to be considered as fully exogenous to the household. The distribution of household incomes, p(y), i s of course the marginal distribution of the joint distribution cp"(y, T) : f "(y) = JJ[p"(y,T)dT. It can therefore be rewritten as f"(y)= JJJg"(ylV,W~"(V,W)dVdW, where gc(y V,W) denotes the distribution of y conditional 1 on V and W, and $'(V, W) i s the joint distribution on all elements of T in country C. Given the distinction made above between the "semi-exogenous"39 household characteristics V and the "truly exogenous" characteristics W, this can be further rewritten as: In (l), joint distribution of all elements of T = [V,W] has been replaced by the product of v the conditional distributions and the joint distribution of all elements in W, $(W). Each conditional distribution h, is for an element of V, conditioning on the v-n elements of V not yet conditioned on, and on W. The order n = { 1,...v) obviously does not matter for the product of the conditional distributions. (1) is an identity, invariant in that ordering. However, the order does matter for the definition of each individual conditional distribution h,(vnlV-l,...,,, W), and therefore for the interpretation of each decomposition defined below.40 Once we have written the distributions of household incomes for countries C = A, B as in (l), one could investigate how fB(y) differs from fA(y) by replacing some of the observed conditional distributions in the ordered set kA= {gA,hA}by the corresponding conditional distributions in the ordered set kB= {gB,hB}. Each such replacement generates a counterfactual (ordered) set of conditional distributions ks,the dimension of which is u+l,(like kAand kB)whose elements are drawn either from kAor kB.It i s now possible to define a counterfactual distribution PA,&; k', $) as the marginal distribution that arises from the integration of the product of the conditional distributions in kSand the joint distribution function $(W), with respectto all elements of W. As an example, the counterfactual distribution PA+B(y; gA,hlB, h.lA,$) is given by: 39This terminologyi s motivated by the fact that we do not pretendthat our modelsof V shouldbeinterpretedcausally, and makeno claims to be endogenizingthese variablesina behaviouralsense. 40Shorrocks (1999) proposes an algorithmbasedon the Shapley Value in order to calculate the correct "average"contribution of a particular h,( ) or of g( ), over the set of possible orderings, to the overall difference across the distributions. Rather than constructingthese values in this paper, we present our results by showing a number of different orderings explicitly in Sections 5 and6 below. 34 such counterfactual distributions i s the number of possible combinations of elements of the set k, i.e. the dimension of its sigma-algebra:' For each counterfactual distribution, it i s possible to decompose the observed difference in the income distributions for countries A and B as follows: f (Y1- f (Y = (Y1- f (Y " lf A ,I+lf (Y>- f s (Y11 where the first term on the right-hand side measuresthe "explanatory power" of decomposition s, and the second term measures the "residual" of decomposition s:* Since these are differences in densities, they can be evaluated for all values of y. Furthermore, any functional of a density function can be evaluated for fA, fB or B, and similarly decomposed, according to its own metric. So, we have the same decomposition relationship as (2) for the cumulative distribution Y 1F,'(q+l) F C ( y )= [f`(x)&. Likewise, for the mean income of quantile q: p,"(y) =- [yf` (y)dy , we 0 Q F,`(q) have: And we have analogous decompositions for any inequality measure I(f(y)) or poverty measureP(f(y); 2). Inthe applications discussedinSections 5 and 6, the results arepresentedexactly inthis form: Tables 5 and 7 contain inequality and poverty measures, evaluated for fA(y), fB(y) and for a set of counterfactual distributions P(y), so that the reader can make his own subtractions. Figures 4-8 and 10-14 plot the differences in the (log) mean income of "hundredths" q E [l, 1001, in a graphical representation of Equation (3). In recognition of their parentage, we call these the Generalized Oaxaca-Blinder decompositions. 3- The DecompositionsinPractice: A Specific Model The essence of the approach outlined above i s to compare two actual income distributions, by means of a sequence of "intermediate" counterfactual distributions. These are constructed by replacing one or more of the underlying conditional distributions of A by those imported from B. In practice, this requires generating statistical approximations to the true conditional distributions. This may be done either through parametric models - following the tradition of Oaxaca (1973), Blinder (1973) and Almeida dos Reis and Paes de Barros (1991) - or through non-parametric techniques - as in DiNardo, Fortin and Lemieux (1996):3 Because of the direct economic interpretations of the parameter estimates in our approximated distributions, we find it convenient in this paper to follow (mainly) the parametric route, by 4'When we turn to the empiricalimplementationof these counterfactual distributions,we will see that is also possible, of course, to simulate replacing the joint distribution@(y) by a non-parametric approximationof wB(y). Depending on how each specific conditionaldistributionis modelled, it is also possibleto have more than one counterfactualdistribution per element of k. These matterspertainmoreproperly to a discussionof the empiricalapplicationof the approach, however, andwe retumto themlater. 42A decompositionis defined(by (2)) with respect to auniquecounterfactualdistributions, andi s thus also indexedby s. 43 Although, as noted earlier, these authors too rely on parametric approximations to some conditional distributions, such as the probit for the conditionaldistributionof unionstatus on individual characteristics. 35 approximating each of the true conditional distributions through a set of standard econometric models, with pre-imposed functional forms.44 Inparticular, we will findit convenient to proposetwo (sets of) models: (4) y = G (V, W, E; R) and f.u(S) V = H(W, q; O), where R and O are sets of parameters and E and q stand for vectors of random variables, with &I{V,W}, and qIW, by construction. G and H have pre-imposed functional forms. We can then write an approximationf(y) to the true marginal distribution fc(y) inEquation (1) as: (1') C(V,w,;R)=y where 7cY(&) i s the joint probability distribution function of E and ~"(q)s thejoint probability distribution i function of q. Just as an exact decomposition was defined by (2) for each true counterfactual distribution, we can now define the (actually operational) decomposition s in terms of the approximated distributions f *(y), as follows: Recall that a counterfactual distribution s i s conceptually given by FA+&; kS, (up up),andi s thus definedby and) the simulated sequence of conditional distributions kS,which consists of some original distributions from A, and some imported from B. Analogously, an approximated distribution fi2B(y;RS,W,YA)i s defined with respect to (UP and) the two sets of simulated parameters Rsand Os,which consist of some original parameters from the models estimated for country A, and some imported from the models estimated for country B. The last term in (2') gives the difference between the approximated and the true counterfactual distribution We therefore call it the approximation error and denote it by RA. Clearly, how useful this decomposition methodology i s in gauging differences between income distributions depends to some extent on the relative size of the approximation error. The applications in the next two sections illustrate that it can be surprisingly small. Followingfrom (l'), our statistical model of household incomes has three levels. The first corresponds to model G (V, W, E; R), which seeks to approximate the conditional distribution of household incomes on observed characteristics: g(yl V,W). This level generates estimates for the parameter set R, which we associate with the structure of returns in the labor markets and with the determination of the occupational structure in the economy. The second level corresponds to model H (W, q; 0) which seeks to approximate the conditional distributions hn(vnlV-,,..,,,,, W), for V ={numberof children in the household (rich); years of schooling of individual i (Eih); and total household non-labor income (yo),)} In the third level, we investigate the effects of replacing @(W) with a (non-parametric) estimate of g(W). This largely corresponds to the racial and demographic make-up of the population. This i s an advantage of our approach vis-&-vis, for instance, the hazard-function estimators of Donaldet. al. (2000), who "note that the estimatesof the hazard function for wages, eamings or incomes are difficult to interpret" (p.616) 36 First-level model G (V, W, E; 0) i s given by equations (6-8) below. Household incomes are an aggregation of individual earnings Yhi,and of additional, unearned income such as transfers or capital income, yo.Per capita household income for householdh i s given by: where ILi i s an indicator variable that takes the value 1 if individual iin household h participates in earning activity j, and 0 otherwise. The allocation of individuals across activities (i.e. labor force participation and the occupational structure of the economy) i s modeled through a multinomial logit of the form: j # s where Ps() i s the probability of individual iin household hbeing in occupational category s, which could be: inactivity, formal employment in industry, informal employment in industry, formal employment in services or informal employment in services. Separate but identically specified models are estimated for males and females. The vector of characteristics Z c T i s given by Z = { 1, age, age squared, education dummies, age interacted with education, race, and region for the individual in question; average endowments of age and education among adults in his or her household; numbers of adults and children inthe household; whether the individual is the heador not; andifnot whether the headis active}. As is well known, the multinomial logit model may be interpretedas a utility-maximizing discrete choice model where the utility associated with choicej i s given by ULi=Zhi.Aj+E; .The lasttermstandsfor unobserved choice determinants of individual i,and it i s assumedto be distributed according to a double exponential law in the population. We prefer, however, not to insist on this utility-maximizing interpretation of the multi-logit and to treat it merely as a building block of the statistical model G, defined in equation (4). Turningto the labor market determination of earnings, yLi in (6) is assumedto be log-linear in a, and pj, and the individual earnings equation is estimated separately for males and females, as follows: logy;, = aj + X h i P j +Ei (8) where x c T i s given by x = {education dummies, age, age squared, age * education, and intercept dummies for region, race, sector of activity and formality status}. In the absence of specific information on experience, the education and age variables are the standard Becker - Mincer human capital terms. The racial and regional intercept dummies allow for a simple level effect of possible spatial segmentation of the labor markets, as well as for the possibility of racial discrimination. Earning activities are defined by sector and formality status. To simplify, it i s assumed that earnings functions across activities also differ only through the intercepts, so that the sets of coefficients pj are the same across activities (pj = p). We interpret these p coefficients in the usual manner: as estimates of the labor market rates of return on the corresponding individual characteristics. 37 This first level of the methodology generates estimates for the set R, comprising occupational choice parameters h, and (random) estimates of the residual terms E: 45, as well as for aj and p and for the variance of the residual terms, Om,Oq . 2 2 Inthe secondlevel of the model, H(W,q; a),weestimatetheconditionaldistributionsofV ={number of children in the household (rich); years of schooling of individual i(Ei& and total household non-labor income (YOh)} on W = {number of adults in the household (nab), its regional location (rh),individual age (Aih), race (Rih) and gender (gih)}. This is done by imposing the functional form associated with the multinomial logit (such as the one in Equation 7) on both the conditional distribution of Eihon W: MLE (E A, R, r, g, nab) and on the conditional distribution of the number of children in the household on {E, I w):M L(nch E,A, R, r,g, nh). ~ I Unlike Equation (7), these models are estimated jointly for men and women. The educational choice multilogit MLEhas as choice categories 1-4; 5-6; 7-8; 9-12; and 13 and more years of schooling, with 0 as the omitted category. Estimation of this model generates estimates for the educational endowment parameters, y. The demographic multilogit M k has as choice categories the number of children in the household: 1, 2, 3, 4 and 5 and more, with 0 as the omitted category. Estimation of this model generates estimates for the demographic endowment parameters, w. Finally, the conditional distribution of total household non-labor incomes on {E, W} i s modelled as a Tobit: T (y IE,A, R, r, g, nh).46Estimation of this model generates estimates for the non-human asset endowment parameters, 5. These three vectors constitute the set of parameters@={ y, y ~ , t}. After each of these reduced-form models has been estimated for two countries (Brazil and a comparator nation), the approximate decompositions in (2') can be carried out. Each decomposition i s based on the construction of one approximated counterfactual distribution fi2B(y; Qs,CD',Y '), defined largely by which set of parameters inQA and @A i s replaced by their counterparts in QB and @*. All of our results in the next two sections are presented in this manner. Tables 5 and 7, for example, list mean incomes, four inequality measures and three poverty measures for a set of approximated counterfactual distributions, denoted by the vectors of parameters which were replaced with their counterparts from B. Similarly, Figures 4-8 and 10-14 draw differences in log mean quantile incomes between actual and approximated counterfactual distributions, where these are denoted by the vectors of parameters which were replaced with their counterparts from B to generate them. As an example, consider line 4 of Table 5 (denoted "a, p, and d").It lists the mean income and the inequality and poverty measures calculated for the distribution obtained by replacing the Brazilian a and p in equation (8), with those estimated for the US; scalingup the variance of the residual terms by the 45Fordetailsonhow the latter maybedetermined,see Bourguignon, FerreiraandLustig(1998). 46We also experimented with an alternative approximationfor the conditional distributionof non-labor incomes. This was a (non- parametric)rank-preservingtransformation of the observeddistributionof yo,conditionalon earned incomes in each country. In practicalterms, we ranked the two distributionsby per capita householdearned income ye = yh --.Yo If p = FB( y e ) ` h was the rank of householdwith income yein country B, then we replaced yOpwith the unearnedincome of the householdwith B the same rank (by earned income) in country A, after normalizingby mean unearnedincomes: y, which are available from the authors on request, were similar in direction and magnitude to those of the parametric exercise reportedin the text. 38 ratio of the estimated variance in the US to that of Brazil; and then predicting values of Yih for all individuals in the Brazilian income distribution, given their original characteristics (I@). The density function defined over this vector of predicted incomes i s fi2B(y; ,a',Y ) R' A for Rs={d,pB,02B,AA,??A}=OA. and Os Whenever ABE as,individuals may be reallocated across occupations. This involves drawing counterfactual E"'S from censored double exponential distributions with the relevant empirically observed variances.47The labor income ascribed to the individuals who change occupation (to a remunerated one) i s the predicted value by equation (8), with the relevant vector of parameters, and with E'S drawn from a Normal distribution with mean zero and the relevant variance. And when 0' # BA,so that the values of the years of schooling variable and/or the number of children in households may change, these changes are incorporated into the vector V, and counterfactual distributions are recomputed for the new (counterfactual) household characteristics. As the discussion in the next two sections will show, the interactions between these various simulations are often qualitatively and quantitatively important. The ability to shed light on them directly and the ease with which they can be interpreted are two of the main advantages of this methodology. The third and final level of the model consists of altering the joint distribution of the truly exogenous household characteristics, yfc(W). The set W i s given by the age (A), race (R), gender (g) of each adult individual inthe household, as well as by adult household size (n*) andthe region where the household i s located (r).Since these variables do not depend on other exogenous variables inthe model, this estimation i s carried out simply by re-calibrating the population by the weights corresponding to the joint distribution of these attributes inthe target country.48 Inpractice, this is done by partitioning the two populations by the numbers of adults inthe household. To remain manageable, the partition i s in three groups: households with a single adult; households with two adults; and households with more than two adults. Each of these groups i s then further partitioned by the race (whites and non-whites) and age category (six groups) of each ad~lt.4~The number of household in each of these subgroups can be denoted Man;:,where a stands for the age category of the group, for the Y the structure from country A (population of households p)to country B (population of households F), race of the group, n for the number of adults in the household, and C for the country. If we are importing we then simply re-scale the household weights inthe sample for country B by the factor: Results for this final level of simulations are reportedin Tables 5 and 7 under the letter $. 4- The Brazil-US Comparison. The decompositions described in the previous section were conducted for differences in distributions between Brazil in 1999 and the United States in 2000. The estimated coefficients for equations (7) and 47The censoring of the distributionfrom which the unobservedchoice determinantsare drawn is designed to ensure that they are consistent with observedbehaviourunderthe altemative vector 1.See Bourguignon,FerreiraandLustig (1998) for details. 48The spiritofthis procedurei s very muchthe same as inDiNardoet. al. (1996). 49In the case of households with more than two adults, this is done for two adults only: the head and a randomly drawn other adult. In this manner, the group of single adult households is partitionedinto 12 sub-groups, and the other two groups into 144 sub- groups each. 39 (S), as well as those for the multinomial logit models for the demographic and educational structures and the tobit model of the conditional distribution of non-labor incomes are included in Tables A1 - A5, in the Appendix. Table 4 - at the end of the paper - presents the results for importing the parameters from the US into Brazil, in terms of means and inequality measures for the individual earnings distributions, separately for men and women. Table 5 displays analogous results for household per capita incomes, and includes also three poverty measure^.^' Figures 4 to 8 present the full picture, by plotting differences in log incomes between the distributions simulated in various steps and the original distribution, for each percentile of the new di~tribution.~' Looking first at individual earnings, the observed differences between the Gini coefficients in Brazil and the US are nine points for men, and ten for women. Brazil's gender-specific earnings distributions have a simulating Brazilian earnings with the US ct and p parameters) account for half of this difference. As we Gini of 0.5, whereas those of the US are around 0.4. Roughly speaking, price effects (identified by shall see, this i s a much greater share than that which will hold for the distribution of household incomes per capita. Among the different price effects, the coefficient on the interaction of age and education stands out as makingthe largest difference. Differences in participation behavior are unimportant in isolation. Importing the US participation parameters only contributes to reducing Brazilian earnings inequality when combined with importing US prices, as may be seen by comparing the rows a,P (viii) and the row h,a,P. Educational and fertility choices are more important effects. The former raises educational endowments and hence both increases and upgrades the sectoral profile of labor supply. The latter leads to increased participation rates by women. This effect accounts for nearly all of the remainingfour to five Gini points. As one would expect, demographic effects are particularly important for the female distribution, where, in combination with the prices. Reweighing the purely exogenous endowments - including race - has no effect. effect of education, it reduces the Brazilian Gini by a full five points even before any changes are made to Table 5, which reports on the simulations for the distribution of household incomes per capita, can be read in an analogous way. The first two lines present inequality and poverty measures for the actual distributions of household per capita income by individuals in Brazil (in 1999) and the US (in 2000). In terms of the Gini coefficient, the gap we are trying to "explain" i s substantial: it i s twelve and a half points higher in Brazilthan inthe US. The difference i s even larger when the entropy inequality measures E() are used. 50In order for the poverty comparisons to make sense across two countries as different as the US and Brazil, the US earnings distributionswere scaleddown so as to have the Brazilianmean.This was done by appropriately adjustingthe estimate for a'', as can be seen from the means reported in Tables 4 and 5. Accordingly, counterfactual poverty measures are not reported for simulations which do not include an CY estimate. The same procedure was used in Section 6, to rescale the Mexican earnings distributions to havethe Brazilianmeans. 51Analogous figures for differences in log incomesby percentiles rankedby the original distribution- which show the re-rankings inducedby each simulation- are available from the authorson request. 40 Table 5 : SimulatedPoverty aidInequalityfor Brazil in1999,Using2000 USA coefficients. Mean Poverty PfC Inequality 2= m e d i dpermonth Income Gm w 41) EIZ) PIO) P(1) P(2) 1 Brad 2948 0569 0597 0644 1395 2623 1010 536 2 USA 294.8 0.445 0.391 0.349 0.485 25.02 10.19 5.92 294.9 0,516 0.486 0.515 1.049 20.32 733 3.92 294.9 0.530 0,517 0.545 1.119 21.92 8.39 4.46 277.9 0.579 0.632 0.653 1.313 255.4 0.535 0.536 0.542 1.022 28.06 11.58 6.46 255.5 0.548 0.565 0,572 1.093 29.59 12.50 7.06 454.0 0.505 0.489 0.460 0.719 283.9 0.480 0.425 0.425 0.732 18.81 7.12 3.75 283.9 0.494 0.453 0.452 0.786 20.33 7.84 4.18 469.0 0.511 0.514 0.467 0.711 274.2 0.490 0.450 0.445 0.780 21.1s 8.36 4.54 274.2 0.505 0.480 0.474 0.837 22.73 9.19 5.07 295.2 0,576 0.613 0.663 1.449 464.6 0.505 0.493 0.454 0.686 287.1 0.486 0.437 0.434 0.746 19.31 7.31 3.85 287.1 0.499 0.464 0.459 0.794 20.85 8.09 4.35 507.2 0.500 0.492 0.441 0.641 299.2 0.481 0.433 0.423 0.709 18.14 7.00 3.75 299.2 0.495 0.462 0.448 0.755 19.59 7.77 4.24 317.5 0.534 0.531 0.551 1.144 20.58 7.97 4.32 356.3 0.428 0.353 0.315 0.416 11.17 4.33 2.38 404.7 0.585 0.637 0.683 1.496 387.7 0.511 0.490 0.489 0.874 14.35 5.43 2.88 23 Y, a,@e2;t5 4364 0.432 0.359 0.325 0.448 8.14 3.11 171 Source: PNAD 1999 andCPS March2000 The first block of simulations suggests that differences in the structure of returns to observed personal characteristics in the labor market can account for some five of these thirteen points?* When one disaggregates by individual os, it turns out that returns to education, conditionally on experience - as for individual earnings - play the crucial role. Overall, it can thus be said that difference in returns to schooling and experience together explain approximately 40 per cent of the difference in inequality between Brazil and the US. The order of magnitude i s practically the same with E(l) and E(2) but it i s higher with E(O), suggesting that the problem is not only that returns to schooling are relatively higher at the top of the Brazilian schooling 52The relative importance of each effect vanes across the four inequality measures presented, but the orders of magnitude are broadly the same, and the main story could be told from any of them. All are presentedin Table 5, but we use the Gini for the discussionin the text. 41 scale but also that they are relatively lower at the bottom. This is confirmed by the fact that importing US prices lowers poverty in Brazil, even though (relative) poverty i s initially comparable in the two countries. Importing the US variance of residuals goes in the opposite direction, contributing to an increase of almost 1.5 Gini points in Brazilian ineq~ality.~~Two candidate explanations suggest themselves: either there i s greater heterogeneity amongst US workers along unobserved dimensions (such as ability) than among their Brazilian counterparts, or the US labor market is more efficient at observing and pricing these characteristics. This is an interesting question, which deserves further investigation. In the absence of additional information on, say, the variance of I Q test results or other measures of innate ability, orthogonal to education, we are inclined to favor the secondinterpretation. It may be that the lower labor market turnover and longer tenures that characterize the US labor market translate into a lesseneddegree of asymmetric information between workers and managers in that country, with a more accurate remuneration of endowments which are unobserved to researchers. We thus consider the o2effect as a price effect, which dampens the overall contribution of price effects to some 3.5 to 4 points of the Gini. The next block shows that importing the US occupational structure (A) by itself, has almost no impact on Brazilian inequality, but lowers average incomes and raises poverty. This i s a consequence of the great differences inthe distribution of education across the two countries, as revealed by Figure 3 below. Since education i s negatively correlated with inactivity, and positively with employment in industry and with formality in the US, when we simulate participation behavior with US parameters but Brazilian levels of education, we withdraw a non-negligible number of people from the labor force, and 'downgrade' many others. Figure 5 shows the impoverishingeffect of imposingUS occupational choice behavior, combined with its price effect, on Brazil'soriginal distribution of endowments. Turning to the second-level model, H(W, q, O),we see further support for the aforementioned role of education in determining occupational choice. When U S educational parameters are imported by themselves, this raises education levels in Brazil substantially, thus significantly increasing incomes and reducing poverty. Education endowments increase more for the poor (as expected by the upper-bounded nature of the education distribution), and inequality also falls dramatically. The y simulation alone takes six points of the Gini off the Brazilian coefficient and, crucially, takes the impoverishing effect away from the occupational structure simulation. The latter result suggests that the most important difference in the distribution of educational endowments between Brazil and the US might actually be in the lack of minimumcompulsory level inBrazil-see figure 3. At this stage, it might seem that almost all of the difference in inequality between the US and Brazil is explained by education-related factors. Six points of the Gini are explained by the differences in the distribution of education and five points by the difference inthe structure of earnings by educational level (that is, the coefficients of the earning functions). Yet, when these changes - i.e. a, p and y - are simulated together, as in row 8a in table 5, it turns out that their overall effect i s not the sum of the two effects (eleven points), but only eight points. The two education-related effects, distribution and earnings structure, are therefore far from being additive. The same i s true of the decomposition of earnings inequality in Table 6. The explanation for this non-additivity property i s straight-forward. As can be seen in figure 3, only a tiny minority of US citizens have fewer than 9 years of education, whereas practically 60% of the Brazilian population do. At the same time, the structure of US earnings for the few people below that minimum level of schooling i s approximately flat, possibly because of minimum wage laws. In Brazil, on the 53This result i s in line with the earlier findings of Lam and Levinson (1992), who noted that the variance of residuals from earnings regressionssuch as these was considerably higher in the US than inBrazil. 4L contrary, earnings are strongly differentiated over that range. People with less than full primary education earn on average 70% of the mean earnings of people with some secondary education.54This proportion i s 95 per cent for the few people with such a low level of schooling in the US. Thus, importing the earnings structure from the US to Brazil contributes to a drastic equalization of the distribution when the demographic structure of education of Brazil remains unchanged. Many people with less than secondary education are then paid at practically the same rate as people with completed secondary. Doing the same exercise with the US demographic structure of education has much less effect, because there are very few people in that country with less than secondary. This appears clearly in table 5 when comparingrows 1and 3 on the one hand, and rows 8 and 8a on the other. The basic effect of switching to USearnings when the US demographic educational structure is used comes from the fact that the relative earnings of college versus high school graduates i s substantially higher inBrazil. The question which remains is: how much of the excess inequality in Brazil with respect to the US i s due to the distribution of education, and how much i s due to the structure of schooling returns.55The foregoing argument makes it tempting to place greater weight on the distribution of education effect. This is becausethe structure of educational returns at low schooling levels is relevant to very few people in the US, and yet it has such an important effect when imported to Brazil. One may also hold that the structure of returns actually reflects the educational profile of both populations. There are positive returns at the bottom in Brazil because many people inthe labor force have zero or a very low level of schooling, whereas this i s exceptional in the US. There are also larger returns in Brazil at the top of the schooling rangebecausethere are relatively fewer people with a college education. Figure 3: Distribution of education across the countries 0 y.s, 1 to 4 y.s. 5 to 6 y.s. 7 to 8 y s . 9 to 12 y.s. 13 or more Group of years of schooling loBrazil .U.S. OMexicol Sources: PNAD/IBGE 1999, CPS/ADS 2000, ENIGH 1994 54These figures refer to meanearningsby educationallevelanddiffer from what maybeinferredfrom the regressioncoefficientsfor schoolingintable A2. 55This is not a new question. Infact, it was at the heart of the public debate about the causes of increasinginequalityin Brazil during the 1960s. See Fishlow(1972) andLangoni(1973) for differentviews on the matter at that time. 43 Moving on to demographic behavior, we observe a similar role for education. As with occupational structure, importing y~ alone hardly changes inequality - it would even increase it slightly. However, fertility is negatively correlated with educational attainment, particularly of women. If the change in fertility were taking place in the Brazilian population with US levels of schooling and participation behavior, inequality would drop by 1percentagepoint of the Gini coefficient and poverty would fall. This seems to mean that fertility behavior differs between the two countries mostly for lowest educated households. When the effects of some of the "semi-exogenous" endowments (embodied in the approximations to the educational and demographic counterfactual conditional distributions) are combined with occupational structure and price effects (as inthe row for y ~A, , y a, p, d),we see an overall reduction of seven points in the Gini. Most of this (around five points) seems to be associated with adopting the US endowments of education, either directly or indirectly, through knock-on effects on participation and fertility. The remainder i s due to the price effects.56This still leaves, however, some additional five Gini points - a rather substantial amount - inthe difference in inequality between the two countries unexplained. Figure 7 illustrates the results of the combined simulations for the entire distribution: while the simulated line has moved much closer to the actual (log income percentile) differences, it i s not yet a very good fit. Of the various candidate factors we are considering, two remain: the differences in the joint distributions of exogenous observed personal endowments: @(W) and @(W); and non-labor incomes. The two final blocks of simulations show that it i s the latter, rather than the former, that accounts for the remaining inequality differences. While reweighing the households in accordance with Equation (9) actually has an overall simulation by about one and half Gini points - importing the conditional distribution of non-labor increasing effect on Brazilian inequality (see line 21) - thus weakening the explanatory power of the incomes has a surprisingly large explanatory power. As may be seen from line 20 of Table 5, it actually moves the simulated Ginicoefficient for Brazilto within 1.7 Ginipoint of the true US Gini. When reweighing the joint distributions of exogenous observed personal endowments i s combined with all the previous steps, in line 23, the difference i s further reduced to 1.3 Gini points It also does differences for two different counterfactual distributions with non-labor incomes - one with and the other remarkably well by all other inequality measures in Table 5. Figure 8 shows the simulated income without reweighing. The fit with regard to the actual differences i s clearly much improved with respect to the preceding simulations, and it is evident that reweighing the exogenous endowments has a limited effect. The fact that the curve for simulated income differences now lies much nearer the actual differences curve graphically illustrates the success of the simulated decomposition. This suggests that the approximationerror RAi s very small, at least inthis application. Inorder to identify the relative importance of the various components of non-labor income, we considered the effect of each source ~eparately.5~Private transfers are responsible for a drop in the Gini coefficient equal to 0.7 percentage points, certainly not a negligible effect. But most of the effect of unearned income is in effect due to retirement income. Retirement income i s strongly inequality-increasing in Brazil, whereas it would be (mildly) equalizing in the US. This can be seen in Figure 9, which shows the mean retirement pension income for each hundredth of the distribution of household income. Apart for some outliers in the middle of the distribution, retirement income clearly concentrates among the richest households in Brazil, whereas it i s the largest in the deciles just below the median in the US. The explanation of that difference i s simple. Retirement income in Brazil concentrates among retirees of the 56 This allocation of the various effects is made difficult by the fact that their size depends on the other effects already being accountedfor. The figures mentionedhere are obtainedas averagesover the variouspossibleconfigurations appearingintable 5. 57This analysis i s availablefrom the authors on request. formal sector who tend to be better off than the rest of the pop~lation.~~the US, on the contrary,In retirees are more evenly distributed in the population. When summing up all income sources, they tend to be around the median of the distribution. Hence the switch from Brazilian to US retirement income i s very strongly equalizing, reflecting first of all the universality of retirement in the US and the privilege that it may represent inBrazil. Overall, the bottom line seems to be that differences in income inequality between Brazil and the United States are predominantly due to differences in the underlying distributions of endowments in the two countries, including among endowments the right to retirement income. Of the almost thirteen Gini points difference, almost ten can be ascribed to endowment effects. Among these, the data suggest almost equally important roles to inequalities in the Brazilian distribution of human capital (as proxied by years of schooling), and other claims on resources, measuredby flows of non-labor income. The remaining three points of the Gini are due to price effects and, in particular, steeper returns to education in Brazil than in the US. Combined to the more unequal distribution of educational endowments themselves, this confirms the importance of education (prices and quantities) in driving Brazilian inequality, as previewed by the Theil decompositions reported in Section 2. While human capital remains firmly at the center-stage, our results suggest that it i s joined there by the distribution of non-labor incomes and, inparticular, of post-retirement incomes. Figure9: incidenceof Retirement Pensions in Bradl andthe US 0 i o 20 30 40 50 60 70 80 90 100 1 +US Relirement income +Brazilian Retirementincome I Sources: PNAD/IBGE1999,CPS/ADS 2000 58See Hoffman (2001) for an interesting analysis of the contribution of retirement pensions to Brazilian inequality. His findings confirmthe importanceof this income source to the country's high levels of inequality, but he shows that this effect is particularly pronouncedin the metropolitanareas of the poorer Northeastemregion, as well as in the states of Rio de Janeiro, Minas Gerais and EspiritoSanto.The effect appears to be much weakerin ruralareas. 45 5- The Brazil Mexico Comparison - The differences between the distributions of household income per capita in Brazil and Mexico are much smaller than those between either country and the US.The two Latin American countries are at roughly the same level of development, and both are high inequality countries in international terms. Nevertheless, urban Brazil i s much poorer than urban Mexico, and more unequal by any of the four measures reported in Table 7 below. The Lorenz curve for urban Brazil, in Figure 1, lies everywhere below Mexico's. The estimated coefficients for equations (7) and (8), run now so as to be strictly comparable between Brazil and Mexico, as well as those for the multinomial logit models for the demographic and educational structures, are includedinTables A6 -A9, inthe Appendix. Interms of the Gini coefficient, Brazil's excess inequality amounts to some seven points. Price effects account for 1.2 of these, with the variance of the residuals making no contribution at all to differences between Mexico and Brazil. Participation behavior and occupational structure also account for about a Gini point, but its interaction with the price effects is more-than-additive. The combined impact of all price and participation effects i s of more than three points of the Gini. price effects i s less-than-additive. Joint simulation of Mexican 1,y, a, p and o2account for some four and Education alone also accounts for some three Gini points, but its interaction with occupational choice and a half of the seven-point difference. Interacting demographic effects takes away another Gini point from Brazil's measure, but again only once the Mexican approximated conditional distribution of education has been imported too. As inthe case of the US, the educational structure of the population seems to be, either directly or indirectly, a powerful explanatory factor of the difference in household income distribution between Brazil and Mexico. exogenous characteristics such as age, race and household type i s the same as Mexico's - has a small Replacing fl(W) by @(W) - i.e. reweighing the Brazilian population so that its make-up in terms of inequality-reducingeffect: the Gini coefficient falls by 0.7 percentagepoint. This effect i s slightly bigger when these new exogenous endowments are interacted with Mexican ("semi-exogenous" endowments of) education and fertility, as well as its price and occupational choice effects. They also help subtract a Gini point. Altogether, the preceding effects account for almost all the difference observed between Brazil and Mexico, interms of the Gini coefficient. This i s not true, however, of the other inequality measures or of poverty, as shown in table 7. Inparticular, it can be seen that very little of the excessive relative poverty in Brazil is explained by the decomposition methodology, when it is limited to price, occupational structure and endowment effects, a feature that also appears quite clearly in Figure 12. As in the comparison with the US, it may thus be expected that what i s left unexplained actually corresponds to the factors behindunearnedincome. 46 Table7 : Simulated Povertv and Ineaualitvfor Brazil in1993.U s i n e 1994 M e x i c o coefficients. Mean Povetty P/C Inequality 2 = m e d i d permonth Income Gini Eo E(1) Eo V(l0g) P O P(l) P O 1 Brad 294.8 0,569 0.597 0.644 1.395 1.101 26.23 10.10 5.36 2 Mexico 294.8 0.498 0.420 0.495 1.028 0.703 14.98 333 139 3 M , P 294.8 0.556 0,567 0.610 1.303 1,059 24.50 9.33 4.90 294.8 0.557 0,570 0.613 1.314 1.063 24.62 9.39 4.94 289.5 0.557 0.567 0.608 1.229 1,053 25.47 9.56 5.00 281.3 0.535 0.518 0.552 1.079 0.977 23.64 8.68 4.45 375.3 0,537 0.544 OS32 0.908 1.112 18.04 6.87 3.62 399.2 0.535 0.540 0.525 0.889 1.108 16.47 6.12 3.18 285.1 0.522 0.500 0.513 0.950 0.981 22.95 8.56 4.44 285.1 0,524 0.502 0.516 0.957 0.985 23.09 8.61 4.46 275.5 0.579 0.619 0.671 1.496 1.133 29.94 11.90 6.44 348.0 0,537 0.550 0.529 0.891 1.144 20.48 8.13 4.41 389.7 0.532 0.538 0.514 0.844 1.125 17.41 6.61 350 282.6 0,514 0.490 0.493 0.887 0.991 22.85 8.82 4.70 282.6 0.515 0.491 0.494 0.888 0.992 22.88 8.81 4.69 291.9 0,529 0.488 0.554 1.216 0.848 M.6 6.3 2.8 279.9 0.447 0.348 0.356 0.539 0.678 14.75 4.40 1.87 284.5 0,562 0,579 0.625 1.330 1.074 26.51 10.17 539 269.2 0.506 0.471 0.473 0.834 0.955 23.39 8.92 4.72 19 d;b 283.7 0.522 0.475 0.535 1.138 0.832 20.9 6.5 2.8 20 6. +,A, Y. a.6. 2 ,b 268 6 0.437 0.331 0.337 0.496 0.650 14.94 4.43 1.X8 Source PNAD 1999 andENIGH 1994 Unlike in Section 5, the conditional distribution of non-labor incomes in Mexico was approximated by a non-parametric method, described in footnote 18. As Figure 13 illustrates, the impact of this approximation i s powerfully equalizing. By itself, it subtracts four points from the Brazilian Gini, and six points from the headcount index (see row 16: 50, in Table 7). Tellingly, it almost halves the distribution- sensitive poverty measureFGT(2). At the same time, it may also be seen that, when combined with all the preceding changes, importing the structure of Mexican unearned incomes overshoots the observed difference between the two countries - see also figure 13. This means that the approximation error RAfor this decomposition is negative -and larger inmodulethan inthe previous section.59 In any case, however, the results obtained so far suggest that the Brazilian urban poor are at a only to their US counterparts - which might not be so surprising- but also when compared to the Mexican disadvantage interms of access to non-human assets and to public or private transfers when compared not urban poor. This i s an issue of clear relevance for the design of poverty-reduction policy in Brazil. Identifying more precisely the reasonsof the difference with Mexico deserves further investigation. 59 Inaddition, the Brazil - Mexico decompositions appear, on the whole, to be less additively separable than the Brazil - US ones. The sum of individualeffects inTable 7 is further away from the correspondingcombinedeffects than inTable 5. 47 6- Conclusions This paper proposed a micro-econometric approach to investigatingthe nature of the differences between income distributions across countries. Since a distribution of household incomes i s the marginal of the joint distribution of income and a number of other observed household attributes, simple statistical theory allows us to express it as an integral of the product of a sequence of conditional distributions and a (reduced order) joint distribution of exogenous characteristics. Our method i s then to approximate these conditional distributions by pre-specified parametric models, which can be econometrically estimated in each country. We then construct counterfactual approximated income distributions, by importing sets of parameter estimates from the models of country B into country A. This allows us to decompose the difference between the density functions (evaluated at any point) of the two distributions - or any of their functionals, such as inequality or poverty indices -into a term corresponding to the effect of the imported parameters, a residual term, and an approximation error. The decomposition residual can be reduced arbitrarily by combining the sets of parametersto be imported into a given simulation. The approximation error i s shown to be small for the applications considered. The sets of counterfactual income distributions constructed in this paper were designed to decompose differences across income distributions into effects due to three broad sources: differences in the returns or pricing structure prevailing in the countries' labor markets; differences in the parameters of the occupational structure of the economy; and differences in the endowments of age, race, gender, education, fertility and non-labor assets, broadly defined. By comparing the counterfactual distributions corresponding to each of these effects andto various combinations of them, we shed light on the nature of the inter-relationships between returns, occupations, and the underlying distributions of endowments. These can lead to interesting findings, such as a quantification of the impact of educational expansion on inequality through a specific channel: its effect on women's fertility behavior and labor force participation. We applied this approach to the question of what makes the Brazilian distribution of income so unequal. Inparticular, we considered the determinants of the differences between it and the distributions of two other large American nations: Mexico and the UnitedStates. We found that differences inthe structure of occupations account for little inboth cases. Prices were not insubstantial inexplaining difference between the US and Brazil, with this being due largely to steeper returns to education in Brazil. But the most important source of Brazil's uniquely large income inequality i s the underlying inequality in the distribution of its human and non-human endowments. Inparticular, the main causes of Brazil's inequality - and indeed of its urban poverty - seem to be poor access to education and claims on assets and transfers that potentially generate non-labor incomes. The importance of these non-labor incomes was one of our chief findings. Income distribution in Brazil would be much improved if only the distribution of this income component was more similar to those of the US or Mexico - themselves hardly paragons of the Welfare State. If this i s due to public transfers, which needs to be investigated further, it i s possible that our findings would vindicate those who have argued for a speedier public approach to the reduction in inequality than that which would be available from educational policies alone. 48 References. Aghion, Philippe, Eve Caroli and C. Garcia-Pefialosa (1999): "Inequality and Economic Growth: The Perspective of the New Growth Theories", Journal of Economic Literature, XXXVII (4), pp.1615- 1660. Almeida dos Reis, Jost G. and Ricardo Paes de Barros (1991): "Wage Inequality and the Distribution of Education: A Study of the Evolution of Regional Differences in Inequality in Metropolitan Brazil", Journal of Development Economics, 36, pp. 117-143. Atkinson, Anthony B. (1970): "On the Measurement of Inequality", Joumal of Economic Theory, 2, pp.244-263. Banerjee, Abhijit and Andrew Newman (1993): "Occupational Choice and the Process of Development", Journal of Political Economy, 101(2), pp.274-298. BCnabou, Roland (2000): "Unequal Societies: Income Distribution and the Social Contract", American Economic Review, 90 (l), pp.96-129. Blau, Francine and Lawrence Khan (1996): "Intemational Differences in Male Wage Inequality: Institutions versus Market Forces", Journal of Political Economy, 104 (4), pp.791-837. Blinder, Alan S. (1973): "Wage Discrimination: Reduced Form and Structural Estimates", Journal of Human Resources,8, pp.436-455. Bourguignon, Franqois(1979): "Decomposable IncomeInequality Measures", Econometrica, 47, pp.901-20. Bourguignon, Franqois, Francisco H.G. Ferreira and Nora Lustig (1998): "The Microeconomics of Income Distribution Dynamics in East Asia andLatinAmerica", World BankDECRA mimeo. Buhmann, B., L. Rainwater, G. Schmaus e T. Smeeding (1988): "Equivalence Scales, Well-being, Inequality and Poverty: Sensitivity Estimates Across Ten Countries usingthe Luxembourg Income Study database", Review of Income and Wealth, 34, pp.115-42. Cowell, Frank A. (1980): "On the Structure of Additive Inequality Measures", Review of Economic Studies,47, pp.521-31. Cowell, FrankA. and StephenP. Jenkins (1995): "How much inequality can we explain? A methodology and an applicationto the USA", Economic Journal, 105, pp.421-430. DiNardo, John, Nicole Fortin and Thomas Lemieux (1996): "Labor Market Institutions and the Distribution of Wages, 1973-1992: A Semi-parametric Approach", Econometrica, 64 (5), pp.1001- 1044. Donald, Stephen, David Green and Harry Paarsch (2000): "Differences in Wage Distributions between Canada and the United States: An Application of a Flexible Estimator of Distribution Functions in the Presence of Covariates", Review of Economic Studies, 67, pp.609-633. Elbers, Chris, Jean 0.Lanjouw, Peter Lanjouw and Phillippe G. Leite (2001): "Poverty and Inequality in Brazil: New Estimates from Combined PPV-PNADData", World Bank, DECRG Mimeo. Ferreira, Francisco H.G., Peter Lanjouw and Marcel0 Neri (2000): "A New Poverty Profile for Brazil usingPPV, PNADand censusdata", PUC-Rio, Department of Economics, TD#418. Fishlow, Albert (1972): "Brazilian Size Distribution of Income", American Economic Review, 62 (2), pp.391-410. Foster, J., J. Greer, and E.Thorbecke, 1984, "A class of decomposable poverty measures.'' Econometrica, 52, pp.761-65. 49 Henriques, Ricardo (2000): Desigualdade e Pobreza no Brasil, (Rio de Janeiro: PEA) Hoffman, Rodolfo (2001): "Desigualdade no Brasil: A Contribuiqgo das Aposentadorias", UNICAMP, Instituto de Economia, mimeo. Juhn, Chinhui, Kevin Murphy and Brooks Pierce (1993): "Wage Inequality and the Rise in Returns to Skill", Journal of Political Economy, 101(3), pp. 410-442. Lam, David and Deborah Levinson (1992): "Age, Experience, and Schooling: Decomposing Earnings Inequality inthe United States and Brazil", Sociological Inquiry, 62 (2), pp.218-145. Langoni, Carlos G. (1973): DistribuiGo da Renda e Desenvolvimento EconGmico do Brasil, (Rio de Janeiro: Express50 e Cultura). Legovini, Arianna, CCsar Bouillon and Nora Lustig (2001): "Can Education Explain Income Inequality Changes inMexico", IADB Poverty and Inequality Unit,mimeo. Machado, JosC A.F. and JosC Mata (2001): "Earning Functions in Portugal 1982-1994: Evidence from Quantile Regressions", Empirical Economics, 26 (l), pp.115-134. Oaxaca, Ronald (1973): "Male-Female Wage Differentials in Urban Labor Markets", International Economic Review, 14, pp.673-709. Sacconato, AndrC L. and NaCrcio Menezes Filho (2001): "A Diferenqa Salarial entre os Trabalhadores Americanos e Brasileiros: Uma Aniilise com Micro Dados", Universidade de S ~ Paulo, Instituto O de PesquisasEconBmicas, TD No. 25/2001. Shorrocks, Anthony F.(1980): "The Class of Additively DecomposableInequality Measures",Econometrica, 48, pp.613-25. Shorrocks, Anthony F. (1999): "Decomposition Procedures for Distributional Analysis: A Unified Framework Based on the Shapley Value", University of Essex, Department of Economics, mimeo. SzCkely, Miguel (1998): The Economics of Poverty, Inequality and Wealth Accumulation in Mexico, (New York: St. Martin's Press). World Bank (2002): Building Institutions for a Market Economy: World Development Report, 2002, (New York: Oxford University Press). 50 "to q 3 3 0 0 0 0 0 0 n a.q~o0 9 O N O * m qm m 0 rm. Y ) r . - t - o w 04 h 00 04 oq oq h h r a m u m r . o o a o o o o W - r U w - NnPm a a o o! oq 4 h 9 0, h Y) I a o o o o - o N N N N N N N N N N N N N N m m m m m m m w w w w w w w m m m m m n m w w w w w w w w w w w w w w w w w w w w w Figure4: Brazil US Differencesin the Logarithmof HouseholdIncome Per Capita, - Actual and Simulated,Steps 1 and 2 0,4 I ."2 0,2 3cn U 0 -0,2 -0,4 0 - 10 20 30 40 50 EO 70 EO 90 100 Percantiles A -a,D -a,!% 02 1 Figure5: Brazil US Differencesin the Logarithmof Household Income Per Capita, - Actual and Simulated,Step 4 0,4 I gt! 0,2 3m 0 -0,2 -0,4 0 10 20 30 40 50 60 70 EO 90 100 Percentiles 53 Figure 6: Brazil US Differencesin the Logarithmof HouseholdIncome Per Capita, - Actual and Simulated,Step 6 0,a 0,6 0,4 8 .-P kg, 0,2 3m 0 -02 -0,4 0 10 20 30 40 50 60 70 80 90 100 Percentiles Figure7: Brazil US Differencesin the Logarithmof Household Income Per Capita, - Actual and Simulated, Step 8 Of 0,E O A 1 a, gg, E 0,z 3m 0 -0,z -0,4 0 10 20 30 - 40 50 60 70 80 90 100 - Percentiles A -I-a.b,o2,h.u a.b.02.h.r.wi 54 Figure8: Brazil US Differencesin the Logarithmof HouseholdIncome Per Capita, - Actual and Simulated 0 10 20 30 - 40 50 60 70 80 90 100 Percentiles A -"-a.~.~~h.v.w.~ - a.B.02,lv.w.E.b Figure 10: Brazil Mexico Differencesin the Logarithmof HouseholdIncome Per Capita, - Actual and Simulated, Steps 1 and 2 0,s 0,s 0,4 e 0 g 0.2 U 30 0 4 2 -0,4 0 10 20 30 - 40 50 60 70 80 90 100 Percentiles A -a.D- a.B.02 1 55 Figure11: Brazil Mexico Differences inthe Logarithmof HouseholdIncome Per Capita, - Actual and Simulated, Step4 0 3 0,6 0 4 0,2 0 -0,2 -0,4 0 10 20 30 40 50 60 70 80 90 100 Figure12: Brazil MexicoDifferencesinthe Logarithmof HouseholdIncomePer Capita, - Actual and Simulated, Step 6 0 10 20 30 40 ,- 50 60 70 80 90 100 Percentiles A --)-a.B.uZ.h'-a.B.uZ.h.v1 56 Figure 13: Brazil MexicoDifferences in the Logarithmof Household incomePer Capita, - Actual and Simulated, Step 8 -0.4' - 0 10 20 30 40 50 60 70 80 90 100 Percentiles A -Ca.D,02.X.r -aD. 02,~u.w I Figure14: Brazil -Mexico Differences in the Logarithmof HouseholdIncomePer Capita, Actual and Simulated, Step 12 0 10 20 30 - 40 50 60 70 80 90 100 - Percentiles A -""--a,5.~2.h,v.w.$ a ,5, 02.w. h. r.$:E 57 0 0 0 0 0 0 0 0 0 0 0 0 0 0 9 9 9 9 9 9 9 0 0 0 0 0 0 0 Pa s R a R E 3 R -1111 2 2 0 0 0 0 0 0 2VI 00 O W N h Y l V I ? n ? n P Y I8 8 9 9 9 9 9 9 9 0 0 0 0 0 0 0 I ' I :I 0 3- INEQUALITY OF OUTCOMES, INEQUALITY OFOPPORTUNITIES AND INTERGENERATIONAL EDUCATION MOBILITYBRAZIL. IN Franqois Bourguignon*, Francisco Ferreira**, and MartaMenCndez*** Abstract This paper departsfrom Roemer's (1998) theoretical formulations of the concept of equal opportunity and analyzes,for the Brazilian case, the general relationship between inequality of outcomes, inequality of opportunities and intergenerational educational mobility. Our main purpose is to study, both in a re- gression framework and through a micro-simulation decomposition technique, what part of observed (outcome)inequality may be attributed to 'circumstances', orfamily background, and what is due to the 'effort' of individuals, given the variables available in our data set. In particular wefocus on intergen- erational educational mobility and the way in whichparents' education affects, directly or indirectly, the earnings of their ofSspring. Data arefrom the 1996Brazilian household survey (PNAD), where informa- tion about parental education is available. The analysis is conducted by 5-year cohorts, which permits following the long-run evolution of the inequality of opportunity and intergenerational mobility over time. Results show that among observed variables, parental educationproves to be the major source of inequality of opportunities in Brazil. It is not only a powerjul determinant of the education of the chil- dren, but also an important independent determinant of individual earnings. The same conclusion ap- plies to household income per capita, though now observed circumstances do not operate only through the individual earnings, but also through other channels:fertility in particular, and to a lesser extent, labor-force participation, non-labor income and matching behavior. We also observe that intergenera- tional educational mobility has increased over time, especially at the bottom of the distribution. How- ever, even after correcting for the inequality of observed opportunities, Brazilian inequality remains at high levels by international standards, which means that observed opportunities may not be enough to explain the excessive inequality observed in Brazil in comparisonwith other countries in the world. JELclassificationcode: C13, D31, D63,121,531,015 Introduction Inequality of "outcomes" and inequality of "opportunities", or chances, have long been considered as corresponding to very different views on socialjustice in the literature on economic inequality. The first definition refers to the distribution of the joint product of the efforts of a person and the particular circumstancesunder which this effort was or is made. It i s mostly concerned with income inequality. The seconddefinition refers to the heterogeneity inthose circumstances that are out of individuals' control but that nevertheless significantly affect the results of their efforts, and possibly the efforts themselves. This distinction, the formulation of which i s borrowed from Roemer (1998), buildingon earlier work by John Rawls, Amartya Sen and others, i s well illustrated by the standard opposition between inequality and mobility. The US are often presented as more unequal than European societies but at the same time more * Worldbank andDELTA, Paris. ***PUC, I* Rio de Janeiro. DELTA, Paris.48, Bd Jourdan, Paris, France.Tel: (0)-1-4313-6311.Fax: (0)-1-4313-6310.E-mail: menendez@delta.ens.fr. 65 mobile from a generation to the next. The latter feature is sometimes taken as the sign of a more equal distribution of chances or opportunities in the US.' Despite the obvious relevance of the concept of inequality of opportunities and implicitly of the question of social mobility, limited empirical work has been done in this area in comparison with the huge literature on the inequality of outcomes. 2 The main reason for this i s probably to be found in the conceptual difficulty of separating out `circumstances' and `efforts', in the limited availability of variables that could satisfactorily describe `circumstances' or in scarce data sources on mobility. All these problems are still more acute in developing countries. Yet, knowingwhat part of observed outcome inequality may be attributed to circumstances, and in particular to family background, is as important there as in richer countries. Such knowledge should help define the actual scope for redistribution policies and inparticular the choice between redistributing current income or expanding the opportunities of the poor through makingthe accumulation of human capital among children less dependent on parents. Inview of the very highlevelof (outcome) inequalityinBrazil, thequestion arisesoftheproportionthat i s due to opportunities that individuals inherit from their parents and the proportion that i s due to the heterogeneity in their efforts and in the results of these efforts. There are various ways to estimate these proportions. The first one consists of studying how much parents do invest intheir children conditionally on the characteristics of the parents. That part of the schooling inequality that i s explained by parents' characteristics corresponds to the inequality of opportunities, whereas the remainder may be attributed to heterogeneous individual efforts. The latter may also be interpreted as an index of mobility across generations as in the study of Behrman, Birdsall and Skelezy (2000) for Latin American countries. Inthe case of Brazil, this line of analysis has been followed by Lam (1999). Because it i s based on observed current schooling decision, a problem with that approach i s that it only permits to study future social mobility, that is, the relation between the education of children, when they will be adults, and that of their parents. Because the (future) income of the children is not observed, this kindof analysis does not permit to disentangle the actual contribution of the inequality of opportunities to overall (outcome) inequality. The approach taken in the present paper is of a different nature. It is basedon direct information given by survey respondents about the education and occupational position of their parents in the 1996 Brazilian household survey (PNAD). That information permits measuring not only the extent of intergenerational educational mobility but also the way in which parents' characteristics and other circumstance variables may affect the earnings or income of their children, directly rather than indirectly through the education of the children. Also, by controlling for the year of birth, it i s possible to see how this influence of parents and background changed over time and whether opportunities account for an increasing or decreasingproportionof total inequality. This analysis reveals a sizeable inequality of opportunities inBrazil. On the one hand, parents' education proves to be a powerful independent determinant of individual earnings, besides the schooling of respondents. On the other hand, parent education i s a strong predictor of children's schooling. Estimated coefficients suggest that for older cohorts the relationship between the number of years of schooling of the parents and that of the children is very high(coefficients close to 0.8). Inother words, the distribution of schooling is close to be fully reproduced - up to some increase in average schooling - across generations. There are signs that this situation is changing for the youngest cohorts, but the evolution i s very slow. For comparisons of mobility between the US and European countries, see Burkhauser et al. (1998),or Checchi et al. (1999). B y contrast, social mobility has always been a leading theme of the sociological literature. However, it i s not clear whether that literature translates easily into standard economic inequality concepts. 66 The paper i s organized as follows. Section 1 shortly discusses the theoretical background for the estimation work undertaken in this paper, that is, the general relationship between inequality of outcomes, inequality of opportunities and intergenerational educational mobility, given the variables available in the database being used. Section 2 discusses the regression results used to measure the preceding concepts. Section 3 analyzes the inequality measures associated with the concepts discussed above for the distribution of individual earnings. It also shows how the proportion of income inequality that may be attributed to opportunities has changed over time. Section 4 generalizes this analysis to the case of household income per capita. The concluding section draws the implications of these results for our understanding of anti-inequality policy inBrazil. Opportunities discussed in this paper focus mostly on those relatedto the education of the parents. There are other dimensions in the space of opportunities. We are able to capture some of them while others are unobserved in the data. Race and regions of origin are in the first group and are of obvious importance in the case of Brazil. 1-Theoreticalbackground. Among the determinants of the earnings of an active individual at some point of time, one may distinguish characteristics that are independent of the individual's will, which we shall call circumstances, following Roemer (1998), and characteristics that, on the contrary, reflect the `efforts' made by the individual to increase hisher productivity and earnings. Let denote C the first set of variables and E the second set. C typically includes fixed socio-demographic attributes like race, region of origin, and the individual's family background. E corresponds essentially to the human capital accumulated by the individual once free to make decisions for himselfherself. This may include the last part of formal schooling, but also on thejob training, past decisions to changejob or region of residence, or current efforts at work. Formally, let the following simple equation represent the interaction between circumstances, efforts and current (log) hourly earnings, w, for an individual i: Ln(wi)= Ci.a Ei.P ui + + (1) where a and$ are two vectors of coefficients and ui i s a residual term that accounts for unobserved circumstance and effort variables, sheer luck, measurement errors, and temporary departures from the permanent level of income. All these factors are assumed to be independent of the variables actually included in C and E. They are also assumed to have zero mean and to be identically and independently distributed across individuals. If inequality were to be measuredby the variance of the logarithm of earning, and if it were justified to assume that circumstances and efforts are mutually independent, then we would have the following simple decomposition of total inequality: v(Lnw) = a'.V(C).a p'V(E)P v(u) + + (2) where v( ) stands for the variance of the variable in bracket and V( ) for the covariance matrix of all the variables in bracket. In other words, total inequality could be explained simply as the sum of the inequality of observed opportunities (first term on the RHS), the inequality of observed efforts (second term) and the inequality due to unobserved earning determinants. A more general description of the role of these various components in shaping the distribution of individual earnings may be obtained by simulating the effects of equalizing C or E across individuals. Such a decomposition is shown for Brazil inthe empiricalpart of this paper. 67 Complications arise in the precedingframework if one assumes that there i s no independence between circumstances and efforts, or between unobservables and observable wage determinants. Consider first that efforts are partly determined by circumstances. For instance, formal schooling i s supposed to be partly determined by family background. Assuming reasonably that unobserved effort determinants, vi, are orthogonal to observedcircumstances, this i s equivalent to specifyinga second model for efforts. Let that model be: where b i s a matrix of coefficients and vi stands for a vector of unobserved effort determinants - one component for each component of the vector Ei. As usual the vi's are supposed to be iid across individuals and with zero mean. Substitutingin (1) yields: Ln(wi)= Ci.(a+ fib) + (Ei- Ci.b).p + ui (4) Inthis expression, it may be seen that circumstances now have a double effect on the wage rate. They affect it directly, for given efforts, through the set of coefficients a.They also affect it indirectly through their influence on efforts, the size of this second effect being given by the scalar product p.b. This restatement of the original model modifies the variance decomposition formula (2) and more generally any decomposition of the distribution of individual wages into components associated with observable circumstances and efforts. Accounting for the possible correlation between observed efforts and circumstances and relying on the joint estimation of the earning and effort equations (1) and (3) i s therefore important. The precedingdecomposition is easy to implement, provided that one can rely on unbiased estimators of the various sets of coefficients, 4 ,8 and b. Some precaution must be taken when the required assumption that u in equation (1) i s orthogonal to C and E i s open to doubt. The problem i s not too serious for the circumstance variables. One may not be so much interested in the 'true' effect of the variables included in C but intheir overall impact once their correlation with unobservable circumstances are taken into account. For instance, say that C include parents education but not their wealth. Then estimating (1) through standard regression techniques will lead to a bias in the estimation of the coefficient of parental education that will depend on the unobserved correlation between parents' education and parents' wealth, and on the effect of the latter on children's earnings. The coefficients a will be biased in the corresponding direction and there will be a doubt in the decomposition of total inequality as to what is the actual role of parentaleducation. It i s simply a matter of being aware of it. Things are more serious when unobservables in the earning equations cannot be assumed to be independent of the effort variable. Again, imagine that the wealth of parents i s important to determine both the schooling and the current earnings of their children, independently of their own education. This correlation between u and E, or equivalently between u and v, i s introducing some bias in the estimation of the p coefficients and therefore in the decomposition of the total inequality into circumstance and effort components. One way out of this difficulty would be to observe instrumental variables, Z that would influence efforts but not earnings. Equation(3) would then be replaced by: Ei = Cj.b Zi.d + + vi (5) with the vector Z, being orthogonal to ui.Then instrumenting the effort variables in (1) through (5) would yield an unbiased estimator of p and then an unbiased decomposition of total inequality into inequality of observed opportunities, or circumstances, and inequality of efforts. Models of this type have been 68 extensively used in the return to education literature. In the standard Mincerian equation, for instance, it was thought that instrumenting education by family background would correct for obvious endogeneity biases of education. It was checked in a few countries that this was indeed the case. Then family background was considered as an independent earning determinant too, which required using additional instruments. Ability tests taken while attending school often played that role. Few data sets come with all that information, however, this problembeing still more acute indeveloping c~untries.~ In the absence of adequate instrumental variables, 2, the only solution is simply to explore the likely effect of the potential bias in the estimation of ,O due to the correlation between u and v, and then to decide on that basis what i s the most reasonable range of estimates. This i s what we shall do in the case of Brazil. When circumstance variables include characteristics of parents, very much of the preceding analysis has to do with intergenerational mobility. A direct measure of income mobility would be provided by the preceding model if parents' income was among the variables C. But other types of mobility may be behind equations (3). For instance, if parental education i s among variables C and individuals' schooling i s among the effort variables E, then part of system (3) actually describes intergenerational educational mobility. The schooling of observed individuals is simply explained by that of their parents and the corresponding coefficient b gives an indication of the extent of intergenerational mobility. For example, ifeducationismeasuredinnumberofyearsofschoolingforbothparentsandchildren,thentheextentto which b i s less than unity would describe how fast differences in education tend to systematically lessen across generations. It can be seen on equation (4) that the degree'of intergenerational mobility determines at the same time the extent of the share of current earning inequality due to individuals' circumstances or opportunities, provided of course that schooling has a positive effect on earnings - i.e. the first term inbracket on the RHS of (4) i s an increasingfunction of b when pispositive. Another source of intergenerational educational mobility could be found in the residual term, v. It corresponds to the non-systematic part of mobility and i s orthogonal to the concept of inequality of opportunities. As a matter of fact, if this residual i s taken to represent the role of individual efforts in schooling achievements, equation (4) shows that it contributes to increasing the share of earning inequality not due to the inequality of observed or unobserved opportunities. The problem, however, i s that, by definition, nothing i s known of the phenomena behind this residual term, v. Because of this, it i s of lesser interest inthe present context? 2-Opportunitiesandthe distributionof individualwages. The preceding methodology to decompose the inequality of outcomes into various components due to the inequality of observed opportunities and that due to other factors i s now applied to Brazilian data. This section first describes the data and the nature of the variables being used. It then discusses the various estimates obtained for the earning equations and for the equation describing intergenerational educational mobility. 'Earliercontributions include Bowles ( 1972), Griliches ( 1972), Taubman (1976 ). For a survey of all models of returns to schoolingbasedon this kindof instrumentationsee Card(2001). 4Note that this term i s the focus of the analysis in Behrman, Birdsall and Szekely (2000), which interpret its contribution to the varianceof individuals'schoolingas a measure of intergenerationaleducationalmobility. 69 Dataandvarzables Data are from the 1996 wave of the Pesquisa Nacional por Amostragem a Domicilio (PNAD), the Brazilian Household Surveys conducted by the Instituto Brasileiro de Geografia e Estadistica (IBGE)'. For that year, information about parental education of all surveyed household heads and spouses i s available. Information i s also available on the occupation of the parents. The analysis i s restricted to urban areas because of the general imprecision of earning and income measurement in rural areas. It i s also restricted to individuals 26 to 60 years old, in an effort to concentrate on individuals having finished schooling andpotentially active inthe labor market. The analysis described in the preceding section i s conducted by 5-year cohorts - from individuals born between 1936-40 up to those born between 1966-70. This permits not only to measure the role of the inequality of opportunities in shaping the inequality of observed earnings at a point of time, but also to study how this role may have changed over time.. An important question is indeed whether the increase in the educational level of successivecohorts was accompanied by more or less educational mobility and a reduction in the inequality of opportunities or whether it corresponded to a uniform upward shift in schooling achievements with constant inequality of opportunities. Comparing various cohorts observed at a single point of time permits to answer this question in a simple way. We shall first focus on individual earnings, measured as "all jobs real hourly earnings", in agreement with most of the intergenerational mobility literature. This might not be the most satisfactory concept to use if one i s interested inthe contribution of the inequality of individual opportunities to the inequality of individual 'welfare', though. This is the reason why the analysis will be conducted at a second stage on the income per capita in the households where observed individuals belong to. This clearly makes more prominent the role of labor supply behavior and fertility as a channel for the intergenerational transmission of inequality. The vector of circumstance variables, C, includes race dummies, parental education expressed in numbers of years of schooling - usingthe mean schooling achievement of the father and the mother and the difference between them - the occupational position of the father (a nine-level occupational status variable), and dummies for the regions of origin.7 The vector of effort variables i s restricted to the schooling achievement of the individual, measured in years of schooling', squared-years of schooling, to capture possible non-linearities, and a migration dummy, defined as whether the observed municipality of residence was different from the one where born. Note, however, that this migration might have been done by the individual him/herself when adult or by hidher parents when he/she was a child. It should be taken as a circumstance variable in the second case and as an effort variable in the first case. Unfortunately, there was no way this distinction could be made. Results obtained are more consistent with the effort interpretationof that variable. The same information i s available in both the 1982 and the 1988 surveys. Serious biases seem to plague the observation of earnings in the former survey however. The latter was used to check the robustness of some o f the results reported in the present paper. Parental education i s given in discrete levels. They were converted into years o f schooling (here in brackets) using the following rule. No school or incomplete 1st grade (0); incomplete elementary (2); complete elementary, or complete 4Ihgrade (4); incom- plete 1" cycle o f secondary or 5'h to 7Ihgrade (6); complete 1'' cycle o f secondary or complete 8Ihgrade (8); incomplete 2"d cycle '(9.5);variable complete 2"dcycle o f secondary (11); incomplete superior (13); complete superior (15); master or doctorate (17). A that was used in a first stage as a 'circumstance' variable was whether the individual was forced to work as a child - i.e. before 14 - or not. This variable proved to be too closely related to the number o f years o f schooling to be o f very much in- dependent interest. * The number of years of schooling directly provided in the PNAD i s bounded at 15. For consistency with the scale used for parents' schooling, this variable was changed to 17 for individuals reporting a master or a doctorate degree. 70 In addition to the preceding list of variables, a labor force participation equation has been estimated for married women in order to correct for the well-known selection bias in estimating the earning equation (1). The standard Heckman 2-stage procedure was used with the composition of the family, the number of children and household income per capita -excluding own earnings - as instrumental variables in the first stage. Summary statistics for all variables used in the analysis are shown intable 1. Table 1.Descriptive statistics Cohort b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 Meanmonthlyeamings (Reais, alljobs) 559.8 762.5 848.9 823.7 753.7 682.1 553.2 Mean number of years of schooling 4.7 5.7 6.6 7.2 7.7 8 7.8 Mean father's number of years of schooling 2.5 2.7 2.8 2.9 3 3.2 3.2 Meanmother'snumber of years of schooling 2.2 2.4 2.6 2.7 2.9 3 3.1 Race (Percents) Branca (Whites) 61.5 60.4 61.1 61.1 60.3 61.1 59.6 Preta(Blacks) 6 5.9 5.9 6 5.1 4.8 5.1 Amarela (Asians) 0.5 0.8 0.6 0.5 0.4 0.3 0.3 Parda (MR) 32.1 32.3 32.4 32.5 34.2 33.8 35 Regions( Percents) North 5.8 6.7 6.3 7.3 7.6 7.4 7.6 North East 24.9 26.3 24.4 22.2 22.9 23.3 24.3 South East 39.7 37.5 39 38.5 37 36.5 33.8 South 21.1 20.3 20.7 22 21.8 21.7 21.5 Center-West 8.5 9.2 9.6 10 10.7 11.2 12.9 Migrants (Percents) 69.4 68.4 68.4 65.5 63.1 58.9 57.5 Agricultural workers 16.4 12.8 8.5 6.5 5.9 5.4 5.6 Numberof individuals 2300 3378 5256 7132 8080 8192 6401 Earnzngequation Earningequations were estimated separately for men and women, and by cohort, using simple OLS for men and usingthe two-stage selection bias correction procedure for women '. Results are shown in tables 2.a. and 2.b. Note that, unlike with the standard Mincerian specification, age or imputed experience does not appear among the regressors. This i s because cohorts are homogeneous at this respect. The Heckmancorrectionwas initially appliedto men as well, but availableinstrumentsprovedunsatisfactory. 71 Table 2.a: Wage equations by cohort using OLS, for MEN. b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 Race Branca(omitted) Preta -0.449 ** -0.305** -0.293** -0.301 ** -0.265** -0.185 ** -0.250** (0.09) (0.08) (0.06) (0.06) (0.06) (0.05) (0.05) Amarela 0.009 0.600** 0.677** 0.054 0.341 * 0.153 -0.333 (0.32) (0.20) (0.18) (0.17) (0.17) (0.19) (0.24) Parda -0.224** -0.294** -0.278** -0.243** -0.191 ** -0.199 ** -0.237** (0.05) (0.04) (0.04) (0.03) (0.03) (0.03) (0.03) Parentalschooling Meanparentalsch. 0.039** 0.037 ** 0.051** 0.034** 0.033 ** 0.033 ** 0.035** (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.01) Diff.Parentalsch. 0.005 0.018* 0.005 0.005 0.000 -0.001 0.007 (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.00) Regiondummies SouthEast (omitted) North -0.223 -0.078 -0.099 -0.052 -0.191 ** -0.113 -0.131* (0.14) (0.12) (0.09) (0.08) (0.06) (0.06) (0.07) North East -0.202** -0.142 -0.096** -0.209 ** -0.169** -0.234** -0.175** (0.05) (0.05) (0.04) (0.03) (0.03) (0.03) (0.03) South -0.194 ** -0.155 ** -0.068 -0.035 -0.044 -0.103** -0.073 * (0.07) (0.06) (0.04) (0.03) (0.03) (0.03) (0.03) Center-West -0.206 -0.129 0.015 -0.207** -0.090 -0.128 * 0.009 (0.16) (0.10) (0.09) (0.07) (0.06) (0.05) (0.05) Years of schooling 0.087 ** 0.069** 0.083** 0.073 ** 0.051 ** 0.027** 0.029 ** (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Years of schooling-squared 0.002 * 0.003 ** 0.002** 0.003 ** 0.004** 0.005 ** 0.004** (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) Migrant dummy 0.010 0.190 ** 0.140** 0.085 ** 0.114** 0.124** 0.160** (0.05) (0.04) (0.03) (0.03) (0.02) (0.02) (0.02) Father'soccupational status Status 1(omitted) Status2 0.056 -0.021 0.034 -0.016 -0.044 -0.011 -0.082 * (0.05) (0.05) (0.04) (0.04) (0.03) (0.03) (0.04) Status3 0.008 0.071 0.121* 0.011 0.040 0.055 0.093 * (0.09) (0.08) (0.06) (0.05) (0.04) (0.04) (0.04) Status 4 -0.042 0.060 -0.017 0.088 0.082 0.043 0.129** (0.10) (0.09) (0.07) (0.06) (0.05) (0.04) (0.05) Status5 0.337** 0.050 0.095 0.093 0.116* 0.115** 0.112" (0.10) (0.09) (0.06) (0.05) (0.05) (0.04) (0.04) Status 6 0.099 0.177* 0.132* 0.158 ** 0.252** 0.125** 0.140* (0.11) (0.09) (0.07) (0.06) (0.05) (0.05) (0.06) Status 7 0.183 0.111 0.046 -0.004 0.069 0.082 0.148** (0.13) (0.11) (0.07) (0.06) (0.05) (0.05) (0.06) Status 8 0.079 0.165 0.126 0.191** 0.173** 0.206** 0.131 (0.11) (0.10) (0.08) (0.07) (0.06) (0.06) (0.07) Status 9 0.176 0.235 0.261* 0.143 0.036 0.211** 0.219** (0.22) (0.14) (0.10) (0.09) (0.08) (0.07) (0.08) Constant 0.210** 0.236** 0.182 ** 0.321 ** 0.274** 0.238** 0.162** (0.06) (0.06) (0.05) (0.05) (0.04) (0.04) (0.05) Sample size 1520 2158 3200 4192 4635 4807 3846 F-test 52.91 80.87 143.7 158.86 167.45 190.4 116.94 R-sauared 0.414 0.431 0.475 0.432 0.421 0.443 0.379 AdjR-squared 0.406 0.426 0.472 0.430 0.418 0.441 0.376 OLS estimates, standard errors in brackets; *=significant at the 5% prob.Level; **=significant at the 1% prob. Level. 72 Table 2.b: Wage equations by cohort using Heckman correction (2SLS), for WOMEN. b1936-40 b1941-45 b1946-50 b1951-55 b19.56-60 b1961-65 b1966-70 Race Branca(omitted) Preta 0.119 -0.080 -0.145 -0.192** -0.200 ** -0.124 0.001 (0.17) (0.10) (0.09) (0.07) (0.08) (0.08) (0.08) Amarela 0.296 -0.259 -0.077 0.193 0.390 0.191 -0.026 (0.50) (0.31) (0.25) (0.25) (0.24) (0.31) (0.30) Parda -0.140 -0.165** -0.196** -0.246** -0.105 * -0.097** -0.129** (0.09) (0.06) (0.05) (0.04) (0.04) (0.04) (0.04) Parentalschooling Meanparentalsch. 0.052** 0.057** 0.057** 0.033** 0.045 ** 0.049** 0.040** (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Diff.Parentalsch. 0.018 -0.009 0.004 -0.002 0.009 -0.003 0.002 (0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.01) Regiondummies SouthEast (omitted) North -0.253 -0.062 0.012 -0.017 -0.088 -0.073 -0.094 (0.25) (0.11) (0.10) (0.07) (0.08) (0.07) (0.08) North East -0.156 -0.267** -0.231 ** -0.215** -0.284 ** -0.265 ** -0.243 ** (0.10) (0.06) (0.05) (0.04) (0.04) (0.04) (0.05) South -0.188 -0.068 -0.034 -0.068 -0.068 -0.015 -0.027 (0.11) (0.07) (0.06) (0.05) (0.05) (0.04) (0.05) Center-West -0.365 -0.281 * -0.232* -0.183* -0.081 -0.122 -0.129 (0.20) (0.12) (0.10) (0.08) (0.07) (0.06) (0.07) Years of schooling 0.053 0.056** 0.018 -0.002 0.020 -0.012 -0.007 (0.04) (0.02) (0.02) (0.01) (0.02) (0.01) (0.02) Years of schooling-squared 0.003 0.004** 0.007 ** 0.007** 0.007 ** 0.008 ** 0.008 ** (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) Migrant dummy 0.120 0.134* 0.143 ** 0.103 0.093 ** 0.100** 0.143** (0.08) (0.05) (0.04) (0.03) (0.03) (0.03) (0.03) Father'soccupationalstatus status 1(omitted) status 2 0.018 0.013 -0.004 -0.049 0.037 0.037 -0.015 (0.10) (0.07) (0.06) (0.05) (0.04) (0.04) (0.06) status 3 -0.010 0.111 -0.065 0.004 0.079 0.092 0.047 (0.15) (0.10) (0.08) (0.07) (0.06) (0.06) (0.07) status 4 0.146 0.018 0.083 0.070 0.079 0.208 ** 0.078 (0.17) (0.12) (0.09) (0.07) (0.06) (0.06) (0.07) status 5 -0.010 -0.232* -0.049 0.034 0.119* 0.107 0.078 (0.16) (0.12) (0.08) (0.06) (0.06) (0.06) (0.07) status 6 0.083 0.240* -0.004 0.251 ** 0.182** 0.204** 0.140 (0.17) (0.10) (0.09) (0.07) (0.06) (0.06) (0.08) status 7 0.294 -0.058 -0.136 0.054 0.023 0.136* 0.101 (0.17) (0.13) (0.10) (0.07) (0.07) (0.07) (0.08) status 8 0.138 0.113 0.272** 0.126 0.177 0.230** 0.250** (0.18) (0.12) (0.11) (0.08) (0.08) (0.08) (0.10) status 9 0.292 -0.003 0.012 0.102 0.217" 0.249** 0.402 ** (0.26) (0.17) (0.12) (0.10) (0.09) (0.08) (0.11) Constant -0.330 -0.175 -0.088 0.135 J-J&O** -0.408** -0.413 ** (0.24) (0.15) (0.13) (0.10) (0.12) (0.09) (0.13) Millslambda 0.539 0.105 0.203 -0.206 0.940 0.703 0.496 Rho 0.617 0.144 0.260 -0.269 1.023 0.855 0.636 Sigma 0.874 0.729 0.781 0.767 0.919 0.822 0.780 Numberof obs 870 1341 2233 3149 3723 3737 2988 Censored obs 675 1097 1888 2748 3251 3212 2424 Uncensoredobs 195 244 345 401 472 525 564 Two-stageHeckmanestimates, standarderrors inbrackets; *=significant at the 5% prob. Level; **=significant at the 1%prob. level. 73 The first part of these two tables refers to circumstance variables. They all have the expected effect on earnings. Racial discrimination coefficients are significant and negative for black and `pardo'. They are generally positive, but not always significant for people with an Asian origin. For women, however, it i s interestingthat discrimination i s strong and significant for the first group (black) only in two cohorts. The extent of discrimination i s also less pronounced than for men. Regional differences are also important. Compared to the omitted region (South East), beingborn in any other region has a negative effect for men, though not always significant - note that part of this effect might be captured by the migration variable. For women this negative effect i s also present though less often significant. The region with the worst effect on earnings, other things beingequal, is the North East. The estimated effect of parental education on individual earnings i s always positive, significant, and relatively stable across cohorts. It i s also sizable since it amounts to 3 to 6 per cent earning difference for each year of schooling of the parents. The difference in the education of the father and the mother i s meant to detect a possible asymmetry in the role of the two parents. Butno such asymmetry seems to be systematically present. Turning now to the vector of "effort" variables, individuals' own education has the usual positive and significant effect on earnings for men. This effect i s decreasing as one considers younger male cohorts. This i s consistent with the negative coefficient generally found for the squared imputed experience term -i.e. age minus number of years of schooling minus first schooling age - in the standard Mincerian specification. In effect, this implies that the return to schooling increases with age. This i s exactly what i s found here. lo Note also that the coefficient of schooling i s sometimes insignificant - it i s almost always the case for women. The reason why i s that the overall education effect i s captured by the squared of the number of years of schooling variable, which i s positive and significant both for men and women. It is interesting that the order of magnitude obtained for the return to schooling in the preceding equations seems to fall slightly below what was obtained in comparable studies for Brazil. For instance, Ferreira and Paes de Barros (2000) found that the marginal return on a year of schooling i s in the range 12to 15 per cent for both men and women in 1999. In table 2, marginal returns at 5 years of schooling range from 7 to 11per cent for menand from 11to 13 for 10years of schooling A possible explanation of that difference may again lie in the specification being used here, which i s not strictly comparable to the Mincerian model. More fundamentally, however, this difference i s consistent with the probable over-estimation of the returns to schooling in an earnings equation that does not include family background variables. Indeed, excluding positive earnings determinants which are themselves positively correlated with the number of years of schooling leads standard OLS estimation to over-estimate the role of schooling. The preceding intuition is fully confirmed by the data used in this study. Inunreported regressions, though available from the authors upon request, we have re-estimated the preceding wage equation with years of schooling instrumented by parents' schooling achievements and the other exogenous variables of the model, and with parents' education excluded from the regressors. The coefficients of the number of years of schooling turn out to be substantially higher than in the previous case becausethey now partly account for the direct influence of parental education on individual earnings. Their order of magnitude is also comparable to what has been found in other earnings equations estimated for Brazil (see, for example, Ferreiraand Paes de Barros, 1999). Migration has a significant and positive effect on earnings, both for men and women. This sign would be consistent with a human capital interpretation of migration. Because the coefficient i s rather large, amounting loThe conventionalMincerian specificationi s such that: Lnw = a.S b.Exp - c.Exp*where Exp = Age - S - 6. Expandingthe Exp term leads to: Lnw = (a-b-I2c).S+ + 2cAge.S - c.S2 + terms in Age or Age squared. If this equation is estimated within groups with constant age, one should indeed observe that the coefficient of S is higher in older cohorts. Note that the present specification also includes an independentS2term. 74 to a 10/20 per cent increase in earnings, it i s temptingto consider that variable among effort variables. But, it may also reflect the decision of parents to move to an area with better income opportunities when the surveyed individual was still a child, in which case this variable should be taken as indicative of circumstances. If this were true, however, the size of the estimated coefficients would suggest very much persistence in the earnings differential that might have motivatedthe migration of the parents." Effort equations..Intergenerational'EducationaZMobikiy The analysis now moves one step forward by examining the effort equations (schooling, schooling-squared and migration). Because no significant model for migration was found, we concentrate here on schooling. In particular, we will concentrate on the relationship between schooling of individuals and that of their parents, so as to be able to measure the extent to which that variable results from circumstances or true efforts. As noted above, this i s partially equivalent with measuring the degree of intergenerational educational mobility. Table 3 gives, for each cohort, the mean number of years of schooling for various levels of the education of the father or the mother. It was seen in table 1that the mean schooling of the Brazilian society increased steadily over time up to the youngest cohorts. Inthis respect, note that the apparent drop in schooling achievement for that cohort i s artificial. It i s due to the fact that some people in that cohort are still going to university. More interestingly, the mean number of years of schooling by parental schooling levels across cohorts shown in table 3 suggests some progressive increase in intergenerational educational mobility. Indeed, it may be noticed that the number of years of schooling of those individuals with low educated parents increased much more over the four last cohorts than that of individuals whose parents had a medium or high level of education, as if there were some kindof saturation effect among the latter. By father's education level By mother's education level Individual's birthcohort None Low Medium High None Low Medium High 1936-40 2.1 4.62 7.48 11.18 2.27 4.48 7.83 11.23 1941-45 2.55 5.46 8.37 12.32 2.8 5.46 8.7 12.28 1946-50 3.3 5.98 9.01 12.66 3.3 5.85 9.28 12.64 1951-55 3.99 6.66 9.23 12.52 4.05 6.55 9.41 12.61 1956-60 4.35 7.02 9.33 12.52 4.3 6.84 9.4 12.45 1961-65 4.53 7.17 9.47 12.31 4.68 6.94 9.35 12.28 1966-70 4.78 7.03 8.97 11.57 4.84 6.82 8.85 11.62 A more direct evaluation of intergenerational educational mobility i s provided by a regression of type (3) where individuals schooling achievement i s explained by all circumstance variables, includingthe number of years of schooling of the parents. Regressions for the various cohorts are shown in table 4a and 4b separately for men and women. They call for several interestingremarks. Intergenerational educational mobility i s measured, negatively, by the coefficient of the number of years of schooling of parents. The higher that coefficient, the stronger i s parental education in determining the schooling of their children, and therefore the less mobility there is. Because education is measured for both parents and children in years of schooling, a unit value for that coefficient is a convenient reference. It would correspond to the perpetuation of differences in years of schooling across generations -thus, being consistent with an increase in mean schooling. On the contrary, a coefficient less than unity means that educational differences tend to l1Note that we are considering migration across municipalities. 75 diminish across generations. From that point of view, a striking feature in tables 4a and 4b is that intergenerational mobility has been increasing monotonically across cohorts - with the exclusion of the oldest cohort, where sample selection might be biasing results. Overall the gain i s substantial. For people born in the early 1940s, a one-year difference in the schooling of their parents resulted in a difference of .7 years in their own schooling. For those born in the late 1960s, the same initial difference in parental education resulted in a little less than half year of schooling. From an educational point of view, these results suggest that the inequality of opportunities may thus have decreasedsignificantly in Brazil over time. Seeing in the preceding evolution essentially a mere image of the general spreading of education over time would not be totally justified. If children are now going to school for 5 years whereas they were going to most school only for 3 years 20 years ago, it may seem natural that the influence of parental education declined with time. This is not necessarily true, however. This 2-year addition to schooling achievement might very well hold for the whole population, whatever their family background. If this were the whole story, then only the constant in the regressions reported in tables 4a and 4b would be increasing across successive cohorts, and the coefficients of all variables would remain approximately constant. This i s clearly true of race for which no clear trend seems to be present. Black people have the same quantitative disadvantage ineducation inthe 1960s - 1to 2 years of schooling - as they hadinthe 1940s or the 1950s. Likewise, the disadvantage of being born in the North-East for men has remained approximately constant, the same being true of the father being a farmer. Somewhat surprisingly, only the coefficient reflecting the disadvantage of being born from parent with a low level of schooling seems to have been falling regularly over time. In other words, the equality of educational opportunities seems to have remained approximately constant over time except with respect to educational family background. For a given race, region of birth and occupation of the father, the schooling gap between children of families with high and low schooling achievements has narrowed substantially between 1940 and 1970. 76 Table 4.a: Educational mobility OLS regressionsby cohort: Men's years of schooling b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 Race Branca(omitted) Preta -1.078** -1.208** -1,145** -1.298 ** -1.690** -1.538** -0.761 ** (0.37) (0.32) (0.28) (0.24) (0.24) (0.22) (0.24) Amarela 4.972 ** 2.079** 2.136** 2.433 ** 1.350 2.971 ** -0.057 (1.26) (0.75) (0.77) (0.72) (0.74) (0.86) (1.03) Parda -0.787 ** -1.303** -1.113 ** -1.111 ** -1.165** -1.269** -0.767 ** (0.21) (0.18) (0.16) (0.13) (0.12) (0.11) (0.12) Parentalschooling Meanparental sch. 0.740 ** 0.777"" 0.741 ** 0.673 ** 0.615** 0.586** 0.526** (0.05) (0.04) (0.03) (0.02) (0.02) (0.02) (0.02) Diff. Parentalsch. 0.093 * 0.027 0.067 * 0.010 -0.004 0.018 0.006 (0.04) (0.04) (0.03) (0.02) (0.02) (0.02) (0.02) Regiondummies South East (omitted) North -1.629** -0.858 -0.657 -0.382 -0.252 -0.374 -0.307 (0.55) (0.49) (0.39) (0.33) (0.28) (0.28) (0.29) NorthEast -0.973 ** -0.895 ** -0.651 ** -1.087** -0.959** -0.764** -0.967 ** (0.21) (0.18) (0.16) (0.14) (0.13) (0.12) (0.13) South 0.313 -0.458 * -0.641 ** -0.505 ** -0.329* -0.148 -0.289 * (0.26) (0.23) (0.18) (0.15) (0.13) (0.13) (0.14) Center-West 0.298 0.437 -0.356 0.303 -0.291 0.317 -0.208 (0.65) (0.43) (0.37) (0.31) (0.25) (0.23) (0.22) Father'soccupational status status 1 (omitted) status 2 0.513* 0.348 0.407 * 0.705 ** 0.801** 0.925 ** 0.552** (0.22) (0.19) (0.17) (0.15) (0.14) (0.15) (0.16) status 3 2.101 ** 1.598** 1.826** 1.560** 1.578** 1.377** 1.347 ** (0.37) (0.34) (0.26) (0.21) (0.19) (0.18) (0.19) status 4 2.057 ** 2.059 ** 2.732** 2.456 ** 2.483** 1.976** 2.077** (0.39) (0.35) (0.29) (0.24) (0.20) (0.19) (0.21) status 5 2.398 ** 2.528 ** 2.629 ** 2.767 ** 3.143** 2.484** 1.707 ** (0.41) (0.35) (0.26) (0.22) (0.19) (0.18) (0.19) status 6 4.290 ** 3.355 ** 4.015** 3.388** 3.426** 3.071 ** 2.667** (0.42) (0.34) (0.28) (0.24) (0.23) (0.21) (0.25) status 7 4.196 ** 2.723 ** 3.768** 2.951 ** 3.094** 2.509** 1.954 ** (0.52) (0.43) (0.32) (0.27) (0.23) (0.22) (0.25) status 8 2.185** 2.725 ** 2.324** 2.754** 2.511** 2.017** 1.735 ** (0.44) (0.39) (0.33) (0.29) (0.27) (0.28) (0.31) status 9 3.640** 2.546** 2.515** 2.705 ** 2.421** 2.091 ** 1.878** (0.86) (0.57) (0.45) (0.39) (0.34) (0.32) (0.36) Constant 2.782** 3.581 ** 3.872** 4.518 ** 4.712** 4.937** 5.080** (0.19) (0.18) (0.15) (0.14) (0.13) (0.14) (0.15) Sample size 1557 2199 3253 4253 4682 4859 3877 F-test 76.73 ** 114.92"" 164.12** 188.93** 232.72** 221.87** 131.47** R-squared 0.459 0.473 0.463 0.43 1 0.459 0.438 0.367 Adj R-squared 0.453 0.468 0.460 0.429 0.457 0.436 0.364 OLS estimates, standard errorsin brackets; *=significant at the 5% prob. Level; **=significant at the 1%prob. level. 77 Table 4.b: EducationalmobilityOLSregressionsby cohort: Women'syears of schooling b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 Race Branca(omitted) Preta 0.482 -1.289 ** -1.866** -0.870** -1.577** -1.425** -1.360** (0.46) (0.42) (0.34) (0.29) (0.27) (0.29) (0.29) Amarela 2.603* 0.979 3.372** 1.034 2.600** 2.081 * 1.523 (1.12) (1.01) (0.80) (0.95) (0.84) (0.88) (1.22) Parda -0.437 -0.696** -1.290** -1.309** -0.757 ** -0.987** -1.027 ** (0.26) (0.26) (0.19) (0.16) (0.14) (0.14) (0.15) Parental schooling Meanparental sch. 0.787** 0.789** 0.759 ** 0.658 ** 0.628 ** 0.548"" 0.499** (0.05) (0.05) (0.04) (0.03) (0.02) (0.02) (0.02) Diff. Parentalsch. 0.107* -0.024 0.084* 0.088 ** 0.094** 0.048 * 0.074** (0.05) (0.05) (0.03) (0.03) (0.02) (0.02) (0.02) Regiondummies SouthEast (omitted) North 0.162 -0.421 -0.142 0.550 0.363 -0.219 0.054 (0.69) (0.59) (0.45) (0.35) (0.33) (0.31) (0.34) North East -0.277 -0.235 -0.145 0.000 -0.557** -0.225 -0.703 ** (0.26) (0.26) (0.19) (0.17) (0.15) (0.15) (0.16) South 0.051 -0.094 -0.469 * -0.366" -0.399 * -0.540** -0.735 ** (0.31) (0.30) (0.22) (0.18) (0.16) (0.15) (0.16) Center-West 0.065 -0.192 -0.398 1.086 ** 0.233 0.177 -0.142 (0.61) (0.61) (0.43) (0.36) (0.29) (0.27) (0.26) Father'soccupationalstatus status 1(omitted) status 2 0.510 0.679 * 0.827** 0.859** 0.985** 1.194 ** 1.048 ** (0.27) (0.27) (0.21) (0.18) (0.17) (0.17) (0.19) status 3 1.707 ** 2.598 ** 2.295 ** 2.059 ** 1.699 ** 2.098** 1.836** (0.47) (0.43) (0.33) (0.28) (0.23) (0.22) (0.23) status 4 1.656** 2.360** 3.119** 2.843 ** 2.275 ** 2.510** 2.098 ** (0.47) (0.46) (0.35) (0.29) (0.26) (0.24) (0.25) status 5 3.536** 3.332** 3.365** 2.737 ** 3.204** 3.114** 2.736** (0.50) (0.48) (0.32) (0.26) (0.23) (0.22) (0.23) status 6 4.670** 3.354** 4.691 ** 4.088 ** 4.142** 4.109** 3.323 ** (0.50) (0.44) (0.34) (0.29) (0.25) (0.25) (0.27) status 7 2.066 ** 2.577 ** 3.237** 3.161 ** 3.198** 3.152 ** 2.694** (0.57) (0.61) (0.39) (0.32) (0.28) (0.27) (0.28) status 8 2.242** 2.420 ** 3.229** 2.881 ** 2.710** 3.007** 2.991 ** (0.55) (0.60) (0.43) (0.36) (0.32) (0.32) (0.37) status 9 3.626 ** 3.364** 3.OOO ** 2.862** 2.679** 2.680** 2.106** (0.84) (0.81) (0.52) (0.44) (0.37) (0.35) (0.42) Constant 1.809** 2.942** 3.597** 4.309** 4.725 ** 4.929** 5.600** (0.24) (0.25) (0.19) (0.17) (0.15) (0.16) (0.18) Sample size 905 1396 2310 3229 3805 3813 3035 F-test 52.35** 54.22** 116.3** 133.01 ** 159.73** 158.85** 108.82** R-squared 0.501 0.401 0.463 0.413 0.418 0.416 0.380 Adj R-squared 0.491 0.393 0.459 0.410 0.415 0.413 0.377 OLS estimates, standarderrors inbrackets; *=significant at the 5% prob.Level; **=significant at the 1% prob. level. 78 Another view at educational mobility consists of examining the relative importance of unobservables, including personal efforts and sheer luck, in determining educational attainments. This may be measured by the complement of the familiar RZstatistic to unity. Indeed, the higher the variance of the residual term, the more -mobilitywhen one downplays the sudden drop the R2for the youngest cohort, which, as seen before, may be there is'*. Tables 4a and 4b that a slight upward trend when one moves from older to younger cohorts even in due to people still studying in that cohort. This trendseems to be rather irregular, though. Finally, an interesting feature of intergenerational educational mobility i s that it seems to be influenced by intra-household decision mechanisms for women but not for men. Inthe case of women, the transmission of education from parents to children i s higher when the relative schooling advantageof the mother with respect to the father i s high. This effect i s significant and persistent across cohorts. This i s not true for men, however. Mothers' education seems to weigh more than fathers' but the difference i s significant only for a single cohort. Inbothcases, itisdifficult tofindatrendintheevolutionacrosscohorts. Ifany, itwould seemtobenegative. Another way of looking at intergenerational educational mobility i s through the usual transition matrices. Matrices for four cohorts are shown in tables 5a-d, where the number of years of education i s expressed as deviations from the mean, in order to correct for the general increase inthe number of years of schooling across generations - and across cohorts. Intergenerational educational mobility has clearly increased for the less educated individuals. For the older cohorts, around 80% of those individuals whose parents had two or less years of education less than the mean, stayed at a similar position with respect to their own mean, while for the youngest cohort this percentage has fallen up to 55%. For the higher-educated people, however, mobility first decreased and then remained more or less constant. Based on transition matrices, it would thus seem that the increase in intergenerational educational mobility depends very much on the weight given to the various educational groups. Another interesting feature in tables 5a-d i s the fact that a significant intergenerational downward mobility (with respect to the cohort mean) i s observed for any cohort, though at a decreasing rate. These transition matrices strongly suggest that there exist important non-linearities in the relation between parental education and that of the sons. To understand intergenerational educational mobility in Brazil well would require a much more detailed analysis. Inparticular, very much of the preceding discussion i s based on measuring education in terms of the number of years of schooling. One might prefer a more general approach where `human capital' i s what matters in intergenerational transmission mechanisms, human capital being measured by the cost of education, including foregone earnings, or possibly by the earnings that a given schooling level actually commands. The two approaches are equivalent when it i s assumedthat the rate of return to the number of years of schooling i s constant. They are not ifthe marginal rate of return to an additional year of schooling depends on the level of schooling, as done in the earning equations above. Also, the quality of schooling i s totally ignored in the preceding description of intergenerational educational mobility. But it cannot be ruled out that taking into account the quality of education so as to get closer again to a concept of human capital would modify the preceding conclusion of an increasing educational mobility in Bra~i1.l~These conclusions must thus be taken with very much care. See Behrman,Birdsall and Szekely (2000). l3 For some referencesto the role of educationalquality in shaping inequalitiesin Brazil see the motivation of the theoretical model in Ferreira(2000). 79 Table 5a: Intergenerational EducationalMobility: cohort 1936-1940 Individuals' years of schooling, in deviations from the mean Parental years of schooling, in devia- tions from the mean -2 or less Between 2 and -2 2 or more Total (mean of two parents) -2 or less 92.75 4.16 3.10 100.00 37.90 1.70 1.27 40.87 Between 2 and -2 66.61 13.78 19.61 100.00 32.98 6.82 9.71 49.52 2 or more 15.57 15.38 69.05 100.00 1S O 1.48 6.64 9.62 Total 72.39 10.00 17.61 100.00 Table 5b: Intergenerational EducationalMobility: cohort 1946-1950 Individuals' years of schooling, in deviations from the mean Parentalyears of schooling, in deviu- tionsfrom the mean -2 or less Between 2 and -2 2 or more Total (mean of two parents) -2 or less 85.13 7.88 6.99 100.00 25.76 2.39 2.11 30.26 Between 2 and -2 53.30 15.19 31S O 100.00 29.95 8.54 17.70 56.19 2 or more 6.80 10.17 83.03 100.00 0.92 1.38 11.25 13.55 Total 56.63 12.30 31.07 100.00 Table 5c: Intergenerational EducationalMobility: cohort 1956-1960 Individuals' years of schooling, in deviations from the mean Parental years of schooling, in devia- tions from the mean -2 or less Between 2 and -2 2 or more Total (mean of two parents) -2 or less 74.46 15.42 10.13 100.00 17.35 3.59 2.36 23.31 Between 2 and -2 38.73 24.91 36.36 100.00 22.65 14.57 21.27 58.49 2 or more 5.06 12.26 82.68 100.00 0.92 2.23 15.05 18.20 Total 40.93 20.39 38.68 100.00 80 Table 5d: IntergenerationalEducational Mobility: cohort 1966-1970 Individuals'years of schooling, indeviations from the mean Parental years o f schooling, in devia- -2 or less Between2 and -2 2 or more Total tions from the mean (mean of two parents) -2 or less 64.83 22.44 12.73 100.00 12.11 4.19 2.38 18.67 Between 2 and -2 32.59 34.30 33.11 100.00 19.42 20.44 19.74 59.60 2 or more 6.89 17.93 75.18 100.00 1S O 3.89 16.33 21.72 Total 33.03 28.53 38.45 100.00 Thezkmeof theendogenez2ybits Before puttingtogether the preceding wage and effort equations to measure the inequality of opportunities and i t s evolution across cohorts, it i s necessary to discuss the implications of the bias in the earning equation that could arise from the endogeneity of the effort variables, that i s their correlation with unobserved earnings determinants. As said above, there is no variable in the data source being used for Brazil that permits instrumenting satisfactorily the effort variables in equation (1) so as to test for the existence of such a bias and to correct for it. Instead, various experiments were made on the basis of the precedingmodels, which permitted defining useful benchmarks for the rest of the analysis. Practically, the problem comes from a possible correlation between the variables which are behind the residual, ui, of equation (1) and the variables, schooling (Si), schooling-squared (S2J and migration (Mi) included inEi. Let pus,pUs2andp , be the coefficients of correlation between the residual and these three effort variables, and ~ assume reasonably that the residual term, ui, i s orthogonal to all other circumstance variables in the earning equation (l)14. Let ~ M S ,~ M S Z be the known correlations between the effort variables and q the standard deviation of any variable X. Consider the covariance matrix C between circumstance variables (C), effort variables (E)andthe residual term, u, after normalizing all variables. E= C'E E'E E'u 1 C'C E'C 0 0 u'E u ' ] This matrix must be definite positive. Given that E'u in L'is not zero, OLS estimates of the regression of w on V=(C,E)i s biased. This bias writes: where N i s the number of observations and Exp i s the expected value operator.. The problem i s that pus,pus2, puMand o, are unknown. Let us concentrate on a, for a moment. OLS gives a biased estimate of it. By definition: 81 0,"= [w-V() + +b)]'[w-V(B+b)] where )is the vector of OLS estimates of the regression coefficients of w on V. Expanding this expression yields: 0,' =6: ++b'V'Vb (7) where a,' i s the variance of the OLS residuals. Substituting the definition of the bias b in (6) into (7) leads to : 0,' = 6,'+KO,' where K is given by: = ('9 pus 9 pus2 * O S 2 9PuM *O M)N(V'V)-' (O,PUS* 3pus2 9puM * )' It follows that the whole bias vector, b, and a, are perfectly known, once the correlation coefficients pus,pusz and puM,are known. As these coefficients are not known, the idea i s to do a Monte-Carlo analysis of the bias they imply by drawing them from uniform distributions over an arbitrary range. Of course, these drawings cannot be independent, since they must satisfy that the matrix Xis positive definite15. The resulting conditions imply inparticular that the coefficient K above i s less than unity. Thus we may write: 0,2 =- e 1-K This is of course what one expects. The true a, is larger than the OLS estimate, which has the property of minimizing the sum of squaredresiduals. The preceding methodessentially i s sensitivity analysis. Practically, 300 drawings were made and the inequality of opportunities evaluated for each permissible drawing. Calculations reported in the next section involve the mean of all these values and extreme values as intervals of confidence. 3-Simulatingthe effects of the inequalityof opportunities on earnings. The preceding models provide a simple way of measuring the effect of the inequality of observed opportunities upon the inequality of current earnings. To see how this may be done, the two basic equations (1) and (3) above are first rewritten with all circumstance and effort variables now being made explicit: Ln(wi)=ao RiaR GR~Lx;;+MPEi.aP DPE,.aD FOi.aF +Si.,Bs + + + + +S2i.,Bs2+Mi.pM+ui (8) E, =ao RiaR GR,.aG MPE,.ap+ DPE,.a, + + + +FOi.aF+vi (9) l5 Analitically, coefficientspus,pUs2and p uare drawnin three uniform distributionsanddrawingsthat do not satisfy the condition that Z ~ Ipositivedefinitearesimplydiscarded.pus2willbedirectlyestimatedfrompus,since psZu- E,whereE - = + os2Ios 2's'psu will also be drawn from a uniform distribution.In particular, we have imposed that own schooling,parentalschoolingand migration must not have a nega- tive effect on wages. 82 where E i s the vector {S,S2,M} and R, GR, MPE, DPE and FO stand respectively for the race dummies, the regional dummies, mean parental schooling, the mothedfather difference in schooling, and father occupation, whereas S is the number of years of schooling of the surveyed individual, S2 is the squared number of years of schooling and M hisher migrant status. Q, a%, UR,and aGare vectors of coefficients whereas other parameters are scalars. An appealing way of measuring the role of inequality of opportunitiesingenerating earnings inequality consists of evaluating what would be the distribution of earnings with the preceding system of equations if all the inequality due to the circumstance variables had been eliminated. Thus one may simply equalize all the circumstance variables across all the population and then, use (8) and (9)to figure out what the distribution of earnings would then be. Comparing with the actual distribution permits then evaluating the role of opportunities. Yet, a decision must still be taken with respect to the two residual terms ui and vi. If they are both interpreted as pure circumstance variables, inequality with respect to the effort variables should be considered as pure inequality of opportunities. Thus, equalizing opportunities would be equivalent to equalizing earnings. On the contrary, if the two residual terms were taken to reflect pure efforts, they must be retained when evaluating the inequality of opportunities. There clearly i s something arbitrary in deciding that the residual terms reflect inequality of opportunities or inequality of efforts, or some combination of both. Because of this ambiguity, measuring the `total' contribution of the inequality of opportunities to observed inequality might simply be impossible or totally arbitrary. Only the inequality of observed opportunities may actually be evaluated. This i s done in what follows through three different types of simulations: (i) Equalizingallcircumstancevariables,aswellastheeffortvariablesschoolingandmigration.Theresid- ual vi of the effort equations i s then considered to be full circumstance, but ui as pure efforts. (ii)Equalizingallcircumstancevariablesin(8) and(9).Schoolingandmigrationarenowconsideredpar- tially as circumstances. Inother words, bothui and vi are taken to be efforts. (iii)Equalizingallcircumstance variables with schoolingandmigrationconsideredaspureefforts.Thus, equations (9)are in (8) ignored and ui taken as pure effort. Comparing with the actual distribution of earnings, (i)i s equivalent to considering that all earnings determinants are circumstances; (ii) assumes that schooling and migration are only partly circumstances; (iii) postulates that there are no circumstances behind the effort variables. Inall cases the residual ui term i s taken as pure effort. It i s in that sense that evaluation of what follows refers to "observed" opportunities. It must also be clear that the preceding scenarios are mostly aimed at fixing some `bounds' for the role of this specific set of opportunities in shaping the actual distribution of earnings. Results are shown in figures la-b and 2a-b. Figure l a presents the contribution of inequality of opportunity to the total inequality of male earnings under the preceding scenarios using the Gini inequality measure. The top line represents observed total inequality for the various cohorts. Mean Ginis (as well as the minimum and maximum) resulting from the permissible Monte-Carlo simulations are then provided for the three scenarios above. When schooling and migration are considered pure efforts, the Gini coefficient drops by around 5 percentage points on average. When schooling and migration are considered partially circumstances, the Gini drops by around 10 points. When they are considered fully circumstances, the fall i s of 12-15 points. Interpreting the scenarios as providing bounds, it can thus be said that the inequality of observed opportunities represents at least 5 percentage points of the actual Gini, but most probably around 10 points as in the intermediate case and 12-15 if one i s ready to accept that there i s no effort nor chance in observed schooling. Of course, it could be more if other opportunity variables were observed (income and wealth of parents, land ownership,...). Figure l b gives the results using the Theil inequality measure, instead of the Gini. The Theil measure i s more sensible to the upper tail of the distribution. Equalizing circumstance variables then implies 83 larger drops of total inequality (around 25-30 point). Results for women (see figures 2a and 2b) show larger inequality drops. Dotted curves associated with the 3 scenarios correspond to the extreme values generatedby the Monte-Carlo experiment described above. It may be seen that the eventuality of a bias due to endogeneity of effort variables does not modify radically the conclusions derived from observing mean estimates, although estimated intervals of confidence are slightly higher for women. Figurela: Contributionof Inequality of Opportunity to Inequality of IndividualWages: 5- year cohort. Gini measuresfor men. 0.700 , 0.650 0600 1 ._...---. 0.550 1 0.500 1 0.450 0.400 0.350 ~ bl936-40 bl941-45 bl946-50 bl951-55 bl956-60 b1961-65 bl966-70 --t TotalInequality +Schooling -+-Schooling&Migrationfullycircumstances &Migration partially circumstances -8- Schooling&Migrationpureefforts Figurelb: ContributionofInequality of Opportunity to Inequality ofIndividualWages: 5- year cohort. Theilmeasuresfor men. 1.100 , 1.000 0.900 0.800 - 0.700 0.600 44 0.500 - 0.400 -I I 0.300 ~ bl936-40 bl941-45 bl946-50 bl951-55 bl956-60 b1961-65 bl966-70 -+- Total Inequality Schooling &Migration fully circumstances -8- Schooling&Migration partially circumstances Schooling & Migration pure efforts 84 Figure2a: Contribution of Inequalityof Opportunity to InequalityofIndividualWages: 5- year cohort. Gini measuresfor women. 0.650 0.350 I bl936-40 bl941-45 bl946-50 bl951-55 bl956-60 b1961-65 bl966-70 +Total Inequality -It-Schooling&Migrationfullycircumstances +Schooling &Migration partially circumstances -M- Schooling & Migration pure efforts Figure2b: Contribution ofInequality ofOpportunity to Inequalityof IndividualWages: 5- year cohort. Theilmeasuresfor women. 0.900 0.800 0.700 0.600 - 0.400 0.300 - 0.200 ~ bl936-40 bl941-45 bl946-50 b1951-55 bl956-60 bl961-65 bl966-70 --C TotalInequality Schooling &Migration fully circumstances --CSchooling&Migration partiallycircumstances --ESchooling&Migrationpureefforts What do we conclude from these experiments? Essentially that, even after discounting for inequality of observed opportunities, inequality in Brazilian earnings remains extremely high, that is, with a Gini of around 0.40 and a Theil of 0.35 in the most extreme of all scenarios. Two interpretations are possible. First, observed 85 circumstance variables (parents' education, father's occupation, region of origin, race) actually are a limited part of actual circumstance variables. Presumably, parents' income and wealth could explain much more of actual inequality. This i s an empirical issue which could be solved only by observing more circumstance variables or by getting an estimate of what those particular variables observed in Brazil actually represent with respect to other variables as observed in other countries. Unfortunately, we could not come up with the right comparison Brazil needs, that is, a country where a wider set of circumstance variables would be available, that would include the variables observed in Brazil. The second interpretationwould be that non-opportunity related earning inequality in Brazil i s very high, and presumably higher than in other countries, because of structural circumstances in the labor market, which remain to be identified. Another important conclusion i s that the proportion of inequality due to observed opportunities in actual inequality seems rather stable over cohorts, that i s over time, whether we look at Ginis or at Theils. Unfortunately, this conclusion may not be fully consistent with the two explanations given above. We have also analyzed isolated effect of each particular observed circumstance variable, for the intermediate case where schooling and migration are considered partially circumstances. Results with Gini and Theil measures are shown in figures 3a-b for men and 4a-b for women. Of all circumstance variables, parental education i s the one that plays the most important role in determining inequality. In this respect, it may be underlined that results are not very different when parents' schooling i s not equalized as above but a lower bound i s imposed as if schooling were compulsory until a certain age. In other words, it i s the inequality of education at bottom of the distribution that really matters. Interestingly enough, race alone seems to account for very little, when parental occupation and education are already controlled for. These results suggest that the most efficient policies for reducing inequality of opportunities in Brazil are those that may weaken the role of parents' education in own schooling and earnings. Figure3a: Oneby One Contributionof Inequality of Opportunity to Inequality of IndividualWages: 5-year cohort. Gini measuresfor men. 0.7 I I 0.55 1 I 0.5! bl936-40 bl941-45 bl946-50 b1951-55 bl956-60 b1961-65 bl966-70 +Total Inequality -#- Equalizlngonly race +Lower-bounding Equalizingonly regon Equaliangonly parentaleducation parental educationat 6 yschl Equalinngonly father's employmentsector 86 Figure3b: One by One ContributionofInequalityof Opportunityto Inequalityof IndividualWages: 5-yearcohort. Theil measuresfor men. 1.15 1.05 7 0.95 1 I 0.85 - 0.75 1 OX5 i 0.55 1 0.45 ~ bl936-40 bl941-45 bl946-50 bl951-55 bl956-60 bl961-65 bl966-70 +Total Inequality +Equalizing only race -4-Equalizingonlyregion +Lower-bounding +Equalizing only parentaleducation parentaleducationat 6 yschl. +Equalizing only father's employmentsector Figure 4a: One byOne Contribution of lnequalityof Opportunityto lnequalityof Individual Wages: 5-year cohort. Gini measures for women. 1 10.65 -I 0.6 I 0.55 - bl936-40 b1941-45 bl946-50 bl951-55 bl9.56-60 b1961-65 bl966-70 +Total Inequality +Equalizing only race I Equalizingonly region +Equalizing only parentaleducation +Lower-bounding parental education at 6 yschl. Equalizingonly father's employment sector 87 I Figure 4b: One by One Contributionof lnequalityof Opportunity to lnequalityof Individual II Wages: 5-year cohort. Theil measuresfor women. 10.85 I 0.8 -I 0.6 - ~0.55 ' 1 0.54 0.4 10.35 1 1 bl936-40 bl941-45 bl946-50 bl951-55 bl956-60 bl961-65 bl966-70 I -Total Inequality +Equalizing only race Equalizingonly region +Equalizing only parentaleducation I +Lower-boundingparentaleducation at 6 yschl. Equalizingonly father's employmentsector I 4-Theeffectsof the inequalityof opportunities on the distributionof householdincome The previous analysis refers to earnings and active individuals. The same type of analysis may also be conducted with exactly the same logic on households. The idea i s then to measurethe effect of the inequality of opportunities faced in the past by household heads and spouses on today's distribution of welfare within the whole population. The welfare measure used in this section i s the monetary household income per capita and the current distribution of welfare i s estimated by weighing all households by their size. Thus, the distribution of welfare i s defined over all living individuals. With this new definition, the inequality of opportunities faced by the parents now passes not only through their earnings as before, but also through participation behavior, fertility, non-labor income, and, of course, the matching of individuals within couples. Inan effort to capture these effects the previous earning model was re-estimated usinghousehold per capita income (hpcy) as the main left-hand side variable. In effect, three models were estimated. Inall cases, income per capita inhousehold i s given by the identity: yi' Y,, +Y,; +Yo; ni where yj stands for earnings of memberj (=h for household head ,s for spouse), yo i s non-labor income and n i s the number of persons in the household. Inthe first model, household per capita income inhousehold i i s specified as a function of the characteristics of both the head (h)and the spouse (s) : 88 As before, effort variables, E are taken to be functions of circumstances, C : E,, = Cjiaj vi with j = h,s + (11) In the second model, individual earnings for each main earner in the same household - head and spouse are - specified separately as a function of the same variables as inmodel (I). (ii) In simulations, household income per capita Yi is then computed using identity (0) with family size n being taken as exogenous. Inthe thirdmodel, the family size itself is beingmadeendogenous.Itis specified accordingto the multi-logit model: Pr{n, =k}= eztLk m with Zi = /Chi, C,i, Ehi, Esi}and k = /2, 3, .., 7 and more). Then household income Yi i s simulated considering changes in both earnings and family size. Based on these models, the effect of equalizing circumstances of household heads and spouses on household income per capita may be simulated in various ways. Comparing these ways permit identifying the role of circumstances on the following determinants of household income: individual earnings, participation and fertility. Consider the three following simulations: i. yi*:=yiI 11. .. yi*" :=(y,", +y; +YoJn, ... 111. where the left-hand side variable i s household income per capita simulated by equalizing circumstance variables according to model (I),(n),y**, or model (ID), that i s (II)and (12), y***. y*, y* i s simulated household per capita income obtained from equalizing circumstances simultaneously in all possible household income determinants. In other words, model (I) i s a reduced form where the role of circumstances of both household heads and spouses in fertility, non-labor income, participation and the earnings of participants i s simultaneously taken into account. y ** is obtained from equalizing circumstances only insofar as the earnings of participants i s concerned. In other words, participation behavior, fertility and non-labor income are maintained constant. Comparing the distribution of y* and y** thus indicates the role of circumstances in the determinants of household income per capita other than individual earnings. Comparing y*** and y** clearly permits identifying the role of circumstances in household income per capita inequality that goes through fertility. 89 Tables 6.1-3 present the simulations for the household per capita income models, considering schooling and migration respectively as fully circumstances -table 6.1, partially circumstances - table 6.2, and pure efforts - table 6.3. Inall these tables, cohorts are definedby the age of the householdhead. The drop of inequality that can be attributed to the inequality of observed opportunities i s roughly of the same order of magnitude for household income as for individual earnings. Interms of the Gini coefficient, inequality falls by some 14 to 18 percentage points when circumstances are equalized and both schooling and the migration status of the household heads and spouses are taken as fully circumstances. As before the remaining inequality i s still high by international standards, amounting to a Gini coefficient around .44. What i s more interesting i s that the comparison between the y* and y** simulation suggests that it is not only through the earnings of labor-force participants that the inequality of observed opportunities affect the inequality of current welfare levels but also through the other determinants of household income - i.e. non-labor income, participation, fertility and matching. Thus concluding from the similarity of overall effects that the inequality of opportunities plays the same role among individual earnings and individual welfare levels would seem erroneous. To be sure, comparing the first two rows in tables 6.1-6.3 to the fifth and sixth row, show that equalizing the role of observed circumstances in individual earnings would have an effect on the overall inequality of household income per capita that amounts to 7/13 percentagepoints of the Gini, that i s roughly 5 points below the effect obtained with all household income determinants. The second interesting result i s the important role played by fertility. Comparing the third and fourth blocks of table 6.1 shows that the role of the observed inequality of opportunities that goes through fertility behavior may account for 3/5 percentagepoints of the Gini in actual household income per capitainequality. It mustbe noted, however, that this effect i s purely mechanical in the sense that the induced effects of fertility on labor-force participation i s not taken into account. Thus, the fertility effect shown in table 6.1 may probably be considered as a lower bound for the role of inequality of opportunities that goes through family size and simultaneously participation decision. The overall effect of fertility would be bigger if this induced effect were accounted for. However, one can see by comparingthe fourth and the secondblock in table 6.1 that not so much i s really to be gained in including these additional effects. In effect, the inequality of opportunities that goes through participationand non-labor income i s extremely limited, except for the two older cohorts. The comparison between tables 6.1 and 6.3 shows that circumstances remain an important determinant of current inequality, even in the case where schooling and migration are considered as partially or completely as the result of individual efforts. The drop in the Gini coefficient coming from equalizing opportunities i s around 10 percentage points when all household income determinants are taken into account. in table 6.3. Another interesting finding i s that circumstances affect fertility directly rather than indirectly throught the schooling variable. Intable 6.3 where the equalizing of circumstances has no effect on schooling, their differential effect on fertility remains as high as in table 6.1 where schooling i s equalized at the same time as all circumstances. The same remark can be done for the role of opportunities' inequality that goes through individual earnings. It i s certainly less important than when schooling i s taken as pure or partial circumstance but it still amounts to approximately 5 points in the Gini coefficient. The same conclusion holds with the direct role of circumstances on the inequality of householdincome due to labor-force participation, non-labor income and matching. table 6 90 Table6.1 Contributionof Inequalit! If Opportunityto Inequalityof FamilyPer CapitaIncome: 5-year cohortSchoolingandMigrationas I illy circumstances. I1936-40 b1941-45 b1946-50 bl951-55 b1956-60 b1961-65 b1966-70 TotalInequality Gini 0.62 0.626 0.614 0.613 0.618 0.616 0.591 Theil 0.803 0.805 0.75 0.779 0.767 0.753 0.684 Joint Earnings, non-labor income, fertility & participation effects Gini 0.428 0.451 0.445 0.451 0.444 0.439 0.438 Theil 0.329 0.388 0.37 0.394 0.381 0.365 0.353 Earnings effects (non-labor in- come, participation & fertility keptGini 0.507 0.504 0.489 0.48 0.48 0.475 0.469 constant) Theil 0.484 0.472 0.454 0.449 0.459 0.43 1 0.409 Joint Earnings & fertility effects (non-labor income & participationGini 0.487 0.478 0.457 0.445 0.442 0.433 0.427 kept constant) Theil 0.432 0.41 0.389 0.37 0.378 0.357 0.343 Table6.2 Contributionof Inequalityof Opportunityto Inequalityof FamilyPer CapitaIncome: 5-year i b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 TotalInequality Gini j 0.62 0.626 0.614 0.613 0.618 0.616 0.591 Theil 0.803 0.805 0.75 0.779 0.767 0.753 0.684 ~ fertility & participation Earnings, non-labor income,Gini j 0.479 0.508 0.499 0.505 0.501 0.491 0.49 effects Theil j 0.44 0.5 11 0.489 0.499 0.489 0.459 0.455 Earnings effects (non-labor in- come, participation & fertilityGini , j 0.54 0.542 0.522 0.518 0.52 0.514 0.508 kept constant) Theil j 0.572 0.561 0.526 0.52 0.534 0.507 0.487 Earnings& fertility effects (non- labor income & participationGini 0.532 0.524 0.497 0.483 0.48 0.469 0.462 kept constant) Theil j 0.549 0.514 0.463 0.43 1 0.442 0.422 0.403 91 Table 6.3 Contributionof Inequalityof Opportunityto Inequalityof FamilyPer CapitaIncome: 5- year cohort.Schooling andMigrationas PureEfforts. /b1936-40 b1941-45 b1946-50 b1951-55 b1956-60 b1961-65 b1966-70 Total Inequality Ginij 0.62 0.626 0.614 0.613 0.618 0.616 0.591 Theil 0.803 0.805 0.75 0.779 0.767 0.753 0.684 Earnings, non-labor income fer- tility & participationeffects Gin{ 0.507 0.516 0.507 0.525 0.518 0.51 0.501 Theil 0.513 0.511 0.497 0.561 0.537 0.503 0.478 Earnings effects (non-labor in- come, participation & fertility ~ i n l0.592 0.583 0.56 0.567 0.564 0.561 0.537 kept constant) Theil 0.726 0.655 0.601 0.638 0.627 0.607 0.544 Earnings & fertility effects (non- labor income & participation Ginii Odo3 0.58 0.549 0.536 0.528 0.518 0.486 kept constant) Theilj 0.75 0.645 0.572 0.537 0.543 0.514 0.446 Insummary, the picture one gets from considering the role of observed opportunities on household income rather than individual earnings inequality i s quantitatively comparable but qualitatively very different. In particular, it seems to be the case that the proportion of the inequality of opportunities of household heads and spouses that i s transmitted to actual household income per capita through individual earnings i s important, but not exclusive More seems to go through other channels, fertility inparticular, but also labor-force participation, non-labor income, matching and particularly, fertility. For all these additional channels, the evidence reported inthis section suggest that circumstances play an important direct role intransmitting inequalities, even though the indirect effect through schooling i s far from negligible. 5-Summaryandconclusion In this paper, we tried to quantify the role of the inequality in observed opportunities of individuals - as summarized by their race, their region of origin, the education and the occupation of their parents - in generating inequality in current earnings or household income per capita inBrazil. There may be some biases in the econometric technique beingused, essentially due to the lack of adequate instruments for correcting for the endogeneity of some income determinants. Yet, a simple sensitivity analysis reveals some strong conclusions. Altogether, the inequality of opportunities that go through parents' schooling may be responsible for a very substantial proportion of total inequality in Brazil. Parents' schooling andor own schooling are jointly responsible for 12 percentage points in the Gini coefficient of individual earnings, on average across cohorts and genders. This percentage i s higher for older cohorts and for women. Out of this, 60 to 80 per cent may be attributed to parents' schooling alone. The same conclusion applies, although for other reasons, to household income per capita. In particular, it turns out that the role of the inequality of observed opportunities in shaping the distribution of household income per capita i s not only through the individual earnings of household heads or spouses but also through other channels : fertility in particular, and to a lesser extent, labor-force participation, non-labor income and matching behavior. 92 I Even though the role of the inequality of observed opportunities in shaping the inequality of current earnings or income i s a major one, what they leave unexplained remains very substantial. In effect, correcting for observed disparities in opportunities would leave the Gini coefficient of individual earnings or household income per capita certainly above .42. To the extent that no strictly comparable figure i s available in other countries, it may be difficult to see the bearing of such conclusion. Yet, one thing i s sure. Even after correcting for the inequality of opportunities, Brazil would still be high in an international ranking of inequality. Inparticular, it would necessarily be above all the countries with a Gini coefficient for actual incomes at .42 or below. On that basis, it i s tempting to conclude that the inequality of observed opportunities may not be enough to explain the excessive inequality observed in Brazil in comparison with other countries inthe world. Parents' education was shown to be the major source of inequality of opportunities in Brazil. It affects earnings either directly or indirectly through own schooling. The exercise undertaken inthis paper suggeststhat the latter effect represents at least 60 per cent of the overall effect of parents' schooling. Eliminating the influence of parents' schooling on the schooling of their children could have reduced the Gini coefficient of individual earnings by at least 3 percentage points on average across past cohorts and gender. Ignoring possible general equilibrium effects, this would clearly leave inequality inBrazil at a very highlevel. Successive cohorts faced different situations in terms of inequality of opportunities. Intergenerational educational mobility has increased over time, especially at the bottom of the distribution. Equivalently, the education of the parents became a less powerful predictor of the education of their children. For the moment, this evolution does not reflect itself completely in the evolution of earning inequality across cohorts. This is because the rate of returnto schooling and therefore the inequality of earnings i s strongly age dependent. Yet, it i s to be expected that because of that evolution earnings inequality along the life cycle will be smaller for younger cohorts. This point has to be checked by comparing different cohorts at the same age and at different points of time, which requires usingadditional data. 93 References Bowles, S. (1972), Schooling and inequality from generation to generation, Journal ofPolitica1 Economy, v80, n3, S219-S51 Burkhauser R., D. Holtz-Eakin, and S. Rhody (1998), Mobility and Inequality in the 1980s: A Cross-National Comparison of the United States and Germany, in S. Jenkins, A. Kapteyn and B. M. S van Praag (eds.) The distribution of welfare and household production: International perspectives, Cambridge University Press, 111-75, Checchi, D., A. Ichino, A. Rustichini (1999), More Equal but Less Mobile? Education Financing and Intergen- erational Mobility in Italy and inthe US, Journal of Public Economics v74, n3, p. 351-93 Ferreira, F. (2000), Education for the masses? The interaction between wealth, educational and political ine- qualities, Mimeo, PUC Ferreira, F.H.G. and R. Paes de Barros (1999): "The Slippery Slope: Explaining the Increase in Extreme Pov- erty inUrban Brazil, 1976-1996", Brazilian Review ofEconometrics, 19 (2), pp.211-296. Goux, D. and E. Maurin (2001): "La Mobilitd Sociale et son Cvolution: le r61e des anticipations rdexamind", Annales d'Economie et de Statistique, Vol:62, forthcoming Griliches, Z. and W. Mason (1972), Education, Income and Ability, Journal of Political Economy, v80, n3, S74-S103 J. R. Behrman, N.Birdsall, and M. Szekely, (2000), "Intergenerational mobility in Latin America : deeper mar- kets and better schools make the difference", in Birdsall, N. and c. Graham (Eds), New Markets, New Opportunities,Brookings, Washington Lam, D. (1999), Generating extreme inequality : schooling, earnings, and intergenerational transmission of hu- mancapital in South Africa and Brazil, Population Studies Center, University of Michigan, Report 99-439 Roemer, J. E. (1998): Equality of Opportunity, (Cambridge, MA: Harvard University Press) 94 4. INDIRECT TAXATIONREFORM: SEARCHING FOR DALTON-IMPROVEMENTS INBRAZIL Carlos EduardoVClez, Salvador Wemeck Vianna, Fernando Gaiger Silveira and Luis Carlos Magalges' Abstract Should Brazil reform indirect taxation? The considerable magnitude of the burden of Brazilian indi- rect taxation -three times that of direct taxation-, its dominant regressive effect on after-tax-income inequality and the heterogeneity of tax rates across goods and services,justify the consideration of indirect tax reform. Thispaper determines the set indirect tax changes that can improve Brazilian's welfare. It uses the "Dalton-Improving Tax Reforms" -DITAR- criterion proposed by Yitzhaki and Slemrod (I991), Mayshar and Yitzhaki (1995), Yitzhaki and Lewis (1996).Dalton Improving Tax Re- form simulationsfor Brazil reveal considerable room for improvement to move indirect taxation to- wards less regressivity and lower efJiciency cost. Out of 21 potential pairs of tax changes, 13 are found to be DITAR. One of the two "best" tax reforms -vehicles and personal expenses- concen- trates the benefits on the lowest quartile, while the other -housing and personal expenses- spreads the benefits more evenly across all income groups. This suggests that a tax reform trio -raising taxes for housing and vehicles, and lowering them for personal expenses- would dominate any other tax reform combination. JEL classification code: D31, D63, H22, H23, H24 Keywords :Iindirect taxation, inequality, tax reform. INTRODUCTION Should Brazil reform indirect taxation?There i s relative consensus about the necessity of reforming, or at least revising indirect taxation in Brazil. Well known arguments for such consensus are on one hand, the lost of competitiveness of national firms due to extremely highand inefficient taxation, and, on the other hand, federative issues, such as the "fiscal competition" among Brazilian states, coming from specific legal matters of taxation in sub-national levels. This paper shows that there i s a third important reason: reducing significant consumption-efficiency costs and moderating the regressive effects that characterize the current Brazilian tax system. The tax burden in Brazil has risen significantly along the 199O's, mainly because of the creation of cumulative indirect taxation; mostly at the federal level. The Constitution of 1988 had delegated to states and local administrations great autonomy on several issues, including fiscal issues16. In addition to rising state shares in federal tax revenues - Income Tax and federal VAT (PI,tax on industrial products, therefore called Federal VAT) -, they were allowed to set their own tax policy. Their most important tax instrument became the state VAT (ICMS, tax on goods and services, therefore State VAT). Although the state VAT existed before 1988, its tax base was more restricted- excluding services- and tax rates were determined at the federal level. Duringthe last decade, Brazil had to increase tax revenue vigorously in order to reduce the fiscal deficit associated to economic stagnation and hyperinflation. Simultaneously, with the weakening of * C. E. VBlez from World Bank and S.W. Vianna, F. G. Silveira and L.C. Magalhges from IPEA, Brazil. Special acknowl- edgement to Schlomo Yitzhaki (Hebrew University) and Rosane Siqueira (FGV-Brazil) for valuable comments. We thank the helpful comments from participantsin the seminar "A Desiguuldude no Brasil: Dimensoes, Peculiaridades e Politicus Publicus" in Rio de Janeiro (August 2001). Valuable research assistantship was provided by Taizo Takeno and Juanita Riafio. l6 The vigorous decentralizationmovement present in the Constitution of 88 has been interpreted as a natural reaction to more than 20 years of strong power concentrationin the federal government, during 1964-1985 military dictatorshipperiod in Brazil. For more details of the discussionsof tax reform inthat context see Wemeck Vianna (2000) andVarsano (1996). 95 Federal Government revenue sources produced by fiscal decentralization during the 1990s, Brazil was facing a severe macroeconomic crisis. The most accepted explanation was the persistent public sector deficit, which was associatedto the national development process carried out since the 1950's, mostly through considerable public investments and incentives. Hence, the correction of public sector imbalance, specially at the federal level required two steps: imposing heavy controls on public expenditures, and raising public revenues. Therefore, new taxes were created in order to increase public revenues and in most cases the preferred taxes were those which could generate great revenues with the lowest administrative costs. Efficiency and equity concerns were not the guiding principles for public administrators; because their most immediate goal was to raise public revenues promptly. Year after year, since the beginning of the 1990's, and in fact untiltoday, new changes have been introducedinto the national tax system, sometimes altering the rules of Income Tax - for households and firms - sometimes raising the rates of payroll taxes and contributions, other times just introducing new taxes - such as CPMF (tax on financial transactions). The common feature of this process was the following: the legal changes of the tax systemalways aimed to increase (federal) public revenues. T h i s paper shows that Brazilians could improve their welfare with a Dalton improving reform of indirect taxation. The best tax reform should combine a reduction of tuxes on personal expenses while raising them for private vehicles and housing. Moreover, the potential gains from such tax reform are significant when compared to the total tax burden. This paper is organized as follows. Section one characterizes Brazilian tax structure, its composition and magnitude with respect to the economy. This section presents the incidence of taxation on the distribution of secondary income of the population: total incidence of direct and indirect taxation on income distribution; second, it presents the separate redistributive effect of direct and indirect taxation; and third, it presents redistributive effects of indirect taxation decomposed by types of goods and services. The bias of decomposition i s evaluated by re-ranking households in the after tax income distribution. Section 3 presents the methodology and section 4 pre-identifies the best candidates for reducingand raising taxes according to efficiency and equity considerations. The last section simulates the DITAR reforms, explores the properties of the DITAR set and ranks them according to first and seconddominance criteria. 1- TaxationinBrazil:recent evolution,trendsandissues Increaszngburdenof zhdzkecttaatzonzn theZPPQx In the 2000 year, national tax burden reached its historical record, almost 33 percent of GDP17.Two macroeconomic programs were successful in rising tax collection inBrazil. The first one was the Tax Reform of 1967. It was implemented in the initial years of military dictatorship period - which started in 1964 and lasted 21 years -,and conceived in the context of radical transformations of the economy as a whole. This tax reform raised dramatically tax collection after no more than three years of implemented: the country's tax burdenjumped from 20.5 percent of GDP in 1967 to 26 percent in 1970, when it stabilized untilthe beginningof the 1990s (Aracjo (2001), Varsano et a1 (1998)). The second one was the Real Plan implemented in 1994, which produced considerate expansion of the tax burden by nearly 3.5 percentage points of GDP. Economic stabilization provided by the Real Plan along with the intensification of tax collection on good and services raised total tax revenue almost 30 percent of GDP in that year, from 26 in 1993 and 25 percent in 1992, respectively. Since 1994, total tax burden has been nearly 30% of GDP, until 1999, when it reached 31.6 percent, to hit a year after, the historicalrecord mentioned above of 32.6 percent. ~ l7According to our calculationsusingthe IBGE'snational accounts (still a preliminary result) 96 Most of the raise of tax revenues in the 1990's was produced by indirect taxation. Table 1 shows that in last two years there was a persistent growth of indirect taxation vis-a-vis direct taxation; in fact, almost the whole increase of tax burden in this period was due to that. The main explanation for this, i s the implementation in 1998 of a higher tax rate for Cofins -3 percent-, which finances Social Security. Created in the Constitution of 1988, its initial rate was 0.5%. Then, in 1990 it was increased to 2%, until 1998, when the raise of public revenues became imperative for the government to face the effects of the Russian crisis. Regardless its apparent low tax rate, Cofins i s a cumulative tax, with its incidence occurring at all stages of production chain. Hence, its final effective rate can get much higher, depending on the product. That i s why it representedlast year 3.5 percent of GDP and almost 11percent of total public revenues". FederalGovernmenttar collectionzi-predomznant,and22hasbeen zncreaszngzn recentyearx Table 2 shows that in recent years the federal share in total tax revenue increased pari passu with reductions in the share of state and local governments. From 1996 to 2000, federal tax collection as share of the total tax collection, increased 3.5 percentagepoints. This rise i s explained almost entirely for the rise in Cofins tax collection, 3.3 percentage points. Since its creation, collection of the tax on financial transactions -CPMF- has rapidly increased, 1.4 percentage points during three years. It is worth noticing that during this period, revenues of the federal VAT declined almost two percentage points. At the state level, tax collection decreased almost 2 percentage points duringthe same period. Localtax collection remained almost unchanged. Table 1. Tax BurdeninBrazil, 1996-2000, as percent of GDP and of total revenues (TR), by its main taxes and contributions Taxes and 1996 1997 1998 1999 2000 contributions % to % to % to % to % to GDP %toTR GDP %toTR GDP %toTR GDP %toTR GDP % to TR Indirect (a) 12,3 42,2 11,7 39,4 373 40,7 41,7 St.VAT 7,3 24,9 6,9 23,2 22,5 22,3 23,2 Fd.VAT 1,9 6,6 1,9 ($3 5 8 5 2 4,9 PIS' 0 3 3,1 0 3 2 3 2,6 3,1 2,7 Cofins' 292 7 6 291 7,1 6,6 10,l 10,9 Direct (b) 10,o 34,3 9,8 33,O 353 33,8 31,7 Income Tax 4,O 13,6 3,8 12,7 15,l 14,8 13,5 Pensions 5,2 17,8 5,l 17,2 17,2 16,O 15,2 IPTU~ 0,4 1 3 0,4 1,4 1,6 1,5 13 IPVA3 0,4 1 4 0,s 1,7 1,6 1,5 195 Subtotal (a+b) 22,3 76,s 21,5 72,4 73,O 743 73,4 Other taxes4 6,8 23,5 8,l 27,6 27,O 25,5 26,6 Total 29.1 100.0 29.6 100.0 29.6 100.0 100.0 100.0 Source:Arai?jo(2001), andVarsano et a1 (1998). Notes: 1. Contributionsto socialfunds. Cofins specificallyfinances Social Security. 2. Tax on urbanarea property; local level. 3. Tax on vehicles property; state level. 4. Constitutedmostlyby indirect taxes. Cofins i s an example of the modus operandi of the government for tax policy, in which financial objectives matter the most while distortions and inequality issues remain ignored. (According to calculations of Aratijo (2001), nearly 20% of total tax revenuecomes from cumulativetaxation). 97 Table 2. Tax BurdeninBrazil, 1996-2000,as percentageof total revenues, by levels of government Taxes 1996 1997 1998 1999 2000 Federal level FederalVAT 5,2 4,9 PIS 3,l 2,7 Cofins 10,l 10,9 IncomeTax 14,8 13,5 Pensions 16,O 15,2 IOF' 1,6 1,9 ImportTax 2,6 2,4 CPMF* 2,6 4,l CSLL3 2,2 2,4 Subtotal 58,2 58,O State level State VAT 22,3 23,2 IPVA 1,5 1,5 Subtotal 23,8 24,7 Local level IPTU 1ss4 1,5 1,5 1,8 1,8 Subtotal 3,3 3,3 Other taxes 15,8 16,7 15,9 14,7 14,O Source: Arafijo (2001) and Varsano et a1 (1998). Notes: 1.Tax on Finance Operations. 2. Tax on Finance Transactions. 3. Social Contribution on net profits. 4. Tax on services of any nature. T ucompet2ionacrossstates&nares constzfutionmandatesandproducestuheterogenezfy acrossstates. According to the Constitution, Brazilian states could only change their VAT rates throughout unanimous decisions at a special forum created exclusively with this purp~se.'~ practice, this In Council has lost its original function and, specially in recent years, states have been competing among themselves using this tax to attract new firms and investments, in a process known as "fiscal war". Effective rates of state VAT may differ from its nominal rates because of several kind of tax benefits and pure exceptions for specific goods. There are about 30 different taxes and contributions and some of them are not considered in the tax reform simulations of this paper because of the difficulties in calculating its impact over households. Such i s the case of import taxes and local level Tax on Services. For the first one the problem i s that Income-Expenditure Survey does not report about purchases of import goods, except for cars, and for the second one, the problem i s the vast number of rules applied in each city. fndirect taesarequzfeheterogeneousacrossgoodsduetoaccumuZation uf taesat dyferent Zevec'sof government. In Brazil, indirect taxation rates are quite heterogeneous. As Table 3 shows, indirect tax rates by types of goods diverge from 88 percent for tobacco to 4 percent for housing. After tobacco, personal ~~ l9The National Council of Finance Policy (CONFAZ), which i s constituted by all state finance secretaries, plus the Ministry of Finance. 98 expenses are the items most heavily taxed (33 percent), followed by leisure activities (30 percent), clothing (26 percent), as well as medication and health expenses (22 percent). Among goods and services with lower tax rates are vehicles (18 percent), food items (18 percent), and transportation services (17 percent). In terms of tax revenues, taxation of food items provides the biggest contribution to total tax revenues (28 percent), followed by vehicles (16 percent), transportation (12 percent), and clothing (12 percent). On the contrary, taxation of personal expenses produces the smaller contribution to tax revenues (4 percent), followed by medications and health expenses (5 percent), housing (5 percent), leisure activities (8 percent) and tobacco (8 percent). Table 3: Indirect taxation inBrazil: rates and revenues, 1999 Magnitude Share Tax (R$bill) Rates Goods and Services Vehicles 3.5 16% 18% Leisure 1.8 8% 30% Transportation 2.7 12% 17% Clothing 2.1 12% 26% Housing 1.2 5% 4% Personal expenses 0.9 4% 33% Medication 1.2 5% 22% Food 6.3 28% 18% Tobacco 1.8 8% 88% Source: Arafijo (2001) andVarsano et a1 (1998) IncomeExpenditure Service 1996, Authors' Calculations. Zssuesof tmatiun zn BraziL.tar competitiunamongstates, znterna?ionaZcompetzhkzess,equity andefficiency As mentioned before, the main aspects that have been regarded inBrazilian tax reform debate rest on two concems: the lost of international competitiveness of Brazilian firms, which results from high and inefficient taxation, and federative issues, such as the fiscal competition among states. However, up to now, issues of equity and consumption efficiency at national level have had not the required attention for a more efficient tax policy. Some recent studies examine distributive incidence of taxation in Brazil, all of them using IBGEs Income-Expenditure Survey of 1995-96. Magalhges et al (2001a and 2001b) are pure empirical studies which present estimations of indirect taxation on food products and medications, respectively. The results obtained point to a high regressivity of this kind of taxation, and to a promising social range to tax policies that give exemptions or subsidies to products which have great importance inthe expenditures of low income households. Barbosa (2001), applying an Almost Ideal Demand System (estimated by Asano & Fiuza, 2001), to a Ramsey model of optimal indirect taxation. 2- What is the impactof taxationonthe distribution of secondaryincome? Figure 1presents tax incidence as percentage of household income. Although income taxation has a moderate progressive effect on income distribution, its effect i s quite modest as compared with the regressive effect of indirect taxation. Thus, the regressive incidence of indirect taxation though small, dominates the progressive impact of direct taxation making total taxation regressive. While individuals in the lowest decile spend 30 percent of their income in taxes, individuals in the top decile, spend 17 percent. As Figure 1shows, the regressive impact of indirect taxation i s evident: in 99 the lowest decile, the tax burden associated to direct taxation i s 2.8 percent, while the tax burden associated to indirect taxation i s 10 times larger - 29 percent-. In contrast, in the top decile, tax burden is almost equally distributed, 9 percent associatedto direct taxation and 7 percent to indirect taxation. Figure 1:Tax incidence as percentageof householdincome, BrazilMetropolitan areas, 1999 1 2 3 4 5 6 7 8 9 10 Income deciles Source: Werneck Vianna eta/ (2001) TotaZtarationhasamoderateregressiveeffect on zncomedzktrz8utzon Table 4 presents the redistributive impact of direct and direct taxation by components inMetropolitan areas of Brazil. The distributive effect of total taxation on income distribution in metropolitan areas in Brazil is equal to 0.7 percentage points of the Gini coefficient (or after re-ranking household by after tax income). Applying Kakwani's (1975, 1976, 1983 and 1986) additive decomposition, we found the redistributive impact of taxation in the following way2': Indirect taxation i s regressive and it increases the Gini coefficient by 1.6 percentage points, while direct taxation is progressive and it reduces the Gini coefficient by 1 percentage point. Table 4 also shows tax collection by type of tax. While indirect taxation represents almost two thirds of total collection, direct taxation only represents one third. Components of indirect taxation by levels of government are relatively similar in terms of regressiveness. ICMS, PIS and PIhave similar concentration coefficients and GIEs. The progressive effect of direct taxation i s explained almost entirely by the Income Tax. Components of direct taxation are not homogeneous in their redistributive impact: income taxation represents two thirds of direct taxation and i s very progressive, it has a concentration coefficient of 0.85; urban real state tax *"Decompositions are based on pre-tax income household ordering. Moreover, they are not exact and are path dependent. However, redistributed bias due to income was found to be negligible, accordingto our calculations with the post-tax in- come distribution.See detailed explanation of the decompositionformulasin the appendix. 100 represents nearly 15 percent of direct taxation and i s almost neutral and it has a concentration coefficient of 0.61; and state tax on private vehicle ownership i s approximately 10percent of direct tax revenue and it i s also progressive with a concentration coefficient of 0.74. From Table 4 it i s obvious that the regressive impact of indirect taxation on income distribution dominates the progressive impact of direct taxation. This reflects the fact that indirect taxation falls more than proportionally on the income of the middle and low income households (i.e. the relative size of the concentration coefficients with respect to the Gini) and the larger magnitude of indirect taxation with respect to other sources of tax revenue. Table 4: Redistributive impact of direct and indirect taxation by components. Brazil, Metropolitan areas, 1999 Magnitude Targeting Redistributive Relative (R$bill) Share Concentratio Shareof Effect Delta Redistriutive n CoefficientTop 20% Poorest40% Gini % points Efficiency (*) IndirectTaxation 22.0 70% 0.417 50% 15.5% 1.6% 3.5 ICMS GoodsandServices 14.3 45% 0.424 49% 16.7% 1.O% 3.4 PIS 3.4 11% 0.402 49% 16.3% 0.3% 3.8 IPI 4.4 14% 0.406 50% 15.8% 0.4% 3.8 Total Indirect Direct Taxation 9.5 30% 0.735 96% 0.2% -1.0% -4.8 Incometax 6.8 22% 0.853 67% 7.0% -0.9% -6.0 UrbanRealState tax 1.6 5% 0.608 80% 2.1% 0.0% -0.7 State privatevehicletax 1.1 3% 0.735 89% 1.6% -0.1% -3.4 Total 31.6 100% 0.798 62% 12% 0.7% 1 Source: Arafijo (2001) andVarsano et a1 (1998) IncomeExpenditureService 1996, Authors'Calculations. Note: (*) RelativeRedistributiveEfficiency is equal to the ratio of the redistributiveeffect of each tax relativeto its share on aggregatetax revenue. Positivei s regressive andnegativei s progressive. Akhoughzndzkecttarationhasaregressiveihpact on zncomedzstrzhtion of metropoh2an householdsof Brazzl notaZlzndzkecttaxesareequal4regressive. Table 5 shows the decomposition of the distributive effect of indirect taxation by type of goods. Taxation of vehicles has the most progressive effect on income distribution, (concentration coefficient i s 0.72). As one could imagine, taxation of tobacco, food items and medications have the most regressive impact, its concentration coefficients are 0.19 0.31, and 033, respectively. Taxation of leisure activities has a neutral impact on income distribution (concentration coefficient 0.54). The remaining items displayed on Table 5 have regressive impact on income distribution: transportation (concentration coefficient 0.44), clothing (concentration coefficient 0.42), housing (concentration coefficient 0.40) and personal expenses (concentration coefficient 0.37). 101 Table 5: Decomposingthe distributive effect of indirect taxation by type of goods. Brazil Metropolitan areas, 1999 Magnitude Targeting Redistributive Relative (R$bill) Share Concentratio Shareof Effect Delta Redistriutive ~ Goods and Services Vehicles 3.5 16% 0.719 87% 2% -0.23% -0.9 Leisure 1.8 8% 0.544 60% 8% 0.03% 0.2 Transportation 2.7 12% 0.440 51% 14% 0.17% 0.9 Clothing 2.7 12% 0.418 49% 15% 0.20% 1.o Housing 1.2 5% 0.400 27% 29% 0.10% 1.1 Personalexpenses 0.9 4% 0.367 44% 17% 0.09% 1.3 Medication 1.2 5% 0.331 41% 19% 0.13% 1.6 Food 6.3 28% 0.311 40% 21% 0.78% 1.7 Tobacco 1.8 8% 0.186 29% 26% 0.34% 2.5 Total 22.0 100% 0.417 15% 51% 1.61% 1 Source: Arafijo (2001) andVarsano et a1 (1998) IncomeExpenditureService 1996, Authors'Calculations. Note: (*) RelativeRedistributiveEfficiency is equalto the ratio of the redistributiveeffectof eachtax relativeto its share on aggregatetax revenue. Positiveis regressiveandnegativeis progressive. 3- Daltonimprovingtax reforms: analyticalframework Dalton welfare improvements are weaker than Pareto improvements. While a Pareto improvement requires that welfare weakly improves for all individuals, Dalton Improvement allows for income of one individual to fall provided some poorer individual's income increases equally (Le.: "the rich are less deserving than the poor"). Inthis section we present the analytical framework to identify Dalton Improving Tax Reforms -DITAR thereafter- following Yitzhaki and Slemrod (1991), Mayshar and Yitzhaki (1993, Yitzhaki and Lewis (1996) and Yitzhaki (2001). Therefore, Dalton's welfare evaluation suppose a decreasing social marginal utility of income.2' That is, if welfare function W(.) i s defined as function of individuals' welfare Vh, h=1, ...H; W = W { V1 (q), ... VH(q)/;and changes inwelfare are givenby dW = 4 h )/3h.MBh where /3h i s the social marginal utility of income for individual h, and MBh i s the marginal benefit (in income units) for individual h, then the marginal social benefits of income decreases with income level of each individual, Le., >/32 >/33 >...>/3h >0. Definition: A Revenue Neutral Tax Reform i s a set of tax revenue changes for each type of good 6' = (61, 62, ..., 6n)-(ai > or e 0)- such that total change inrevenue i s zero ai)& = 0. The impact of the RNTR on welfare i s given by dW (a)= 4 h )Ph. MBh(6), where MBh(8) i s a function of behavioral parameters of LLLeconsumer anG the characteristics of the prevailing tax system. Assuming individuals welfare i s represented by the indirect utility function Vh(p+t, y), by Roy's identity the marginal change in utility of the h individual, produced by changing tax rate i by dti is equal to dVWdti=-xih. dVW& 102 then for individual h all tax changes imply that the total marginal benefitinincome unitsi s MBh = qi=I..N) xih dti Besides, it i s identically true that the change intax revenue for each tax is equal to marginal revenue times the tax rate change 6i = MRi. dti hence dti = 6i /MRi then substituting and adding for all goods (i=l..N) on the left hand side, and all individuals (h=I..H) on the right hand side we obtain Z(h=I..H)MBh = -Z(i=l..N) (Xi/MRi). 6i where Xi i s equal to the aggregate consumption of good i. Since Xi /MRi i s equal to the Marginal Efficiency Cost of Public FundsMECFi, the welfare impact of tax reform can be rewritten as qh=l..H)MBh = -qi=I..N) (MECFi). Si The intuition behind this expression i s the following, in income units, the welfare impact of the tax changes on each good i s larger than the direct impact of the tax reduction on Xi, Si, because MECFi are larger or equal to unity. This follows from the fact that Xi 2MRi becauseMRi =Xi + Z + ti. dXi/aqi in j+i (tj. dXj/dqi)) and the difference between them grows with the efficiency cost of taxation which i s directly dependent on the price effects on consumer's demand, dXi/dqi and aXj/dqi. In summary,,the MECFi = F(z' , d, lei/) i s a function of the tax rate vector, z', the price elasticity vector-own and cross-, {@i/,and the expenditure shares of all goods, d. In particular, the MECFi increaseswith the tax rate and with its own price elasticity. Obviously, the marginal benefits for each individual h is proportional to its share on consumption of each good ( x i m i ). That is, the marginal benefit for individual h i s equal to MBh = -qi=I..N) ( x i m i ) (MECFi).6i LznkzngtheDakonImprovzngReform toSecondOrderStochasticDomznancc uszng Concen- tration Curves. Proposition: Second Order Stochastic Dominance of post-reform disposable income relative to pre- reform disposable income is a sufJicientcondition for a Dalton Improving Marginal Reform 6' = (61, 62, ..., 6n)whereqi)Si = 0. &f The Welfare effect of the reform i s given by dW (6) = qk=l, ...H) ph MBh(6) (1) Second Order Stochastic Dominance of post-reform disposable income relative to pre-reform reform are non-negative for every percentileh= 1, ...H. That is, disposable income implies that Cumulative Marginal Benefits -CMB (6, h)-- produced by the CMB (6, h) = q k = l , ...h) MBk 2 0for all h = l...H (2) Which i s caused by concavity of the social welfare function and/or concavity of the utility function with respect to income, 103 Note that MBk(6) = CMB (6, k) - CMB (6,k-1) Therefore, by substitution in (1) we obtain dW (6) CMB (6, H) (PH - P(H-I))*CMB (8, H-1) + + ... + (P2 - Pl)* CMB (8, 1) and since the Daltoncriterion implies (Pk - P(k-1))20 and from (2) CMB (6, k) 2 0 for all k then welfare change i s non-negative dW(6)2 0 Definition: CC(i, h) i s the vertical coordinate of the concentration curve of good i at household h, therefore 1. CC(i, h) i s equal to the share of expenditure of good i of all household up to the hth. Therefore (2) can be rewritten as CMB(6, h) = 2(i=l,...n) [(MECF i)*(- &)* CC(i, h) J 2 0for all h = l...H (3) A&urz?hm andcharacterzzatiunuf the DATORsuZutiun Basically the problem i s to determine a vector of revenue neutral of tax changes 6' = (61, 62, ..., 6n) that satisfy equation (3) for all percentiles of the distribution of income. Following Yitzhaki and Lewis (1996) we chose one commodity as the numeraire give it a value 6i= 1or 6i= -1 percent of total indirect tax revenue. Satisfying equation (3) i s equivalent to finding a tax reform vector that makes the following equal to zero Min Zk{Max [-CMB (a), OJ], subject to 2&1,..,,)6i=O; 61zO22 6 The numerical algorithm was programmed in Excel for 200 percentiles of the income distribution. And it was executed for all 21 potential pairs of tax reform. According to Yitzhaki and Lewis (1996) the whole set of infinite DATOR can be characterized by those extreme solutions.23 4- Identifyingpotentialcandidates for raisingandreducingtaxes. The first step to understand condition (3) i s to explore its graphical interpretation. That is, the easiest way to identify the existence of a DITAR for two goods i s to plot the two concentration curves times their corresponding MECF. If their difference i s positive for every percentile -or Second Order Stochastic Dominance holds- then, a DATOR exists for a reduction of the tax on the good corresponding to the curve above, and a tax increase for the good corresponding to the curve below. This implies that the best candidates for tax reduction (increase) are for goods subject to larger price distortions -efficiency cost- and which tend to be consumed more than proportionally by lower income groups (and vice versa). 22As stated in Yitzhaki and Lewis (1996),p. 547. 23All infinite solutions in the DATOR set are convex combinations of the extreme solutions 104 while luxury goods are indicated by larger Concentration coefficients. In other words, the best candidates for tax increase are those goods with lower coordinates of the concentration curves and lower marginal efficiency cost of funds. That implies, raising taxes for luxury goods or goods for which the rich have a large share of expenditure, and/or goods for which tax rates are lower, and/or their own price elasticities are smaller. According to this criterion, the best candidates for tax reduction are on one hand, goods which exhibit large price distortions. Inother words, goods that when taxed, have higher efficiency costs becauseof the associated distortions on individuals' behavior. Larger price distortions are indicated by larger Marginal Efficiency Cost of Funds -MECF-, which are increasing on tax rates and Mashallian price elasticities of demand. That i s larger MECF indicates higher efficiency costs. On the other hand, goods which account for important shares in poor individuals' baskets. In other words, goods that when taxed have higher costs in terms of equity. Concentration coefficients indicates the share poor individuals spend on certain good. Lower Concentration coefficients indicates higher shares and vice versa. Therefore, from all goods displayed in Table 6, the best candidates on efficiency grounds for tax reduction because of the large price distortions associated to taxing its consumption -higher MECF- are tobacco24 (MECF equal to 1.5l), personal expenses (1.5 l), (1.33), and medications clothing (1.27). On the other hand, the best candidates on equity grounds for tax reductionbecause of the large share they represent on poor individuals' expenditure -lower Concentration coefficients- are, non- luxury goods such as food (Concentration coefficient equal to 0.31), medications (0.33), and personal expenses (0.37). Analogously, from Table 6 it i s shown that the best candidates on efficiency grounds for tax increase because of the small price distortions associated to taxing its consumption -lower MECF - are housing (MECF equal to 1.04), transportation (1.19), and food items (1.2). While on equity grounds the best candidates for tax increase because of the large share they represent on rich individuals' expenditure, are vehicles (Concentration coefficient equal to 0.72), leisure activities (0.54), transportation (0.44), and clothing (0.42). Table 6: Equity and efficiency of indirect taxation Equity Efficiency Concentration Gini income coefficient elasticity Tax rates MECF Vehicle 0.719 1.25 18% 1.23 Leisure 0.544 0.94 30% Transportation 0.440 0.76 17% 1.19 Clothing 0.418 0.72 26% 1.33 Housing 0.400 0.69 4% 1.04 Pers. Expenses 0.367 0.63 33% 1.51 Medications 0.331 0.57 22% 1.27 Food 0.311 0.54 18% 1.20 Tobacco 0.186 0.32 88% 1.75 Source: IncomeExpenditureService 1996, Authors'calculations By comparing the concentration curves of the burden of each tax we identify potential DATOR pairs. Inorder to check condition (3) we plot the burden of tax by income deciles (Figure 2) for food, personal expenses, medications, and vehicles. Goods with the higher tax burden are the candidates for tax reduction, while goods with the lowest tax burden are the candidates for tax increase. As Figure 2 24Tobacco is an special case becauseof the healthbenefitsobtainedby reducingconsumptionof this addictive substance. 105 shows expenditure on vehicles i s highly concentrated on top income deciles. Conversely, personal expenses, food items and medications, representan important share of lower income deciles. In other words, we can win on equity basis if we reduce taxes for personal expenses, food items or medications. Figure 2 also shows that whereas taxing vehicles generates the lower marginal burden per Real of tax, taxing personal expenses generates the higher tax burden. That is, we can win in efficiency basis if we increase taxes for vehicles. Joining efficiency and equity gains, we can see in Figure 2 that one DATOR i s the reduction of taxes for personal expenses and the increase of taxes for vehicles. By visual inspection, we can see that reducing taxes for medications instead of reducing it for personal expenses would reduce the gains on efficiency grounds. That i s because taxing medicines generates a lower marginal burden than taxing personal expenses, which means that reducingtaxes for medicines has lower efficiency gains than reducingit for personal expenses. Figure 2: Distribution of marginal burdenof indirect taxation on selected group of goods 1.6 1.4 X c 1.2 0 a 5t 0.8 E =I 0.6 .-c $0.4 2 0.2 0.0 0 1 2 3 4 5 6 7 8 9 10 Income deciles Sen weyarezndexhe&for thez2entfication of candiidateesfor trw reform Sen's welfare function might be a useful instruments to preidentify candidates for reform, because it provides the necessary conditions for stochastic dominance. Under a Sen (1973) welfare function W = p (1-G), where ,u is mean after-tax income, and G is the Gini coefficient of after-tax income; a reform with two tax changes 61 and 62 will bring the following effects to total welfare. dW(6) = -[A,Ul (I-G 71)61 + Ap2 (1-G 772)) 621 = MECFl (1- Cl)61 MECF2 (1-C2)62 + Since movements towards optimum taxation would require changes towards equalization of the marginal social burden of each tax, that i s towards MECFi (1- Ci)6i = K for all good i Therefore, for a tax reform, the goods with higher burden should reduce their tax rates and (vice versa). In the graph below (Figure 3) we plot the equity Ci and efficiency cost MECFi for each good. Each curve represents different combinations of tax efficiency and tax progressiveness that produce the same tax burden. Tax burden increases towards the origin. That is, from the three plotted curves 106 in Figure 3, the blue curve represents the combinations of tax efficiency and tax progressivenessthat produce the lower tax burden, while the red curve represents the combinations that produce the higher tax burden. The green curve plots combinations of efficiency and progressivenessthat produce the average marginal tax burden. Goods below the red line have the associated highest tax burden and thus, are the candidates for tax reduction, e.g. food, medications and personal expenses. Goods above the blue line -housing and vehicles- correspond to the opposite case, they have the lowest tax burden, and thus they are the candidates for tax increase. FromFigure 3 we can see that one DATOR alternative i s the reduction of taxes for personal expenses and a corresponding tax raise for vehicles. The reason i s that personal expenses represent an important share in poor individuals' expenditure and its taxation i s not very efficient; and vehicles because its taxation i s progressive and it i s highly efficient. Figure 3: Efficiency and equity of indirect taxation under Sen Welfare Index. Brazil Metropolitan Areas, 1999 8 Concentration Coefficient (Taxation Progressivity) 5- The pairsof tax changesthat satisfy the Daltonimprovementcondition The main questions of this section are the following: i s there room for improvement? How large i s the DATOR set? How dominant are the simulated reforms when compared to the status quo? Which taxes should be increased or reduced? How do tax reforms rank among themselves? And finally, how robust i s the set of DATORreforms to the variance of the parameter estimates of the demand system? Our simulations of tax reforms as pairs of tax changes reveal abundant opportunities for improvement. Out of a total of 21 potential pairs of tax reform25, 13 are Dalton Improvements (Second Order Dominant) relative to the status quo. Table 7 displays those reforms as "tax raise-tax reduction" pairs: housing-personal expenses, transport-personal expenses, vehicles-personal expenses, clothing-vehicles, housing-medication, clothing-housing, vehicle-medication, housing- 2521 i s the total number of pair combinations of 7 taxes: Personal Expenses,housing, transport, vehicles, medication,food, and clothing. 107 food, clothing-personal expenses, transport-food, clothing-transport, personal expenses-medication and medication-transport. Table 7: Tax rates changeand welfare effects for alternative pairs of Dalton improving indirect tax reforms, Brazil. Tax Reform Pair Tax rate changes by goods and Services Ejjkiency Gain* Winners* Personal Bottom Food Expenses Medication Housing Clothing Transpon Vehicles All 50% Housinevs Clothme 06% -18% 29% 5% 100% Direction of tax change D D DIU U UID U D U Initial conditions. Tax rate 18% 33% 22% 4% 26% 17% 20% Tax revenue R$ (Billion) 6.3 0.9 1.2 1.2 2.7 2.7 3.5 Note, D: DecreaseinTax rete fromthe initial level. U:Increaseintax rate fromthe initial level. U D : Increaseor decrease inthe tax rate. Max changein revenue of any indirect tax was limited to 1%of total tax revenue. (*) 100%mans the first order stochastic dominance relatlve to status quo. *As percentageof the revenueswift (15%of total indirect tax revenue) Moreover, the amplitude of the room for improvement i s confirmed by the fact that 6 of those 13 reforms are First Order Dominant relative to the status quo. That is, 100 percent of the households affected by those reforms are winners. The reforms are: housing-personal expenses, transport- personal expenses, housing-medication, clothing-housing, clothing-transport, and personal expenses- transport. According to our simulations taxes should be increased for housing, vehicles and transport and reduced for personal expenses and food. The magnitude of those changes varies across goods: increased by 0.6 percent for housing, 1.1percent for vehicles and 1.2 percent for transport, while it should be reduced by 6.4 percent for personal expenses and by 0.5 percent for food. The bigger increase i s for transport (1.2 percent) while the bigger reduction i s for personal expenses (6.4 percent). Tax changes for medication and clothing do not seem relevant because they increase for some reforms and decrease for others.26 26 See Yitzhaki and Lewis (1996),and Yitzhaki and & (19&&) 108 Table 8: Welfare rankingbetween the DALTONimproving tax reforms FirstOrder SecondOrder Third Order Stochastic Stochastic Stochastic Dominance Dominance Dominance II I I Housing vs PersonalExpenses 5 1 2 Transport vs PersonalExpenses 3 II 3 3 Vehicles vs Personal Expenses II 1 I 4 I 0 Clothingvs Housing I I I c c 0 1 4 Vehiclevs Medication Housingvs Food 0 I 2 I 2 Clothingvs Personal Expenses 0 2 3 TransDort vs Food III 0 1I 0 II 0 Clothingvs Transport 1 0 1 I 1 PersonalExpenses vs Medication I 0 I 0 I 1 Medicationvs Transport I 0 I 0 I 0 Source: See Table A2 in the appendix Table 8 presents the welfare ranking between the Dalton improving tax reforms. Best reforms are those that dominate -in welfare terms- a larger number of tax reform alternatives. Ranking tax reforms among themselves according to its first order dominance shows that the "best" i s housing- personal expenses followed by transport-personal expenses, and vehicles-personal expenses. However, according to its second order dominance, of the thirteen reforms listed in Table 8, the most dominant i s vehicles-personal expenses, followed closely by transport-personal expenses and housing medications, then by housing-food, clothing-personal expenses, and finally by housing-personal expenses, clothing-vehicles, clothing-housing, vehicles-medication, and clothing-transport. Under the third order dominance criterion, housing-medication and clothing-housing are the dominant, followed in first place by transport-personal expenses, and transport-food, in second place by housing-personal expenses, and housing-food, and in final place by clothing-vehicles, clothing- transport, and personal expenses-medication. As Table 8 shows, neither the transport-foodreform nor the medication-transport reform presents any improvement on welfare under these criteria. DATOR reforms can bring significant efficiency gains relative to the status quo. However, their rankings varies with the distribution of those gains. According to the gains for the average Brazilian household, the largest gains are for the tax reform involving housing and personal expenses (46 percent), followed closely by a second group of four reforms that includes vehicles-medication (35 percent), transport-personal expenses (32 percent), vehicles-personal expenses (29 percent) and housing-clothing (29 percent)27.The second group includes two reforms which gains are 23 percent: housing-medication and personal expenses-medication. The third group includes the following reforms: clothing-personal expenses (18 percent), housing-food (17 percent), clothing-transport (14 percent), and clothing-vehicles (12 percent). The group of reforms that presents the smaller gains for the average Brazilian household includes medication-transport (9 percent) and transport-food (2 percent). However, according to efficiency gains for the poorest 50% priorities are different. Preferred tax reforms involves taxing vehicles in exchange for tax reductions of personal expenses, medication and *'Theefficiency gain as a percentageof the revenueshift, 1 percent. 109 clothing. The maximum benefits for the top 50% of the income distribution would be 32% of the revenue shift i s produced by reducing taxes for Personal Expensed and raising them for housing. In summary, different a distributional perspective make a difference about the preferred good to tax more but not about the preferred good to tax less. How robust are the reform simulations relative to errors in estimates of the MECF? In order to answer this question we rerun all simulations for DATOR condition after modifying the MECF in order to reduce the chances of satisfying the conditions given by equation (3). MECF were increased and reduced by 10% corresponding to expected increased or reduction in taxes in the previous simulations of the DATOR set. The set of DATOR reforms shows clear robustness under those demanding conditions. The number of feasible DATOR was reduced to nearly half the size -7 instead of 13- and the persistent candidate for tax reduction i s still the personal expenses group (see Table && in the appendix). Moreover, the set of FOD tax reforms i s still not empty -3 instead of 6 options-. 6- Conclusions The considerable magnitude of the burden of Brazilian indirect taxation -three times that of direct taxation-, its dominant regressive effect on regressive direct taxation and the heterogeneity of tax rates across goods and services, clearly justify considering the possibility of indirect tax reform in order to improve Brazilian's welfare. This paper has identified the set of pair-wise indirect tax reforms that could improve their welfare, according to he Dalton improving criterion. The best indirect tax reform should combine a reduction of taxes on personal expenses while raising them for private vehicles andhousing. The potential gains from such tax reform are significant when compared to the total tax burden. Efficiency gains can reach as high as 46% of the tux shqt for the average Brazilian and over 30% of the tax shift for othe poorest 50 percent. And those estimates are relatively robust. Sensitivity analysis relative to the estimates of the marginal efficiency cost of funds shows that results are persistent and the main conclusion hold. 110 References Ahmad, E. and N. Stern. 1984. The Theory of Reform and Indian Indirect Taxes. Journal of Public Economics 25(3): 259-98 Castro, P. F., Magalhiies L.C. Recebimento e dispendio das famlias brasileiras: evidCncias recentes da pesquisa de orqamentos familiares (POF) - 1995/1996. Brasilia - PEA, 1998. (Texto para Discussiio Interna, no614) Creedy, J. Are Consumption TaxesRegressives?,WP 20197, Melbourne Institute Working Paper Se- ries, University of Melbourne, Melbourne, 1997. Dalton, Hugh. 1949. "The Measurement of the Inequality of Income." In The Inequality of Income. 1920. Reprint.London: Routledge & Kegan Paul, Ltd. Decoster, A., Schokkaert, E.,, Van Camp, G. I s redistribution through indirect taxes equitable? EuropeanEconomic Review, vo1.41, 1997. Feenberg, D.,Mitrusi, A. W., Poterba, J. M. Distributional efsects of adopting a national retail sales tax.NBERworking paper No. W5885,1997. Feldstein, Martin. 1972. "Distributional Equity and the Optimal Structure of Public Spending.'' American Economic Review 62(1):32-36. Hoffmann, R. Distribuigiio de Renda: medidas de desigualdade e pobreza. Siio Paulo, Editora da Universidadede Siio Paulo.1998. IBGE.Pesquisade OrgamentosFamiliares 1995'96. Rio de Janeiro, IBGE, 1998. Midia digital. IBGE. Pesquisa de Orgamentos Familiares 1995/96.Rio de Janeiro: IBGEDepartamento de hdices de Preqos, 1999. v.1. Despesas, recebimentos e caracteristicas das famlias, domicilios, pessoas e locais de compra. Kakwani, Nanak C. 1977. "Applications of Lorenz Curve in Economic Analysis." Econometrica 45(3, April)1719-27. .................... . 1975. "Applications of Lorenz curves in economic analysis". Washington, D.C.: InternationalBank for Reconstructionand Development. _____--__--____--_-_"Redistribution . 1983. effects of income tax and cash benefits in Australia". Mimeo. Kensington, NSW : University of New South Wales. Centre for Applied Economic Research. .................... . 1986. Analyzing redistribution policies :a study using Australian data. Cambridge [Cambridgeshire] ;New York :Cambridge University Press. Lambert, Peter, and Shlomo Yitzhaki. 1995. "Equity, Equality and Welfare." European Economic Review 39574-82. Lerman, Robert, and Shlomo Yitzhaki. 1994. "The Effect of Marginal Changes in Income Sources on U.S. Income Inequality." Public Finance Quarterly 22(4):403-17. Newbery, D. and N. Stem (eds), 1987. The theory of taxationfor developing countries. Oxford: Ox- ford University Press. Mayshar, Joram, and Shlomo Yitzhaki. 1995. "Dalton-Improving Indirect Tax Reforms." American Economic Review 84(4, September):793-808. -1 996.. "Dalton-Improving Tax Reforms When Households Differ in Ability and Needs." Jour- nal of Public Economics. 111 Petti, R. H. V. ICMS e agricultura: da reforma tributdria de 1965/67 ci sistemdtica atual, Itaguai, Rio de Janeiro, 1993, DissertaqBo (Mestrado em Planejamento e Politicas de Desenvolvimento Agricola e Rural na Ame`rica Latina e Caribe) - Instituto de Cihcia Humanas e Sociais - Uni- versidade Federal Rural do Rio de Janeiro. Rezende, F. 0 peso dos impostos no custo da alimentaGBo: anAlise do problemae propostas de redup Bo, Rio de Janeiro, mimeo, 1991. Rodrigues, J. J. Carga tributdria sobre os saldrios, Brasilia - Secretaria da Receita Federal: Coorde- na@o Geral de Estudos EconGmico e TributArio. 1998. (Texto para Discussiio no1) Siqueira, R.B., Nogueira, J.R., Souza, E.S. Uma andlise da incidkia final dos impostos indiretos no Brasil, Departamento de Economia, Universidade de Pernambuco, Recife, mimeo.1998. Siqueira, R.B., Nogueira, J.R., Souza, E.S. Impost0 sobre consumo no Brasil: a quest60da regressi- vidade reconsiderada, Departamento de Economia, Universidade de Pernambuco, Recife, 1999. Stem, N. 1987. Aspect of the Grneral Theory of Tax Reform. InNewbery, D. and N. Stem (eds), 1987. The theory of taxation for developing countries. Oxford: Oxford University Press and World Bank. Yitzhaki, Shlomo, and Joel Slemrod. 1991. "Welfare Dominance: An Application to Commodity Taxation." American Economic Review 81(3,June):480-9 Souza, M. C. S. TributaGBo indireta no Brasil: eficihcia versus equidade, Revista Brasileira de Economia, vo1.50, no1, p. 3 - 20. 1996. Vianna, S. T. W. TributaG6o sobre renda e consumo das familias no Brasil: avaliapio de sua in- cidgncia nus grandes regi6es urbanas - 1996.Rio de Janeiro, 2000, Dissertagiio (Mestrado em Economia). Instituto de Economia, Universidade Federal do Rio de Janeiro. 112 Appendix 1- Measuringincidenceofindirecttaxation2' A good example of the heterogeneity of the state VAT taxation across states is the case of food products, which differ on both the number of food items covered, as well as on tax rates. Some states have special tax treatment to groups of commodities defined as essentials so-called "basic food baskets". The "basic food basket" i s defined at the state level. Table A1 presents the data of all states related to their VAT taxation of these goods. The bigger basic food basket, in terms of number of items included, i s the one defined in the state of Santa Catarina, which includes 44 food items, while the smallest -17 items- i s defined in the state of Acre. The "mean" basket comprises 26 items, and it i s defined in the states of CearA, Espirito Santo, Minas Gerais, and Mato Grosso do Sul. Table A1 illustrates how, while nominal tax rates varies between 17 and 18 percent, with the exception of Espirito Santo -12 percent-, effective tax rates varies between 7 and 20,5 percent. Sao Paulo and Rio Grande do Sul have the minimumtax rate, 7 percent, while states such as Acre, Paraiba, Rio Grande do Norte, Sergipe, and GoiAs have the maximumtax rate, 20,5 percent. Table Al. State levelVAT incidence nominalandeffective onfood productsof statebasic - - food basketsfor allBrazilianstates. Modal Aver- RegiodState Nominal Tax Foodproducts Rate (%) inthe basket Tax Benefit or Credit age Tax Bur- den (%) NORTH Acre 17 17 no 20,5 Amazonas Not available Not available Not available Not available Par6 17 30 Yes 15,3 Rond6nia 17 22 Yes 13,6 Roraima Not available Not available Not available Not available Tocantins 17 21 Yes 10,4 NORTHEAST Alagoas 17 24 Yes 15,3 Bahia 17 21 Yes 795 Cear6 17 26 Yes 15,3 Maranhao 17 27 Yes 12,o Paraiba 17 21 no 20,s Pernambuco 17 29 Yes 7,9 Piaui Not available Not available Not available Not available Rio Grande do Norte 17 22 no 20,s Sergipe 17 24 no 20,s SOUTHEAST Espirito Santo 12 26 Yes 11,2 S5o Paulo 18 32 Yes 7,O Rio de Janeiro 18 25 Yes 16,2 Minas Gerais 18 26 Yes 11,2 '*There have not been great difficulties estimating of direct taxation incidence, because in this study we are working just with the payments related by families in Income-Expenditure Survey. 113 Table A1 (continued) Modal Aver- RegiodState NominalTax Foodproducts Rate(%) in the basket Tax Benefitor Credit age Tax Bur- den (%) SOUTH Paranfi 17 42 11,2 Rio Grandedo Sul 17 32 7,O SantaCatarina 17 44 11,2 CENTER-WEST Goitis 17 21 20,5 Mato Grosso 17 30 15,3 Mato Grossodo Sul 17 26 10,4 Distrito Federal 17 34 15,3 Source: Wemeck Vianna et a1 (2000). 2- Calculating the effect of taxation onincomeinequality -Ginicoefficient-by Kakwani's method A Gini = Gini ({Yi - Ti, i=l..n}) - Gini ({Yi, i=l..n }) where {Yi, i=l..n }: income distribution {Yi - Ti, i=l..n } : after tax income distribution and Ti = qj.=I..k) Tij,where Tij denotes the taxes paid by individual ion goodsj. A Gini = CC ({Yi - Ti, i=l..n}) - Gini ({Yi, i=l..n }) +Gini ({Yi - Ti, i=l..n}) - CC ({Yi - Ti, i=l..n}) A Gini = CC ({Yi - Ti, i=l..n}) - Gini ({Yi, i=l..n }) H + (1) where H i s the change in inequality due to the re-ranking effect caused by taxation (which should be non-negative) H = Gini ({Yi -Ti, i=l..n}) - CC ({Yi - Ti, i=l..n}) Kakwani showed that (1) can be re-written as an additive decomposition by different type of taxes (j =l..k) as follows: A Gini = aj =l..k) AGini (j) H + where the impact of each tax i s equal to AGini (j) = - [CC(Tj) - Gini(Y)] @. B/(l-B) and $ji s the share of tax j in total tax revenue and 8 i s the ratio of tax revenue to total household income -1 3 percent-. Equation (2) implies that any tax with a Concentration Coefficient smaller that the Gini of income (0.578) will produce a regressive effect, and viceversa. This condition is equivalent to the Gini income elasticity for tax j GlEj = CC(Tj)/ Gini(Y ) being below (or above) unity. Which in terms of the Lorenz curve and the Concentration Curves means that taxes with concentration curves inside the Lorenz curve for income increase inequality and viceversa. 114 Applying equation (2) to the Brazilian data - tax revenue by tax type and incidence by tax type from the incomeexpenditure survey- we obtain Table 4 and Table 5. 3-StatisticalAppendix TableA2.a: Welfare rankingof alternativeDALTONimprovingindirecttax reforms(pair of Bl& inupperdiagonalmeansthe pairofreformd m not dominateareformlistedina correspondingrow. Blankinlower diagonalmeansthe pairofreformarenot domintedby areformlistedina correspondingcolumn. FOD :First Order Stochastic Dominance. SOD : Second Order Stochastic Dominance. TOD :Third Order Stochastic Dominance.(Small inequality aversionis sufficient conditionfor this result) nd :Rankingwas not determined 115 TableA2.b: Welfare rankingof alternativeDALTONimprovingindirecttax reforms (pair Blaniinupperdiagonal meansthe pair of reform does not dominatea reformlistedina correspondingrow. Blank inlower diagonal meansthe pair of reform are not domintedby areform listed ina correspondingcolumn. FOD :First Order StochasticDominance. SOD :SecondOrder Stochastic Dominance. TOD :Third Order StochasticDominance.(Small inequality aversioni s sufficient condition for this result) nd :Rankingwas not determined 116 5. SCHOOLINGEXPANSIONINDEMOGRAPHIC TRANSITION: A TRANSIENT OPPORTUNITY FORINEQUALITY REDUCTION INBRAZIL Carlos Eduardo Velez, Marcel0 Medeiros, and Sergei Soares * Abstract Thispaper explores the connection between Brazil's schooling inequity and deficiency and demographic transition. The most general versions of the Kuznets curve hipothesis highlight the role of supply side factors -demographic transition and the expansion of education- in redistributing assets and modifying the relative prices of skills, with clear efsects on income inequality and growth. The dynamics of this process might be perverse, and unless education opportunities are vigorously equalized, developing economies might converge high inequality equilibria (Kremmer et al, 2002 and Lam, 1999). This paper tries to understand how demographic transition in Brazil modifies the time lag required to extend to the whole labor force the educational improvements enjoyed by younger cohorts -the stock-to-cohort time lag- More formally, we want to answer how time correlation between educational efforts and demo- graphic transition affects long term inequality. This paper simulates what would have happened if the educational expansion of the 90s had occurred one decade earlier -before the demographic transition started. Results show that in the long run taking advantage of this window of opportunity to expand edu- cation reduced the stock-to-cohort time lag from 25 to 20 years, and long term inequalities of schooling and labor income. However, in the short run schooling inequality overshoots temporarily -induced by rising between inequality-. Another lesson is that, even very strong improvements above the current trend of schooling attainment, take more than two decades to show-up as higher educational endow- mentsfor the whole working age population. That is, by taking demographic inertia into account, policy makers conviction about education should be reinforced with patience and the will to monitor educa- tional policy outcomes with a clear long termperspective. JEL classification: 03, IO, 12, 02. Keywords: Education: Inequality, Policy evaluation, Kuznets Curve. Introduction How long will improvements in schooling of younger cohorts take to change the distribution of educational endowments of the total labor force and, in turn, change the distribution of labor income in Brazil?When rates of return to schooling are significant, as they are in Brazil, the size and distribution of educational endowments determines to a large extent the distribution of labor income. However, improvements in the educational attainment of younger cohorts do not translate immediately into proportional improvements for all cohorts of the economy. Demographics might play an important role in that process. This paper attempts to develop a demographic model linking the educational profiles of successive cohorts of individuals entering the labor force with the level and inequality of educational endowments of the whole labor force. We ask how the demographic transition might affect the impact of cohort educational profiles on the level and inequality of educational endowments of the entire labor force. For example, an aggressive education policy to improve high school completion should result in large differences between the educational profiles of younger versus older population cohorts in the labor 117 force, with obvious effects on the distribution of the educational stock. However, the size and speed of that effect will depend on the pace of demographic transition in Brazil. Presumably, if demographic transition has not been completed and the fertility rates remain high, the effect will be larger and faster. The profile of population growth for Brazil will show how large those effects will be, and how long it will take to observe them. Moreover, it will show whether Brazil's position inthe demographic transition provides an opportunity (or makes it more difficult) to reduce the inequality, or improve the level, of the educational stock of the labor force. Hopefully this model could be a device to show how much time i s required to recover the full social benefits of sustained investments in education. That is, to understand the links and the lags between current policy actions and future outcomes. In other words, it would enhance the value of current policies in terms of the equity improvements for present and future generations. Considered within the Latin American context, Brazilian educational outcomes are lacking and the allocation of resources towards education i s low. From the outcome side, educational attainment i s too low and the differentials in access to education are significant both across regions, and income groups. From the allocation of resources side, public expenditure in Brazil i s too biased towards public pension subsidies, while the share of education is small. These characteristics of the educational system of Brazil become particularly worrying given the fact that Brazilhas one of the highest levels of income inequality in the world and it is clearly linked to inequality of education. Bourguignon, Ferreira and Leite (2002) show that nearly 30% of excessive inequality relative to the US i s explained by educational inequities. The original version of Kuznets Curve hypothesis emphasizedthe role of demand-side forces to shape the relation between income inequality and economic development. Basically, the dynamics interaction of technological change and the induced demand for capital skills where supposedto explain why inequality first rise and the fell with development. A more general version adds the role of supply side determinants, namely arguing the impact of demographic transition forces could flood the market with young unskilled workers reducing the rise of inequality. Hence the age-earning curve would be flatten once those fat cohorts reachthe pick earning age. 29 Demographics i s particularly relevant for developing countries because as Higgins and Williams (1999) show, the demographic transition in these countries "has generated much more dramatic changes in relative cohort size than did the baby-boom in OECD countries. (p4). Evidence from Higgins and Williams provides support for a linkbetween cohort size aggregate inequality. The estimated quantitative impact i s considerable: a one standard deviation increase in the fraction of population in peak earnings would increase inthe fraction of population in peak earnings would lower a country's Gini coefficient by 6.5. 30 However, decomposition exercises show that the demographic transition has two opposite effects on inequality. The first effect which increases inequality, i s the change in the composition of the labor force and while the secondone, the change in the age-earning profile, reduces inequality. A more sophisticated perspective by Kremmer and Chen (2002) shows that on its own course, the dynamics of education inequality could lead to a perverse cycle of increasinginequality in a country such as Brazil. Nevertheless, according to their model the timely enhancement of educational opportunities for the poor i s critical to avoid this outcome, and position the economy on path leading to a steady state with a more balanced distribution of skilled and un-skilled workers. Empirical evidence -by Kremer and Chen *'The analysis of adverse supply effect on the relative wages of the baby-boom cohort in the United States is presented in East- erlin 1980; Freeman 1979; Welch, 1979; Lam, 1997; Murphy and Welch, 1992; and Murphy and Katz, 1992. 30Over time the relation is becoming weaker: stronger in the 1970sand 1980s but non existent in the 1990s. 118 (2002)- suggests that the fertility differentials between educated and uneducated parents i s stronger in more unequal countries -like Brazil-.31 If children of uneducated parents are less likely to become educated, the fertility differential will induce an increasing proportion of unskilled workers in the next generation. Which in turn tend to depress their wages and increase their chances of having more children and so on. 32 Based on a dynamic markovian framework of fertility and education inequality across generations, Kremer and Chen (2002) show depending on the initial conditions the economy might converge to high or low inequality scenarios. "If the initial proportion of skilled workers i s too low, inequality will be self reinforcing and the economy may approach a steady state with a low proportion of skilled workers and greater inequality between the skilled and unskilled." These findings have the most important implications for the timing and efficiency of educational policy. According to their estimates, in middle income economies like Brazil a temporary increase in schooling opportunities for the children of the poor that raise the share of skilled workers above a certain critical value, would induce a virtuous dynamics of education equalization across generations. The key question then is, could the window of opportunity for this policy intervention been expiring?: as time passes, have Brazil reached the point in which for producing the desired outcome the required effort i s too large? Moreover, given the fact that lower fertility rates reduces the demographic weigh of younger cohorts, i s the leverage of current educational policies to modify the distribution of the whole labor force still available? Kremer et a1(2002) show that any effort to reduce the unitcost of taking the children of the poor to reach high educational attainment has the same consequences. Hence a whole range of policy instruments can produce the desired effect: from improvements in nutrition and childcare to the incentives to reduce unit cost and improve quality in the allocation of public educational funds. Moreover, they also show that if fertility is endogenous to skill wage differentials temporary policy interventions can have even larger multiplier effects. This paper i s organized as follows, section 1describes the demographic background of Brazil, section 2 presents the methodology and data used in this paper, section 3 depicts the evolution of education between cohorts, section 4 defines the stock-to-cohort lag of educational attainment, section 5 shows the results of the simulations, and finally, section 6 presents the conclusions. DemographicBackground According to the 2000 Census, the Brazilian population amounts to 170 million people, most of which live in the coastal urban area. Spatial differences are strong. Brazil i s divided by geographers into five regions: South, Southeast, Center-west, North, and Northeast. The first two are the most developed and rich, the last one i s the poorest. Population density can be considered highinthe metropolitan areas of all regions, medium in non-metropolitan areas of the Southern and Southeastern regions and low in rural areas of the Northern and Center-westem regions. As in other countries, the Brazilian demographic history of the last hundred years can be divided into three periods. The first ranges from the early 1900's to the late 1930's, when birth and death rates were high. However, as mortality balances birthrate, a good part of population growth was due to international immigration. The second period begins after the 1930, when international migration i s reduced and the falling mortality together with high fertility became the main reason for rapid population growth. Rates were at their peak, around 2.9% per year, during the decades of 1950 and 1960. The third period begins 31Most Latin American countries show very high fertility differentials.Well above the predictedlevel conditional on inequality. Those error terms are the highest in the case of Colombiain the 1990s. 32Assuming the substitutioneffect dominates the incomeeffect. 119 at the late 1960's, with a rapid fall in fertility rates and, therefore, of the population growth. Mortality keeps falling duringthe period but its level i s not enough to undo the effects of the reduction in fertility. Population growth rates are estimated at 1.3% inthe inthe late 1990's. The story above happened in all Brazilian regions, but not at exactly the same time. Except for migration, the demographic patterns of all regions followed, with some delay, what happened in the Southeast. Furthermore, in the last decades the demographic patterns of all regions became much more homogeneous than before, although we can still identify clear differences among them. Figure 4.0% a, 6 3.0% .-C - 0 gj 2.0% Q a 0 1.O% 0.0% During the 1990's infant mortality has fallen significantly, but not enough to compensate reduced fertility. As a result, during this decade younger cohorts are smaller than their predecessors.Although the pressures for the supply of schooling caused by total population growth are reduced, other factors of pressure such as short distance migration and the increase of school enrollment are still ineffect. The net effect of the demographic transition of Brazil on the age composition of the labor force i s shown on Figure 1, below. This Figure shows the composition of the whole working-age population every 10 years from 1950 to 2000, and provides some forecast for the period 2010 to 2040., namely all those between 16 and 70 years of age in any given calendar year, in terms of cohorts. It i s clear, that the decreasing demographic weight of the youngest cohort of the labor force -16-20 year olds- that started in the 1980s. It i s notable that for the labor force of 1970 the demographic share of the youngest cohorts peaks for the labor force borne in 1954 (4.2 percent) and then falls significantly for the labor force in 1990 and 2000. That is, for the cohorts born in 1974 and 1984 the demographic share fell to 3.5 and 3.2 percent respectively. Up to the year 2000 population, younger cohorts are always more numerous than older ones. From year 2010 onwards, the functions are no longer monotonously increasing and there i s a point from which newer cohorts become less numerous. Our thesis i s that the strongest educational expansion should coincide with the peak of demographic "replacement" -when young workers have the largest population share-. This window of opportunity for education should not be missed if there i s a concern for closing the educational gap and reach more equitable access to education. 120 Methodology And Data The methodology to be used in this paper will be the simplest possible capable of providing an answer to the questions on the interplay between the educational level and inequality of each cohort and the educational level and inequality of the population as a whole in a given calendar year. Let: T index calendar years (year of observation) t index cohorts S, be the final average educational level of cohort t. 1, be a decomposible measure of final educational inequality of cohort t. 8 (T,t) be the weight of cohort t inthe population aged 16to 70 inyear T. Then: ST = zte(T,t)st the final educational level of the 16-70 population in year T i s a e weighted average of the final educational level of each cohort Weights vary over time due to demographic transition. IT=Ztf(T,t) It Ztg(T,t)St + the final educational inequality of the 16-70 population. We adopt decomposable entropy inequality measureE2. The methodology consists in (i) Estimating the final educational attainment --mean and inequality- of each cohort (including those younger cohorts which have not converged yet). (ii) Measuring the stock- vs-cohort time lag for the calendar years 1970-1998. And (iii) Measuring the impact of the timing of educational: Simulating temporary deviations from the path of cohort education expansion in periods of higher demographic growth and establish its impact on the stock-vs-cohort time lag and on aggregate inequality. The decomposable inequality measure we decided to use i s one of most common: one-half of the squared coefficient of variation. According to Shorrocks (1980), this measure corresponds to the member of the generalized entropy class with an inequality aversion parameter of 2. This inequality measure can be decomposedinto within and between components by usingthe following decomposition weights: w(Tt) = e(rt). ( S J S ~ ) ~ Henceforth, we will always refer to this inequality measureas 12. Keep in mindthat the inequality of education i s linked to the inequality of labor income var (log y): var (logy) = p2 var (E) + var (p), where pisthe returnto education in a linear Mincerian equation and p the error term, 33 and var (E) = I 2 mean (E) . The data we use are all from the Pesquisa Nacional por Amostragem de Domicilio (PNADs) from 1977 to 1999. These are surveys covering the whole nation, except for the rural area of the Northern region, where the vast distances make a yearly survey too costly. The sampling scheme has been the same - stratified and clustered - but the strata change every time the Census Bureau Grid changes, which 33 For a general version see Lam(1999). 121 happens every 10 years with the national Census. The questionnaire has changed considerably over time, but schooling and age, the only variables important in this study, have been largely spared. The PNAD imposes two shortcomings upon our analysis. The first i s that the same people are not followed over time. This means that we do not have real cohorts but pseudo-cohorts. In principle, this should not be a problem, if we believe in the PNAD sampling scheme. The second problem i s that the PNADsexist only from 1977to 1999. This means that any one cohort was followed only duringa part of its evolution. The Evolution Of Education Between Cohorts: MonotonicallyIncreasing Mean, Decreasing Ine- quality And InvertedU-ShapedMean-Variance Schedule Figure 2 shows educational progress in Brazil. On the horizontal axis is the year of birth of each successive cohort from 1900 to 1983, on the vertical axis i s the estimated final average educational level of the cohort. We will explain exactly how this estimate i s made later on, but for now what i s important i s that average education i s a monotonically increasing function of cohort date of birth but the rate of increase i s not fixed. Figure 2 shows that the education of each cohort increased at a more or less steady rate until about the 1940 cohort, accelerated for those born between 1940 and 1960, slowed its rate of growth for the 1960's cohorts and then acceleratedagain for the cohorts bom after 1970. Rgure2- AverageYearsofEducationbyCohort ~ , 0 20 40 60 80 100 ~ Cohort Figure 3 shows the same for the Z2 measure of education of the cohort. Once again, the most important fact i s a monotonic relation - each successive cohort has less internal inequality the previous one. It i s important to note that the Z2 measure does not bear a linear relation to the amount of income inequality explained by education, and this i s due to the highly nonlinear returns to education in Brazil. 122 I Figure 3 Inequality in EducationalAttanment by Cohort - .-"6 3 1.4 1 1.2 - 1.0- c a 0 0.8 - 2s 0.6 - $ - 0.4 - 0.2 - 0.0 -T- 0 20 40 60 80 100 Cohort As a r sult of this evolution the educational attainment by Cohorts has taken an interesting shap n inverted U of the mean-variance schedule , with decreasing variance since the early 1960's. Figure 4 shows a wide range for both mean education and variance and wide oscillations in the variance. The latter would be approximately 65% of the level achieved in 1959. This trend over the last four decades i s due to the reduction in education inequality much faster than the increase of average education (square). Demographic simulations suggest that it takes more than two decades to see the benefits of increasing educational attainment for the younger cohorts reflected on the whole labor force. One way of looking at how contemporary educational policy affects the distribution of educational endowments of the whole labor force i s to measure how many years it takes for the whole labor force to reach the level of educational attainment of one cohort. Our observations show that the labor force of 1970s had the same number of years of education of the cohort born in 1940, which on average was finishing school in 1951 (entering school at 7 years of age and attaining nearly 4 years of schooling). Resulting in a gap of 19 years between 1951 and 1970. That gap grew over time to a maximum of 25 years at the end of the century. That is, the labor force of 1998 had 6.5 years of schooling, which was the same educational attainment obtained by the cohort born in 1960-that on average was leaving school in 1973-74. The fact that the gap -between the cohort and the whole labor force- grew more that the marginal increase in schooling -6 more years for the gap versus 2.5 years of mean school attainment- i s associated with the demographic transition. That is, the decreasing demographic weight of the youngest cohort of the labor force of 1990 and 2000 vis a vis the labor force of 1970and 1980. 123 Figure 4 EducationalAttainment by cohort: Mean-variance schedule. Ob- served and simulated. Brazil 1920-2009 Path of Educational Attainment by Cohort: Mean and Variance. Observed and Simulated. Brazil 1912-2009. 11 , 10 9 6 5 I I I I I 0 2 4 6 8 10 12 Mean Educational Attainment Note:Year of graduationis equal to cohort year plus t2. That is, the year in which a 12 years old individual graduatesfrom 5th grade ilentered schwl when 7 yearsold. Before going into the interactions between the aggregate educational level of the whole labor force and educational inequality of each cohort, its weight in the population and the education levels and inequalities of the whole population by year, it i s important to note how education levels and inequality converge within each cohort over time. In principle, all cohorts are born with zero average education, and over time this number increases up to the point at which there i s no one in the cohort that still in school and then stabilizes. Basically, early observations of younger cohorts underestimate mean educational attainment and overestimate inequality of education. Cohorts below 30 years of age are still changing, hence we model theirfinal convergence levels for mean and coefficient of variation. In order to associate a unique pair of distributional parameters to each cohort.34 The IncreasingStock-To-Cohort Time LagOf Educational Attainment Figures 5 and 6 show how educational levels and inequality have evolved from one calendar year to the next from 1969 to 1999, in addition to a projection to 201335.What i s shown does not correspond exactly to what i s observed in more recent years because those cohorts still increasing their education are imputed their final educational levels and inequality as explained above. These figures illustrates the increasing stock-to-cohort lag of educational attainment. That is, the time lag required to extend to the whole labor force the educational improvements enjoyed by younger cohorts. For comparison, the education of each cohort i s also shown on the same graph. These Figures can be thought of as depicting the "permanent education levels and inequality" of people aged 16 to 70 in each year, even if they have not yet achieved these levels. Some interesting things are apparent from these graphs. The first i s that educational levels appear to be increasing in very monotonous, slow, and linear fashion. This i s not really surprising, given the monotonous and slow increase ineducational levels of each cohort. 34 SeeAppendix 1explains the convergenceof educationalattainment within cohorts. 35 The projectionis very simple: the education of cohorts bom after 1984is linearly projectedbased on those bom from 1961to 1983. 124 Figure 5: Observed Schoolingby Cohort and the whole labor force each calendar year 1940 1950 1960 1970 1980 1990 2000 2010 202 , Cohort and Calendar Year The secondinteresting fact i s that inequality, a measuredby the Z2 measure, i s falling continuously. This fall i s mostly due to within cohort inequality, as between cohort inequality i s much smaller.36This i s surprising, as it i s not evident that within cohort inequality dominates total inequality. It i s important to note that this fall does not necessarily mean that income inequalities due to education are falling - given Brazil's highly convex returns to education, the two may well go indifferent directions. Figure 6: Inequality inEducation by Cohort and Calendar Year Figure 15 Inequality in Mucdion by Cohort and Calendar Year - 0.8 0.7 0 Between OTotal 0.6 0.5 a 0.4 z -N0.3 0.2 0.1 1 1940 1955 1970 1985 2000 201 I Cohort and Calendar Year 36We have assumedconstant within cohort inequality for cohorts born after 1987. 125 Interestingly, the shape of the Educational attainment -mean and variance schedule- for the whole labor force (Figure 7) has the same inverted U shape of the corresponding schedule for cohorts (Figure 4). However important differences are evident, the stock-to-cohort time lag -approximately 25 years- and has been increasing duringthe last decade. Moreover, the range of variation of variance and the mean i s much smaller. While the variance for cohorts goes from 7 to 10 in a period of 15 years ending in 1959, the variance of the stock increases from 9 to 10.1 in a similar period culminating in 1986. Figure7: Evolutionof EducationalAttainmentof the Labor Force -MeanandVariance. I Observed. Brazil,1962-2013 10.3 ......................... ...................... ........ .................................. , """"'I 10.1 - - - 9.9 - - - 9.7+ - - - I > 9.3 - - - 9.1 - - - Simulations: Permanentandtemporary accelerationof educationalattainment Permanentaccelerationof educationalexpansion Figures 8 and 9 are identical to Figures 5 and 6, except that they show a simulation as well. In this simulation, we increase the education of all cohorts born after 1959 by (t - 1959)/10 years, where t i s the cohort's year of birth. The final impact on average education of this very large increase in education of each i s an increase of 1.74 years in 2013. On the other hand, the increase in inequality i s very small - in 2013, the Z2 measure would be 0.162, and not 0.148, due to the smaller rate of reduction in between cohort inequality. 126 Figure8: Observedand Simulated Schoolingby Cohort and Calendar Year I Figure 16 Observedand Simulated Schooling by Cohort and - 1 Calendar Year c 'P 0 i V L e c E USimulaton n A SimulatedCohorts 1940 1955 1970 1985 2000 2015 I Cohort and Calendar Year Figure9: ObservedandSimulatedInequalityinEducationby Cohortand Calendar Year figure 17 Observedand Simulated Inequality in Educetion by - Cohort andcalendar Year 0.7 +Sirmlation BEtw een 0.6 0.5 J 0.4 I -N 0.3 0.2 0.1 1940 1955 1970 1995 2000 2015 Cohortand Calendar Year The fact that even a very large intervention, operated on more recent cohorts, does not significantly affect educational inequality i s due to the dominant effect of within cohort inequality. Within inequality, as measured by Z2, i s not independent from between cohort inequality because average cohort education composes the weights, but this effect i s almost insignificant. On the other hand, average income in- creases, but relatively slowly, given the dramatic nature of the simulation at the cohort level. Both of these effects suggest the existence of strong demographic inertia. Tempuraryaccekrafiunuf educafiunal'expansiun&@re demugraphzcfransifiun The simulation above supposes very strong and incremental improvements in the educational system, which ignore reasonable fiscal constraints. A second -more interesting- simulation i s temporary 127 acceleration of educational expansion before demographic transition, which basically (i) Maintains the accelerated expansion during the 1970's and 1980's. Which i s equivalent to accelerating through the Mean-Variance path of educational attainment for cohorts (Figure 4) when demographic growth of schooling cohorts i s at its maximum. Inparticular, keeping the acceleration observed for cohorts born in the 1940-60period for the cohorts 1961-72period (policy period 1973-84). This policy i s equivalent to anticipating by nearly one decade the improvements of the educational system enjoyed by younger cohorts during the 1990s - the light blue line versus the yellow line. We should remark that the deviation of the expansion path i s temporary and after 1998 graduation year (cohort years) simulated and observed paths coincide again. Because of the fact that the demographic transition has not yet been completed in Brazil, providing a rationale for an "educational push" in the first decade of the XXI century, one would expected that the simulation shows firstly that heavier cohorts have higher mean educational attainment. Secondly, lower within inequality. The reason of why within inequality falls much faster than between inequality i s because the largest cohorts receive the lowest levels of inequality. Thirdly, an overshooting of between inequality, with a reduction under the historical trend after two or three decades. This behavior i s explained because the large cohorts where already more educated, therefore closer to the mean. Inthe short run, the expected effect of the simulations on the overall inequality would depend on which component dominates the other. While in the long run, one would expect a lower overall inequality. The expected effect on variance depends on whether inequality falls enough to compensate the growth of mean educational attainment. Finally, as labor inequality changes are proportional to changes in the variance of education (approximation under linear Mincerean equation). Hence, long term inequality i s expected to fall (General equilibrium could reduce B the wage skill gap and reduce inequality even further). The results from the simulation confirmed our expectations about faster achievement of mean educational attainment goals, overshooting of the inequality and the variance of education with lower longterm levels for both and consequently. The magnitude of shift in the mean educational attainment i s considerable, seven years to the left and 0.6 years upwards for the year 2002, which means that the time-gap falls from 25 to 18 years. The results also confirmed the anticipated temporary overshooting of the inequality and variance of education from mean 5.3 to 7.4 years of schooling, maximum 4% at mean 6.5. The long term levels reached by these variables by 2013 are 15% ininequality and 6% in variance of education. Figure 10 shows the temporary deviations in the cohort-path as the simulation of policy changes that anticipate by nearly one decade the improvements of the educational system enjoyed by younger cohorts during the 1990s -the light blue line versus the yellow line-. That is, maintaining the rate of growth of educational attainment enjoyed by the cohorts bome in the 1940 and 1950's for the generations borne in the 1960's and early 1970's. For example, the simulated average education of the individual borne in 1970 -which was leaving school in 1984- was nearly seven years of schooling, the same as the value observed for the person borne in 1980 -which was leaving school by the year 1994-. The consequences for the school attainment of whole labor force i s a permanent North-West shift -the blue line versus the orange line-. The magnitude of the shift i s considerable, seven years to the left and 0.6 years upwards for the year 2002, which means that the time-gap falls from 25 to 20 years. 128 Figure 10: Observedand SimulatedSchoolingby Cohortand Cal- endar Year. I .-ul : e 0 0 B5 ! 1940 1950 1960 1970 1980 1990 2000 2010 2020 Cohort and Calendar Year Since the transformations are much less dramatic than those in the first simulation, the impact on average education i s also much less dramatic. The educational level of the 16 to 70 population in the final calen- dar year we look at, 2013, rises from 7.7 years to 8.1. The effects on inequality are also not very impres- sive: there i s a small increase as the cohorts whose education was increased come into adult age, but it wears out by the year 2000. Figure 11 illustrates the temporary overshooting of the inequality of education along with its long term reduction. The black line illustrates the observed the mean educational attainment and the inequality of education, while the gray line depicts the relation for the simulation case. relation between the observed path. I I 1 I 1 I 2013 I I I I I I I 129 Figure 12: Evolutionof EducationalAttainment of the Labor Force: Meanand Variance. Observed and Simulated. Brazil 1969-2013 Evolutionof EducationalAttainment of the Laborforce: Mean andVariance. Observed and Simulated.Brazil, 1969-2013 Since the demographic transition will be almost completed in the first two decades of the XXI century there i s an opportunity to reduce the lag between current educational policy and its impact on the whole labor force. Therefore, delaying a vigorous "educational push" beyond demographic opportunities would have a significant permanent cost in terms of extending the benefits of current educational policy to the whole populations and catch-up with the rest of the world. Conclusions The main conclusion i s that demographic inertia i s strong in Brazil and improvements in the school sys- tem today will take long to translate into more education for the population as a whole. Between cohort inequality will always increase because increasing the educational level of younger cohorts i s equivalent to giving more education to those cohorts that already have the most. On the other hand, total inequality would relatively unaffected becausewithin cohort inequality of older cohorts dominates total educational inequality within the population. The paper gives a response to the implicit question raised in the title: There i s a transient window of op- portunity for long term inequality reduction via schooling expansion in the short and medium term. Part of that temporary window has already been lost. However, given the slow nature of demographic transi- tion, this window i s really wide and there i s still time to take advantage of it by expanding education vig- orously. The main conclusions of this report are: 0 Previous educational policy -specially in the 1980s- did not take full advantage of demographic opportunities. The stock-to-cohort time lag could have been reduced substantially -25%- if a country like Brazilhad maintainedthe historical rate of growth of schooling during the 1980's. 130 0 Educational Policy makers should maintain a patient long term perspective, because the stock- to-cohort time lag is considerable (from two to three decades). Therefore, policy makers should put in place monitoring mechanisms to follow both the performance of younger cohort and the demographic transition. 0 The appropriate time correlation between educational efforts and pre-demographic transition would reduce long term inequality. However there might be a labor income inequality overshoot in the short run, the period in with inequality between cohorts might run dominant. Additional reductions labor income inequality would follow if the wage skill gap responds to the supply changes, and even further if income-fertility differentials persist (Kremmer et al, 2002).Demographic opportunities are still available, but they are smaller and transient. Since the demographic transition will be almost completed in the first two decades of the XXI century there i s still an opportunity to reduce the stock-vs-cohort time lag via educational expansion. 0 Delaying "educational push" would have a permanent long term equity cost for Brazil. There- fore, delaying a vigorous "educational push" beyond demographic opportunities would have a significant permanent cost in terms of extending the benefits of educational policy to the whole populations and catch-up with the rest of the world. Perhaps there i s a corollary in terms of the specific role of foreign credit (vis a vis domestic credit) to finance education expansion before demographic transition takes place. Fiscal constraints might be too binding precisely when countries like Brazil face optimal demographic opportunities. Own country resources are the most limited at this point in time because dependency ratios are the highest and productivity per worker i s very low. Hence this constraint might be efficiently relaxed via foreign credit from multilateral institution. In summary, the level of schooling of the labor force in Brazil i s clearly insufficient and efforts to make educational attainment higher and more equitable should be emphasized. Nevertheless policy makers and policy observers should be aware that any expected impact of education on inequality of income will not be immediate. Hence they should be willing to establish monitoring systems can follow to young cohorts of students trough the different stages of the educational ladder and evaluate educational outcomes with an explicit long term perspective. 131 Bibliography Anand, S. and S. Kanbur. 1993. "Inequality and Development: A Critique," Journal of Development Economics 41, 19-43. Ahluwalia, M. 1976. "Inequality, Poverty and Development,'' Journal of Development Economics 3, 307- 42. Barros, Ricardo and David Lam (1996) "Income and Education Inequality and Children's Schooling At- tainment in Brazil," in Nancy Birdsall and Richard Sabot (eds.), Opportunity Foregone: Education inBrazil, Washington: Inter-AmericanDevelopment Bank. Birdsall, Nancy, and Richard Sabot (eds.) (1996) Opportunity Foregone: Education in Brazil, Washing- ton: Inter-American Development Bank. Bloom and Willimason. 1998. "Demographic Transitions and Economic Miracles in Emerging Asia," World Bank Economic Review 12,419-55. Case, Anne, and Angus Deaton (1999) "School Inputs and Education Outcomes in South Africa." Quar- terly Journal of Economics, forthcoming Deaton, Angus and Paxson, Christina. "The effects of economic and population growth on national sav- ing and inequality" inDemography, Vol34, No 1, February,l997. Deninger, Klaus, and Lyn Squire (1996) "A New Data Set Measuring Income Inequality," The World Bank Economic Review, lO(3): 565-91. Duryea, Suzanne and Miguel SzCkely. 1999. "Decomposing Schooling Differences in Latin America." Mimeo IDB. Higgins, Matthew and Jeffrey G. Williamson. 1999. "Explaining Inequality the World Round: Cohort Size, Kuznets Curves, and Openness." NBERWorking PaperNo. 7224 Katz, L. and K. Murphy (1992), "Changes in Relative Wages, 1963-1987: Supply and DemandFactors," Quarterly Journal of Economics , 107,35-78. Kremer, M., and D. Chen. 2002. "Income Distribution Dynamics with Endogenous Fertility." Journal of Economic Growth 7(3):227-58. Kuznets, S. "Growth, population and income distribution, selected essays" New York; Norton, 1979. Lam, David. "Generating Extreme Inequality: Schooling, Earnings, and Intergenerational Transmission of Human Capital in South Africa and Brazil." Report No. 99-439 Research Report. Population Studies Center at The InstituteFor Social Research University Of Michigan Lam, D. 1986. "The Dynamics of Population Growth, Differential Fertality and Inequality." Economic Review 76(5) 1103-1116. Levison, D., and D. Lam. (1991). "Declining Inequality in Schooling inBrazil and Its Effects on Inequal- ity inEarnings" in Journal of DevelopmentEconomics, Vol. 37, 1992. Mookherjee, D. and A. Shorrocks. "A decomposition analysis of the trend in UK income inequality" in TheEconomic Journal, Vol. 92, No 368. Ram, R. "Educational expansion and schooling inequality: international evidence and some implications" in TheReview of Economic and Statistics, Vol. 72, No 2. May, 1990 Shorrocks, A. F. "The class of additively decomposable inequality measures" in Econometrica, Vol. 48, No 3. April, 1980. 132 Appendix: Convergeestimationfor educational attainment (mean and inequality) within cohorts In principle, all cohorts are born with zero averageeducation, and over time this number increasesup to the point at which there i s no one in the cohort that still in school and then stabilizes. InBrazil, after 30 years of age, very few people are still in school. In 1999, only 2.9% of the population 30 or over were still in any kind of regular learning. The year of 1999 may be used as an upper bound, given that each successivecohort i s completing more education than its predecessor^^^. FigureA1 Figure 4 Wte of Growth of Average Education Within Each Cohort - OverthePeriod ofObsermtion i 1 0.97 0.8 0.7 0.6 0.51 0.4 0.3 0.2 0.1 0.0 1920 1940 1960 1980 Cohort Since we observed cohorts from 1977 to 1999, this means that any cohort born previous to 1947 should no longer show any increasesin education over the period of observation and even those born previous to 1952 (those 25 and older in 1977) should show very little. This is indeed what we observe, Figure A1 shows the slope linear trend line showing the increase in education of each cohort over the 1977 - 1999 period. It i s flat for cohorts born from 1900 to the mid-1940s and then increases strongly and monotonously up the last cohort observed -the one born in 1987. A related pattern can be seen inthe evolution of inequality. FigureA2 shows that the 12 measure does not change significantly until the 1965 cohort and then there i s a strong downward trend for all successive cohorts. In other words, cohorts aged 12have already achieved their final inequality, as measured by one half of the coefficient of variation squared. This i s not as intuitive as the effect on average education. The Z2 measure i s one-half of the variance divided by the square of the mean. When each cohort comes into the world none of its members has any education, the mean i s zero, and Z2i s not even defined. Once at least one child finished one year of schooling, Z2 becomes defined and then increases very quickly as a part of the cohort acquires some education, yielding a positive denominator, but the numerator, average years of education, remains very low. Z2 then falls mostly because this denominator i s increasing. Since we only observe each cohort after it i s 10 years old, we do not observe the increasing part of the curve, only the downward part. We will also see later on that it becomes stable within each cohort before average educational level does. This i s why we observe changes in the Z2 measure over the 1977-1999 37 For example, in1995 only 1.6% of people 30 or older were involved in education. 133 period only with the 1965 cohort while the average changes over the same period for all cohorts after the one born in 1945. Figure A2 I Figure 5-Rate ofGrowthofInequality Within EachCohortOverthe Period of Observation I b I -0.1 Cohort Another way to observe the evolution within cohorts and over time i s to look at the average education of each cohort from 1977 to 1999. Figures A3 through A5 show this for ten cohorts born from 1920to 1975. Figure A3 shows that the cohorts born in 1920, 1930, and 1940 show no increase at all in average education over the period; Figure A4 shows very slight increases for cohorts born in 1945, 1950, and 1955; and finally Figure A5 shows the large increases in the education of cohorts born from 1960 onwards, whose members were still overwhelmingly in school during the observation period. Finally, each successive cohort attains a final educational level superior to that of its predecessors Figure A3 Figure 6 Evolution of Cohorts Born Before 1940 - -Pd inbnio(194Q 75 80 E5 90 95 100 I Year of Observation 134 8 - .-6 - 7 - 3 6 if c0 5 - m 2 4 - 3 - - a, Polinamio (19%) p 2 - - Polindmio (1955) 2a, 1 - 0 -r , FigureAS I Figure 8 Evolution of Cohorts BDrn from 1960 to 1975 - 8 1 .-6 7 - au i% 6 - 5 - $ 4 - -1 965 s - 0 - 1 970 3 - -1 975 Q -Polirdnio(l965) 8 P 2 - -PoIirddo(l960) < 1 - 5 ---Poiirbnio(l970) 0 4 75 80 a5 90 95 100 Year of Observation Figures A6 through A8 show the ZZ measure for the same cohorts as Figures A3 through A5. The message i s again clear: the inequality of education i s stable from 1977 to 1999 for cohorts born until 1965 but falls over the observation period for those born after 1965. Of course, the final value of Zz for each cohort i s lower than for its predecessor. 135 Figure A6 Figure 9 Evolution of Cohorts Born Before 1940 - 09 - 08 - 0.7 - 0.6 - *I93 -Pd inhio (1940 -Pdinhio(193q -Pd inhio (1920 , 80 85 90 95 100 Year of Observation Figure A7 Figure 10 Evolution of Cohorts Born from 1945 to 1955 - 1 - 0.9 - 0.8 - 0.7 - 0.6 - 0.5 - -6-1955 1950 0.4 - 0.3 - - Polinhio(1955) 0.2 - Polinhi O(lR5) 0.1 - 0 4 , 75 80 85 90 95 100 Year of Observation 136 Figure AS 0 9 ' ,!1 0 8 ' I 0 7 - 't Y 0 6 - D 61365 18 I 05. 1370 0 4 - --b1975 0 3 . z - Polin8nio(1930) II 0 2 . Poiinbnio(l970) 0 1 " 8 75 80 85 90 95 100 Yea of Observation Finally, if we shift the curves on graphs A5 and A8, we can see how the level and inequality in education vary as different cohorts age. This i s what i s shown on Figures A9 and A10, which may be the most important figures thus far. Let us start with FigureA9. FigureA9 ~ Figure 12 -Evolution of EducationLevels of Cohorts Born from 1955 to 1975 -Polindmio(65) -Polindmio(60) -PolinBmio(70) -Po linBmio(75) -Po IinBmio(55) " I 10 15 20 25 30 35 40 Age Figure A9 appears to show that successive cohorts, at least those born from 1960 to 1975, have higher education levels at any given age and appear to level off at more or less the same age. This i s equivalent to saying that most educational improvement involves advancing further in the educational ladder in the same time rather than staying longer in school. In other words, kids are doing better because they are repeating less. This i s coherent with most analysis in the education literature inrecent years. 137 The practical impact of this upon our analysis i s on how we will model final educational level of the cohorts born from 75 to 83, whose years of schooling had not yet reached its final value in 1999. What we do i s take the 1974cohort as a baseline and see how much more education a given younger cohort has at each observed age and then attribute to these newer cohorts the 1974 cohort final value multiplied by the average percentage difference between the two over the years of observation. Figure AZO I figure 13- Evolution of Inequality of Cohorts Born From 1955to 1975 I I I Q 0.6 - 1 % - 1 0.5 a %0 Poiir8mio (65) Polir8mio (60) 1 -N 0.4 - l___l Polir8 mio (70) I -Poiin3mio I (75) 0.3 Poli16mio (55) L 1 1 I ~ 0.2 I 10 15 20 25 30 35 40 I Age Figure A10 i s easier to interpret. By age 20, educational inequality, as measuredby the Z2 measure, levels off. This means that additional increases in the variance of education are matched by equal increases in the square of average education, leaving this inequality measureunchanged. 138 6. EX-ANTE EVALUATION CONDITIONAL CASHTRANSFER OF PROGRAMS: THE CASEOF BOLSA ESCOLA3* Franqois Bourguignon, Francisco H.G.Ferreira and Phillippe G. Leite3' Not to be quoted Abstract: Cash transfers targeted to poor people, but conditional on some behavior on their part, such as school attendance or regular visits to health carefacilities, are being adopted in a growing number of developing countries. Even where ex-post impact evaluations have been conducted, a number of policy- relevant counterfactual questions have remained unanswered. These are questions about the potential impact of changes in program design, such as benefit levels or the choice of the means-test, on both the current welfare and the behavioral response of household members. This paper proposes a method to simulate the effects of those alternative program designs on welfare and behavior, based on micro- econometrically estimated models of household behavior. In an application to Brazil's recently intro- ducedfederal Bolsa Escolaprogram, wefind a surprisingly strong effect of the conditionality on school attendance,but a muted impact of the transfers on the reduction of currentpoverty and inequality levels. JEL Codes:I38,513,522,524 Key Words:Conditional Transfers; Demandfor Schooling, Child Labor Introduction During the 1990s, a new brand of redistribution programs was adopted in many developing countries. Although local versions varied, programs such as Food for Education in Bangladesh, Bolsa Escola in Brazil, and Progresa in Mexico are all means-tested conditional cash transfer programs. As the name indicates, they share two defining features, which jointly set them apart from most pre-existing programs, whether in developing or developed countries. The first of these i s the means-test, defined in terms of a maximum household income level, above which households are not eligible to receive the benefit.@The second i s the behavioral conditionality, which operates through the requirement that applicant households, in addition to satisfying the income targeting, have members regularly undertake some pre- specified action. The most common such requirement i s for children between 6 and 14 years of age to remain enrolled and actually in attendance at school. In Mexico's Progresa, additional requirements applied to some households, such as obligatory pre- and post-natal visits for pregnant women or lactating mothers. The implementation of these programs have generated considerable interest, both in the countries where they took place and in the international academic and policy-making communities. Accordingly, a great deal of effort has been placed in evaluating their impact. There are two types of approach for evaluating the effects of these programs on the various aspects of household welfare that they seek to affect. Ex-post approaches consist of comparing observed beneficiaries of the program with non-beneficiaries, possibly 38 Paper to bepresentedat the WBLJNICEFIILOconferenceon Chile Labor, Oslo 28-29 May, 2002 39Respectively, Delta andWorld Bank, Paris, PUC, Rio andPUC, Rio. 40For verification and enforcement reasons, the means-test is often specifiedin terms of a score based on responses to a ques- tionnaire and/or a home visit by a social worker. In some countries, the score is `calibrated' to be approximatelyequivalent to a pre-determinedlevelof householdincome per capita. See Camargo and Ferreira (2001) for a discussionof the Brazilian case. 139 after controlling for selection into the first or the second group if truly random samples are not available. An important literature has recently developed on these techniques and many applications to social programs have been made in various countries.41 Ex-ante methods consist of simulating the effect of the program on the basis of some model of the household. These models can vary widely in complexity and coverage. Arithmetic simulation models simply apply official rules to detennine whether or not a household qualifies for the program, and the amount of the transfer to be made, on the basis of data commonly available in typical household surveys. More sophisticated models include some behavioral responseby households. Ex-ante and ex-post evaluation methods are complements, rather than substitutes. To begin with, they have different objectives. Ex-post methods are meant to identify the actual effects of a program on various dimensions of household welfare, by relying on the direct observation of people engaged in the program, and comparing them with those same dimensions in a carefully constructed comparison group, selected so as to provide a suitable proxy for the desired true counterfactual: "how would participants have fared, had they not participated?". Insome sense, these are the only "true" evaluations of a program. Even when comparison groups are perfectly believable proxies for the counterfactual, however, ex-post evaluations leave some policy-relevant questions unanswered. These questions typically refer to how impact might change if some aspect of the program design - such as the level of the means-test; the nature of the behavioral conditions imposed; or the level of the transfer benefits - changes. It i s difficult enough to obtain an actual control group to compare with a single program design inreality. It i s likely to be impossible to "test" many different designs in experimental conditions. Ex-ante methods are valuable tools exactly because it i s easier to experiment on computers than on people. These methods are essentially prospective since they rely on a set of assumptions about what households are likely to do when faced with the program. They also permit direct counterfactual analysis of alternative programs for which no ex-post data can be available. Thus, they are indispensable when designing a program or reforming existing ones. Simulation models of redistribution schemes based on micro data sets are widely used in developed countries, especially to analyze the effect of the numerous and often complex cash transfer instruments found in those countries. Given the progress of direct cash transfers in developing countries, buildingthe same type of models in developing countries may become necessary.42 However, the specific behavioral conditionality that characterizes these programs requires modifications, and a focus on different aspects of household behavior. The present paper takes a step in that direction by proposing a simple ex-ante evaluation methodology for conditional means-testedtransfer programs. We apply the method to the new federal design of Bolsa Escola, in Brazil, and we are concerned with both dimensions cited by the program administrators as their objectives: (i) the reduction of current levels of poverty and inequality; and (ii) the provision of incentives for the reduction of future poverty, through increased school enrollment among poor childrentoday. The paper is organized as follows. Section 2 describes the Bolsa Escola program, as it was launched at the federal level in Brazil in 2001. Section 3 presents the simple econometric model used for simulating 41 This literature relies heavily on matchingtechniques, and draws extensivelyon the early work by Rubin (1977) and Rubin and Rosenbaum(1985). For a survey of recent applications,see Heckmanand Vytlacil (2002). For a study of the effects of the Food for Ed~icationprogramin Bangladesh, see Ravallionand Wodon (2000). A number of importantstudies of Progresa were under- taken under the auspices of the International Food Policy Research Institute (IFPRI). See, in particular, Parker and Skoufias (2000) and Schultz (2000). 42 See, for instance, Harding(1996). On the need for and difficulties with building the same type of models in developing coun- tries, see Atkinson and Bourguignon (1991). 140 the effects of the program. Given the conditionality of Bolsa Escola, this model essentially deals with the demand for schooling and therefore draws on the recent literature on child labor. The estimation of the model i s dealt with in Section 4, whereas the simulation of program effects and a comparison with alternative programdesigns are discussedin Section 5. Section 6 concludes. Mainfeatures of the BolsaEscolaprogram The Brazilian national Bolsa Escola program, created by a law of April 2001 within the broader context of the social development initiative known as Projeto Alvorada, i s the generalization at the federal level of earlier programs, which were pioneered in the Federal District and in the city of Campinas (SP) in 1995, and later extended to several other localities.43 The law of April 2001 made these various programs uniform in terms of coverage, transfer amounts and the associated conditionality. It also provided federal funding. Yet, the monitoring of the program itself i s left under the responsibility of municipal governments. The rules of the program are rather simple. Households with monetary income per capita below 90 Reais (R$)44 per month - which was equivalent to half a minimum wage when the law was introduced - and with children aged 6 to 15 qualify for the Bolsa Escola program, provided that children attend school regularly. The minimumrate of school attendance i s set at 85 per cent and schools are supposedto report this rate to municipal governments for program beneficiaries. The monthly benefit is R$15 per child attending school, up to a maximum of R$45 per household. Transfers are generally paidto the mother, upon presentation of a magnetic card that greatly facilitates the monitoring of the whole program. The management of the program i s essentially local. Yet, control will be operated at two levels. At the federal level, the number of beneficiaries claimed by municipal governments will be checked for consistency against local aggregate indicators of affluence. In case of discrepancy, local governments will have to adjust the number of beneficiaries on the basis of income per capita rankings. At the local level, the responsibility for checking the veracity of self-reportedincomes i s left to municipalities. It i s estimated that some ten million children (in six million households) will benefit from this program. This represents approximately 17 percent of the whole population, reached at a cost slightly below 0.2 percent of GDP. The latter proportion i s higher in terms of household disposable income: 0.45 percent when using household income reported in the PNAD survey and 0.3 per cent when using National Accounts. Of course, this figure i s considerably higher when expressed in terms of targeted households. Even so, it amounts to no more than 5 percent of the income of the bottom two deciles. A simple frameworkfor modelingandsimulating BolsaEscola The effects of such a transfer scheme on the Brazilian distribution of income could be simulated by simply applying the aforementioned rules to a representative sample of households, as given for instance by the Pesquisa Nacional por Amostra de Domicilios (PNAD), fielded annually by the Brazilian Central Statistical Office (IBGE). This would have been an example of what was referred to above as 'arithmetic' simulation. Yet, for a program which has a change inhousehold behavior as one of its explicit objectives, this would clearly be inappropriate. After all, Bolsa Escola aims not only to reduce current poverty by 43Early studies of these original programsinclude Abramovay et. al. (1998);Rocha and Sab6ia(1998) and Sant'Ana and Moraes (1997). A comprehensive assessment of different experiences with Bolsa Escola across Brazil can be found in World Bank (2001). There is much less written on the federal program, for the good reason that its implementationin practice i s only just beginning. The description given in this section draws on the official MinistCrio da EducaGIo website, at http://www.mec.gov.brhome/bolsaesc. 44ApproximatelyUS$30, at August 2002 exchange rates. 141 targeting transfers to today's poor, but also to encourage school attendance by poor children who are not currently enrolled, and to discourage evasion by those who are. Any ex-ante evaluation of such a policy must therefore go beyond simply counting the additional income accruing to households under the assumption of no change in schooling behavior. Simulating Bolsa Escola thus requires some structural modeling of the demand for schooling. This section presents and discusses the model being used in this paper. There i s a rather large literature on the demand for schooling in developing countries and the related issue of child labor. The main purpose of that literature i s to understand the reasons why parents would prefer to have their kids working within or outside the household rather than going to school. Various motives have been identified and analyzed from a theoretical point of view,45 whereas numerous empirical attempts have been made at testing the relevance of these motives, measuring their relative strength and evaluating the likely effects of policies.46 The empirical analysis i s difficult for various inter-related reasons. First, the rationale behind the decision on child labor or school enrollment i s by itself intricate. In particular, it i s an inherently intertemporal decision, and it will differ depending on whether households behave as a unitary model, or whether internal bargaining takes place. Second, it i s difficult to claim exogeneity for most plausible explanatory variables, and yet no obvious instrument i s available for correcting the resultingbiases. Third, fully structural models that would permit a rigorous analysis of policies are complex and therefore hard to estimate while maintaining a reasonable degree of robustness. Inlightofthesedifficulties, our aimsaremodestandour approachisoperational: ratherthanproposinga new, more complete structural model of the demand for schooling and intra-household labor allocation, we aim simply to obtain reasonable orders of magnitude for the likely effects of transfer programs of this kind. We thus make the choice to limit the structural aspects of the modeling exercise to the minimum necessary to capture the main effects of the program. Inparticular, we makefour crucialsimplifying assumptions.First,weentirely ignoretheissueofhowthe decision about a child's time allocation i s made within the household. We thus bypass the discussion of unitary versus collective decision-making models of household. Instead, we treat our model of occupational choice as a reduced-form reflection of the outcome of whichever decision-making process took place within the household.47 Second, we consider that the decision to send a child to school i s made after all occupational decisions by adults within the household have been made, and does not affect those decisions. Third, we do not discuss here the issue of various siblings inthe same household and the simultaneity of the corresponding decision. The model that i s discussed thus i s supposed to apply to all children at schooling age within a household. Fourth, we take the composition of the household as exogenous. Under these assumptions, let Si be a qualitative variable representing the occupational choice made for a child in household i.This variable will take the value 0 if the child does not attend school, the value 1if she goes to school and works outside the household and the value 2 if she goes to school and does not work outside the household. When Si=O, it will be assumed that the child works full time either at home or on the market, earnings being observed only in the latter case. Similarly, Si=2 allows for the possibility that the child may be employed in domestic activities at the same time he/she goes to school. The 45 See the well-known survey by Basu (1999) as well as the recent contribution by Baland and Robinson (2001). 46 Early contributions to that literature include Rosenzweig and Evenson (1977), as well as Gertler and Glewwe (1990). For more recent contributions and short surveys of the recent literature see Freije and Lopez-Calva (2000), Bhalotra (2000). On pol- icy see Grootaert and Patrinos (1999). 47 For a discussion of how intra-household bargaining affects the occupational choice o f members, see Chiappori (1992). See also Bourguignon and Chiappori (1994)and Browning et. al. (1994). 142 occupational choice variable Si will be modeled using the standard utility-maximizing interpretation of the multinomial Logit framework, so that: where sk( )i s a latent function reflecting the net utility of choosing alternative k (=O, 1 or 2) for deciders in the household. Ai is the age of the child I;Xi is a vector of her characteristics; Hi, is a vector of the characteristics of the household she belongs to - size, age of parents, education of parents, presence of other children at school age, distance from school, etc.; Y.i i s the total income of household members other than the child and yij i s the total contribution of the child towards the income of the household, depending on her occupational choice j. Finally, vij i s a random normal variable that stands for the unobserved heterogeneity of observed schooling/participation behavior. If we collapse all non-income explanatory variables into a single vector 2,and linearize, (1) can be written as: T h i s representation of the occupational choice of children i s very parsimonious. In particular, by allowing the coefficients yj and q to differ without any constraints across the various alternatives, we are allowing all possible tradeoffs between the schooling of the child and hisher future income, and the current income of the household. Note also that the preceding model implicitly treats the child's number of hours of work as a discrete choice. Presumably that number i s larger inalternative 0 than inalternative 1becauseschooling is taking some time away. This may be reflectedinthe definition of the childincome variable yij as follows. Denote the observed market earnings of the child as wi. Assuming that these are determined in accordance with the standard Becker-Mincer human capital model, write: Log wi = Xi .6 + m*Ind(Sj=l) +ui (3) where Xi i s a set of individual characteristics - including age and schooling achieved - and where ui i s a random term that stands for unobserved earnings determinants. Assumptions on that term will be discussed below. The second term on the right hand side takes into account the preceding remark on the number of hours of work. Children who attend school and are also reported to work on the market presumably have less time available and may thus earn less. Based on (3), the child's contribution to the household income, yij, inthe various alternative j i s defined as follows: where it i s assumed that yi, covers both market and domestic child labor. Thus domestic income i s proportional to actual or potential market earnings, wi, in a proportion K for people who do not go to school. Going to school while keeping working outside the household means a reduction in the proportion 1-Mof domestic and market income. Finally, going to school without working on the market means a reduction in the proportion 1-Dof total child income, which in that case i s purely domestic. The proportions K and D are not observed. However, the proportion M i s taken to be the same for domestic and market work and may be estimated on the basis of observed earnings. Replacing(4) in (2) leads to with : Po= oh K ; PI= a,MK; P2= a2DK 143 We now have a complete simulation model. If all coefficients a, y are known, as well as the actual or p, potential market earnings, wi and the residual terms vij ,then the child's occupational type selected by household iis: k* =Arg max[Ui(j)] (6) Equation (5) represents the utility of household iunder occupational choice j [Uiu)] in the benchmark case. If the Bolsa Escola program entitled all children48going to school to a transfer T, (5) would be replaced by: Ui(j)= &.rj (Y-I+BEij).aj pj.wi + + + vijwith BEio=Oand BEil =BEi2= T (7) Under the assumptions we have made, equation (7) i s our full reduced-form model of the occupational choice of children, and would allow for simulations of the impact of Bolsa Escola transfers on those choices. All that remains i s to obtain estimates of p, y, a,wiand the vij's, Assuming that the vi, are iid across sample observations with a double exponential distribution leads to the well-known multi-logit model. However, some precautions must be taken in this case. It i s well known that the probability that household iwill select occupational choice k i s given by: Exp(Z,.yk Y-iak wi + + .Pk pik =cExp(Zi.yj + Y p j +wi.Pj) j The difficulty i s that the Multinomial logit estimation permits identifying only the differences (aj-%), (pj-po), and (%-yo)for j = 1, 2. Yet, inspection of (6) and (7) indicates that - since the Bolsa Escola transfer i s state-contingent, meaning that the income variable i s asymmetric across alternatives - it i s necessary to know all three coefficients % ,al and a2in order to find the utility maximizing alternative, k*. This is where the only structural assumption made so far becomes useful. Call 6, and b, the estimated 1. coefficients of the multilogit model corresponding to the income and the child earning variables for alternatives j = 1, 2, the alternative 0 being taken as the default. Then (5) implies the following system of equations: It will prove simpler to discuss the estimation problem under this simplifying assumption. We reintroduce the means test, without any loss of generality, at the simulation stage. 144 a,-ao=GI a2-ao=a2 A (a,M-ao).K=b, A (a,D-ao)K= i2 M is known from equation (3). It follows that arbitrarily setting a value for K or for D allows us to identify @, al and a2 and the remaining parameter in the pair (K,D). The identifyingassumption made in what follows is that kids working on the market and not going to school have zero domestic production, Le. K = 1. Inother words, it i s assumed that the observed labor allocations between market and domestic activities are comer solutions in all alternati~es.4~ Itthen follows that : 6, -b,A a;=-1-A4 and a2=a,+ii2-6, Of course, a test of the relevance of the identifying assumption i s that both al and a2 must be positive. One could also requirethat the value of Dobtainedfrom system (9)with K=l be inthe interval (0,l). For completeness, it remains to indicate how estimates of the residual terms vij-vio may be obtained. In a discrete choice model these values cannot be observed. It is only known that they belong to some interval. The idea i s then to draw them for each observation in the relevant interval, that is: in a way consistent with the observed choice. For instance if observation ihas made choice 1, it must be the case that : 8.~1 + Y-j. 6, +bl .wj + (vjl-vio) > SUP[O, A 8 . ~ 2+Y.j. 62+b2.wj + (vjz-vio)] A The terms vij-vio must be drawn so as to satisfy that inequality. All that is missing now i s a complete vector of child earnings values, wi. Estcinahnofputential earnzitgs The discrete choice model requires a potential earning for each child, including those who do not work outside the household. To be fully rigorous, one could estimate both the discrete choice model and the earning equation simultaneously by maximum likelihood techniques. This i s a rather cumbersome procedure. Practically, a multinomial probit would then be preferable to a multinomial logit in order to handle simultaneously the random terms of the discrete choice model and that of the earning equation. Integrating tri-variate normal distributions would then be required. Also, other issues which are already apparent with a simpler technique would not necessarily be solved. We adopt a simpler approach, which has the advantages of transparency and robustness. It consists of estimating (3) by OLS, and then to generate random terms Q for non-working kids, by drawing in the distribution generated by the residuals of the OLS estimation. There are several reasons why correcting the estimation of the earning function for a selection bias was problematic. First, instrumenting earnings with a selection bias correction procedure requires finding instruments that would affect earnings but not the schooling/labor choice. No such instrument was 49 Ineffect, this assumption may be weakened using some limited information on hours of work available in the survey. 145 readily available. Second, the correction of selection bias with the standard two-stage procedure i s awkward in the case of more than two choices. Lee (1983) proposed a generalization of the Heckman procedure, but it has been shown that Lee's procedure was justified only in a rather unlikely particular case.5oFor both of these reasons, failing to correct for possible selection bias in (3) did not seem too serious a problem. On the other hand, trying to correct using standard techniques and no convincing instrument led to rather implausible results. Szmulatingprogramsof theBolsaEscolatype As mentioned in footnote 11, the model (6)-(7) does not provide a complete representation of the choice faced by households in the presence of a program such as Bolsa Escola. This i s because it takes into account the conditionality on the schooling of the children, but not the means-test. Taking into account both the means-test and the conditionality leads to choosing the alternative with maximum utility among the three followingconditional cases: ui(0)=z,.yo+a0Y-/ +powi +vio ui(1)=zi.yl +a,(Y-, +T )+p,wi+vi, if Y-/ +Mw, IY O Ui(l)=zi.yl +a,Y-, +plwi + V i l if Y-,+ M W i >Y" ui(2)=zi.y,+a,(Y-, +T)+P2Wi +vi, if Y-/ I Y " ui(2)=zi.y,+a,Y-, +p2wi+vi2 if Y-, >Yo where Yo stands for the means test. Of course, as mentioned above, only the differences between the utility corresponding to the three cases matter, so that one only need to know the differences (&Po), (1"~- yo) and (vij- vio)- but the three coefficients aj.Inthis system, one can see how the introduction of Bolsa Escola might lead households from choice (0) - no schooling -to choices (1) or (2), but also from choice (1) to choice (2). Inthe latter case, a household might not qualify for the transfer T when the child both works and attends school, but qualifies if she stops working. A wide variety of programs may be easily simulated usingthis framework. Both the means-test and the transfer T could be made dependent on characteristics of either the household or the child (X and H).In particular, T could depend on age or gender. Some examples of such alternative designs are simulated and discussedin Section 5. Before presenting the model estimations results, we should draw attention to two important limitations of the framework just described. Both arise from the set of assumptions discussed in the beginning of this section. The first limitation i s that we can not take into account the household transfer ceiling of R$45 per household. The reason i s that by ignoring multi-children interactions in the model, it i s as though we had effectively assumed that all households were single-child, from a behavioral point of view. Inthe non-behavioral part of the welfare simulations which are reported in Section 5 below, however, each child was treated separately, and the R$45 limit was applied. The second limitation has to do with the exogeneity of non-child income Y+ This exogeneity would clearly be a problem when there are more than one child at schooling age. But it i s also unrealistic even when only adult income i s taken into account. It i s clearly possible that the presence of the means-test might affect the labor supply behavior of adults, since there are circumstances in which it might be in the interest of the family to work slightly less in order to qualify for Bolsa Escola. Note, however, that this 5"See Bourguignonet al. (2001). 146 might not be so sharply the case if the means-test is based, not on current income, but on some score- basedproxy for permanent income, as appears to be the case inpractice. Descriptivestatistics and estimation results The model consisting of equations (3) and (12) was estimated on data from the 1999 PNAD household survey. This survey i s basedon a sample of approximately 60,000 households, which i s representative of the national pop~lation~~.Although all children aged 6-15 qualify for participation in the program, the model was only estimated for 10-15 year-olds, since school enrollment below age 10 i s nearly universal.52At the simulation stage, however, transfers are of course simulated for the whole universe of qualifying 6-15 year-olds. Table 1 contains the basic description of the occupational structure of children aged 10-15 in Brazil, in 1999. In this age range, 77% of children report that they dedicate themselves exclusively to studying. Some 17% both work and study, and 6% do not attend school at all. This average pattem hides considerable variation across ages: school attendance declines -and work increases-monotonically with age. Whereas only 2.5% of ten year-olds are out of school, the figure for fifteen year-olds i s 13%. Whereas 90% of ten year-olds dedicate themselves exclusively to studying, fewer than 60% of fifteen year-olds do so. From a behavioral point of view, it i s thus clear that most of the action i s to be found among the eldest children. Table 2 presents the mean individual and household characteristics of those children, by occupational category. Children not going to school are both older and less educated than those still enrolled. As expected, households with school drop-outs are on average poorer, less educated and larger than households where kids are still going to school. Dropping out of school and engaging in child labor are relatively more frequent among non-whites and in the North-East. Both forms of behavior are least common in metropolitan areas, but proportionately more common in non-metropolitan urban areas than inrural areas. Interestingly, households where children bothwork and go to school are inan intermediate position, along all dimensions, between those whose children specialize, but are generally closer to the group of drop-outs. A remarkable feature of Table 2 is the observed amount of children's eamings, when they work and do not study. Ranging from around R$80 to R$120 per month, children's earnings represent approximately half the minimum wage, an order of magnitude that seems rather reasonable. These amounts compares with the R$15 transfer that i s granted by the Bolsa Escola program for children enrolled in school. Note, however, that the R$90figure i s not a good measure for the opportunity cost of schooling, since school attendancei s evidently consistent with some amount of market work. Tables 3 and 4 contain the estimation results. Because of the great behavioral variation across ages even within the 10-15 range - as revealed, for instance, in Table 1 -we estimated the (identically specified) model separately for each age, as well as for the pooled sample of all 10-15 year-olds. The simulations reported in the next section rely on the age-specific models, but in this section we focus on the joint estimation, both for ease of discussion and because the larger sample size allowed for more precise estimation inthis case. 51Except for the rural areas of the statesof Acre, Amazonas, Pari, RondGniaandRoraima. 52We know that school enrolment is nearly universalfrom answers to schoolingquestions in the PNAD. An additionalreasonto limit the estimation of the behavioral model to children aged ten or older is that the incidence of child labor at lower ages is probably measured with much greater error, since PNAD interviewersare instructedto pose labor and incomequestions only to individualsaged ten or older. 147 Table 3 shows the results of the OLS estimation of the earnings function (3), both for the pooled sample and for the 15 year-old Geographical variables54,race and gender have the expected sign, and the same qualitative effect as for adults. So does (the logarithm of) the average earnings of children in the census cluster, which i s included as a proxy for the spatial variation in the demand for child labor. The effect of previous schooling i s best described as insignificant. Even though the coefficient of the squared term i s positive and significant, the influence of the (negative and insignificant) linear term implies that earnings decline with schooling in the range relevant for 10-15 year-olds. It should be noted that our separate specifications mask the main determinant of earnings for children, namely age. Inan alternative (unreported) specification for the pooled sample, when age was included as an explanatory variable, an additional year of age increased earnings by approximately 40 per cent. But there was a clear non- linearity in the way age affected earnings, which i s reflectedin changes in the coefficient estimates when the model i s separately estimated. These non-linearities and interactions between age and other determinants are the reason why the separatespecification was preferred. The estimate for m- the coefficient for "dummy WS" in Table 3 -reveals that, as expected, the fact that a child goes to school at the same time as she works outside the household reduces total earnings in comparison with a comparable child who dedicates herself exclusively to market work. If one interprets this coefficient as reflecting fewer hours of work, then a child going to school works on average 40 per cent less than a dropout (for the pooled sample), or just under a quarter less for fifteen year-olds. These seem like reasonable orders of magnitude. The results from the estimation of the multinomial logit for occupational choice also appear eminently plausible. They are reported in Table 4 (for the pooled sample) and Tables 4a and 4b for 10-12and 13-15 year-olds, respectively. The reference category was "not studying" (i = 0), throughout. As expected, household income (net of the child's) has a positive effect on schooling, whereas the child's own to the alternative^.^^ Previous schooling at a given age has a positive (but concave) effect. Race has an (predicted) earnings have a negative effect. Householdsize reduces the probability of studying, compared insignificant effect on occupational choice, unlike gender which reflects the usual asymmetry between market work for males and domestic work for females. Parents'education has the expected positive effect -ontopoftheincomeeffect-onchildren'sschooling. Inview of this generalconsistency ofboththe earnings andthediscrete occupational choice models, the question now arises of whether the structural restrictions necessary for the consistency of the proposed simulation work -positive a1 and a2, and 0 < D < 1- hold or not. For the pooled sample and using (1I), we find that: 8, -b^1 - 0.0004 0.0075 + + = l-M-1 -Exp(-0.4118) =0.023 and a2=al 8* -8, =.024 The coefficients of income in the utility of alternatives j = 1and 2 i s thus positive, which i s in agreement with the original model. This i s also true of the utility of alternative j =O since it may be computed that Q = 0.023. The value of the parameter Dmay also be derived. Under the identifying assumption that K =1, it is given by : 53 Analogous resultsfor the 10, 11, 12, 13 and 14year-old samples are available from the authors on request. 54 With the Southbeinginsignificantly different from the referenceSoutheast region, as expected. 55 To the extent that household size reflects a larger number of children, this is consistent with Becker's quantity-quality trade- off. 148 D=--b, h +ao- -0.0074+0.023 a2 0.024 =0.6609 This figure means that children who are going to school but do not work on the market are estimated to provide domestic production for approximately two-thirds of their potential market earnings. Note that this is almost identical to the estimated value for M [= Exp (-0.4118) = 0.66251. Since M denotes the average contribution to household income from children both studying and working, as a share of their potential contribution if not studying, this implies that the estimated value of non-market work by children studying (and not working in the market) i s approximately equal to the market value of work by those studying (and working in the market). If there was little selection on unobservables into market work, this i s exactly what one would expect. Overall, the estimates obtained from the multinomial discrete occupational choice model and the earning equation seem therefore remarkably consistent with rational, utility-maximizing behavior. We may thus expect simulations run on the basis of these models and the identifying structural assumptions about the parameter K to yield sensible results. We can now turn to our main objective: gauging the order of magnitude of the effects of programs such as Bolsa Escola. An ex-ante evaluation of BolsaEscola and alternative program designs Bolsa Escola - and many conditional cash transfer schemes like it - are said to have two distinct objectives: (i) to reduce current poverty (and sometimes inequality) through the targeted transfers, and (ii) reducefuturepoverty, byincreasingtheincentives for today's poorto investintheirhuman to capital. Later on in this section, we will turn to the first objective. We begin by noting, however, that, as stated, the second objective i s impossible to evaluate, even in an ex-ante manner. Whether increased school enrollment translates into greater human capital depends on the trends in the quality of the educational services provided, and there i s no information on that inthis data set.56Finally, whether more "human capital", however measured., will help reduce poverty in the future or not, depends on what happens to the rates of return to it between now and then. This i s a complex, general equilibrium question, which goes well beyond the scope of this exercise. What we might be able to say something about i s the intermediatetarget of increasing school enrollment. While the preceding remarks suggest that this i s not sufficient to establish whether the program will have an impact on future poverty, it i s at least necessary.57An ex-ante evaluation of impact on this dimension of the program thus requires simulating the number of children that may change schooling and working status because of it. p, estimated from (9) - (1l), policy parameter values (T and Yo)taken from the actual specification of This i s done by applying the decision system (12) - with behavioral parameter values (a, y, M and D) and Bolsa Escola - to the original data. Equation (12) i s then used to simulate a counterfactual distribution of occupations, on the basis of the observed characteristics and the restrictions on residual terms for each individual child. Comparing the vector of occupational choices thus generated with the original, observed vector, we see that the program leads to some children moving from choice Si= 0 to choices Si=l or 2, 56 There i s limited information in other data sets, such as the Education Ministry's Sistema de Acompanhamento do Ensino Bisico (SAEB), but not for sufficiently long periodsof time. See Albemaz et. al. (2002). 57 One could argue that it i s not even necessary, since the transfers might, by themselves, alleviate credit constraints and have long-termpositive impacts, e.g. through improved nutrition. We focus on whether the conditional nature of these transfers actu- ally have any impact of the children's occupationalchoices (or time allocationdecisions). 149 and from Si= 1to choices Si= 2. The corresponding transition matrix i s shown in table 5 for all children between 10and 15, as well as for all children in the same age group living in poor households. '* Despite the small value of the proposed transfer, Table 5 suggests that one in every three children (aged 10-15) who are presently not enrolled in school would get enough incentive from Bolsa Escola to change employed on the labor market . The other three quarters would actually cease work outside their occupational status and go to school. Among them, just over a quarter would enroll, but remain household. This would reduce the proportion of children outside school from 5.8% to 3.9%. The impact on those currently both studying and working would be much smaller. Barely 2% of them would abandon work to dedicate themselves exclusively to their studies. As a result of this small outflow, combined with an inflow from occupational category 1, the group of children both studying and working would actually grow in the simulated scenario, albeit marginally. The impacts are even more pronounced, as one would expect, among the poor - who are the target population for the program. According to the poverty line being used, the incidence of poverty in Brazil i s 30.5%. However, because there are more children in poor households - this being one of the reasons why they are poor - the proportion of 10-15 children in poor households i s much higher: 42%. The second panel in Table 5 shows that dropouts are much more frequent among them-9.1 instead of 5.8 per cent for the whole population. It also shows that Bolsa Escola i s more effective in increasing school enrollment. The fall in the proportion of dropouts i s one-half, rather than one-third. As a result, the simulation suggests that Bolsa Escola could increase the school enrollment rate among the poor by approximately 4.4 percentage points. Once again, this increase comes at the expense of the "not studying category", whose numbers are halved, rather than of the "working and studying" category, which actually becomes marginally more numerous. A 50% reduction in the proportion of poor children outside school is by no means an insubstantial achievement, particularly in light of the fact that it seems to be manageable with fairly small transfers (R$15 per child per month). This i s partly due to the fact that the value of the current contributions of children who are enrolled in school i s a sizable proportion of their potential eamings when completely outside school. Those proportions are exactly the interpretation of the parametersM (for those who work on the market as well as study) and D (for those who work at home as well as study), which we estimated to be of the order of 0.66. Applying that factor to R$100, as a rough average of the earnings of children in category j = 0 (see Table 2), we are left with some R$33 as the true opportunity cost of enrolling in school. Consequently, those children who change occupation from that category in responseto the R$15 transfer must have average personal present valuations of the expected stream of benefits from enrolling greater than R$18. Those who don't, must on average value education at less than that. Because our simulations suggest that Bolsu Escola, as currently formulated, would still leave some 4% of all 10-15 year-olds (4.7% among the poor ones) outside school, it i s interesting to investigate the potential effects of changing some of the program parameters. This was, after all, one of the initial motivations for undertaking this kind of ex-ante counterfactual analysis. Table 6 shows the results of such a comparative exercise in terms of occupational choice, usingtransition matrices analogous to those in Table 5, once again both for all children and then separately for poor households only. Table 7 compares the impact of each scenario with that of the benchmark program specification, in terms of poverty and inequality measures. Four standard inequality measures were selected, namely the Gini ~ A household was considered poor if its (regionally price-deflatedand imputed rent-adjusted)per capita income was less than R$74.48 in the reference month of the 1999 PNAD survey. For the derivation of the poverty line, see Ferreiraet al. (forthcom- ing). 150 coefficient and three members of the Generalized Entropy Class: the mean log deviation, the Theil-T index and (one half of) the square of the coefficient of variation. For poverty, we present the three standard FGT (0, 1, 2) measures, with respect to the aforementioned Ferreira et. al. (forthcoming) poverty line. This later table allows us to gauge impact in terms of the first objective of the program, namely the reduction of current poverty (and possibly inequality). In both tables, the simulation results for six alternative scenarios are presented. In scenario 1, the eligibility criteria (including the means test) are unchanged, but transfer amounts (and the total household ceiling) are both doubled. In scenario 2, the uniform R$15 per child transfer i s replaced by an age- contingent transfer, whereby 10 year-olds would receive R$15, 11 year-olds would receive R$20, 12 year-olds would receive R$25, 13 year-olds would receive R$35, 14 year-olds would receive R$40, and 15 year-olds received R$45.59In scenario 3, transfer amounts were unchanged, but the means-test was raised from R$90 to R$120. Scenario 4 combines scenarios 1 and 3: the transfer was doubled, and the means-testraised to R$120. Scenario 5 combines scenarios 2 and 3 in the same way: an age-progressive transfer with a R$120 means-test. Scenario 6 simulated a targeted transfer exactly as inBolsa Escola, but with no conditionality: every child in households below the means-test received the benefit, with no requirement relating to school attendance. Table 6 gives rise to three main results. First of all, a comparison of Scenario 6 and the actual Bolsa Escola program suggests that conditionality plays a crucial role in inducing the change in children's time-allocation decisions. The proportions of children ineach occupationalcategory under Scenario 6 are to enroll in order to receive the benefit - rather than the pure income effect from the transfer - which i s almost identical to the original data (i.e. no program). This suggests that it i s the conditional requirement the primary cause of the extra demand for schoolingevident inthe Bolsa Escolacolumn. Second, scenario 1reveals that the occupationalimpact of the program i s reasonably elastic with respect to the transfer amount. The proportion of un-enrolled children drops another percentage point (Le. some 25%) in response to a doubling of the transfers. The proportion of children in the "studying only" category rises by the same percentage point. Scenario 2 suggests that it doesn't matter much, in aggregate terms, whether this increase in transfers i s uniform across ages, or made to become increasing in the age of the child. Finally, scenario 3 (and the combinations in scenarios 4 and 5) suggest that occupational effects are less sensitive to the means-testthan to the transfer amount. Results are considerably less impressive in terms of the program's first stated objective, namely the reduction in current poverty (and inequality) levels. Table 7 suggests that the program, as currently envisaged, would only imply a one percentage point decline in the short-run incidence of poverty in Brazil, as measured by P(0). However, there i s some evidence that the transfers would be rather well targeted, since the inequality-averse poverty indicator P(2) would fall by proportionately more than P(O), from 8% to 7%. This i s consistent with the inequality results: whereas the Gini would fall by only half a point as a result of the scheme, measures which are more sensitive to the bottom, such as the mean log deviation, fall by a little more. Overall, however, the evidence in column 2 of Table 7 falls considerably short of a ringing endorsement of Bolsa Escola as a program for the alleviation of current poverty or inequality. The situation could be somewhat improved by increases in the transfer amounts (scenarios 1 and 2). Nevertheless, even a doubling of the transfer amount to R$30 per month would only shave another 1.3 percentage points off the headcount.60An increase in the means-test would not help much, as indicated 59 The householdceiling was also doubledto R$90in this case. 6o The simulated one-percentage-pointfall in P(2) is, once again, morerespectable. 151 by Scenario 3. This i s consistent with our earlier suggestion that the program already appears to be well- targeted to the poor. If it fails to lift many of them above the poverty line, this i s a consequence of the small size of the transfers, rather than of the targeting. These results contrast with the arithmetic simulations reported by Camargo and Ferreira (2001), in which a somewhat broader, but essentially similar program would reduce the incidence of poverty (with respect to the same poverty line and in the same sample) by two-thirds, from 30.5% to 9.9%. This was despite the fact that the absence of a behavioral component to the simulation weakened its power, by excluding from the set of recipients those households whose children might have enrolled in response to the program. The reason i s simple: Camargo and Ferreira simulate much higher transfer levels, ranging from R$150 to R$220 per household (rather than child). Conclusions Inthis paper, we proposedamicro-simulationmethodfor evaluating andexperimentingwithconditional cash-transfer program designs, ex-ante. We were concerned with the impacts of the Brazilian Bolsa Escola program, which aims to reduce both current and future poverty by providing small targeted cash transfers to poor households, provided their children are enrolled in and in actual attendance at school. We were interested inassessingtwo dimensions of the program: its impact on the occupational choice (or time-allocation) decisions of children, and the effects on current poverty and inequality. For this purpose, we estimated a discrete occupational choice model (a multinomial logit) on a nationally representative household-level sample, and used its estimated parameters to make predictions about the counterfactual occupational decisions of children, under different assumptions about the availability and design of cash transfer programs. These assumptions were basically expressed in terms of different values for two key policy parameters: the means-testlevel of household income; and the transfer amount. Because predicted earnings values were needed for all children in the simulation, this procedure also required estimating a Mincerian earnings equation for children in the sample, and using it to predict earnings in some cases. Also, because the income values accruing to each household were not symmetric across different occupational choices, standard estimation procedures for the multinomial logit were not valid. An identification assumption was needed, and we chose it to be that children not enrolled in school work only in the market, and have a zero contribution to domestic work. Under this assumption, the estimation of the model generated remarkably consistent results: marginal utilities of income were always positive, and very similar across occupational categories. Time spent working by those enrolled in school, as a fraction of time spent working by those not enrolled, was always in the (0, 1) interval and was basically identical-and equal to two-thirds - whether work was domestic or inthe market. When this estimated occupational choice model was used to simulate the official (April 2001) design of the federal Brazilian Bolsa Escola program, we found that there was considerable behavioral response from children to the program. About one third of all 10-15 year-olds not currently enrolled in school would - according to the model - enrol in response to the program. Among poor households, this proportion was even higher: one half would enter school. The proportion of children in the middle occupational category ("studying and working in the market") would not fall. In fact, it would rise, marginally. Results in terms of the reduction of current poverty, however, were less heartening. As currently designed, the federal Bolsa Escola program would reduce poverty incidence by one percentage point only, and the Gini coefficient by half a point. Results were better for measures more sensitive to the bottom of the distribution, but the effect was never remarkable. 152 Both the proportion of children enrolling in school in response to program availability and the degree of reduction in current poverty turn out to be rather sensitive to transfer amounts, and rather insensitive to the level of the means-test. This suggests that the targeting of the Brazilian Bolsa Escola program i s adequate, but that poverty reduction through this instrument, although effective, i s not magical. Governments may be transferring cash in an intelligent and efficient way, but they still need to transfer more substantial amounts, if they hope to make a dent inthe country's high levels of deprivation. 153 References Abramovay, M.; C. Andrade and J.J. Waiselfisz (1998): Melhoria Educacional e ReduqBo da Pobreza, (Brasilia: EdiqBes UNESCO). Albernaz, Angela; Francisco H.G. Ferreira and Creso Franco (2002): "Qualidade e Eqiiidade na Educaq Bo Fundamental Brasileira", Discussion Paper #455, Departamento de Economia, Pontificia Uni- versidade Catblica, Rio de Janeiro. Atkinson, Anthony and Franqois Bourguignon (1991): "Tax-Benefit Models for Developing Countries: Lessons from Developed Countries", in J. Khalilzadeh-Shirazi and A. Shah, (eds.), Tax Policy in DevelopingCountries, (Washington, DC: The World Bank). Baland, Jean-Marie and James A. Robinson (2000): "Is Child Labor Inefficient?", Journal of Political Economy, 108, pp.663-679. Basu, Kaushik (1999): "Child Labor: Cause, Consequenceand Cure, with Remarks on International La- bor Standards", Journal of Economic Literature, XXXVII, pp.1083-1119. Bhalotra, Sonia (2000): "Is Child Work Necessary?', Mimeo, Cambridge University. Bourguignon, Franqois; M. Fournier and M. Gurgand (2001): "Selection Bias Correction Based on the MultinomialLogit Model", Working Paper#20023-04, CRESTANSEE, Paris. Bourguignon, Franqois and Pierre-Andre Chiappori (1994): "The Collective Approach to Household Behavior", in Blundell, Preston and Walker (eds.), The Measurement of Household Welfare, (Cambridge: Cambridge University Press). Browning, Martin; Franqois Bourguignon; Pierre-Andre Chiappori and Valerie Lechene (1994): "Income and Outcomes: A Structural Model of Intra-HouseholdAllocation", Journal of Political Economy, 102 (6), pp.1067-1096. Camargo, JosC Mircio and Francisco H.G. Ferreira (2001): "0Beneficio Social Onico: uma proposta de reforma da politica social no Brasil", Discussion Paper #443, Departamento de Economia, Pontifi- cia Universidade Catblica, Rio de Janeiro. Chiappori, Pierre-Andre (1992): "Collective Labor Supply and Welfare", Journal of Political Economy, 100, pp.437-467. Ferreira, Francisco H.G., Peter Lanjouw and Marcel0 Neri (forthcoming): "A Robust Poverty Profile for Brazil Using Multiple Data Sources", Revista Brasileira de Economia. Freije, Samuel and Luiz F. Lopez-Calva (2000): "Child Labor and Poverty in Venezuela and Mexico", mimeo, ElColCgio de Mexico, Mexico City. Gertler, Paul and Paul Glewwe (1990): "The Willingness to Pay for Education in Developing Countries: Evidence from Rural Peru", Journal of Public Economics, 42, pp. 251-275. Grootaert, Christiaan and Harry Patrinos (eds.) (1999): The Policy Analysis of Child Labor: A Compara- tive Study, (New York: St Martin's Press). Harding, Ann (ed) (1996): Microsimulation and Public Policy, (Amsterdam: Elsevier). Heckman, James and E.Vytlacil(2002), Econometric evaluation of social programs, inJ. Heckman and E. Leamer (eds), Handbook of Econometrics, vol. 5, (Amsterdam: North-Holland) Parker, Susan and Emmanuel Skoufias (2000): "The Impact of Progresa on Work, Leisure and Time Al- location", IF'PRIFinal Report on Progresa, IFPRI, Washington, DC. 154 Ravallion, Martin and Quentin Wodon (2000): "Does Child Labor Displace Schooling?Evidence on Be- havioral Responsesto an Enrollment Subsidy", Economic Journal, 110, pp.Cl58-Cl75. Rocha, SGnia and Joiio Sab6ia (1998): "Programas de Renda Minima: Linhas Gerais de uma Metodologia de Avaliaqiio", DiscussionPaper#582, PEA/UNDP, Rio de Janeiro. Rosenzweig, Mark and Robert Evenson (1977): "Fertility, Schooling and the Economic Contribution of Children inRuralIndia: An Econometric Analysis", Econometrica, 45 (5), pp.1065-1079. Rubin, Donald (1977): "Assignment to a Treatment Group on the Basis of a Covariate", Journal of Edu- cational Statistics, 2, pp.1-26. Rubin, Donald and Paul Rosenbaum (1985): "The Bias Due to Incomplete Matching", Biometrica, 41 (l),pp.103-116. Sant'Ana, S.R. and A. Moraes (1997): Avaliagiio do Programa Bolsa Escolado GDF, (Brasilia: Fundagiio Grupo EsquelBrasil). Schultz, T. Paul (2000), "The Impact of Progresa on School Enrollment", IFPRI Final Report on Pro- gresa, IFPRI,Washington, DC. World Bank (2001), Brazil: An Assessment of the Bolsa Escola Programs, Report 20208-BR, Washing- ton, DC. 155 Table 1:Schoolenrollment and occupationof childrenby age (10-15 years old) 10 11 12 13 14 15 Total Not Studying 2.5% 2.3% 3.3% 5.6% 8.0% 13.0% 5.8% Workingand Studying 8.1% 10.9% 14.0% 18.3% 22.6% 27.3% 16.9% 89.4% 86.8% 82.7% 76.1% 69.4% 59.6% 77.3% Total 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% Source: PNAD/IBGE 1999and author's calculation Table 2 :Samplemeans. Characteristics of children and the householdthey belongto (10-15 years old only) Notstudying WorkingandStudying Studying Total Age 13.5 13.2 12.3 12.51 Years of schooling 2.9 3.9 4.1 3.97 Householdper capitak o m 80.9 104.5 202.0 178.25 Eamhg'schikiren(observed) 10 118.4 34.2 38.0 11 98.3 44.6 50.4 12 100.7 50.8 57.0 13 76.8 66.9 68.5 14 100.5 83.8 87.8 15 127.6 109.3 113.9 Years of schoohgof the rmst educatedparent 3.2 4.0 6.4 5.79 Age of the oldestparent 46.4 46.1 44.5 44.89 Nurnberof householdmrrbers 5.8 5.9 5.2 5.39 Race (White) 36.9% 40.9% 51.6% 48.9% Gender (Male) 53.0% 65.2% 46.9% 50.3% North 6.1% 5.6% 6.0% 5.9% Northeast 40.4% 45.6% 29.9% 33.2% Southeast 32.8% 26.1% 43.5% 39.9% south 14.1% 15.9% 13.7% 14.1% Center-West 6.6% 6.7% 6.9% 6.9% Metropolitanarea 18.2% 12.8% 30.9% 27.1% Urbannonmtroplitan 34.0% 49.2% 16.0% 22.7% Ruralareas 47.8% 38.0% 53.0% 50.2% Proportionof universe 6.1% 16.8% 77.1% 100.0% Population 1,208,3 13 3,345,075 15,329,237 19,882,625 Source: PNAD/IBGE1999and author's calculation 156 10to 15years old 15 years old Coeficient Std P>lzl Coeficient Std Pqzl nobs 2444 1010 R' 0.43 0.54 Dummy W S -0.4118 0.0324 0.0000 -0.2285 0.0385 0.0000 Years of schooling -0.0136 0.0198 0.4920 -0.0409 0.0244 0.0930 Years of schoohg' 0.0110 0.0021 0.0000 0.0077 0.0025 0.0020 Male 0.1746 0.0283 0,0000 0.1349 0.0355 0.0000 W t e 0.0658 0.0295 0.0260 0.0600 0.0355 0.0910 North -0.2329 0.0447 0.0000 -0.1515 0.0748 0.0430 Northeast -0,2054 0.0379 0,0000 -0,1529 0.0472 0.0010 %Uth -0.0461 0.0422 0.2750 -0.0165 0.0475 0.7290 Center-West -0.1082 0.0426 0.0110 -0.0801 0.0490 0.1020 Urbannonmetroplitan -0.0284 0.0408 0.4870 0.0472 0.0538 0.3810 RUfd 0.0042 0.0327 0.8980 0.0507 0.0393 0.1970 Logofmeans e&s by cluster 0.3788 0.0148 0.0000 0.4756 0.0199 0.0000 Intercept 3.5266 0.0751 0,0000 2.7600 0.1176 0.0000 Source: PNADlIBOE1999 andauthot`s calculation Table4: MultinomialLogitCoefficients WorhngandStudylrw Shrdyq Pseudo-R2 #obs Coeficicnt Std P+ Coeficient Std P>lrl 10to 15years old 0 1586 42153 Totalhouseholdincome 0 0004 0.0001 4.6300 0.0006 0.0001 7.8200 chddren(What) -0 0075 0.0018 -4.2100 -0.0074 0.0015 4.9300 Totalpeople byhousehold -0.0343 0.0169 -2 0300 -0.1751 0.0157 -11.1400 Years of schooling 0.6635 0.0407 16.3100 0.8338 0.0378 22.0500 Years of schooling' -0 0383 0.0051 -7.5300 -0.0837 0.0048 -17.6000 w h i t e 0.0138 0.0628 0.2200 0.1613 0.0566 2.8500 Male 0.7447 0.0567 13.1400 -0.1841 0.0503 -3.6600 Max parent's education 0.0371 0.0104 3 5700 0.1300 0.0093 13.9100 Max parent'sage -0 0035 0.0027 -1.2900 0.0023 0.0024 0.9600 Number o fchildrenbelow7 -0.0108 0.0362 -0.3000 0.0875 0.0332 2.6400 Rank of child 0.6433 0.0538 11.9500 0.9099 0.0504 18.0500 North 0.5673 0.1113 5.1000 0.0980 0.0998 0.9800 Northeast 0.7086 0.0789 8.9800 0.2854 0.0717 3.9800 south 0.1901 0.0867 2.1900 -0.3569 0.0778 -4.5800 Center-West 0.2757 0.0977 2.8200 -0.1013 0.0869 -1.1700 Urbannonmetroplitan 1.1803 0.0807 14.6200 -0.4999 0.0728 -6.8700 Rural 0.2030 0.0735 2.7600 -0.1835 0.0628 -2.9200 Meansofeamings by cluster 0.0118 0.0073 1.6200 0.0022 0.0057 0.4000 Intercept -2.3104 0.2071 -11.1600 04412 0.1846 2.3900 Source. PNADnEGE1999 and authox'c calculation 157 Table4a: MultinomialLogit Coefficients W o r m and Studymg Studying Pseudo-R2 #obs Coefieient Std P+ Coeficient Std P>lzI 10 years old 0.2393 6853 Total householdmcome 0 0001 0 0004 0 8570 0 0006 0.0003 0 0760 Earmng'schtldren(What) -0 0711 0 0273 0 0090 -0 0460 0.0251 0 0670 Total people by household -0 0072 0 0769 0 9250 -0 0766 0.0679 0 2590 Years of schoohg 2 5342 0 2466 0 0000 2 8347 0.2138 0 0000 Years of schoohg2 -0 4023 0 0599 0 0000 -0 4993 0.0513 0 0000 Whlte -0 2006 0 2611 0 4420 -0 1311 0.2375 0 5810 Male 0 6865 0 2057 0 0010 -0 2596 0.1803 0 1500 Maxparent's education 0 0235 0 0396 0 5530 0 0621 0.0343 0 0710 Maxparent's age -0 0030 0 0094 0 7460 -0 0037 0.0079 0 6430 Number of children below 7 0 1721 0 1294 0 1840 0 0682 0.1145 0 5510 Rank o f child 0 1935 0 1354 0 1530 0 0982 0.1179 0 4050 North 18948 0 4854 0 0000 0 7064 0.4214 0 0940 Northeast 17310 0 3279 0 0000 0 8418 0.2865 0 0030 south 0 5136 0 3755 0 1710 -0 2513 0.3263 0 4410 Center-West 15302 0 4621 0 0010 0 8179 0.4202 0 0520 Urbannon metropktan 3 1158 0 3732 0 0000 0 5128 0.3077 0 0960 RUrd 10942 0 3324 0 0010 0 1258 0.2610 0 6300 Means of cmnrngs by cluster 0 3847 0 1175 0 0010 0 1872 0.1090 0 0860 Intercept -3 4075 0 7173 00000 14643 0 5963 0 0140 11Years old 02610 7022 Totalhouseholdmcome -0 0001 0 0002 0 7180 0 0002 0 0002 0.3690 Ea."& chtldren(What) -0 0247 0 0313 0 4310 0 0481 0 0296 0.1050 Totalpeople by household 0 1202 0 0750 0 1090 0 1143 0 0698 0.1020 Years of schookng 18700 0 2440 0 0000 19526 0 2194 0.0000 Years of schookng' -0 2545 0 0500 0 0000 -0 2714 0 0451 0.0000 Whlte 0 0327 0 2585 0 8990 0 0935 0 2424 0.7000 Male 0 3583 0 2115 0 0900 -0 4660 0 1970 0.0180 Max parent's education 0 0057 0 0416 0 8910 0 0850 0 0381 0.0260 Max parent's age -0 0061 0 0094 0 5180 0 0020 0 0085 0.8180 Number of children below 7 -0 1829 0 1392 0 1890 -0 2591 0 1310 0.0480 Rank of c u d -0 0341 0 1468 0 8160 -0 2566 0 1372 0.0610 North 12805 0 4554 0 0050 0 5387 0 4270 0.2070 Northeast 0 8725 0 3029 0 0040 -0 1828 0 2794 0.5130 south 14466 0 4633 0 0020 0 3018 0 4 6 3 0.4990 Center-West 0 4704 0 3925 0 2310 -0 5806 0 3546 0.1020 Urbannonmetroplltan 16909 0 3100 0 0000 -0 0622 0 2874 0.8290 RUral -0 0171 0 2962 0 9540 -0 1303 0 2621 0.6190 Means of eanun,es by cluster - . 0 0277 0 0313 0 3750 -0 0778 0 0313 0.0130 Intercept -2 4141 0 6731 00000 16659 0 6096 0 0060 12 years old 02258 7196 Total householdincome 0 0000 0 0002 0 8790 0 0003 0 0002 0 1610 Earmng'schddren(What) -0 0093 0 0084 0 2680 -0 0150 0 0084 0 0730 Totalpeople by household -0 0005 0 0581 0 9940 -0 0769 0 0554 0 1650 Years of schoohg 13963 0 1728 0 0000 15883 0 1572 0 0000 Years of school& -0 1405 0 0305 0 0000 -0 1787 0 0278 0 0000 whlte 0 1590 0 2030 0 4330 0 2339 0 1907 0 2200 Male 0 9392 0 1726 0 0000 0 0547 0 1580 0 7290 Maxparent'seducation 0 0072 0 0319 0 8220 0 0795 0 0289 0 0060 Maxparent'sage -0 0023 0 0089 0 8010 0 0001 0 0082 0 9920 Number of chlldren below 7 -0 0121 0 1164 0 9170 0 0150 0 1082 0 8900 Rank of chdd 0 6002 0 1712 0 0000 0 4909 0 1601 0 0020 North 12716 0 3599 0 0000 0 6064 0 3377 0 0720 Northeast 0 8998 0 2481 0 0000 0 3845 0 2312 0 0960 south 0 0463 0 2760 0 8670 -0 5530 0 2496 0 0270 Center-West -0 0045 0 3113 0 9890 -0 2569 0 2818 0 3620 Urbannon metrophtan 2 5243 0 2654 0 0000 0 1413 0 2319 0 5420 RUrd 10872 0 2437 0 0000 0 2634 0 2035 0 1960 Means of e m s by cluster 0 0214 0 0184 0 2440 -0 0046 0 0188 0 8080 Intmcept -4 0732 0 6442 0 0000 -0 1458 0 5756 0 8000 Source PNADflBGE 1999 and authois calculahon 158 Table 4b: MultinomialLogitCoefficients Workulg andStud= Studyuw Pseudo-R2 #obs Coeficient Std P+ Coeficimt Std Pqzl 13years old 0.1813 7077 Totalhouseholdmcome 0 0003 0.0002 0 1390 0 0004 0.0002 0.0280 Earrnng'schddren (What) -0 0211 0.0078 0 0070 -0 0143 0.0078 0.0660 Totalpeoplebyhousehold 0 0422 0.0434 0 3310 -0 0561 0.0402 0.1630 Years of schookng 0 7544 0.1192 0 0000 0 9879 0.1135 0.0000 Years of schoohg' -0 0431 0.0184 0 0190 -0 0737 0.0176 0.0000 Whtte 0 0422 0.1606 0 7930 0 1379 0.1492 0.3560 M a l e o 8550 0.1365 0 0000 -0 0430 0.1270 0.7350 Maxparent'seducation 0 0097 0.0250 0 6990 0 0798 0.0232 0.0010 Maxparent'sage -0 0022 0.0064 0 7260 -0 0020 0.0059 0.7300 Number of ciddrcnbelow7 -0 1093 0.0921 0 2350 -0 0676 0.0856 0.4300 Rank ofchdd 0 1376 0.1495 0 3570 0 0841 0.1422 0.5540 North 0 7935 0.2676 0 0030 0 4388 0.2477 0.0760 Northeast 10844 0.1923 0 0000 0 7627 0.1812 0.0000 south 0 4987 0.2313 0 0310 -0 2157 0.2142 0.3140 Center-West 0 4728 0.2452 0 0540 0 1034 0.2218 0.6410 Urbannonmetrophtan 11527 0.2210 0 0000 -0 5803 0.2061 0.0050 RlUd 0 2636 0.2008 0 1890 -0 1524 0.1824 0.4030 Meansofe a " - .by cluster 0 0342 0.0131 0 0090 -0 0090 0.0138 0.5140 Intercept -2 4040 0.5116 0.0000 03448 0 4679 04610 14 years old 0 1795 7052 Totalhouseholdincome 0 0002 0.0002 0.2150 0 0004 0.0001 0.0060 Eannng's cMdren(What) -0 0029 0.0039 0.4590 0 0077 0.0049 0.1190 Totalpeoplebyhousehold 0 0431 0.0362 0.2350 -0 0256 0.0349 0.4630 Years of schookng 0 4374 0.0924 0.0000 0 7161 0.0946 0.0000 Years of schookng' -0 0041 0.0132 0.7530 -0 0385 0.0132 0.0040 White -0 0286 0.1310 0.8270 0 1265 0.1233 0.3050 Male 0 6915 0.1151 0.0000 -0 2034 0.1092 0.0630 Maxparent's education 0 0369 0.0233 0.1130 0 1091 0.0218 0.0000 Maxparent'sage -0 0137 0.0060 0.0220 -0 0024 0.0056 0.6750 Number ofchildrenbelow7 -0 1234 0.0769 0.1090 -0 1285 0.0750 0.0870 Rankofchid -0 1313 0.1638 0.4230 -0 2028 0.1582 0.2000 North 0.2363 0.0070 0 4337 0.2236 0.0520 Northeast oo 6328 9830 0.1634 0.0000 0 8621 0.1573 0.0000 south 0 0849 0.1802 0.6380 -0 5569 0.1678 0.0010 Center-West 0 5093 0.2100 0.0150 0 2439 0.1995 0.2220 Urbannonmetroplitan 0 9129 0.1687 0.0000 -0 7278 0.1599 0.0000 RWal 0 2720 0.1529 0.0750 -0 1551 0.1388 0.2640 Meansofearningsbycluster 0 0016 0.0051 0.7620 -0 0397 0.0078 0.0000 Intercept -1.4708 0.4538 0.0010 -0.0760 0.4350 0.8610 15vears old 0 1549 6953 Totalhouseholdincome 0 0002 0.0001 0 0180 0 0004 0.0001 0 0000 Earmng'schddren (What) -0 0029 0.002~3 0 2860 -0.0049 0.0032 0 1290 Totalpeoplebyhousehold 0 0752 0.0294 0 0110 0 0195 0.0291 0 5040 Years of schoolrng 0 2210 0.0719 0 0020 0 3994 0.0735 0 0000 Years of schookng2 0 0109 0.0087 0 2130 -0 0052 0.0088 0 5510 whltc -0 1459 0.1070 0 1730 0 1201 0.1015 0 2370 Male 0 6201 0.0949 0 0000 -0 1786 0.0903 0 0480 Maxparent'seducation 0 0503 0 0173 0 0040 0 1109 0.0162 0 0000 Maxparent's age 0 0103 0.0050 0 0400 0 0214 0.0049 0 0000 Number ofchddrenbelow7 -0 2800 0.0669 0 0000 -0 2619 0.0638 0 0000 Rpnk ofchdd North 0 3019 0 1848 0 1020 0 3707 0 1741 0 0330 Northeast 0 6628 0 1291 0 0000 0 6156 0.1260 0 0000 south -0 0736 0 1440 0 6090 -0 5285 0 1384 0 0000 Center-West 0 1186 0 1635 0 4680 -0.0937 0 1538 0 5420 Urbannonmetroplitan 0.4465 0 1439 0 0020 -0 7331 0.1403 0 0000 Rural -0 1145 0 1298 0 3780 -0 3243 0 1216 0 0080 Means of e m s by cluster 0 0048 0 0038 0 2030 -0 0100 0 0050 0 0440 Intercept -2 2590 0.3516 0.0000 -1.4724 0.3444 0.0000 Source. PNAD/IBGE1999and author's calculation 159 Table 5: Simulatedeffect of BolsaEscolaon schoolingandworking status (allchildren10-15years old) All Households Not Studying Working andStudying Studying Total Not Studying 66.6% 9.0% 24.4% 5.8% WorkingandStudying 98.1% 1.9% 16.9% Studying 100.0% 77.3% Total 3.9% 17.1% 79.1% 100.0% PoorHouseholds Not Studying Working and Studying Studying Total Not Studying 52.0% 13.4% 34.6% 9.1% WorkingandStudying 99.0% 1.O% 23.7% Studying 100.0% 67.2% Total 4.7% 24.7% 70.6% 100.0% Source: PNADnBGE 1999and author'scalculation 160 7. THEDYNAMICS THE SKILL-PREMIUMINBFUZIL:GROWING DEMANDAND OF INSUFFICIENT SUPPLY? Andreas Blomand Carlos Eduardo VBlez61 Abstract Labor market income in Brazil is extremely unequal. Previous literature, such as Ferreira and Paes de Barros (1999) hasfound that thepersistent rise in skill-premium explains a largepart of the increasing wage-inequality. During the last decades, the marginal returns to higher educa- tion have increased dramatically and have generated an increasing "convexification " of the earnings function deteriorating income inequality in two ways: first by increasing the income differences between the skilled and unskilled and by severely weakening the potential income equalization that could be obtained from the higher and more equally distributed educational endowments among new cohorts of workers, Ve`lez, et a1 (2001). This paper uses theframework of Katz and Murphy (1992) and Murphy, Riddell and Romer (1998), to explain the evolution of returns to labor skills in terms of specific supply and demand changes during the last two dec- ades. Our findings suggest the substantial but asymmetric expansion of the education system in Brazil -with weaker growth of tertiary education- combined with a steady skill-biased change in labor demand explains the increasing skill premium. Our simulations suggest that an aggressive long term expansion in the supply, ceteris paribus, reducing the skill-premium to levels observed in developed countries would lead to a 4.5 gini points reduction in the gini coeficient of individ- ual labor market income. Therefore, expanding tertiary education aggressively would not only increase economic growth by investing in high return assets, but it would also mitigate wage- inequality in the long run. The policy challenge is finding ways to expand tertiary education at reasonable marginal costs - well below the current level observed in average Brazilian universi- ties- and with minimum burden on thepublic budget. Introduction Brazil has one of the most skewed income distributions inthe world. The huge disparity between those that have and those that have not, receives growing attention among policymakers and in the public. Multiple factors lie behind the notorious income-inequality. The increasing wage disparity between those who have tertiary education and those that do not appears to be one key piece in the inequality-puzzle. This paper models the evolution of returns to labor skills inBrazil interms of specific supply and demand changesduringthe last two decades. The role of education for wage-inequality dates back to the human capital revolution in economics, Becker (1965) and Mincer (1974). In particular Tinbergen (1975) emphasizes that differences in salary between workers to a large extent can be attributed to differences in attained schooling and that a thorough inequality analysis has to encompass both supply and demand: "Quite often the opinion is held that income diflerences in a rigid way reflect differ- ences in the productive qualities of people. As a consequence, the inequality between human beings is seen as a reasonfor income inequality to persist, not to say to be pre- 61 The authors are grateful for comments and suggestions from Mauricio Santamaria, Serguei Soares and from partici- pants at the World Bank presentation. 163 ordained. Even if we assumefor a while that differences in ability cannot be changed [...Iwhat matters (for inequality) is the difference between qualities available and qualities required by the demand side." Jan Tinbergen (1973, p.15 in "Income Differences: recent research" In the case of Brazil, abundant evidence shows that the distribution and reward of education matters for wage-inequality, for example Birdsall and Salbot (1996). They find that the Brazilian workforce both in the past and in the present has accumulated less education compared to other countries with the same per capita income. Furthermore, they present findings similar to Patrinos (2001) and Psacharopoulos (1993) emphasizing that returns to schooling in Brazil in the 1970s and 1980sexceeded returns to schooling in most other nations inthe world.62 The returns to schooling by education level display large variation between education levels and over time. Ferreira and Barros (1999); Blom, Holm-Nielsen and Verner (2001); as well as Arbache, Green and Dickerson (2001) show that the reward of lower and middle levels of education in the early 1980s substantially exceededthose currently prevailing. While the wage of tertiary graduates relentlessly increased duringthe same period. Blom, Holm-Nielsen and Verner (2001) finds that for the period 1982 to 1998, returns to 4, 8, and 11 years of completed schooling dropped 26 percent, 35 percent and 8 percent, respectively, while returns to tertiary schooling surged 24 percent. Hence, the reward of education became increasingly convex implyingthat the marginal reward of education increased with the years of education. InBrazil, increasedreturnto skills exacerbatesexistingincome-inequality.This isaresult oftwo opposite directed impacts. On the one hand, the decline in returns to schooling for workers with middle levels of education, 5-11 years of schooling, entails that the wage difference between workers with middle levels of education and workers with less schooling decreases. Due to the comparatively small stock of education in Brazil, workers with middle levels of education generally earn above the average wage. Consequently, a decline in returns to middle levels of education diminishes wage-inequality. On the other hand, the surging returns to tertiary income distribution. Therefore, increased convexity also worsens wage-inequality. Ferreira and schooling raises the wage of workers predominantly positioned in the highest deciles of the Paes de Barros (1999) find that for the 1976-1996 period, the former effect dominated the latter. That is, the change in returns to schooling decreased wage-inequality. These findings indicate that (a) education policy i s a powerful tool to reduce wage-inequality and (b) the rise in the returns to tertiary education exacerbates wage-inequality. In addition to the previous direct effect, there is another unequalizing indirect effect of earnings convexification on income inequality. VClez et al(2001) show that beyond a certain level of convexity of the earning function, an increase in mean education produces an (unexpected) un- equalizing effect on income inequality. That is, if rates of return rise with the level of schooling, accumulation of schooling worsens income inequality unless the accumulation of schooling i s sufficiently egalitarian.63Thus, increasing returns to education also weaken the potential income The finding of below-average accumulation of education relative to per capita income and above average returns to schooling, supports the analysis o f Tinbergen (1975). The essence o f Tinbergen's idea i s a race between technology and accumulated education, where returns to education i s interpreted as a price o f education. Specifically, the returns to schooling i s an outcome of supply determined by outcome o f the education system and demand for schooling deter- mined by the marginal productivity o f different types o f labor, which Tinbergen strongly associates with the level o f technology. In the case o f Brazil, the relative low stock o f accumulated education relative to the GDP-level suggests that technology "leads the race" and returns to schooling therefore exceed that found internationally. 63 Ibidem, p.31. 164 equalization that could be obtained from the higher and more equally distributed educational endowments of new cohorts of workers entering the labor market inBrazil. The increased convexity in the human capital earnings function has attracted a fair amount of attention due to the proliferation of the phenomenon. Several major economies in Latin America experienced increased convexity during the 1980s and 1990s; see Beyer et aZ(1999), Santamaria (2000) and VelCz et a1 (2001), Lachler (1998), Galiani and Sanguinetti (2000) for evidence on Chile, Colombia, Mexico and Argentina, respectively. Some low and middle-income countries in other regions underwent the same change, see Foley (1997) as well as Muraimy and Lam (1999) for studies on South Africa and India. Furthermore, the trend i s not confined to developing countries. Katz and Murphy (1992); Romer, Riddle and Murphy (1998); and Card and Lemieux (2000) reveal that the USA, UK and Canada to different degrees all experienced a rise in the income disparity between workers with tertiary education and workers with secondary education. The suggested explanations for increased reward of advanced human capital generally evolve around (a) the relative supply of tertiary graduates or (b) increased demand for skill due to trade- liberalization and the revolution in information and communication technology (ICT). In a seminal paper Katz and Murphy (1992) demonstrate how the relative supply of college graduates to high school graduates combined with a time-linear increase in labor demand for college graduates drive the relative wage of the two education groups. Later studies corroborate and refine this finding for the US, Card and Lemieux (2000) and Romer, Riddle and Murphy (1998). Other studies have examined reasons for increased skill-premium to highly educated workers in middle-income countries, but they mostly concentrate on the role of trade-liberalization. Few studies focus on the role played by relative supply with the exception of Santamaria (2000) for Colombia. The supply-focus i s crucial for policy analysis, since it yields estimates of how policymakers through education policy can impact on long-term wage-inequality. The paper provides policymakers with information on the scope for long term reduction of wage inequality by turning the handles of the education system, and in particular the number of graduates from tertiary education. We estimate the Katz and Murphy-model in order to understand how long-run education policy influences the wage premium and identify policy initiatives that could reduce inequality. The Katz and Murphy model has to our knowledge not been fully applied to the case of Brazil. Arbache, Green and Dickerson (2001) use the framework to investigate how labor demand for advanced skills developed over time. In addition to applying the Katz and Murphy model that links the supply of skilled labor with relative wage, we examine the role of wages of tertiary graduates for wage inequality. We utilize the by now fairly standard technique of wage simulation. First, we express the relative wage of tertiary graduates as returns to tertiary schooling. By usingthe technique of wage-simulation we evaluate how returns to schooling impacts on wage-inequality. Hence, the paper directly links supply of skilled labor that policymakers strongly influence in the medium to the long run, and wage-inequality . The paper i s organized as follows. The subsequent section presents the level and evolution of wage-inequality as well as the distribution, evolution and remuneration of education from 1976 to 1999. In section three we estimate the Katz and Murphy model. The analysis enables us subsequently in section four to perform counterfactual policy analysis to illustrate the impact of education policy on wage-inequality. More specifically, we ask how wage-inequality would have developed if the education policy in the 1980s and 1990s had differed from the observed policy. 165 The final section provides an assessment of the scope for reducing wage-inequality by expanding tertiary education. 1- Wage-inequalityand Education. Inequality manifests itself in numerous ways in a society. In this paper, we focus on income inequality and more specifically on pre-tax wage-inequality. Wage-inequality i s measured for individual wage earners. This provides a well-defined and quantitatively solid starting point. However, this paper does not cover a number of important aspects of inequality, such as government transfers, taxation, distribution of assets and unemployment. The analysis adopts the gini-coefficient as the general measure of inequality. The Brazilian national household survey, PNAD, provides the household information. We consider income from both primary and secondary employment and deflated by the national consumer price index, INPC, to obtain real wages.6465 All wages are expressedas hourly wage in fixed September 1997 prices. We use all existing household surveys from 1976 to 1999 with the exception of 1984 and 1992. Figure 1presents the evolution of wage-inequality in Brazil in the examined period. Figure 1Wage inequality from 1976to 1999 0.63 0.62 .-5 8 3 0.61 --- F5 a 0.6 m 0 0.59 .-0 Q) E8 0.58 .-Y eC 0.57 0.56 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 Source:Authors' own calculationbased on PNAD Note: Hourly wage from principal and secondaryjob for full time workers. From 1976 to 1981, wage-inequality decreased from an internationally remarkably high level of 0.61. The so-called "lost decade" of the 1980s plagued by negative growth and heavy economic distress severely exacerbated wage-inequality that peaked in 1989 at 0.63. Hereafter, the distribution of wages became less unequal. From 1989 to 1999, the gini-coefficient fell 6 gini- points from 0.63 to 0.57. Notably with two substantial drops in 1990 and in 1994. The latter drop @ In 1976, the Brazilian statisticaloffice, IBGE, did not assemble the INPC. Alternatively, we deflate the wage by the IPC price index also collectedby the IBGE. The two price indexesfollow similar paths during the 1980s. Hence, they seemingly measureprice changes in a similar way. The necessary change in deflatorsshould therefore not significantly reducethe comparabilityof the wage series from 1976to 1981. 65Appendix A provides detailedinformation about the computationof each data series presentedin this paper. 166 i s consistently linked to the economic stabilization following the Real plan. The evolution depicted by figure 1 corresponds to similar findings by PEA (2001),Neri and Camargo (1999) as well as Ferreira and Barros (1999).Despite the reduction in wage-inequality in the 1990s,the distribution of wage still remains notoriously inequal in international comparison, Ferreira and Bourguignon(2000). EducationaZcompositionof theZaborforcefrom Z976toZ999 Duringthe 1970s, 1980s and 1990s,policymakers strived to provide universalprimary education and as well as to turn the secondary education system into a mass-system. 66 As a consequence, the schooling system expanded in these decades. Nevertheless, the speed of expansion in 1970s and 1980s was moderate on an international scale, Birdsall and Sabot (1996).The fruits of these efforts are clearly visible today. The educational attainment of the workforce steadily improved. The average number of attained years of schooling increased from 4.8 in 1976 to 6.9 in 1999. Figure 2 shows how the improvement in the education system gradually translated into increased education attainment of the workforce. Figure 2 Educational composition of the workforce 100% 90% 80% 70% 60% 50% 0Some Lowersec 40% mSome Prima 30% 20% 10% 0% I 1976 1982 1985 1987 1989 1992 1995 1997 1999 Source: Authors' own calculation based on PNAD Note: Hoursworked by full time workers In 1976,over 60 percent of the workforce had at most a diploma from primary school (4 years of schooling). As older primarily less educated workers were replaced by younger peers with considerably more schooling, the share of workers with 4 years of schooling or less declined to 38 percent. In 1999,more than half of the workers had graduated from lower secondary school (8 years of schooling). The distribution of schooling became considerably more equitable over the examined period. Inequality of schooling, measured by the gini-coefficient of the years of schooling, decreased 66 The Brazilian education system consists o f four levels with national graduation exams; primary education from 1-4 years of schooling, lower secondary education (sometimes called upper primary) from 5-8 years of schooling, upper secondary education from 9-1 1 years of schooling and, lastly, tertiary education from 12-15/17 years of schooling. 167 every single year from 1976 to 1999. In 1976, the gini-coefficient for years of schooling was 0.49. By 1999, inequality in schooling had dropped to 0.37. Educafionandwages Education continues to be the main determinant of an individual's labor market income. Generally, wage increases monotonically with the level of education. We compute the evolution of wage by education group using a so-called fixed-weight wage method developed by Freeman (1980) and applied by Katz and Murphy among others. The method calculates the wage for a fixed demographic composition of the work force in order for alteration over time in demographic characteristics (age and gender) not to affect the wage-series. Specifically, we divide the labor force into demographic cells by age and gender.67Weights are assigned to each cell on the basis of the average number of workers in each cell during the entire period. The wage of an education group i s given by a weighted average of the average wage of each demographic cell with that level of education. Figure 3 presents the results. VI VI 6 - .-8 h 3 - 8 `C (I m z h -s; m 5 - u .u t= .-X 2.5 - i= X .-C 4 - F al 8 3 - r z -$a z I a 2 - P 2 -. Y L 12 1976 1980 1983 1986 1989 1992 1995 1999 1976 1980 1983 1986 1989 1992 1995 1999 Some Upper secondary year Some Tertiary year lource: Authors' own calculationbased on PNAD Notably, the wages display high sensitivity to economic cycles. The spikes in wages 1986, 1989 and 1994/5 correspond to economic expansions. The average wage of all levels of education decreased during the two decades considered.68However, the wages across education groups 67 We divide the workforce into 9 age-cohorts, 2 genders and 7 education groups (no schooling, incompleteprimary, complete primary, lower secondary, upper secondary, incomplete tertiary and complete tertiary), which yields 126 demographiccells. This does not imply that income per worker decreased over the period, since the size of education groups with sec- ondary and tertiary education increasedover time. Actually, the per capita incomefell during the 1980s, but recovered morethan fully in 1990s. 168 declined by different magnitudes. In particular, the workers with tertiary education endured a small decline in wages. As the wage of secondary graduates and tertiary graduates drifted apart, the relative wage of tertiary graduates increased. In 1976, a typical worker with tertiary education earned in 320 percent and 210 percent more than a colleague with lower secondary education and upper secondary education, respectively. By 1999, the same wage gap had expanded to 450 percent and 240 percent, respectively. Oppositely, the wage of lower secondary graduates and upper secondary graduates declined substantially relative to workers with less education. RETURNSTO SCHOOLING In order to purge the relationship between wage and education further, we estimate returns to schooling. By estimating returns to schooling from Mincerian regressions, we control for additional factors than in the above Katz and Murphy (1992) method. We estimate the following Mincerian regression for every year in our sample. lny, =PschSi +Xi'y+.si * (1) Where yi and Si indicate wage and number of years of schooling for individual i. X stands for a matrix of control variables. y i s the vector of estimated coefficients for control variables. We control for age (quadratic formulation), gender, labor market status (formal employee, informal employee6', self-employed, or employer), region of living (5 major regions) and rural residence. Furthermore, we estimate the returns to each level of schooling separately by adopting a spline specification. The full form of (1) i s then (lb). 1' Yi = P p r i*'i,pri +P l s e c*'i,lsec +Pu s e c*'i,usec +P t e r*'i,ter+Xi 'Y+ Ei (1') Where, Si indicates the number of years of schooling attended at each level of education (primary, lower secondary, upper secondary and tertiary). The ,8s are estimated returns to school to one additional year in primary, lower secondary, upper secondary and tertiary schooling, re~pectively.~'Figure 4 displays the returns to schooling to each education level. 69 The partition of employees into a formal and an informal group is done on the basis of "carteira assinada", signed workcard. Holding a signed workcard entitles an employee to a series of rights and benefits, he or she can therefore meaningfully be classifiedas working in the formal (regulated)sector of the economy. 7" We tested an aitemativeformulation with dummies for completionof each education level. The spline specification provedthe most explanatory indicating that years of schooling in betweengraduation years matter in the wage determi- nationprocess. Although, clear signs of so-called sheep-skineffects exist. 169 Figure 4 Returns to schooling by education level .064 , , , , , , , , 1976 1960 1983 1986 1989 1992 1995 1999 Returns to Lower secondary schooling year .18-I R -. .I4 .184 , , , , , , , , 1 4 1976 1980 1983 1986 1989 1992 1995 1999 Returnsto Upper secondary schooling year Returns to Tertiaryschooling year Source: Authors' own calculation based on PNAD Controlling for additional factors, we confirm the findings from the average wage data. The returns to primary, lower secondary and upper secondary schooling declined over the period, while returns to tertiary schooling persistently increased through out the two decades. These findings correspond to previous findings by Mom, Holm-Nielsen and Verner (2001) based on metropolitan labor market data. Remunerationof educationana' Wage-znequahQ How did the shifts in returns to schooling affect wage-inequality? To answer this question, we apply the technique of wage-simulations. Specifically, the technique consists in estimating a mincerian regression for the year under investigation as done above in (lb). From (lb) we obtain a predicted log wage for each worker, In j j .We subtract the estimated effect of education as given by the estimated vector of returns to schoolingfl,,, . We then add the effect of education A given a simulated (hypothetical) vector of returns to schooling, psi,,,Hence, A . the simulated wage for individual i, ysim,i,i s computed as: 170 The unexplained residual from (l),i s added to the simulated wage, so that unexplained wage- e, variation remains constant.71By analyzing the wage-distribution of the simulated wage, ysim,we assessthe impact of returns to schooling on wage-inequality. Figure 5 depicts how wage inequality would have developed from 1976 to 1999 if only returns to schooling changed. Hence, all other factors including control variables, attained schooling and residual wage remain identical to the 1976 level. Figure 5 Wage-inequality if only returns to schooling changed 0.615 0.610 0.605 0.580 0.585 I 0.580 1976 1981 1 W 1983 1585 1986 1987 1988 1989 1990 1993 1995 1996 1997 1998 Source: Authors' own calculationbasedon PNAD Note: The graph depicts wage-inequalityfrom 1976 to 1999 if only returns to schooling had changed from 1976to 1999. All other factors are kept constant at the 1976level. The decline in returns to lower and medium levels of schooling that took place from 1988 to 1996 accounts for a drop in the gini-coefficient of more than 2 gini-points. Changing the base year of the simulation, the endowments, from 1976 to 1989 or 1999 marginally reduces the bearing that returns to schooling exert on wage-inequality. For all three base years, the drop in wage-inequality exceeds 2 gini-points. Consequently, we conclude that the decline in returns to schooling, ceteris paribus, explains at least a third of the reduction in wage-inequality from 1989 to 1996. TheskzlZ-premzumandwageznequahy Workers with tertiary education experienced as the only education group rising returns to schooling. Since the salary of this education group lies above the national average, the rise in returns to schooling lead to a deterioration of wage inequality. We assess the role of returns to tertiary schooling for wage inequality by observing the change in wage-inequality as the returns to tertiary schooling change. This i s done by the same simulation technique as above where we only change the 4" element of the vector of returns to schooling. That is, we vary the returns to tertiary schooling while keeping returns to the other three education levels constant. Figure 6 71In order for the simulationsnot to be biased, we needto assume that the error termis independentof attained school- ing, see Velez et a1 (2001). 171 presents the findings for the level of wage-inequality in 1999 as a function of returns to tertiary schooling. Figure6 Wage-inequalityin 1999as afunction of returns to tertiary education 27% 25% 23.9% 22% 20% iw0 16% 1 4 ~ 0 12% iocx0 (Actual) Returnsto tertiary schooling Source: Authors' own calculation based on PNAD Note: The graph depicts wage-inequality in 1999 if only returns to tertiary schooling changed to the hypothetical level given by the horizontal axis. All other factors are kept constant. A hypothetical reduction in demand for workers with tertiary education that causes returns to tertiary schooling to decline from the 1999-level of 23.9 percent to 13 percent -the level of returns inthe US- would reduce wage-inequality from 0.575 to 0.530. This simulation shows that a reduction in the skill-premium leads to a substantially more equitable income distribution. The returns to schooling for the other three education levels influence less on wage-inequality, appendix figure 1-4. Nevertheless, the return to upper secondary education still impacts considerably on wage-inequality, whereas returns to lower secondary and primary education matters less for wage-inequality. This finding motivates the rest of the paper, which investigates how policymakerscan influence the skill-premium and thereby reduce wage-inequality. 2- RelativeSupply, RelativeDemandandthe Skill-premium The skill-premium is a price on advanced human capital. As all other prices, the interaction of demand and supply determines the skill-premium. In a seminal work, Katz and Murphy (1992) investigate how demand and supply for tertiary graduates determine the skill-premium in the United States. This section applies the Katz and Murphy model to the case of Brazil. The analysis will help us understand why the skill-premium rose during the last two decades. Furthermore, since output of the education system in the long run dictates the educational composition of labor supply, the analysis yields valuable insight into how education policy, through the impact on supply, affects the skill-premium and wage-inequality. 172 A sghzedreZations& betweendemand supp.(uandwages T h i s sub-section provides the theoretical foundation for the estimated models. Katz and Murphy (1992) develop an empirical model basedon a simple CES production function:72 Where Y, L, arefers to output, labor of the type indicated by the subscripts and the elasticity of substitution between the two types of labor, respectively. The subscripts ter and usec stand for tertiary and upper secondary. Many similar models exist, Murphy, Riddle and Romer (1998), Card and Lemieux (2000), Haskel (2000) and Santamaria (2000). Assuming labor supply to be predetermined in the short run, hence a vertical supply curve, the relative labor demand curve arising from (3) determines the relative wage. These assumptions can be shown to imply the following simplerelationship between wage, demand and supply: Where w, D and S indicate wage, demand and supply, respectively. The key parameter for the analysis i s the economy wide elasticity of substitution, a A highelasticity of substitutionimplies that secondary graduates easily substitute tertiary graduates. The higher the elasticity, the smaller the impact of relative supply on output and the smaller the impact of relative supply and demand on relative wage (and wage-inequality). The economy wide elasticity of substitution i s determined by (a) sector specific elasticity of substitution embodied in each sector's choice of production technology and (b) the ease with which labor flows between sectors with different intensity of skilled labor. The model i s an aggregate labor market model that assumes clearing of the labor market across sectors within each year. While the assumption plausibly holds in the long run, there exist short runbarriers to flow of labor across sectors inthe economy. That is, in the short runfirms face a sector specific upward sloping supply curve due to sector specific experience and different geographical locations of sectors. Sector specific shocks translating into sector specific changes in relative labor demand might therefore create short run deviations from the long run equilibrium. The estimated model i s therefore expected to hold only inthe long run.73 The estimated regression version of (4) i s (5): 72 The usual notations denote two education groups; unskilled and skilled workers. The actual educational attainment of the two reference groups often differs from one study to another. Since we do not perceivegraduates from upper secondary schoolingto be unskilled, we prefer to denote the two groups by their education level, upper secondary and tertiary. 73Card and Lemieux (2000) find that in the case of the US, Canada and UK the rise in the relative wage almost exclu- sively benefited the younger cohorts between 26 and 40 years old. They refine the Katz and Murphy model to take account for imperfect substitution between age-cohorts.Additionally, Santamaria (2000) shows that the relative wage differs betweengenders, which strongly suggests imperfect substitution between these two segments of the labor mar- ket.These shortcomingsof (4) underscorethe aggregate natureof the model. 173 Where the error term, 4, picks up misspecifications, unobserved factors and measurementerrors. From the coefficient estimate of p, we compute the elasticity of substitution between the two types of labor as In order to force the elasticity of substitution to be similar for demand and for supply, we compute the change innet-supply, NS. Model (1) i s the first model that we estimate. The model estimated by Katz and Modelincludes a time trend to capture a skill-biased change in labor demand. The change in demand i s caused by an increase in the productivity of workers with tertiary education relative those with secondary education.74Murphy, Riddle and Romer (1998) provide a theoretical foundation based on skill- biased technological progress for the inclusion of the time trend. We equally estimate this model, -1 henceforthcalled model (2): log[%) =a P,log[ NSSJ + +P2t+E, model(2) wu,t NSU,t Furthermore, we consider an alternative model where we estimate the impact of demand and the impact of supply separately. The motivation i s twofold: (a) For reasons discussed in the following section, the applied demand indicator underestimates the increase in the demand for highly skilled labor; (b) The shock to supply might impact wages differently than shocks to demand. Why? Innovations in supply almost exclusively occur when young cohorts of school- leavers enter the labor market. Since school-leavers possess little sector experience, they are relatively more flexible in terms of sector of employment (and region of employment). Hence, innovations in supply tend to have high elasticity of substitution due to high sector and geographical mobility. Oppositely, sector specific demand shocks affect workers that already hold sector specific experience. The latter reduces mobility across sectors, since workers loose the reward of sector specific experience by switching sector. Therefore, innovations in demand impact relative wage to a larger extent than supply.75Inthis scenario, a negative sector specific 74 The assumptionof a constant rise in the relative productivity of highly skilled labor differs from the assumption of a constant increases in either Total Factor Productivity (TFP) or output per labor, since the two latter conceptsconsider the productivity of the labor force as a whole without distinguishingbetween different types of labor. An increase in TFP could both favor and disfavor highly skilled labor depending upon the sector or education group in which the productivity gains take place. 75 Birdsall and Salbot (1996),Menezeset a1 (2000) and Blom, PavcnikandSchady(forthcoming) document that wages indeedvary across sectors and industriesof employment, even controlling for observablepersonal characteristics.Such high wage-differencesbetweensectorstestify to the importanceof sector specific experience. 174 shock to a skill-intensive sector that decreases the demand for tertiary graduates by 1% would reduce the relative wage of tertiary graduates more than the impact of a 1% increase in the supply of tertiary graduates. Hence, we propose the following model, henceforthmodel (3): log[:rt] -=a+-*log :s [ -+-*log ?r,t] JD [;r,t] -+p*t+&tmodel(3) sec, t sec, t sec,t Where os and oD refer to the elasticity of substitution estimated for supply and demand, respectively. App&zng themodeZtoami2dZe-zncomecountry Like Katz and Murphy (1992), this paper focuses on the wage of workers with tertiary education relative to those with secondary education. However, the educational composition of the U S and that of a middle-income country like Brazil differ dramatically. In the US, the two reference education groups encompass more than 90 percent of the workforce, while only a little bit more than a third have attended upper secondary education in Robbins and Gindling (1999) argue that the cut-off point on the education scale between low skilled and highskilled should be placed lower when analyzing the skill-premium in a middle-income country than in a developed country. As a consequence they examine the wage-difference between graduates from primary education and graduates from tertiary education in the case of Costa Rica. Nevertheless, we uphold the original cut-off point because we explicitly focus on explaining the rising skill- premium to tertiary education. Hence, the only sensible cut-off point on the education scale i s between upper secondary education and tertiary education. Furthermore, the sharp divergence in wages between the two reference groups undoubtedly demonstrate that the labor market distinguishbetween the two education groups. In the case of Brazil, the relative small coverage of the two groups might reduce the overall relevance of the analysis, but the fact that returns to tertiary education substantially matter for wage-inequality, warrants, in your eyes, the choice of focusing on the wage premium to tertiary education only. Computationof reZativewage, reZativesupp& andreZativedemand RELATIVE WAGE We follow Katz and Murphy (1992) and compute relative wage with fixed demographic weights as outlined previously. The computation excludes other education groups than complete upper secondary graduates and complete tertiary graduates. The opposite would affect relative wage to the extent that the relative share or the wage of dropouts changes duringthe period.77 SUPPLY We measure supply as the share of working hours supplied by each education group out of the total number of working hours. Arguably, this choice of method misses some aspects of supply 76Measured in hours of work, figure 2. "Murphy,RiddleandRomer(1998)proposeanalternativetothefixeddemographicweightsmethod.Theyestimatea mincerian regression for each year with dummies for completed college and completed high school controlling for age and gender. The difference between the two coefficient estimates i s used as a measure for relative wage. 175 that might impact wages.78We include both genders in the supply series. Hence, we assume full substitutability between male and female colleagues. Although the assumption i s not fully correct, it i s preferred to the alternative assumption of no substitutability. The analysis focuses on explaining the relative wage of upper secondary graduates to tertiary graduates. However, the evolution in the supply of other education groups likely influences this wage-ratio. For example, an abundant supply of lower secondary graduates potentially adds to the supply of upper secondary due to the higher degree of substitutability with upper secondary graduates than with tertiary graduates. Therefore, the supply of upper secondary graduates and tertiary graduates should be adjusted. Ideally, the adjustment should be substituted for a multi- equilibrium framework estimating elasticities of substitution between all education groups. A such model has so far not been developed, we therefore follow the literature and add a weighted supply of the excluded education groups to the two reference groups. Several suggestions exist for how to determine the weights. We adopt the "wage-level approach" suggested by Card and Lemieux (2000). They base the weights on the relative wage of the excluded education group to the wage level of the two reference groups. For example, the hours worked by dropouts from tertiary education are added to the supply of upper secondary graduates with the weight cp computed completeusec - (Wcomplete - ter-Wincomplete ter)'(Wcomplete ter-wcompleteusec) (7) Applying (7), the hours of work by tertiary dropouts are split 66 percent to the supply of completed upper secondary and 33 percent to completed tertiary. The weights are more important inthe case of Brazilthan inthe case of developed countries, since the excluded education groups contain the bulk of the working force. As a robustness check, we estimate the model (1) - model (3) without adjusting for the supply of lower secondary and primary education. RELATIVE DEMAND Contrary to supply, we do not observe demand for labor. Changes in relative demand for education arise from two sources; (a) Within-sector demand shifts. That is, the same products are produced, but firms have changed production technology. The change in production technology implies a change in relative demand for labor. For example, if shoe manufactures replace cutters or/and sewers ~~ ~ 78 For example, the computed supply series do not reflect the incidence of open unemployment. A differencein unem- ployment level between the two education groups could affect the relative wage. However, if insiders determine the wages, unemploymentexert no influenceon wages. Accountingfor unemployment could therefore introducenoiseinto the supply series and thereby blur the associationbetween supply and wages. A similar argument pertainsto the failure of the supply series to account for the educationalattainment of the peoplenot participatingin the labormarket or part- time employees. Additionally, Katz and Murphy (1992) measure supply series in terms of so-called efficiency units. The idea i s that moreexperienced workers supply more efficient labor.Therefore, the supply series shouldtake account of this difference in efficiency. Katz and Murphy (1992) weigh the supply of each age-cohort by the wage in order for the supply series to reflect possible differencesin effectivesupply. Santamaria(2000) applies both methodsin the case of Colombiaand finds only a marginaldifference. 79 Alternatively, Katz and Murphy (1992) proposesweighting according to wage level and wage-evolution as determinedin the following regression(with no constant): Wincom ter - Pcom fer - *Wcom ter +P c o m upper sec*Wcom uppersec Additionally, Robbinsand Gindling (1999)proposesome a-priori reasonableweights. Appendix table 3 presents the weights arising from the proposedmethods. The alternativeapproaches give weights similar to those of the adopted Card and Lemieux method. Consequently, the chosen methoddoes not affectthe findings. 176 with machines, then the demand for unskilled labor decreases while demand for skilled tech- nicians increases. Within-sector demand shifts closely relate to the installation of skill-biased technology.*' Between-sector demand shifts. Sectors differ in their demand for skilled workers. When some sectors expand and others retract, the economy's relative demand for skilled labor shifts. For example, if the relative size of the agriculture sector decreases, the demand for un- skilled workers declines, since agriculture intensively employs unskilled labor. We measure "between-sector demand shift" by the fixed coefficient "manpower requirements" index developed by Freeman(1975)and ditto (1980): Where AD,,stands for the demand shift for education group e in time period t. pejindicates the employment share of education group e in sectorj. The weight equals the share of employment at the beginning of the period and i s constant. N indicates the absolute number of employees. Intuitively, the demand shift measures the aggregate change in labor demand due to a change in the sectoral composition of the economy. That is, if a sector intensive in skilled labor (brhigh) expands (ANpO), then demand for skilled labor increases (AD,,>O). The estimated demand shift i s biased, since the shift only reflects changes in employment and not in wages (quantity). The bias will be the opposite of the movement in wage. For instance, if demand for tertiary graduates increased causing the wage to rise, firms will respond by hiring a smaller number of graduates. Hence, in this case, the demand indicator underestimates the "true" demand shift. The degree of bias varies from one education group to the other depending on the evolution of wages; Johnson (1992), Katz and Murphy (1992) as well as Santamaria (2000). Appendix table 5 and 6 display the factor intensity and the employment share, respectively, which are the two components needed to construct the demand shifts. Appendix table 7 presents labor demand shifts in sub- periods of 1976-1999, while appendix Figure 1graphs the yearly between-sector demand shifts. The between-sector shifts show that sectors intensive in low human capital labor- predominantly agriculture and mining- decreased consistently from 1976 to 1999. All other education groups experienced rising demand in the period taken as a whole, although demand decreasedfor certain groups in sub-periods. A part from the sub-period 1990-1995, sectors relatively intensive in skilled labor expanded.*l The largest rise in demand occurred for upper secondary graduates. The demand for this group surged 20.8 percent from 1976 to 1999 due to changes in the sectoral composition. The period from 1976 to 1981 stands out as a period with large flows towards skill intensive sectors. 8o Changes in technology lead firms to demand more skilled labor if the installed technology i s skill-biased. These within-sector shifts can be observed by keepingthe productionvolume of each sector constant and computethe change in labor intensityacross sectors. Hence, the opposite of the between-sectorshifts. Katz and Murphy (1992)estimatethe within-sector shifts by observingthe change in labor input intensity for 3 occupationgroups. However,the PNAD sur- vey do not have well-structuredoccupationdata, we therefore choose not to computethis indicator. 81 The period of exception, 1990-1995, coincides with the trade-liberalizationperiod. The simultaneity suggests that trade-liberalization lead to a specializationin sectors intensive in unskilled labor. The observed specializationcorre- sponds to the prediction of the Hekscher-Ohlin model; Countries that are relatively abundant in low-skilled workers specialize in sectors with low skill intensity. The drop in demand for skilled labor could hence be a consequence of falling profits and reductionsin outputs taking place in the skill-heavy sectors as the price of their product fell due to increasedforeign competition following reductionsin tariffs. 177 rigure 7 Relative wage and relative net supply of tertiary to upper secondary grad iates o Rel. Wage (log) A Rel. supply (log) -.78 1.08 -.82 1.04 52 -.86 1 9 n S > .96 -.9 2 .92 -.94 .88 -.98 .-. -1.02 1981 1983 1985 1987 1989 1991 1993 1995 19971999 Year Source: Authors' own calculation based on PNAD Figure 7 presents the net supply and wage series. In Brazil, the log relative wage in 1976 was 0.88, which widened to 1.09 in 1999. The relative wage in Brazil substantially exceeds the relative wage prevailing in developed countries. In the US, log relative wage bottomed out at an all time low at around 0.39 in the late 1970s. Thereafter, the log relative wage grew to 0.52 in 1987, where the Murphy, Riddle and Romer (1998) series end. In Canada, the log relative wage oscillated around 0.48 from 1981to 1994 with no observable trend line. The figure displays the expected negative association between relative net supply and relative wage. Furthermore, as expected from the discussion of the model's characteristics, the supply curve displays less volatility than the wage curve. The theoretical considerations suggestedthree ways to estimate the relationship between relative wage, relative supply and relative demand. Table 1presents the estimation results. 178 Table 1Estimationresults Model (1) (2) (3) Time sample 1981-1999 1981-1999 1981-1999 Constant 0.076 0.063** 0.056*** (0.087) (0.24) (0.16) Net Supply (-lh) -0.98*** -0.21 -0.62*** (0.095) (0.33) (0.18) Time trend 0.015** 0.011*** (0.0056) (0.002) Demand 3.55*** (0.69) RL 0.89 0.92 0.98 Observations 15 15 15 Testsfor misspecif cation Durbin-Watsontest 2.27 2.30 2.00 RESETtest 0.39 0.87 0.93 Implied Elasticity (0) 1.o 4.7 1.61 Note: *,** and *** denotestatisticalsignificanceat the lo%, 5% and 1%level, respectively. The value for RESETtest is the probability of correct specification. Inmodel(l), net-supply statistically significantly affects relativewage with acoefficient relative close to minus one. The inferred elasticity of substitution becomes 1. Given the simplicity of the model, the explanatory power i s unexpectedly high. The R2reaches 0.89, which in part can be explained by the few number of observations. The constant i s statistically insignificant. Inmodel(2), weintroduces time trendto capture aconstant skill-biasedchangeinlabordemand. The addition of a time trend slightly augments the explanatory power from 0.89 to 0.92. The constant equally increases marginally and turns statistically significant. The estimated elasticity of substitution alters completely. The coefficient (numerically) drops from -0.98 to -0.21 and becomes statistically insignificant. Oppositely, the introduced time trend i s statistically significant, but only on a 5 percent level. Hence, the standard errors appear substantially inflated considering the model's high explanatory power. The size of the coefficient of the time trend suggests that, all other things equal, the skill-premium to tertiary education increases by 1.5 percent annually due to skill-biased change in labor demand. The difference between the two models i s striking. Following model (l), the asymmetric development inthe educational composition of labor supply almost fully explains the rising skill- premium to tertiary education. The expansion of lower and upper secondary education duringthe 1980s and 1990s reduced the marginal productivity of these types of labor and therefore decreased the wage relative to that of tertiary graduates.82According to model (2), a constant skill-biased change in labor demand causes the relative wage to rise. In this setting the relative supply played a statistically insignificant role for the rise in the skill-premi~m.~~ ** The decrease in marginalproductivity does not imply that investmentin education at this level is unprofitable. On the contrary, an individual still significantly increases hisher revenue by attending secondary school. However, the marginalincreasein wage derivedfrom the completionof secondaryeducation decreasedover time. 83The lack of relationshipbetweenrelativesupply and wage could be explainedby the Rybczynski effect (as derivedin a Heckscher-Ohlinmodel): In a multi sector economy with a higher number of products than factors, the zero-profit conditions determine relative factor prices. Although each sector faces a downward sloping demand curve, the aggre- gated economy wide demand curve is horizontal. An increase in relative supply of skilled workers causes skill- 179 The disappearance of the impact of supply arises due to multicollinearity between the supply series and the time trend. That is, the evolution of relative supply i s almost identical to a line. The correlation between the relative supply and the time trend is 0.94. The Variance Inflation Factor (VIF) for model (2) i s 16.5, which clearly exceeds the rule of thump value of 10indicating high incidence of multicollinarity. This implies that the separation of the impact of supply from the impact of the time trend hinges on short term (yearly) deviations between the two series. However, in the short run the association between supply and wages i s likely to be weak due to the time-lags involved in wage-setting and barriers to labor flow between sectors. The statistical distinction between the two explanations i s therefore tro~blesome.~~ 85 Model (3) allows for supply and between-sector demand shifts influence wages differently. The model exhibits an extremely high fit, which once again partly derives from the low number of degrees of freedom (10). All coefficients are statistically significant at the 1 percent level and display the expected sign. The estimated elasticity of substitution lies in between the estimate for the other models. The impact of demand on relative wage exceeds the impact of supply by a factor 6, indicating either (a) the demand indicator i s grossly underestimated or/and (b) the labor markets adapts more rigidly to demand shocks than to innovations in supply. The interpretation of the result from model (3) i s less polarized than the two other models. It stresses that both the supply side and the demand side contributed to the rise inthe wage of highly skilled labor. The three models yields widely different estimate of the role that the asymmetric expansion in supply played for the rising skill-premium (and wage-inequality). Froman econometrical point of view model (3) fares the best. It features the highest explanatory power, all variables highly significantly and no signs of misspecification as given by the RESET and Durbin-Watson test. Furthermore, the estimated elasticity of substitution resembles in magnitude that found on other labor markets, table 2. Note: Card and Lemieux (2000) find a larger substitutability between the two labor types once they take into account a large but finite elasticity of substitution between age-cohorts. intensive sectors to expand production and employment whereas skill-extensive sectors retract. Hence, in this setting relative supply does not change relative wage. However, the models predictions do not hold if (a) the number of factors exceed number of products manufactured, (b) the production composition (type of products) alters, and (c) barriers to labor flow exist across sectors, for example sector specific experience; see Haskel (1999). 84Cardand Lemieux (2000) observed an identical problem of colinearity in the case of Canada: "the relative supply essentiallyfollows a linear trend [...I, effect cannot separately be identi3ed from the effect from its the linear trend. The large standard errors reflect this identification problem.", Card and Lemieux (2000) p. 18, foot- note 19. 85Forthe period considered, the relative supply displays an evolution corresponding to an I(1) process similar to the time trend. Co-integration techniques could therefore be appropriate. However, the number of observation does not allow for application o f this technique. 180 International evidence suggests that the elasticity of substitution lies in the interval 1.1-2.1. Hence model (1) and (2) are extremities. Assuming the Brazilian labor market works in a similar way to the Colombian and the North American, model (1) overestimates the influence of supply, while model (2) underestimate the effect. The extreme estimates likely arise from the high collinearity between the supply and the time trend. Omitting the time trend in model (1) causes the coefficient of supply to reflect the steady increase in relative demand and therefore overestimate. The oppositely occurs for model (2), the high and insignificant estimate of the elasticity of substitution likely reflects that the time-trend accounts for the impact of relative supply on relative wage, which then turns insignificant. Hence, three factors -econometrical criteria, international evidence and balanced economic common sense- concurrently indicate that model (3) most appropriately estimates the elasticity of substitution. As a robustness test for the role played both other education groups weighted in the supply series, we carried out an estimation without added the supply of workers with lower secondary and primary education. As shown in appendix table 8, the exclusion of other education groups has no influential bearing on the findings.86Although a precise estimate of the impact of relative supply of tertiary educated labor on relative wage in Brazil is handicapped by the time trend behavior of the supply series, we conclude from the analysis in this section that the economy wide elasticity of substitution between upper secondary graduates and tertiary graduates lies inthe vicinity of 1.6. This implies that the asymmetric expansion in the educational composition of the labor supply accounts for 54 percent of the increase inthe relative wage of tertiary graduates while changes in demand answer for the remaining 46 percent. We found that the change in the sectoral composition decreased the relative demand for workers with tertiary education by 34 percent of the observed increase. However, a constant increase in skill-biased demand for highly skilled labor of 83 percent of the observed wage-change more than fully off-sat the equalizing effect from the between-sector demand shifts.87 3- What if?Alternativepaths for supply and wage-inequality inthe past The estimated model of the relationship between educational composition of the labor supply and relative wage between education groups provides information on the mechanics of the labor market. Inthis section, we take advantage of the previous findings. By rearranging the estimated model, we illustrate how policymakers through the influence on the output of the education system can affect relative wage. Specifically, we ask the question: How much should the tertiary education system had expanded for the skill-premium to remain constant at the 1981 level? 86 The size and significance levels remain unchanged. Only the size of the constant changes, which reflects that the supply or workers with primary or lower secondary education essentially develops similarly to that of upper secondary education. The inclusion of lower levels of education hence amounts to a multiplication of the supply of workers with upper secondary education, by a factor Z. Technically: 87The relative supply of tertiary graduates fell by 0.21 and the relative wage increased by 0.24, figure 7. Additionally, we estimated the coefficient to 0.62. The fraction explained by supply thus equals 0.21*0.62/0.24=0.54. For demand, the between-sector demand shifts reduced the wage gap by 3.55* -0.024/0.24=-0.36 whereas the skill-biased change in labor demand accounts for 0.01 1*18/0.24=0.825. 181 We apply model (2) from the last section, which includes a time trend and restricts demand and supply to impact relative wage in the same magnitude. Our policy variable, the explained variable, i s the supply of hours worked by workers with tertiary education. We turn around the model in order for the policy variable, the (net) supply of tertiary graduates to be expressed as a function of the estimated coefficients, the exogenous variables and the desired level for relative wage. Consequently, we assume explicitly that upper secondary education and lower levels of education expanded as observed. Turningaround the model and performing counterfactual simulationrequires a strong assumption of the exogenous character of the model. That is, the economic relationship estimated between the time-series (the estimated coefficients) i s assumed to remain stable even though the time series counterfactually change.@ Requiredexpansionzn tertiary educationfor aconstantskiZlFremium How much shouldthe labor supply of workers with tertiary education have expanded inorder to keep the skill-premium constant at the 1981-level?Interms of the above model framework, we set the relative wage through out the period to be constant at the 1981-level. In 1981, tertiary graduates earned 14.1 reais per hour compared to 6.1 reais for graduates of upper secondary education, hence a log relative wage of 0.84. This i s around the double of the US-level in 1981, 0.41; Katz and Murphy (1992). We estimate -S,,,,,, *exp[ ~ [ I o g ( R W , , , ),-0.005 t -0.341 * -0.62 1 *'Assessing the seriousness of the bias due this shortcomingis complex. Murphy, Riddle and Romer (1998) perform stability tests on the elasticity of substitution and the time trend of the skill-bias of labor demand. Interestingly, they found two statistically significant breaks in the time trend: A slow-down in skill-biased labor demand in 1976 and a speed-up in 1981.They interpret the breaksto be caused by variation in the introduction of general purpose technolo- gies. If such technologicalbreaks dependon the evolution of supply of education, such as the number of highly edu- cated workers, then the model's parameters are path dependant. Furthermore, the output of secondary education i s highly likely to influence later outputs of tertiary education. Therefore, the supply series of secondaryand tertiary edu- cation might not be independent as assumed in this policy analysis. These considerations emphasize that the findings should serve as educated indications for the impact of education policy, not final exact results. In the case of Brazil, it would make less sense to test for similar breaks since (a) the limited number of observations complicates the test and (b) the very high fit of the model signals that the linear specification is a correct specification. Arbache, Green and Dickerson (2001) find no breaks in demand. Similarly, we find no sign of a trend break, see residual plots appendix figure 6 and 8. 89 Additionally, the estimated equation takes into account the change in demand and the evolution of supply of other educationgroupsthan the two referencegroups; upper secondary and tertiary. 182 Estimating the model yields the following figure. 'igure8 Requiredsupply of tertiary graduatesfor aconstantskill-premiumconstant o Requiredfor constant Skill-Pre. A Observed .14 ! j .12 1 3 J 2 .1 .08 - .06 - Source: Authors' own calculation basedon PNAD Figure 8 shows that the share of working hours supplied by tertiary graduates should have expanded from 6 percent in 1981 to 13.5 percent in 1999 in order for the relative wage to have remained at the 1981-level. As shown the share increased by a meager 2.5 percentage point to 8.5 percent. In 1999, the difference between the two series approximately corresponds to 5 percent of the labor force or a staggering shortfall of four million graduates from the tertiary education system. In 1997, only 254,000 graduated from tertiary education while an impressive 1,3 million graduated from upper secondary education. World Bank (2000b) estimates that in order to keep the current continuation rate from secondary education constant an additional 400,000 extra seats in the tertiary education is necessary. Given labor demand plausibly continues to favor highly skilled labor, as indicated by the significance of the time trend, even such a major expansion would prove insufficient to stop the relative wage and wage-inequality to widen. 183 4- Summary T h i s paper seeks to inform policymakers about the role of the skill-premiumfor wage-inequality and the available policy instrumentsand their impact on the skill-premium.The paper finds that: 0 Graduates from Brazilian colleges are highlyrewarded for their education. The reward, the skill-premium, has risen constantly for two decades now. The returns to tertiary schooling, 23.9 percent, tops the list of returns in Latin America and i s almost the double of the re- turns in the US and Canada. The extremely high wage presentspolicymakers with an op- portunity for economic growth that can be capitalized by expanding the tertiary education system and thereby satisfy the high demandfor highly advancededucated workers. 0 The risingwage of workers with tertiary education exacerbates wage-inequality. The paper shows that if the returns to schooling inBrazil decreasedto levels found inNorth America, the wage-inequality would decline by around 4 gini-points. Additionally, we showed that the decline in the reward of primary and secondary education accounted for half the reduc- tion in wage-inequality that took place after 1988. This fact indicates the importance of the reward of education for wage-inequality. 0 An asymmetric expansion in the Brazilian education system and a constant skill-biased change in labor demand fully explain the rise in the skill-premium. The tertiary education system was unable to accommodatethe surge of graduates from upper secondary education during the 1980s and 1990s. Consequently, the supply of workers with tertiary education dropped 20 percent relative to the supply of workers with upper secondary education. Our application of the Katz and Murphy model in the case of Brazil fully explain the rise in relative wage. The analysis suggests that the elasticity of substitution between the two types of labor i s in the vicinity of 0.61. This implies that the asymmetric expansion in the educational composition of the labor supply accounts for 54 percent of the increase in the relative wage of tertiary graduates while labor demand answer for the remaining 46 per- cent. 0 A maior expansion in the tertiary education system could reverse the rising relative wage that deteriorates wape-inequalitv. We find that in order for the relative wage to have re- mained at the 1981 level, the share of workers with tertiary education should have more that doubled from 6 percent in 1981 to 13.5 percent in 1999. However, the share only reached 8.5 percent. Currently, about 12 percent of an age-cohort enters tertiary education inBrazil. Hence, at the current speed, it would take around 40 years for the share of work- ers with tertiary education to reach a meager 12 percent of the workforce. Policymakers face a daunting task in order to reverse the current trend of rising relative wage. For exam- ple, a doubling of the current rate of enrolment into tertiary education, from 12 percent to 24 percent, demands the creation of 2.5 millions extra seats in the tertiary education sys- tem, Hauptman (1998). A such major expansion in the tertiary education system would not only reduce the skill-premiumand lead to a more equitable income distribution, but simul- taneously increase higher labor market incomes and improved living standards. Further- more, an aggressive expansion potentially could propel Brazil into the knowledge economy by revolutionizing the technological capabilities of the labor market force benefiting do- mestic producers. 184 References Arbache, J.S., Green, F. and Dickerson, A. (2000) "A Picture of Wage Inequality and the Allocation of Labor in a Period of Trade Liberalization: The Case of Brazil", Universidadede Brasilia. Arias, O., K.F. Hollack, and W. Sosa (1999). "Individual Heterogeneity in the Returns to Schooling: InstrumentalVariables Quantile Regression UsingTwins Data," Mimeo, university of Illinois Barros, R. and Ramos, L.(1996) "Temporal Evolution of the Relationship between Wages and Educa- tional of Brazilian Men" in Birdsall, N. and Sabot, R.H. (1996) ed."Opportunity Foregone; Edu- cation inBrazil", Inter-American Development Bank, Washington, DC. Birdsall, N. and Sabot, R.H. (1996) ed."Opportunity Foregone; Education in Brazil", Inter-American Development Bank, Washington, DC. Behrman, J.R., Birdsall, N. and Kaplan, R.(1996) "The Quality of Schooling and Labor Markets Out- comes" in Birdsall, N.and Sabot, R.H. (1996) ed."Opportunity Foregone; Education in Brazil", Inter-AmericanDevelopment Bank, Washington, DC. Blom, Verner and Holm-Nielsen (forthcoming). "Education, Earnings and Inequality in Brazil 1982- 1998; Implications for Education Policy", Policy Research Working Paper, The World Bank, Washington D.C. Bound, J. and Johnson, G. "Changes in the Structure of Wages inthe 1980's: An Evaluation of Alter- native Explanations", The American Economic Review, Vol. 82, Issue 3 (June 1992), pp. 371- 392. Card, D. (1998) "The Causal Effect of Education on Earnings", forthcoming in Handbook of Labor Economics, Vol. 3, Ashenfelter, 0.and Card, D.(eds.) Caselli, F. and Coleman, W. J. (2001) "Cross-Country technology Diffusion: The Case of Com- puters", NBER working paper 8130. Deaton, A. (1997) "The Analysis of household Surveys - A Microeconometric Approach to Devel- opment Policy", Johns HopkinsUniversityPress, Baltimore. Ferreira, F. and R. Paes de Barros (1999) "The Slippery Slope: Explaining the Increase in Extreme Poverty in Urban Brazil, 1976-1996", Policy Research Working Paper No. 2210, The World Bank, Washington, DC. Feenstra, R. and Hanson, G. (1997), "Productivity Measurement and the Impact of Trade and Tech- nology on Wages", NBERWorking Paper No. 6052 Gonzaga, M. G. (1996) "The Effect of Openness on Industrial Employment in Brazil" PEA, Serie Seminarios No 27/96. Hauptman, A. (1998) "Accommodating the Growing Demand for Higher Education in Brazil: A role for the Federal Universities?'' Discussion Paper, World Bank Latin America and the Caribbean Regional Office, The World Bank, Washington DC. Katz, L.and K.Murphy (1992) "The Change in Relative Wages 1963-1987; Supply and Demand Fac- tors", Quarterly Journal of Economics, February 1992, Vol. 107. Juhn, C., K.M. Murphy, and Pierce, B. (1993) "Wage Inequality and the Rise in Returns to Skills", Journal of Political Economy, Vol. 101, No. 3. 185 Lam, D. (1999) "Generating Extreme Inequality: Schooling, Earnings, and Intergenerational Trans- missions of Human Capital in South Africa and Brazil", ResearchReport, No. 99-439, Popula- tion Studies Center, University of Michigan. Lam, D. and Schoeni, R. (1993) "The effects of Family Background on Earnings and Returns to Schooling: Evidence from Brazil", Journal of Political Economy, 101(4): 710-740 Lachler, U. (1998) "Education and Earnings Inequality in Mexico", Policy Research Working Paper No. 1949, The World Bank, Washington, DC. Lee, J. (2001), "Education and Technology Readiness: Prospectsfor Developing Countries" Working Paper, Korea University, 2001. Levey, F. and R. J. Murnane, (1992) "U.S. Earnings Levels and Eamings Inequality: A Review of Re- cent Trends and Proposed Explanations", Journal of Economic Literature, Vol. 30, pp. 1333- 1381. Mankiw, N.G., Romer, D., and Weil, D.N. (1992) "A Contribution to the Empirics of Economic Growth", Quarterly Journal of Economics, Vol. 107, No. 2." Mincer, J. (1974) "Schooling, Experience and Earnings", NBERWorking Paper, New York. Murphy, K.M., Riddle, W.C. and Romer, P.M. (1998) "Wages, Skills, and technology in the United States and Canada", NBER working paper 6638. Neri, M.C. and Camargo, C.M.( 1999) "Distributive Effects of Brazilian Structural Reforms", Texto Para Discussao, No. 406, Departamento de Economia, PUC-Eo. Prichett, L. (2000) "An Economist's Midnight Thoughts on Education: The Puzzle of Government Production", Kennedy School of Government. Rosembaum,D.T (2000) "Ability, Educational Ranks, and Labor Market Trends: The Effects of Shifts in the Skill Composition of the Educational Groups", Department of Economics, University of NorthCarolina at Greensboro. Robbins and Gindling(1999), "Trade Liberalization and the Relative Wages for More-Skilled Work- ers inCosta Rica", Review of Development Economics, 3(2), 155-169. Sachs, J.D. and Shartz,H.J.( 1996). "U.S. Trade with Developing Countries and Wage Inequality", AEA PapersandProceedings, Vol. 86 No. 2 Santamaria, M.(2000) "External Trade, Skill, Technology and the Recent Increase of Income Inequal- ity in Colombia", Dissertationpaper. VClez, C.E., J. Leibovich, A. Kugler, C. Bouillon and J. NG5ez (2001) "The Reversal of Inequality Gains in Colombia, 1978-1995: A Combination of Persistent and Fluctuating Forces". Mimeo. World Bank and IDB. Verner, D. (2000) "The Dynamics of Poverty and its Determinants: The Case of Pemambuco and the Northeast Brazil" Vemer, D. (2000) "Wage Determination in Pemambuco, Bahia, CearA and the Northeast: An Applica- tion of Quantile Regressions", The World Bank policy working paper, Washington, DC. Wodon, T.Q. (2000) "Poverty and Policy inthe Latin America and the Caribbean", World Bank Tech- nical Paper No. 467, The World Bank, Washington, DC. 186 Wood, A. (2000) "Globalization and Wage Inequalities: a Synthesis of Three Theories", Department for International Development, London. World Bank(2000a) "Brazil, Secondary Educationin Brazil", The World Bank, Washington, DC. World Bank(2000b) "Brazil, Sector Study Higher Education in Brazil", The World Bank, Washing- ton, DC. 187 (V9058 V9101)*4 + * 8 9 Year 1976 1981 1982 1983 1985 1986 1987 1988 1989 Age (mean) 32.7 32.9 32.9 33.0 32.9 32.8 33.1 33.3 33.3 Femaleparticipationrate 29.5% 32.3% 33.0% 33.6% 34.2% 34.4% 35.2% 35.4% 35.8% Years of schooling (mean) 4.75 5.24 5.25 5.35 5.57 5.63 5.72 5.80 5.92 Hourly wage (mean) 3.03 3.01 3.00 2.56 2.85 3.56 2.85 3.01 3.57 Monthly wage (mean) 550 527 526 448 500 617 489 515 603 Hours worked weekly (mean) 47.7 46.3 46.3 45.8 46.0 45.7 45.5 45.3 44.3 Ruralresidence 23.8% 17.6% 17.7% 18.3% 17.5% 18.1% 18.1% 18.4% 17.5% % livinginNorth 4.4% 6.9% 7.2% 7.2% 8.0% 8.5% 8.8% 8.5% 8.8% % living inNorthEast 20.3% 26.6% 25.9% 26.7% 27.4% 28.3% 28.2% 28.6% 27.9% % livinginCenter-West 16.6% 12.9% 13.2% 13.3% 11.0% 11.9% 12.1% 12.5% 12.9% % livingin South 13.2% 16.9% 16.9% 16.4% 16.7% 15.8% 16.2% 16.1% 16.2% % livinginSouthEast 45.5% 36.7% 36.8% 36.4% 37.0% 35.6% 34.8% 34.3% 34.3% Number of observations 118,282 154,330 167,105 167,713 179,392 101,834 106,750 107,311 109,224 188 Year 1990 1992 1993 1995 1996 1997 1998 1999 Age (mean) 33.5 33.8 34.0 34.3 34.5 33.9 34.9 35.1 Femaleparticipation rate 36.2% 36.8% 36.6% 37.8% 38.0% 41.3% 38.2% 38.6% Years of schooling (mean) 5.99 5.96 6.17 6.33 6.59 6.82 6.82 6.91 Hourly wage (mean) 2.65 2.10 2.79 3.16 3.26 3.17 3.18 2.96 Monthly wage (mean) 444 349 464 522 528 512 520 481 Hours worked weekly (mean) 44.1 44.5 44.1 44.1 44.4 43.7 44.2 43.9 Rural residence 18.0% 15.2% 15.0% 14.5% 14.3% 14.0% 14.7% 15.0% % livingin North 8.7% 6.6% 6.9% 6.8% 7.0% 6.9% 7.2% 7.0% % livingin North East 28.3% 27.2% 27.3% 28.0% 27.5% 27.4% 28.2% 28.7% % livingin Center-West 12.9% 11.3% 10.9% 11.0% 11.0% 11.4% 11.7% 11.6% % livingin South 16.0% 18.5% 18.0% 18.0% 18.4% 17.9% 18.2% 17.9% % living in SouthEast 34.1% 36.4% 36.8% 36.3% 36.2% 36.3% 34.6% 34.8% Number of observations 111,456 111,136 114,969 123,215 119351 106504 125052 128384 Source: PNAD Appendix table 3 Descriptive statisticsby education level 1976-1989 Year 1976 1981 1982 1983 1985 1986 1987 1988 1989 Hourly wage by education level No schooling 1.3 1.3 1.2 1.o 1.1 1.5 1.1 1.1 1.3 Incompleteprimary 2.0 1.9 1.8 1.6 1.7 2.3 1.7 1.7 2.1 Completeprimary 3.2 2.8 2.8 2.3 2.5 3.3 2.5 2.5 3.0 Completelower secondary 5.9 4.3 4.5 3.7 4.0 4.7 3.6 3.8 4.4 Completeupper secondary 6.8 6.3 6.4 5.3 5.6 6.5 5.5 5.6 6.5 Incompletetertiary 8.7 8.8 9.2 7.6 8.2 9.6 8.1 8.9 9.8 Completetertiary 15.5 14.0 14.1 11.6 13.1 15.7 12.3 13.4 15.3 Share of work force innumbers of workers No schooling 28.6% 24.4% 25.3% 23.7% 22.4% 22.0% 21.5% 20.8% 20.3% Incompleteprimary 25.6% 23.5% 22.4% 22.4% 21.2% 21.3% 20.5% 20.6% 19.5% Completeprimary 28.9% 29.9% 29.5% 29.9% 30.4% 29.9% 30.0% 29.5% 30.1% Completelower secondary 6.6% 8.0% 8.2% 8.5% 9.2% 9.5% 9.8% 9.9% 10.3% Completeupper secondary 5.1% 7.4% 7.7% 8.3% 9.3% 9.4% 10.2% 10.6% 11.2% Incompletetertiary 1.5% 2.3% 2.3% 2.3% 2.3% 2.4% 2.5% 2.6% 2.6% Completetertiary 3.8% 4.5% 4.7% 5.0% 5.3% 5.4% 5.7% 6.0% 6.0% Sum 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% Note: Wages are in fixed September 1997 prices. 189 h e n d i x table 3 cont. Descriptive statistics by education level 1976-1989 Year 1990 1992 1993 1995 1996 1997 1998 1999 Hourly wage by education level No schooling 1.0 1.0 1.1 1.2 1.3 1.2 1.2 1.1 Incomplete primary 1.5 1.4 1.6 1.7 1.8 1.8 1.7 1.5 Completeprimary 2.2 1.8 2.2 2.4 2.4 2.3 2.2 2.0 Completelower secondary 3.1 2.6 3.2 3.4 3.4 3.1 3.1 2.9 Completeupper secondary 4.7 3.6 4.6 5.0 5.0 4.8 4.5 4.3 Incomplete tertiary 7.4 4.8 6.7 7.2 7.2 7.1 7.3 6.5 Completetertiary 10.9 6.6 11.5 12.9 12.8 12.0 12.4 11.5 Share of work force innumbers of workers No schooling 19.5% 19.5% 18.2% 17.0% 16.8% 15.6% 15.3% 14.7% Incompleteprimary 19.3% 18.4% 18.3% 17.9% 16.4% 16.4% 16.2% 15.9% Completeprimary 30.1% 30.2% 30.6% 31.1% 30.3% 30.4% 30.5% 30.6% Complete lower secondary 10.5% 11.0% 11.3% 11.6% 13.1% 12.6% 13.1% 13.0% Completeupper secondary 11.7% 11.7% 12.2% 12.9% 13.6% 14.3% 14.7% 15.4% Incomplete tertiary 2.6% 3.0% 2.9% 2.9% 2.9% 3.2% 2.9% 3.0% Completetertiary 6.3% 6.2% 6.4% 6.9% 7.5% 7.4% 7.5% Sum 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% 100.0% -6.7% Note: Wages are in fixed September 1997 prices. Source: PNAD )pply Average differ- Regression Average wage ence inreturns Average wage (fixed demo- Card andLe- (from mincer (fixed demo- graphic groups) mieux (2000) wage regres- graphic groups) Katz andMurphy sion) (1992) To complete upper sec- ondary Primary 0.47 0.44 0.42 0.26 Lower secondary 0.70 0.64 0.7 0.74 Incompletetertiary 0.66 0.65 0.61 0.98 To complete tertiary Primary 0 0 0 0.07 Lower secondary 0 0 0 -0.02 Incomplete tertiary 0.34 0.35 0.39 0.22 Source: PNAD 190 8 2 x I 1 I N 0 9 9 rd t 9 9 N 0 9N slenp!say sienp!satj 8. DISTRIBUICAO TERRAEAS POLfTICASPUBLICAS DE VOLTADAS A 0 MEIORURALBRASILEIRO* Juliano Junqueira Assun@o** Abstract This paper evaluatespublic policies directed towards rural land use and distribution in Brazil, considering both equity and eficiency criteria. Empirical evidence shows that Brazil is a highly urbanized country where the gap between the living conditions in rural and urban areas is persistent, and where overall agriculture productivity remains low. Only one-third of the total area available has been used byfamily-based farms which employ 77%of the rural laborforce. Most of the agricultural land is constituted by large landholdings, with a much lower pe$ormance per hectare -less than half the average productivity of family based farms-, despite having similar levels of investment per hectare. Brazilian land distribution is found to be not only unequal but ineficient. Based on an equilibrium model of the land market -where land has an alternative non-agricultural use induced by impe~ectionsin the marketsfor credit and/or insurance [Stiglitz e Weiss (1981) [Feldstein (1980), Brand60 e Rezende (1992)l- we find that three necessary conditionsfor ineficiency of land distribution hold in Brazil. Namely: (i) heterogeneity in access to agricultural technologies across landholding types; (ii) non- agricultural benefits derived from land property; and (iii) malfunctioning land rental market. These conditions generate an inverse relationship between farm size and productivity. As a result, equity and eficiency criteria in land use become closely linked. Moreover, some additional symptoms of agricultural ineficiency are identified. Based on these results, the paper analyzes the role of four major federal government programs: the land tax (Impost0 Territorial Rural - ITR), the National Program of Land Reform, the Land Bank and the National Program to Strengthen Family Farming (Programa Nacional de Fortalecimento da Agricultura Familiar - PRONAF). Although these programs constitute alternative ways to combat causes and effects of the agricultural ineflciency, many issues of implementation appear to compromise their effectiveness. Finally, the main policy recommendations are: (i) reformulate the land tax (ITR) scheme to prevent the non-agricultural use of land; (ii) expand of land access via improvements in land rental market ;and (iii) incentives to cooperative formation, strengthening small farms, improving their competitiveness in the land market, and facilitating the implementation of group liability programs such as the Land Bank and micro- lending. Resumo 0 artigo tem como objetivo a avaliapio das politicas phblicas que afetam a distribuipio de terras no Brasil, considerando crite'rios de equidade e eficigncia econ8mica. A se@o I apresenta o cendrio brasileiro, examinando os principais indicadores referentes c i urbanizapTo, condiG6es de vida rural e urbana, agricultura brasileira e a distribuig6o de terras. Os indicadores evidenciam umpais eminentemente urbano, onde o atraso das condi@es de vida no meio rural em relac60 ao urbano e'persistente e a agricultura apresenta uma baixa produtividade, com apenas um terGo da drea total destinada ir agricultura familiar, que emprega 77%dos trabalhadores rurais e apresenta um desempenhopor hectare superior ao da agricultura patronal, que ocupa o restante das dreas agricultdveis. A sepio II mostra que a distribuig6o de terras no Brasil, ale'm de injusta, e' ineficiente, o quejustifica uma inteweng6o. * 0 autor agradece os comentirios de FranGois Bourguignon e dos participantes do Workshop de Economia do Tra- bdho e Bem-Estnr Social, e, em especial, a Francisco Feneira e Daniel Santos. 0 tratamento dos dados foi feito pelo ;sistente de pesquisa Daniel Chrity. Qualquer erro remanescentee' de inteira responsabilidade do autor. Doutorando em Economia pela Pontiffcia Universidade Cat6lica do Rio de Janeiro - PUC-Rio e Pesquisador As- sociado do Centro de Desenvolvimento de Planejamento Regional - CEDEPLMUFMG. 195 0 resultado tem como base um modelo tedrico de equilibrio no mercado de terras a partir do qual s60 derivadas e testadas condi@es necessa'riaspara a inejcidncia. Alkm de testar essas condi@es necessa'rias, a se@o II mostra eviddncias de sintomas da inejcidncia. Com base na justificativa da rlecessidade de politicas pu'blicas voltadas ao meio rural, a seG6o III apresenta quatro linhas de a@o do governo brasileiro: o Impost0 Territorial Rural (ITR), o Programa Nacional de Reforma Agra'ria, o Banco da Terra e o Programa Nacional de Fortalecimento da Agricultura Familiar (PRONAF). Por j m , a se@o IV identijlca as principais licdes do cas0 brasileiro tendo em vista o modelo tedrico e os resultados da sepio II e aponta novos caminhos para a a@o do governo - a reformulap.To do esquema de cobraqa do ITR, a ado@io de medidas que visem o aprimoramento do mercado de arrendamentos e a promog?io de cooperativas de produtores. IntrodugPo 0 Brasil, nos 6ltimos anos, tem intensificado significativamente suas politicas de reforma agraria e combate h pobreza rural. Houve uma significativa aceleraqiio do programa de reforma agraria com base em desapropriaqdes e a criaqiio de novos programas que visam o fortalecimento dos pequenos agricultores e outras formas de acesso 2 terra. As razdes que levariam um governo a empreender intervengdes da magnitude observada no Brasil podem ser diversas. 0 debate sobre o tema, no Brasil, tem girado principalmente em torno da justiqa social e da equidade. Dessa forma, as politicas voltadas ?iquestgo fundiaria teriam um impacto direto sobre redugiio da pobreza rural, determinando uma aceleraqgo na acumulaggo de capital e promovendo o desenvolvimento econbmico [Alesina e Rodrik (1994), Persone Tabellini (1994)l. Deininger e Olinto (2000), utilizando um painel de 60 paises, mostram que a desigualdade inicial na distribuiqiio de riqueza, medida em termos da desigualdade da propriedade de terras, tem umimpacto negativo e significativo sobre o crescimento. Emais, as politicas educacionais empaises com grandes desigualdades na distribuiqiio da riqueza tCm umimpacto menor do que nos demais paises. Portanto, segundo os autores, a desigualdade na distribuigHo de terras nZo apenas afeta o crescimento como tambtm reduz a eficacia de politicas de desenvolvimento. Esses argumentos, estreitamente relacionados com o crittrio de equidade ejustiqa social, j6 S ~ O suficientes para que haja um esforqo para a reduqgo da desigualdade de riqueza. As politicas de redistribuiqiio de terras teriam um efeito direto sobre a reduglo da pobreza e um efeito indireto sobre o crescimento econbmico. Entretanto, ao considerar critdrios de eficiCncia econbmica, a relaggo entre as diferentes formas de desigualdade deve ser analisada com mais cuidado. Por exemplo, o coeficiente de Gini da distribuiggo de terras da Australia C de 0.90, mais alto que o brasileiro (0.85), mas a distribuiqiio de renda C bem menos concentrada, com um coeficiente de Gini de 0.41 contra 0.60 para o B r a d em 1990, segundo a base de dados de Deininger e Squire. Ou seja, a concentraqHo fundiaria em si n8o constitui uma justificativa, sob um critCrio de eficihcia da alocaggo de recursos, para a intervenggo no mercado de terras - a Australia e os EstadosUnidos (que possui um coeficiente de Gini semelhante ao Sul e Sudeste brasileiros) ilustram bem esse fato. Nos dois casos ngo ha uma rela@o estreita entre a distribuiqgo de terras e a distribuigHo de renda. 196 0 artigo mostra que os determinantes da desigualdadede terras tambCm deveriam constituir o foco da intervenqiio do govern0 e que politicas de redistribuiqiio de terras podem tambCm ter um efeito direto sobre o crescimento econbmico, aumentando a eficizncia da alocaqiio de recursos. Ao combater as causas da concentraqiio fundiAria, as politicas de desenvolvimento e reduqiio da pobreza rural teriam efeitos sustentfiveis a longo-prazo sobre a melhoria das condiqaes de vida e sobre a eficiCncia da agricultura. Ap6s uma breve descriqiio das condiqdes de vida no campo e da agricultura brasileira, os resultados empiricos justificam a necessidadeda intervenqiio do govemo no mercado de terras. Em seguida, esses resultados e o arcabouqote6rico apresentado seriio utilizados para analisar a sustentabilidade das principais aqdes empreendidaspelo govemo brasileiro, identificando novas Areas de atuaqiio ainda pouco exploradas pela politica pdblica. 0 artigo est6 organizado em quatro seqdes. A seqiio Icaracteriza o ambiente de atuaqiio das politicas pdblicas voltadas B distribuiqiio de terras, iniciando-se com uma descriqiio do process0 de urbanizaqiio no Brasil. Ao longo do sCculo XX, houve uma rApida concentraqgo da populaqiio nos centros urbanos. Esse fato C um reflex0 de disparidades nas oportunidades dos grandes centros urbanos em relaqiio ao meio rural e, como efeito, implica em uma maior dependencia da populaqgo urbana em relaqiio B produqiio de alimentos, sob responsabilidade de uma populaqiio cada vez menor. Dessa forma, a eficiCncia da produqiio agricola toma-se uma questgo relevante na medida que constitui o dnico caminho intemo para o abastecimento das grandes cidades. Em seguida, siio apresentados indicadores sobre as condiqdes de vida no campo. Os dados revelam, de um lado, que houveram avanqos importantes ao longo das dltimas duas dCcadas mas, por outro, o atraso do meio rural em relaqiio ao urbano C persistente e significativo. As informaqdes sobre a agricultura brasileira demonstram que, alCm de importantes difereqas regionais, existe um dualismo na exploraqiio agricola - mais de 4 milhaes de estabelecimentos familiares, que ocupam menos de um terqo da Area total, convivem com meio milhiio de estabelecimentos patronais distribuidos em dois terqos da Area, com um desempenho por hectare inferior. Esse dualismo reflete-se em uma concentraqiio fundi6ria bastante intensa, cuja din8mica apresentaespecificidades espaciais. A caracterizaqiio da distribuiqiio de terras encerra a descriqiio do cenkio brasileiro, em que apenas 20% da populaqiio distribui-se no meio rural, com uma condiqiio de vida inferior e uma produqiio agricola heteroghea, onde os estabelecimentos familiares empregam 76,9% do pessoal ocupado em apenas um terqo da Area total. A ineficiencia da distribui@o de terras brasileira C analisada na seqiio II. modelo te6rico de Um equilibrio com imperfeiqdes no mercado de terras C apresentado para direcionar a anAlise subsequente. 0 modelo determina trCs hip6teses necessArias e suficientes ao resultado de ineficihcia da distribuiqiio de terras: (i)os individuos devem diferenciar-se no acesso Bs tecnologias produtivas; (ii) existir uma demanda por terra para fins niio produtivos; (iii) deve o mercado de arrendamento deve apresentar imperfeiqaes em equilibrio. Em seguida, cada uma dessas hip6teses C testada e verificada para o cas0 brasileiro. E ainda, para reforqar as evidCncias sobre a ineficiCncia, siio apresentados resultados que caracterizam alguns de sew sintomas, como a existhcia de uma rela690 negativa entre lucro por hectare e tamanho de estabelecimento e a comparaqiio de indicadores intemacionais de produtividade. A seqiio 111concentra-senas politicas pdblicas voltadas ao meio rural brasileiro, considerando o Impost0 Territorial Rural (ITR), o Programa Nacional de Reforma Agr6ria com base em desapropriaqdes, o Banco da Terra e o Programa Nacional de Fortalecimento da Agricultura Familiar (PRONAF). As politicas siio descritas sucintamente e os resultados alcanGados siio apresentados. 197 A tiltima se@o conclui o artigo, sintetizando liq6es obtidas com a experigncia brasileira e apontando novos caminhos para as politicas pdblicas voltadas ao meio rural brasileiro. Basicamente, as propostas envolvem uma reformulaqgo da politica de taxasgo de terras, a promoqgo de um mercado de arrendamento mais eficiente e o incentivo h formaq3o de cooperativas agricolas. 1- Cenitrio Brasileiro 0 Brasil, ao longo do stculo XX, apresentou um processo de urbanizaqgo da populaqgo acelerado. Apesar de sua extensgo territorial e da disponibilidade de recursos agricultiveis, o B r a d atualmente encontra-se entre os 30 paises mais urbanizados do mundo. Entretanto, a produtividade agricola brasileira n3o condiz com aquela observada nos paises com essa caracteristica. Na AustrAlia, que possui uma extensgo territorial e um nivel de urbanizaqlo compar6veis aos do Brasil, a produtividade agricola C 7 vezes maior, mesmo com uma disponibilidade maior de terra agricultAve1per capita. A figura 1.1 mostra a velocidade com que ocorreu a migraqgo da populaqgo rural para as Areas urbanas ao longo do stculo. 0 movimento C tgo forte que a populaqiio rural, a partir da dCcada de 70, sofre uma queda em termos absolutos. 0 Brasil, que havia entrado no sCculo XX com mais de 80% da populaqlo em 6reas ruraisgO,chega ao ano 2000 com o quadro completamente revertido, ou seja, 80% da populaqiio mora em Areas urbanas (figura 1.2). Entretanto, o processo ocorreu de forma bastante heteroghea no interior do Brasil. A figura 1.3 ilustra a magnitude dessas diferenqas. 0 estados brasileiros, ordenados pela taxa de urbanizaqiio de 2000, apresentam, em 1991, a mesma variaqgo observada na mCdia nacional nos dltimos 60 anos. Umviajante que percorresse todos os estados brasileiros, em 1991, iria testemunhar uma variabilidade no grau de urbanizaqiio equivalente B da mtdia nacional no period0 de 1940 a 2000. E ainda, a evoluqiio do processo entre 1991 e 2000 tambtm ngo se deu uniformemente. Pode-se perceber a importfincia das diferenqas regionais no Brasil. Figura 1.1 EvolugPo da PopulagPo (1940-2000) - Fonte: IBGE- Instituto Brasileirode Geografiae Estatistica Notas: 1940, 1950-popula@oresidente; 2000 - dadospreliminares Segundo Cardoso (1975), a taxa de urbaniza@odo Brasil no inicio do stculo XX era de 17.28%. 198 Figura 1.2 EvolucPoda Taxa de Urbanizaciio(1940-2000) - - a . 100% 90% I I I ! I 20% 4 I I I I I I I Figura 1.3 -Taxasde UrbanizagPoEstatuais(1991,2000) 100% 90% 5i 80% .-B 55 70% 60% I- 8 50% 40% 30% 1- 1991 -2000] Fonte:IBGE-InstitutoBrasileiro de Geografiae Estatfstica. Nota:2000-dados preliminares A anilise dos indicadores referentes ?icondiStio de vida revela, de um lado, uma melhoria significativa de virios aspectos da vida rural brasileira nos liltimos anos e, por outro lado, que ainda existem discrepiincias importantes em relaqtio aos indicadores urbanos. E mais, as variaqaes inter-regionais siio bastante significativas. A tabela 1.1 apresenta uma comparaqtio entre indicadores urbanos e rurais nos anos de 1981 e 1999, considerando dados da PNAD (PesquisaNacional por Amostra de Domicilios). Os dados niio consideram a regiiio Norte, cuja zona rural ntio C coberta. Pode-seperceber que a populaqtio que vive nas Leas urbanas C mais idosa do que na zona rural. Entre 1981 e 1999, houve um envelhecimento da populaqiio, o que pode indica uma melhoria nas condiqaes de vida no period0 alCm de mudanqas demogrGficas, mas o percentual de idosos ainda e' menor que a metade da mCdiados paises de alta renda segundo a classificaqiio do Banco Mundial (14.1% em 1998). 199 Diferengas dramaticas silo encontradas no que diz respeito B educagiio. Mesmo com os avanqos ocorridos no periodo, a educagilo nas Areas rurais apresenta uma defasagem enorme em relaqilo Bs Areas urbanas, que j6 estilo bastante aquCm de paises com renda per capita semelhantes ao Brasil". E as diferengas regionais silo ainda maiores. Enquanto as taxas de analfabetismo das Areas rurais de estados da regiiio nordeste superam 25% (Piaui, CearA, Paraiba, Pernambuco, Alagoas, Sergipe e Bahia), em Estados do sul, a mCdia C inferior a 9%. A mCdia de anos de estudo da populagiio com mais de 25 anos apresenta caracteristicas semelhantes. Em alguns estados do Nordeste o indicador niio chega B metade da mCdia dos estados do Sul. Associado a essas diferenqas nos niveis de educagilo rural e urbano, pode-se observar uma variaGilo significativa no valor dos rendimentos mensais. 0 percentual de domicilios com rendimento total inferior a 1 salArio m'nimo por m&sC bem menor nas Areas urbanas. Essas disparidades mantiveram-se praticamente inalteradas ao longo do periodo analisado, contrariamente ao que ocorreu com a diferenga nos salirios mCdios rural e urbano, que aumentou muito. As diferenqas regionais tambCm silo, do mesmo modo verificado em outros indicadores, bastante significativas. Em alguns Estados (Sergipe, Paraiba, Rio de Janeiro e CearA) a renda urbanaC maior do que 2.5 vezes a renda rural. Tabela 1.1- Caracteristicas da PopulaqBo,Educaqiio e Renda segundo a Area de Moradia 25.43 27.30 29.30 4.36 6.23 6.32 9.91 18.78 6.91 5.11 2.76 6.72 75290.03 259.00 587.24 (1.34) (1.o> (2.26) 7.05 15.59 4.74 D)- 1981e Notas: (i) A taxa de analfabetismoconsideraapenas as pessoas com mais de 15 anos. (ii) A mCdiade anosde estudo considera apenas as pessoas commais de 25 anos. (iii) 0valor do salfiriominimode 1981foicorrigido paraconsiderar adesvaloriza@oreal do perfodo (Cr$7397,00). (iv) Os indicadores n8o incluema regib Nortedo Brasil que n5.o 6 tratada pela PNAD. 91Os quatro paises com renda per capita mais pr6ximas B do B r a d apresentam, aproximadamente, as se- guintes taxas de analfabetismo: Chile (4.5%), Croicia (2%), Tnndade e Tobago (6.5%) e Uruguai (2.5%). 200 rbanos 4.39 4.07 3.62 3.35 3.18 2.98 95.29 87.84 98.21 75.51 51.78 94.04 91.46 75.28 99.09 95.52 97.31 99.27 70.18 52.43 89.95 L e 1999 A tabela 1.2 mostra as condiq6es dos domicilios rurais e urbanos nos anos de 1981e 1999. Os indicadores mostram uma persistente defasagem entre a situaqBo dos domicilios rurais em relaqBo aos urbanos. Mesmo havendo uma atenuaqiio ao longo das dltimas duas dkcadas, algumas diferenqas ainda sBo muitograndes, principalmente ao nivel estadual. Os domicilios rurais SBO constituidos por um ndmero de pessoas maior do que os domicilios urbanos. Entre 1981e 1999 observou-seuma queda desse indicador, com maior intensidade nas keas rurais. E ainda, o percentual de maiores de 10 anos em cada domicilio apresentou o mesmo movimento. Esses resultados, condicionados nas demais caracten'sticas das famiias, podem indicar uma melhoria na capacidadede geraqBo de renda dos domicilios. Os indicadores apresentados na tabela 1.2, exceto o percentual de domicflios com fogBo, mostram que a melhoria observada no pen'odo encontra-se ainda longe de equiparar os domicilios rurais e urbanos. E a maior precariedade na vida do campo se distribui de forma bastanteheterogheanas diferentes unidades da federaqBo. Em 1999, apenas 41% dos domicilios do estado do MaranhBo possuiam parede durAvel. E um avanqo quando comparado aos 15% observados em 1981, mas encontra-se num patamar bastante inferior aos 80% referentes 21 kea urbana no MaranhBo em 1999 e aos 73% da mCdia nacional para a k e a rural em 1981. 0 percentual de domicilios com iluminaqBo elCtrica apresentaumgrande salto no periodo. Os dados de 1999 mostram que um terqo dos domicilios rurais ainda nBo possuem iluminaqgo elCtrica contra os 3% da Area urbana. Nos Estados do Piaui, Bahia e Tocantins, esse indicador engloba a metade dos domicflios. E, em 1981, a situaqBo era muito pior com a mCdia nacional dos domicflios rurais com iluminaqBo elktrica em torno de 25% e alguns estados como o Piaui, CearA e ParanA com menos de 10%. Os indicadores de Agua canalizada e geladeira apresentam movimentos semelhantes. 0 finico indicador a apresentar uma converghcia entre a zona rural e urbana C o percentual de domicflios com foggo. 201 Do mesmo modo que os indicadores de condi@o de vida, a agricultura brasileira apresentauma enonne diversidade espacial. Esse fato decorre de diferenqas naturais, sociais, culturais e econ6micas e C ilustrado pela figura 1.4. Figura 1.4 Mapas da Agricultura Brasileira - (a) Lucropor Estabelecimento (b) Lucropor Hectare 4170 IO 7850 (110) 2380 IO 4170 1113) (c) Area Mtdia dos Estabelecimentos (d) PessoalOcupado por Hectare 0 I54 10 I 9 7 3 (119) 0095 to0 I 5 4 (103) 0061 lo0095 (111) 0031 lo0061 104 (e) N6mero de Tratores por Trabalhador (f) % Estab. c/ Adubos/Corretivos 0 1283 80 09651 11101 00'43 la01283 I I I 2 l 00142 ,000543 11121 202 (g) % Estab. com Irrigaqiio (h) % Estab. c/ Assistencia TCcnica Fonte: IBGE- Censo Agropecutkio Quanto ao desempenho econ8mico, existe uma Clara diferenciaqilo regional. 0 lucro por estabelecimento (figura 1.4a) C maior nas Areas do Centro-Oeste, Trihgulo Mineiro e oeste paulista e outras Areas isoladas como o extremo sul. Entretanto, ao considerar o lucro por hectare (figura 1.4b), que C uma medida de intensidade, o retrato C bastante diferente. Os estabelecimentos do Centro-Oeste e Triiingulo Mineiro caracterizam-se por uma exploraqiio lucrativa, mas pouco intensa, de grandes keas, utilizando-se da mecanizaqiio em larga escala. A figura 1.4 mostra que as Areas com os maiores lucros por hectare caracterizam-se por um grande ndmero de trabalhadores por hectare, o que permite uma exploraqilo mais intensa, com ummaior valor adicionado por hectare. E uma parcela significativa dessas fireas caracterizam- se pela agriculturafamiliar, cujo padrilo de produqiio agricola C mais intensivo em milo-de-obra. AlCm disso, observa-se que o us0 de adubos e corretivos, imgaqiio e assistencia tCcnica silo restritos a apenas algumas regi6es. Algumas keas do Norte e Nordeste, principalmente, nilo t&macesso a nenhumdessesinsumos. Em2000, o projeto de Cooperaqilo TCcnicaINCRA/FAO divulgou umrelat6rio que evidencia a qualidade e eficiCncia dos estabelecimentos caracterizados como familiares em relaqiio aos denominados patronais [veja Guanziroli e Cardim (2000)l. 0 relatdrio contCm uma caracterizaqiio bastante abrangente da atividade agrfcola dos produtores familiares. A seguir serilo apresentadas algumas estatisticas que diferenciam os agricultores familiares dos patronais. Foram considerados estabelecimentos familiares aqueles cuja direqilo dos trabalhos C do produtor, a milo-de-obra C predominantemente familiar e a Area total do estabelecimento niio ultrapassaumteto regionalg2. A tabela 1.3 apresenta uma comparaqiio entre diversos indicadores da agricultura familiar e da patronal. Emlinhas gerais, a agricultura familiar constitui a forma de organizaqiio de 85,2% dos estabelecimentosbrasileiros, o que representa apenas 30,5% da Area total. Entretanto, apesar de ocupar nilo mais do que um terqo das terras, C responsAve1 por quase 38% da produqiio, ocupando 76,9% dos trabalhadores rurais. Enquanto o desempenho por estabelecimento dos patronais C superior, os indicadores por hectare silo sempre inferiores aos da agricultura familiar, o que decorre da maior intensidade da exploraqilo familiar, como mencionado anteriormente. Tanto o investimento quanto a renda total apresentam esse padrilo, ou seja, a lucratividade por hectare dos estabelecimentos familiares (que siio menores em Area) C bem mais alta que a daqueles patronais, o que C compativel com o modelo da seqiio seguinte. 921.222 hapara a regifio Norte, 694,5 hapara o Nordeste, 384 ha para o Sudeste, 280,5 ha para o SUI, e 769,5 ha parao Centro-Oeste. 203 Dessa forma, pode-se concluir que a agricultura brasileira caracteriza-se n5o apenas por uma enorme desigualdade espacial como tambCm por um dualism0 na organizaq5o produtiva dos estabelecimentos. Grandes propriedades, caracterizadas por uma exploraqb mais extensiva, convivem com pequenas propriedades familiares, que operam suas ireas de forma mais 11 554.501 (11,4%) 107.768 (303%) 240.042 (67,9%) 18.117.725 (37,9%) 29.139.850 (61,0%) 937.828 (25,3%) 2.735.276 (73,8%) 7 3 673 13.780.201(76,9%) __- 2.535.459 5.108.372 612,5 9.212,6 23,5 21,3 2.717 19.085 104 44 0 Brasil possui apresentaumadas maioresconcentraqdesde terra do mundo. Deacordo como Censo Agropecuirio, o coeficiente de Gini da distribui@o de terras brasileira era 0,856 em 1995. A se@o IIdo artigo iri mostrar que existe uma ineficihcia por tris dessa concentraqgo para o cas0 brasileiro. A seguir serSio apresentadas algumas informaqaes que caracterizam a evoluqgo e a distribuiqgo espacial da concentraqgofundiiria brasileira. 204 Figura 1.5 EvoluqHo do Coeficiente de Gini da DistribuiqHo de Terras - I I 1 `2 0.94 .- 0.88 0.92 0.87 u 0.9 4 0.86 .- 0.88 0.85 14 4 1 Q 0.86 0.84 % 0.84 * 0.82 0.83 0.8 f* 0.82 1 I I I I I I 1950 1960 1970 1975 1980 1985 1995 1950 1960 1970 1975 1980 1985 1995 -BBRASlL - NORTE -BRASlL - NORDESTE .-= .- 0.88 - I B 0.9 5 0.86 - 1 Q 0.88 `0 4 0.86 0.82 - .-25 Q I c0 0.8 0.84 `5 0.76 5 0.78 8 0.82 .- c 0.74 0.8 0.72 I I I 0.78 I I * 0.7j ~ 1950 1960 1970 1975 1980 1985 1995 1950 1960 1970 1975 1980 1985 1995 II I-BRASIL 1 1-BRASlL -SUDESTE I L -CENTRO-OESTE I I / I 1 .- 0.9 0.85 0.855 2 U 0.8 0.85 1.5 .- r ' 0.845 0.75 0.84 1 0.7 0.835 I I 0.5 0.65 0.83 , II I I , 0 1950 1960 1970 1975 1980 1985 1995 - - 1950 1960 1970 1975 1980 1985 1995 BRASIL SUL I-GINI -LOG(INFLACAO MEDIA DO ANO) 1 Fonte: IBGE- CensoAgropecuirio A figura 1.5 apresenta a evoluqiio do coeficiente de Gini da distribuiqiio de terras, de 1950 a 1995, para as cinco regides brasileiras. As diferenqas observadas entre as regiks ocorrem niio apenas em termos dos niveis de concentraqiio como tambCm na dingmica do periodo. A distribuiqb de terras nas regides Sul e Sudeste siio uniformemente menos concentradas que a mCdia brasileira, com dingmicas semelhantes. N o Nordeste, por outro lado, a dingmica aproxima-se da mCdianacional mas com uma concentraqiio maior. As maiores mudanqas siio observadas nas regides Norte e Centro-Oeste, que apresentam uma grande desconcentraqiio entre 1960 e 1995, o que decorre, principalmente, da expansiio da fronteira agricola. Outro fato interessante C a relaqiio entre a evoluqiio do coeficiente de Gini e a inflaqiio. A figura 1.5 mostra que, com o aumento da inflaqiio, observa-se um aumento na concentraqiio. Entretanto, a reduqiio da inflaqiio ocorrida em 1995 niio se associa a uma desconcentraqiiotiio significativa. 205 A figura 1.6 mostra as diferenfas observadas na concentraqgo fundiiiria ao longo do tenitbrio brasileirog3. AlCm de haver diferentes dinhicas de concentraqtio de terras entre as regides brasileira, como mostra a figura 1.5, h6 uma enorme variaf8o espacial na concentraqiio da distribuifiio da terra emumdeterminado momento do tempo. Figura1.6 Mapado Coeficiente de Gini da DistribuigBode Terras em 1995 - Leeend 0.87 to 0.93 (117) 0.851 to 0.87 (106) 0.829 to 0.851 (110) 0.792 to 0.829 (106) L 0.644 to 0.792 (119) I I Fonte: IBGE Censo Agropecuivio - A comparaf8o intemacional do coeficiente de Gini da distribuiqiio de terras revela tambCm grandes variafdes mas, principalmente, mostra que n8o h6 uma relaf8o Claraentre concentraf8o fundi6ria e eficiCncia da agricultura. A tabela 1.4 mostra que paises como a Argentina, AustrBlia, Ithlia, Espanha e Estados Unidos apresentam indices de Gini semelhantes ao do Brasil, mas uma agricultura muito mais produtiva. Os demais paises latinos mencionados na tabela 1.4 apresentamtanto um coeficiente de Gini semelhante ao do Brasil como tambCm uma agricultura de baixa produtividade. Do mesmo modo, existem tanto paises com baixa produtividade agricola (Bangladesh, fndia) quanto paises de agricultura desenvolvida (Alemanha, Reuno Unido, Japgo) que apresentam coeficientes de Gini inferiores ao brasileiro. A seqiio IIdo artigo mostra que a questgo mais importante, no que diz respeito 2 eficiencia econbmica, siio as razdes que levam 2 concentraqBo fundi6ria e, portanto, justificam uma intervenf8o no mercado de terras. y30 coeficiente da distribuiGlo de terra para cada microrregigo foi calculado apartir de umaC u m de Lorenz com6 pontos. Apesar dessa aproximaGHo, os valores calculados slo compativeis com as midias estaduais calculadas a partir dos microdados e servem para mostrar a variaclo espacial da concentraGHo de terras. 206 Tabela 1.4 - la 1listribuiciio de terra^^^ 1988 0.850 1990 0.903 1977 0.419 1989 0.768 1995 0.856 1990 0.774 1990 0.667 1985 0.592 1990 0.739 1990 0.382 1991 0.784 1972 0.911 1989 0.858 1990 0.62 1 1987 0.754 1980 0.803 1971 0.910 2- A IneficiCncia da Distribuiqiio de Terras: aspectos tedricose empiricos 2Z- C'n,sz2era@e,s Te&rica,s Antes de proceder h analise do desempenho dos programas financiados pel0 govern0 brasileiro para o desenvolvimento agririo sera apresentada uma estrutura te6rica que serviri como pano de fundo, delimitando o espaqo de atuaqiio para uma intervenqiio governamental. Essa estrutura tedrica ira concentrar-se numa relaqiio entre equidade e eficigncia da distribuiqlo de terras derivada a partir de imperfeigdes de mercado. Considere uma economia ficticia em que os individuos distribuem-se em duas classes distintas. Uma proporqlo da populagiio, os agricultores, dedica-se inteiramente h produg80 agricola. Para esse grupo, a terra tem como linica finalidade a agricultura. A outra porqiio da populagiio, os empresarios, alCm do acesso B mesma tecnologia de produqgo agricola, demandam terras para outros fins. A literatura tem sugerido tres finalidades bisicas. A demandapor terras para fins de colateral C resultado de imperfeigdes no mercado de crCdito [Stiglitz e Weiss (1981)l e da necessidade desses empresarios de financiar outras atividades. 0 segundo us0 niio agricola da terra resulta da incapacidade do mercado financeiro oferecer um ativo financeiro que reproduza com perfeiqgo as qualidades da terra. Dessa forma, a propriedade de terras torna-se parte da carteira de investimentos dos empresirios ou C utilizada como um mecanismo de seguro contra instabilidades macroeconBmicas [Feldstein (1980), Brandgo e Rezende (1992)l. Finalmente, a propriedade de terras pode gerar beneficios individuais aos empresarios, seja como um forma de exploraggo politica, fonte de prestigio local, ou como um meio de acesso a subsidios governamentais diretos [Deininger e Feder (1998)l. y4Os coeficientes de Gini da distribui@o de terra de diversos paises foram gentilmente cedidos pelo pesquisador PedroOlinto, sendo provenientesde fontes diversas. 207 Impe$ei@o no Mercado de Crkdito Suponha que o empresirio, alCm da produqiio agricola, tenha acesso a uma tecnologia de produqiio industrial. Tanto a atividade industrial quanto a atividade agricola utilizam-se de um contingente de trabalhadores que n b sera tratado explicitamente no modelo. Entretanto, os resultados qualitativos niio dependemdessa hip6tese. Os individuos (agricultores e empreshrios) gastam sua renda com o consumo de produtos agricolas e industrializados, vendidos e produzidos em um mercado perfeitamente competitivo. A renda de cadaindividuo provCm apenas do lucro das atividades empreendidas. A tecnologia agricola disponivel para agricultores e empresirios C idCntica, com a quantidade produzida dependente explicitamente apenas da Area cultivada. 0 custo da atividade agricola envolve o custo da terra e o custo da Area cultivada. 0 custo por hectare de area cultivada niio depende da Area cultivada total e C detenninada exogenamente por uma tecnologia de proporq6es fixas e pelos preqos dos demais fatores necesskios h produqiio agricola. Ou seja, cada hectare cultivado C formado pel0 emprego de proporqdes fixas de outros insumos como trabalho, tratores, arados, fertilizantes, irrigaqiio, etc., cujos preqos estb detednados exogenamente. Suponha ainda que n5o hA mercado de arrendamentog5e, portanto, a area cultivada por cadaindividuo n5o pode superar o tamanho do estabelecimento. Sup6e-se que niio existam dificuldades para o financiamento da safra. A produqiio industrial depende do emprego de trabalhadores, que silo supostamenteabundantes a um salkio fixo. Entretanto, existe um mercado de cridito imperfeito com o qual os empresfirios se deparam para financiar a compra de insumos para a produqgo industrial. Uma motivaqiio para essa imperfeiqiio C a assimetria de informaqiio na relaq8o entre tomador e emprestador [Stiglitz e Weiss (19Sl)I. Como resultado dessas imperfeiqdes, o total de recursos disponiveis para financiar a atividade industrial ir8 depender da propriedade de terras, utilizada como colateral. Dessa forma, os empresarios demandam terras para a produqiio agricola e para o us0 como colateral para o financiamento da atividade industrial. Os agricultores, por outro lado, utilizam a terra apenas para a produqBo. Para que os efeitos realmente ocorram em qualquer equilibrio, serA suposto que a restriqiio de racionamento no mercado de crCdito para os empresarios. A restriqiio imposta pel0 mercado de arrendamento C sempre ativa para os produtores, uma vez que s6 adquirem terras para o cultivo. No equilibrio competitivo descentralizado, os individuos determinam a produqiio em cada atividade e sua demanda por produtos agricolas e industriais considerando como dados os preqos da economia. Uma vez que o custo de cultivo de cada hectare e os salarios industriais siio determinados exogenamente, existem trCs mercados a serem equilibrados: o mercado de produtos agricolas, o mercado de produtos industrializados e o mercado de terras. Sob a hip6tese de que a restriqiio de racionamento de crCdito C ativa para os empresarios e sabendo que a restriqlo colocada pel0 mercado de arrendamento C ativa para os agricultores, a soluq8o descentralizada difere-se da soluqiio determinada por um planejador central que maximiza o bem-estar social, quando a produq2o agricola tem um peso social adicional. A queda na produq5o agricola C compensada por um aumento na produ@o industrial e entiio C necessariauma ponderaq5odiferenciada para a produqgo agricola. Esse peso adicional pode ser motivado pela necessidadede uma oferta minima de alimentos para abastecer os trabalhadores que n2o est20 tratados explicitamente no modelo. Ou pelas externalidades negativas impostas por taxas de urbanizaqgo muito altas. y5 Para osresultadosobtidos, basta que o mercado de arrendamentoseja imperfeito. Ou seja, a urn dado preqo, a de- mandapor arrendamento de terras 6 maior do que a oferta [Otsuka, Chuma e Hayami(1992)l. 208 Nessa versiio em que a terra C utilizada como colateral para a atividade industrial, os presos de equilibrio traduzem toda a escassez relativa da economia. 0 us0 de terra como colateral faz com que os empresfirios tenham terra para financiar a atividade industrial em cada periodo. 0 efeito da produsiio industrial sobre a redusgo na produqfio agricola C resolvido pelos preqos em equilibrio. 0 mecanismo da ineficisncia C o seguinte. A demandapor terras para colateral faz com que o preqo de equilibrio das terras seja aumentado o que, diante de agricultores com potencial produtivo restringido pel0 mal funcionamento do mercado de arrendamento, reduz a quantidade total produzida do bem agricola, que tem umpeso adicional na funqgo de bem-estar social. Esse custo social do us0 especulativo da terra n b C internalizado pelos empresfirios em was decisdes. A introduqiio de umplanejador nessaeconomia, sob essas hipbteses, pode levar a uma melhoria de Pareto. 0 planejador, ao determinar as alocaGdes btimas, considera o efeito do us0 da terra como colateral sobre a produqBo dos agricultores em equilibrio. As formas de atuaqso desse planejador central serHo tratadas abaixo. Impe$ei@o no Mercado Financeiro Se os empresiirios demandam terras pel0 fato de que o mercado financeiro niio oferece umativo adequado, o efeito sobre a eficiencia irfidepender da natureza dessa demanda. A necessidadede uma valorizaqBo diferenciada da atividade agricola depende se o us0 final dessas terras constitui ou n b umbemcomercializado no mercado. Caso essa demanda seja motivada pel0 desejo dos empresiirios em proteger sua renda contra instabilidades macroeconamicas apenas, sem que promova o aumento na produqiio de qualquer bem comercializado no mercado, o resultado de ineficiencia C verificado mesmo com uma funqgo de bem-estar utilitarista tradicional, em que ngo hh uma valorizas8o extra da produqiio agricola. Os empreshrios, ao demandar terras para esses fins, elevam o pres0 da terra e, em equilibrio, restringem a produqiio agricola. Uma vez que n8o consideram esse efeito ao tomar was decisdes em um mercado competitivo, acabampor manter grandes glebas de terra, o que C socialmente ineficiente. Por outro lado, se a demanda decorre da necessidade de diversificasiio de risco com outra atividade produtiva, a diferenciaqgo da produqiio ap'cola torna-se necessfiria. Assim como na versiio em que terra era utilizada para colateral, a queda na produqzo agricola C compensada pel0 aumento na produqBo da outra atividade. E, uma vez que essa outra atividade C comercializada no mercado, o sistema de presos C capaz de promover a alocaqiio eficiente dos recursos, exceto se a agricultura tiver uma importincia destacada. Beneficios Individuais Se a demanda por terras para fins n5o agricolas C gerada pel0 valor da propriedade da terra como mecanismo de promosb social, controle politico ou qualquer outro tip0 de beneficio individual, o resultado de inefici6ncia mantCm-se, independente do peso social adicional da produq8o agricola. Ao desfrutar desses beneficios individuais os empresfirios determinam uma demanda que pressiona para cima o preGo da terra, restringindo o acesso dos agricultores e a produsiio agricola. Implica@es para a Poli'tica EconBmica A solusiio descentralizada de todas as versdes apresentadas acima exibe uma desigualdade excessiva na distribuiqgo de terras. As imperfeisdes de mercado determinam que os estabelecimentos agricolas dos empresfirios sejam excessivamente extensos enquanto os lotes cultivados pelos agricultores s20 excessivamente pequenos, ou seja, hfi uma desigualdade na distribuiQBo de terras que C ineficiente. Portanto, na presensa de imperfeis6es de mercado, C 209 possivel a associaqiio entre os critCrios de equidade e eficiCncia da distribuiqiio de terras, em que a desigualdadepassa a ser ineficiente. Resumindo, essas consideraqdeste6ricas determinam algumas condiqaes necess6riaspara que a distribuiq5o de terras seja ineficiente: (i)os individuos devem diferenciar-se no acesso 2s tecnologias produtivas; (ii) existir uma demanda niio-produtiva por terras; (iii) mercado deve o de arrendamento deve apresentarimperfeiqdes em equilibrio. Se qualquer dessas hip6teses n5o se verificar, o resultado de ineficihcia deixa de ser v6lido. A necessidade da hip6tese (i) C 6bvia. Cas0 todos os agentes sejam idCnticos, n5o haver6 alguns se beneficiando emdetriment0 de outros, sem que haja espaqo para transfersncias mais eficientes. Se n50 houver demandapor terra n2o-produtiva, o preqo da terra ir6 representar o seu valor como insumo produtivo, n5o havendo ganhos de bem-estar com a introduq5o de um planejador central. E por fim, se o mercado de arrendamento for perfeito, a propriedade da terra toma-se irrelevante do ponto de vista de eficiCncia, uma vez que o acesso 2 terra est6 garantido pel0 arrendamento. Ap6s a caracterizaq5o da relaq5o entre desigualdade da distribui@o de terras e ineficiencia agricola, o passo seguinte C a discuss50 das possiveis formas de intervenq5o. De fato, no modelo descrito acima, hB uma indeterminaqiio sobre a melhor forma de intervenqiio do governo. Qualquer politica econdmica que promova o fortalecimento dos pequenos produtores agricolas, iniba o us0 de terras para fins niio produtivos, e/ou transfira terras diretamente de grandes para pequenos proprietfirios aponta na dire@o de uma distribuiG2o de terras mais eficiente. 2.2- Condi@esNecessiriasao Resultadode Ineficicncia :evidbcia empirica A estrutura te6rica acima determinou um conjunto de hip6teses necess6riaspara que haja uma racionalidade de intervenq5o no mercado agrfirio, utilizando o critCrio de eficiCncia econdmica. E mais, na presenqa das imperfeiqdes de mercado mencionadas, uma distribuis5o mais equitativa dos estabelecimentos rurais iria aumentar a produqiio agricola, o que relaciona os conceitos de eficiCncia e equidade. A seguir seriio apresentadas evidCncias de que essas hip6teses s50 verificadas no cas0 brasileiro e que a produq5o agricola encontra-senumpatamar inferior ao que poderia ser. Urbanizaqiio e a Importfincia do Setor Agricola Os dados apresentados na seqiio Imostram que, ao longo do sCculo XX, houve uma mudanqa dr6stica no padriio de urbanizaqiio brasileiro. A partir da dCcada de 30, o desenvolvimento da economia encontrou-se focalizado nos setores urbanos e industriais, baseando-se em uma industrializaqgo voltada para a substituiqiio de importaqdes, e numa urbanizaqiio acelerada pela intensificaqiio do Cxodorural. 0 Cxodo rural, apesar de acompanhado pel0 empobrecimento dos camponeses, refletia o progress0 das tCcnicas agn'colas, o qual permitia um trabalhador rural sustentar, em mCdia, um nlimero maior de habitantes que niio est50 envolvidos na atividade agricola. Esse movimento migrat6rio se intensificou a partir da Grande Depressiio. Combinado com o crescimento vegetativo da populaqiio urbana, o Cxodo rural mostrou-se mais do que suficiente para atender 21 demandapor miio-de-obra das atividades urbanas. Dessa forma, criou-se condiqdes para o atendimento da hip6tese (i)do resultado de ineficiencia da distribuiq5o de terras. De um lado, o empresariado urbano, com acesso a atividades industriais e diante de um mercado de crCdito imperfeito, tem na propriedade da terra uma oportunidade niio apenas produtiva que diversifica sua carteira de investimentos, mas tambCm um 6timo colateral e um mecanismo de proteqiio contra as instabilidades da economia brasileira. Por outro lado, os agricultores remanescentes no meio rural tem seus rendimentos 210 provenientes inteiramente da atividade agricola. De fato, segundo dados da PNAD, 70% da populaqiio rural economicamente ativa est6 envolvida em atividades agricolas enquanto para a populaqiio urbana esse indicador C apenas 7%. E mais, o elevado grau de urbanizaq2o determina uma valorizaq5o social adicional para a produqlo agricola. Impe$eiGdes de Mercado e Demandapor Terra A segunda hipdtese necesshria para o resultado de ineficigncia da distribuisgo de terras C a existbncia de uma demanda por terras para fins n2o produtivos. Os testes empiricos dessa hipdtese iriio considerar as series de preqos mCdios de venda e arrendamento de terras agricolas, divulgadas semestralmente pel0 Centro de Estudos Agricolas do Instituto Brasileiro de Economia (IBRE) da FundaGBo Getdlio Vargas (FGV). Desde de 1966, o CEA/IBRE/FGV faz o levantamento dos preCos de terras agricolas atravCs de entidades vinculadas Bs Empresas Estaduais de Assistbncia TCcnica e Extensgo Rural, que conta com mais de 3600 escritdrios espalhados por v6rios municipios do Brasil. Na Bahia, os dados s50 fornecidos pelos tCcnicos da Comissiio Executiva do Plano da Lavoura Cacaueira e, em SBo Paulo, pel0 Instituto de Economia Ap'cola da Secretaria da Agricultura, referindo-se apenas B mCdiado Estado. Os preqos s50 coletados nos municipios, ao final de cada semestre, e agregadospara microrregides, unidades da federaqb, grandes regides e Brasil. As figuras 2.1 e 2.2 mostram a evoluG8o dos preGos de venda e arrendamento de terras para pastagem (plantadas e naturais) e lavoura (permanente e temporfiria), no periodo de 1966 a 2000. A dCcada de 70 apresenta, em geral, preqos mais altos daqueles que vigoravam na segundametade da dCcada de 60. Emseguida, h6 umperiodo de oscilagdes fortes entre meados dos anos 80 e meados dos anos 90. E, por fim, pode-se verificar uma tendCncia de queda a partir da estabilizaqlo da economia brasileira, no final da dCcada de 90. Os movimentos dos preCos verificados nas sines podem ser explicados pel0 ambiente macroeconamico de cada periodo. Na dCcada de 70, o aumento dos preGos reflete o crescimento observado na agricultura brasileira. A intensificac$io da urbanizasiio, associado ao crescimento da renda per capita, determinaram uma ampliaGlo do mercado interno de alimentos, especialmente dos produtos derivado da pecu6ria. 0 final da dCcada de 60 e o comeqo da dCcada de 70 marcaram tambCm uma revers50 da agricultura brasileira na direqiio de mercados extemos. Enquanto as lavouras de exportaqiio tiveram o seu crescimento acelerado, a produq8o de culturas alimentares teve o seu crescimento reduzido - a dnica exceCBo foi o trigo. Durante o periodo, a diversificaqb de culturas e a rtipida expansso da fronteira ap'cola foram facilitadas pela introduqgo de cridito rural subsidiado para a compra de insumos modernos, investimentos governamentais em infra-estrutura (principalmente estradas para as ireas de fronteira) e em pesquisaagricola. E a partir de 1975, o Pro6lcoolofereceu incentivos para o cultivo de cana-de- aq6car [Goldin e Rezende (1993)l. Goldin e Rezende (1993) mostram que, na dCcada de 80, mesmo com a recessb e a contraqzo do setor industrial, a produq8o agricola manteve o seu crescimento. Entretanto, os preqos de terra parecem refletir fundamentalmente a incerteza presente no ambiente macroeconamico de estagnaqgo e inflaq5o. E, a partir da estabilizaqBo de 1994, verifica-se uma tendCncia de queda bastanteevidente. 211 Tabela 2.1- EstatisticasBgsicaspara as Skries Filtradas dos Precosde Terras para Venda 121.32 1616.10 197.40 2786.73 117.61 1580.95 193.50 2723.70 212.09 4952.82 270.46 8045.96 84.69 574.53 144.39 1267.43 0.16 0.33 0.11 0.30 1.85 3.39 0.77 3.51 Nota: f o r m consideradas as shies de desvios emrela@o itendtncia, obtidaspelo filtro de Hodrick-Prescott, centradasnamtdia original. 0 objetivode centralizi-lasemtorno da mtdia t o de preservw a mtdia original (mesmo asCrie nio sendo estacioniua)para que hajauma no@o de magnitudeutilizadano cilculo do coeficientede varia@o. Figura 2.1- Pregosde Venda e Arrendamento de Terras para Pastagens(em reaisde de- zembro de 2000) REGIA0 NORTE REGIA0 NORDESTE -Pastagens (Verda)- Paslagens(Anendamento) - Pastagens(Venda) - Paslagens(Anendamento) REGIA0 SUDESTE 1- 1.......................................................................... .......,. I - Paslagens(Verda)- Paslagens(Anendamento) -Pastagens (Verda) - Paslagens(Arendamento) REGIA0 SUL BW\SIL -Pastagens (Verda) -Paslagens(Arrerdamento) Fonte:CENIBREFGV 212 Figura 2.2- Preqosde Venda e Arrendamento de Terras para Lavouras (em reais de de- zembro de 2000) REGIAO NORTE REGIA0 NORDESTE 3500 250 1600 350 3000 fpm1400 300 200 1200 2500 250.g0 EI::: (L 1000 P02000 150 0 800 '?a 150 $3 1008 400 100 J 3 50 200 50 O" 5 w 0 ; ; e 2 f E L 8 2 $ 8 8 8 m - - - - - . - - - - . , i ; ; g 0 -~alouras (Venda) ---Lalouras (Anendamento) J I -Lalouras (Venda)---Lalouras (Anendamento) REGIA0 CENTRO-OESTE REGIA0 SUDESTE 16000 I r 500 14000 450 a *(I 6000 12000 400 5000 300 g r: 0 350 > loo00 m g 4000 250 200 g 8000 2505p -3Wo 200 150 33 P 6000 2000 100 J p 4000 150 3 100 8 loo0 a 50 2000 50 REGIAO SUL BRASIL 1ow0 350 9000 12000 500 6000 300w ' 10000 4 wIg 4 7000 250 P Em 8000 B 3M)s 5g m o 200 2 m 0 5000 j m 0 203 0J 3 0 4000 150 8 4000 2000 100 8 2000 1000 50 O" 0 - - - - - 0 g 8 a + v r K 2 07$ ; ; ; ; ; ; ; ; H 7 -. - 1-Lalouras IVenda)-Lalouras IAnendamento) 1 A tabela 2.1 apresenta estatisticas descritivas basicas para as sCries de preqos de venda e arrendamento de terras para lavouras e pastagens. Note que ha uma grande diferenqa entre a volatilidade de cada sCrie. 0 coeficiente de variagHo da sCrie de arrendamento C duas vezes menor que o da sCrie de venda de terras para pastagens e tr&s vezes menor para as terras destinadas 2 lavoura. AlCm disso, as sCries de venda si30 mais assimktricas. De fato, as dries de vendas apresentam grandes pulsos positivos e, pel0 menos nas figuras 2.1 e 2.2, nHo tem nenhuma queda stibita. Os aumentos nos preqos sHo muito mais velozes do que as quedas. A tabela 2.2 mostra o efeito mCdio dos planos de estabilizagFio sobre os pregos de terras agricolas para venda e arrendamento. Foram considerados os planos Cruzado (fevereiro de 1986), Bresser Cjunho de 1987), Veri30 Cjaneiro de 1989), Collor (abril de 1990) e Real Cjunho de 1994). A tabela 2.2 mostra os resultados de uma regressgo linear das skries dos desvios dos preGos em relaG8o B tendhcia sobre uma constante e duas varihveis indicadoras para a ocorrCncia de umplano de estabilizaGi3o no semestrecorrente e no semestreseguinte. 213 A varifivel dependente de cada regressgo C uma sine de desvios em relaq8o B tendCncia, calculada pel0 filtro de Hodrick-Prescott, centralizada na mCdia da se'rie original. Portanto, a constante da regressgorepresenta a mCdiados preqos dado a ocorrCncia ou ngo de umplano de estabilizaqgo no semestre corrente ou no semestre anterior. As linhas (B) e (C) da tabela mostram os coeficientes estimados de cada varifivel indicadora. Esses coeficientes podem ser analisados como o aumento mCdio no prego das terras agricolas provocado pel0 plano de estabilizaqgo, no semestre corrente e no semestre seguinte. Os resultados indicam que os planos de estabilizaggo provocaram umaumento mCdio de 48.2% no prego de venda de terras para pastagens e 41.4% no preqo das terras para lavouras no semestre posterior. A magnitude do efeito sobre os preqos de arrendamento s50 bem menores, mesmo que significativo para o cas0 de pastagens. 0 andncio desses planos esteve associado a momentos de instabilidade macroeconbmica. Os pacotes foram criados ngo apenas como uma resposta 51 conjuntura, mas tambCm introduziram incertezas na economia, como foi o cas0 do plano Collor. Dessa forma, a data de anlincio desses planos constitui um indicador dos momentos de grande instabilidade da economia brasileira. E, assim, os resultados da tabela 2.2 demonstram a importiincia do us0 da propriedade da terra como um mecanismo de seguro contra essas instabilidades. A evid6ncia empirica sobre a existCnciade uma demandapor terra para fins ngo-produtivos ter6 como base uma comparagiio entre as dries de pregos de venda e arrendamento de terras. A idCia bfisica C que a existCncia de imperfeigdes nos mercados financeiro e de crCdito afeta a demanda por propriedade de terras sem modificar a demanda por terras para o plantio, essa dltima determinada pela tecnologia disponivel e pelas perspectivas do mercado de produtos agricolas. Emumambiente onde a disponibilidade de crCdito encontra-sevinculada 21 exigCncia de colateral e a terra constitui um bom mecanismo de proteqiio contra instabilidades macroeconbmicas, h6 uma demanda por titulos de propnedade, o que pode determinar uma dissociaggoentre as decisdes sobre o tamanho do estabelecimento agn'cola e a Area cultivada. Tabela 2.2- Efeitos dos Planos de EstabilizaqBosobre os Preqosde Terras Agricolas para Venda e Arrendamento 1 0.17 I 0.14 I 0.15 t 0.02 Nota: foram consideradas as series de desvios em rela@oitendhcia, obtidas pel0 filtro de Hodrick-Prescott, centradas namediaoriginal. Os p-valores dos coeficientesest6 entreparhteses. Dessa forma, pode-se esperar que as se'ries de preqos de terras para venda e arrendamento apresentem uma componente comum, que representa a tecnologia disponivel e as expectativas sobre o mercado de produtos agricolas. A se'rie de preGos de venda pode conter uma componente especifica determinada pel0 us0 da terra como colateral ou seguro contra instabilidades macroeconbmicas. E a se'rie de pregos de arrendamento pode apresentar uma 214 componente especifica determinada pelas variaqdes no funcionamento do mercado de arrendamento. Devido Bs caracteristicas da economia brasileira, sera suposto que as diferenqas encontradas entre as sdries de venda e arrendamento de terras iriio representar, fundamentalmente, as imperfeiqdes que afetam o preqo de venda da terra. Emoutras palavras, a anilise a seguir estarh sujeita B hip6tese de que as imperfeiqdes no mercado de arrendamento mantiveram-se aproximadamente estacionhrias ao longo, niio afetando na din2mica da sCrie. No period0 analisado, diante das oscilaqdes do ambiente macroeconamico brasileiro, essa hip6tese parece razohvel. Formalmente, essa comparaqiio sera feita atravCs de um teste de cointegraqtio. Segundo a representaqiio de Stock e Watson (1988), se duas sCries siio cointegradas entiio podem ser descritas pela soma de duas componentes: uma componente niio-estacionhria comum a ambas as sdries (tendCncia comum) e uma componente estacioniria especifica a cada sCrie. Dessa forma, dada a hip6tese anterior, cas0 as sdries de arrendamento e venda de terras niio sejam cointegradas, hh uma evidCncia de que existe uma demanda por terra para fins niio-produtivos, seja como ummecanismo de proteqiio contra inflaqtio ou como colateral. Por outro lado, a reciproca desse teste deve ser analisada com cuidado. Se duas sCries si40 cointegradas, ntio hh evidencias, necessariamente, de que os mercados siio perfeitos e que niio h i demanda por terra para fins niio-produtivos. Por exemplo, isso ocorreria quando o incremento do preqo de equilibrio, proveniente de uma demanda por terra para colateral, fosse estacionirio e independente dos demais componentes da sdrie, relacionados com a tecnologia agricola e as perspectivas do mercado de produtos agn'colas. A tabela 2.3 mostra os resultados dos testes de cointegraqiio. Para a mCdia brasileira, verifica-se que as sCries de venda e arrendamento de terras para pastagens ntio siio cointegradas, o que constitui uma evidencia de que hh uma demanda por terras de pastagens para fins niio produtivos. Para as sCries de lavoura, que siio cointegradas, nada pode ser afirmado, uma vez que existe a possibilidade de que o efeito sobre o preqo causado pela falha de mercado seja estacionhrio. Esse C umresultado tambCm verificado nas regides Nordeste e Sul. 215 Na regiiio Sudeste, tem-se que a sCrie do prego de venda de terras para lavoura C estacioniria, enquanto a sCrie de prego de arrendamento tem uma raiz unitiria. Dessa forma, o teste de cointegraGiio niio se aplica mas essa constatagiio indica que as duas dries provCm de processos estocisticos diferentes, o que sugere tambCm que haja demanda por terras para lavoura para fins niio produtivos no Sudeste. Na regiiio Norte tanto as sCries de arrendamento quanto 2s de venda siio estacionArias,indicando que nada pode ser dit0 sobre ahipdtese (ii). Teoricamente, a demanda por terra ociosa decorre de seu valor para outras atividades que niio siio agricolas. Dessaforma, quanto maior for o peso dessas atividades maior seri o efeito sobre o prego de venda da terra. A figura 2.3 mostra o diagrama de dispersiio entre as estatisticas de teste do teste de cointegragiio (h=O) e a raziio entre a renda urbana e renda rural para cada unidade da federagiio, para as terras de pastagens. Foram utilizados apenas as unidades da federaqiio para as quais o teste se aplicava, excluindo-se os estados da regiiio Norte niio coberta pela PNAD. Os demais estados apresentavam pel0 menos 1 sCrie j i estaciondria o que impossibilita o us0 do teste de cointegraGiio. Para as terras de lavoura, niio hduma relaqiio Clara entre os indicadores mencionados. Dessa forma, pode-se constatar que niio apenas hd uma evidCncia de que existe demanda por terra para fins niio-produtivos no Brasil, como tambCm que o mecanismo que a explica se aproxima daquele mencionado no modelo tedrico. Figura 2.3- Estatisticas do Teste de CointegracPox RazPo entre Renda Urbana e Renda Rural (Terras de Pastagens) 30 35, 0 II 25- L 20- 15- 10' I I I I I I I 1.05 1.30 1.55 1.80 2.05 2.30 2.55 Renda Urbana / Renda Rural Impe$ei@es no Mercado de Arrendamento Uma imperfeigiio de mercado ocorre quando, ao prego de equilibrio, existe um excess0 de demanda que niio C atendido pela oferta, o que ocorre por diversas razdes. No mercado de arrendamento, essas falhas podem ser explicadas por uma sCrie de r a z d e ~ ~ ~ . Utilizando a notaqiio de Stiglitz (1974), pode-se entender a escolha de um contrato de arrendamento dentre de um continuo de contratos possiveis. Definindo como Q o valor do produto agn'cola gerado por um campon& tem-se que sua renda final sera dada pela fungiio linear y6 Otsuka, Chuma e Hayami (1992) e Otsuka e Hayami (I988) resenham criticamenteos principaisargu- mentos. 216 onde a e ,8 siio os pariimetros estabelecidospel0 contrato com o proprietario das terras. Essa formulaqiio apresentacomo casos particulares os seguintes contratos: 0 contrato de parceria: 0 < a < 1,p =0; 0 contrato de arrendamento: a = 1, p <0; contrato de salhio fixo: a = 0, p >0. As explicaqdes para que o mercado de arrendamento niio funcione adequadamente remetem a problemas de informaq5o assimktrica, tanto de seleqiio adversa quanto de risco moral. Os modelos que tratam o assunto diferenciam-se pelas fontes de assimetria de informaqiio entre proprietirios e camponeses. Quando o proprietario das terras niio tem uma informaqiio precisa sobre a qualidade dos camponeses com os quais estabelece um contrato, utiliza ummenu de contratos com o objetivo de discriminaqiio. Se por outro lado, o trabalho do campones, mais avesso ao risco, desenvolve aqdes niio observadas pel0 proprietkio de terras, a determinaqiio do contrato de equilibrio envolve a soluq5o de um dilema entre incentivos e compartilhamento de risco. 0 contrato de salkio fixo, apesar de niio impor nenhum risco ao campones, pois n30 depende da produq5o agricola, niio oferece incentivo para a escolha da melhor aq5o. Ao contrario, o contrato de arrendamento oferece um incentivo maxim0 aos camponeses que passam a desfrutar completamente dos beneficios de suas aqdes, mas os transfere tambCm todo o risco de uma quebra de safra [Stiglitz(1974)]. Dessa forma, os contratos de arrendamento seriam menos freqiientes porque impdem muito risco aos arrendatirios ou porque os candidatos ao arrendamento siio heterogheos. Eswaran e Kotwal (1985) argumentam que em cada tipo de contrato os contratantes (proprietirios de terra e camponeses) apresentam diferentes incentivos a tomar aqdes niio contrativeis. Enquanto em um contrato de arrendamento o proprietirio n50 tem incentivo a tomar aqdes que melhorem a produqb agn'cola, o mesmo ocorre com os camponeses no contrato de salirio fixo. Portanto, o estabelecimento de um contrato de parceira seria uma forma de oferecer incentivos para que ambos estejam mais engajados na produqiio. Recentemente, Ghatak e Pandey (2000) consideraram a possibilidade de que o arrendatiirio exerGa duas atividades niio observadas pel0 proprietario - o esforqo na produqiio e a exposiqiio a um risco maior do que o desejivel. A segunda possibilidade decorre da hip6tese de que o arrendatArio tem responsabilidade limitada, ou seja, que suas perdas niio podemultrapassar suas posses. Nesse contexto, se apenas o esforqo empregado na produq3o fosse niio-observado, um contrato do tipo de arrendamento seria o 6timo. Se niio houvesse problemas com o esforqo despendido na produqiio mas o campones pudesse se expor a riscos niio desejiveis, do ponto de vista do proprietirio, um contrato de salkio fix0 seria o mais indicado. Portanto, quando as duas possibilidades ocorrem simultaneamente, os contratos de parceria passama ser indicados. Os argumentos acima motivam, teoricamente, a adoqiio da hip6tese (iii) necessariaao resultado de ineficihcia no mercado de arrendamento. Empiricamente, evidencias da validade dessa hip6tese podem ser obtidas pela constataqiio de uma coexistencia entre terra ociosa e propriedades rurais operando em tamanho sub-btimo. Segundo dados do Censo Agropecuario de 1996, apenas 2.4% da Area dos estabelecimentos agricolas brasileiros est6 sob contrato de arrendamento e 0.9% sob parceria. A tabela 2.4 mostra que a pouca utilizaq2o dos contratos de arrendamento niio decorre de uma baixa demanda por esses contratos, mas de imperfeiqaes nesse mercado. Quase a metade dos estabelecimentos 217 rurais tCm menos de 10 hectares e esses estabelecimentos convivem com grandes glebas de terra ociosa. No Brasil, o total de Areas em descanso e niio utilizadas C maior do que as terras ocupadas por estabelecimentoscom menos de 10hectares. A tabela 2.4 tambkm mostra as difereyas interestaduais observadas nos indicadores. Enquanto em GoiBs apenas 11.2% de pequenos estabelecimentos convive com 2.9% de terra ociosa, no Maranhiio, 73.9% dos estabelecimentos tCm menos de 10 hectares e um quarto das terras encontra-se ociosa. Dessaforma, a hipdtese (iii) justifica-se no B r a d tanto teoricamente quanto empiricamente. e Terra Ociosa I $96) 3962 16.7% 100242 3.1% 953 28.5% 37524 5.4% 43793 52.6% 314409 9.5% 64838 31.4% 1933700 8.6% 17618 22.9% 305847 3.4% 1025 13.7% 143561 4.8% 2614 5.8% 1663026 9.9% 92736 80.6% 195286 9.1% 401734 57.5% 3141730 10.5% 245312 72.2% 1689658 18.8% 272100 73.9% 3057373 24.3% 101435 69.2% 661409 16.1% 186669 72.2% 706830 12.7% 134949 64.8% 2577757 26.7% 57958 63.4% 609488 16.3% 77618 77.8% 75673 4.4% 23492 32.1% 144832 4.2% 169638 34.2% 1764594 4.3% 28439 53.0% 77495 3.2% 65303 30.0% 382660 2.2% 154620 41.8% 649155 4.1% 150679 35.0% 861887 4.0% 72462 35.6% 293858 4.4% 930 37.8% 16444 6.7% 12526 11.2% 803193 2.9% 9170 18.6% 522134 1.7% 9801 12.4% 1940465 3.9% 49.4% 24670230 7.0% 218 Sintomas da InejiciZncia na Produtividade Agri'cola A evidgncia empirica acima mostra que todas as condiqdes necesstiriaspara que a distribuiqlo de terras seja ineficiente. Com o objetivo de reforqar esses resultados, serlo apresentados resultados que evidenciam sintomas dessaineficisncia. A tabela 2.5 apresenta os pargmetros de produtividade total dos fatores (TFP) estimados por Avila e Evenson (1998). A TFP C uma razlo entre umindice de produqiio e umindice agregado de insumos e C geralmente utilizado para comparaqdesde produtividadeem diferentes regides. Tabela 2.5- ProdutividadeTotaldos Fatores Nota: Os indicesde Produtividade Total dos Fatoresforamnormalizadosde modo que o indice 100equivale ?I mCdia dos periodos 1970-75. Os resultados revelam grandes mudanqas no periodo analisado. 0 exemplo mais marcante C o da regilo Centro-Oeste, uma 6rea de fronteira agricola. Em 1970, ocupa a ~5ltimaposiqlo na ordenaqlo dos indicadores agregados mas, em 1985, o quadro C completamente revertido e passa a ocupar o primeiro lugar tanto para lavouras quanto para a pecutiria. 0 dinamismo da regilo Centro-Oeste foi acompanhado apenas pel0 que ocorreu na regiiio Sudeste. Outra caracteristica interessante desses resultados C que, para as regides em que o teste de cointegraqlo da tabela 2.3 pode ser utilizado (Nordeste, Centro-Oeste e Sul), o setor com maior dinamismo foi aquele que nlo apresentou evidbncias das imperfeiqdes de mercado referidas anteriormente. Na regilo Nordeste, o teste da tabela 2.3 indicou que pode-se rejeitar a hip6tese de que nlo hti demanda por terras de pastagens para fins especulativos mas, para as terras de lavouras o teste C inconclusivo. 0 crescimento da produtividade da pecutiria no Nordeste, entretanto, C bem menor que o crescimento observado nas plantaqdes. A regilo Sul apresenta resultados antilogos. Na regilo Centro-Oeste, por outro lado, o teste de cointegraqiio indicou, curiosamente, a presenqa de imperfeiqdes no mercado de terras para lavouras, que apresentaramumcrescimento inferior ao da pecutiria. 219 ies Internac nais 9159 23% 0.66 39523 550 25% 0.31 36889 7682 15% 2.79 30904 377 21% 0.03 30272 349 13% 0.14 22452 294 33% 0.14 20031 2737 11% 0.71 13715 499 23% 0.38 13499 8457 20% 0.33 4081 16889 23% 0.85 2476 2382 41% 0.26 1943 2671 44% 1.93 1450 2973 72% 0.17 406 9326 69% 0.10 307 A tabela 2.6 mostra comparaq6esentre indicadores internacionais de produtividadeagricola. Os paises escolhidos pertencem ao grupo dos 10 paises mais ricos ou ao grupo dos maiores 10 paises. Os linicos paises excluidos foram o Canada e a Inglaterra, cujos indicadores de produtividade agn`cola niio siio reportados pelo World Development Indicators. 0 indicador de produtividade agricola C a raziio entre o valor adicionado da produq5o agricola e o nlimero de trabalhadores na agricultura. Os dados indicam que, com a exceqiio da Federaqiio Russa, todos os paises com produtividade agricola inferior B do B r a d apresentam taxas de urbanizaqiio muito menores que os 20% da economia brasileira. A escassez relativa entre terra e trabalho n5o C capaz de explicar as diferenqas de produtividade. Paises como o Japiio, a Alemanha e a Ithlia, que apresentam uma escassez relativa de terras em relaqiio ao Brasil, tem uma produtividade pel0 menos 5 vezes maior. E, para os Estados Unidos, Australia e Argentina, onde o fator trabalho C maisescasso, a produtividade tambCm C mais muito mais alta. Nenhum pais com dados apresentados no relatdrio acima e com grau de urbanizaqiio maior do que o Brasil tem produtividade agricola menor que o brasileiro9'. Enfim, essas comparaqaes internacionais indicam que a produtividade agricola no Brasil encontra-se em niveis bastante inferiores Bqueles observados em paises com algum grau de similaridade, seja na extensiio do territdrio, no tamanho da economia ou no grau de urbanizaqiio. E essa constataqiio pode ser interpretada como um sintoma dos problemas de ineficihcia da distribuiqiio de terras mencionados anteriormente. Impactos da Distribuiciio de Terras sobre a EjiciCnciaAgricola e a Pobreza 0 resultado de ineficiencia da distribuiq5o de terras enunciado teoricamente e constatado no B r a d determina, dentre outros fatores, que pequenos e grandes proprietfirios operam com diferentes produtividades marginais. Isso ocorre porque os grandes proprietfirios, ao determinarem o tamanho de suas propriedades, consideram um beneficio niio produtivo e, ''Esses paises slo, em ordem crescente de produtividade agn'cola: Venezuela, Chile, Uruguai, Arabia Saudita, Argentina,Alemanha, Lihno, Australiae Singapura. 220 diante de uma hip6tese rendimentos marginais decrescente, acabamproduzindo a umnivel mais baixo de produtividade marginal. Os pequenos proprietirios, por outro lado, niio acessam esse beneficio extra da propriedade da terra e, portanto, escolhem extensdes de terra menores, o que implica em uma produtividademarginal da terra maior. Figura 2.7 Lucratividadee Tamanho de Estabelecimento - Sergipana Oeste Potiguar CentralPotiguar SertbeAgresteLagoano c 200 100 160 3 g 120 += 80 8 -k540 0 0 0 20 40 60 SO 100 0 10 20 30 40 50 Tamanhodo estabelecimento(hectares) Tamanho do estabelecimento (hectares) - h n o porhectare ~`-Porcentagemde estaklecimentos -Lucro por hectare --Porcentagem de estabelecimentos MapaAlagoana %rtbParaibano Fonte: Barrosef al. (2000) Nesse contexto, a ineficiCncia C uma consequCncia da possibilidade de transferencia de terra dos grandes proprietirios, menos produtivos marginalmente, para os pequenos agricultores. Portanto, um outro sintoma dessa ineficisncia C a existCncia de uma relaqiio sistemitica entre lucratividade da terra e tamanho do estabelecimento. Os resultados a seguir foram extraidos do artigo "Impactos da Distribuiqgo da Terra sobre a EficiCncia Agricola e a Pobreza no Nordeste", escrito por Ricardo Paes de Barros e colaboradores, para o seminirio Desajios para a Pobreza Rural no Brasil. 0 artigo, alCm de estimar a relaqgo entre lucratividade e tamanho para o Nordeste, apresenta simulaqaes de mudaqas na estrutura fundiiria, mostrando que h i espaqo para melhorias na eficiCncia agricola, usando dados do Censo Agropecufirio de 1985. A figura 2.7 mostra os resultados obtidos para 6 das 21 mesorregiaes analisadas pelos autores. Exceto no caso da regia0 Central Potiguar, h i uma relaggo negativa entre tamanho de 22I estabelecimento e lucro por hectare. 0 tamanho 6timo dos estabelecimentos C relativamente baixo (variando de 5 a 20 hectares) e, mesmo assim, existe uma porqHo muito grande de estabelecimentos que operam em tamanhos sub-6timos. Em todas as regides analisadas em Barros et al. (2000) mais da metade dos estabelecimentostem tamanho inferior ao 6timo. Para algumas regides da Bahia, os autores encontram uma relaqao positiva entre lucro por hectare e tamanho de estabelecimentos. Emcada diagrama da figura 2.7, o circulo preto mostra o lucro mCdio e o tamanho mCdio dos estabelecimentos em cada mesorregigo. 0 circulo cinza mostra o lucro por hectare e o tamanho dos estabelecimentos cas0 a propriedade da terra fosse distribuida entre as fam'lias que trabalham nos estabelecimentosagricolas. A distiincia vertical entre os circulos indica o ganho de eficigncia causado pela redistribuiqgo das terras. Os resultados das simulaq6es realizadas pelos autores indicam que, a distribuiqao de terras entre os individuos que j6 possuem terras C mais eficiente que a distribuiqiio entre todos aqueles envolvidos na agricultura. 0 ganho no lucro por hectare C de 81.2% para a distribuiqb entre todos os agricultores e 167.5% para a distribuiCBo de terras entre aqueles quej6 s30 proprietfirios. Dessa forma, os resultados apresentados demonstram que, n5o apenas h i evidhcias de que as condiqdes necess6riaspara a ineficigncia da distribuiqb de terras sgo atendidas para o Brasil, como tambCm os sintomas dessa ineficigncia podem ser mensurados. E, com base nessa eficiencia, a prdxima seqb avalia os programas financiados pel0 govemo brasileiro. 3- Politicas P6blicas voltadas ao Meio RuralBrasileiro Os dados apresentados nas seqdes anteriores mostraram que a distribuiqiio de terras (e demais meios de produqiio agn'cola) no Brasil n2o 6 apenas injusta, como tambim, ineficiente. Dessa forma, as politicas pfiblicas voltadas ao meio rural brasileirojustificam-se tanto por critCrios de equidade como tambCm de eficigncia. Portanto, dada a necessidade de intervenqiio, essa se@o analisa a abrangencia e a efic6cia de quatro instrumentos utilizados pel0 govemo brasileiro atualmente. 0 primeiro instrumento a ser analisado C o Impost0 Territorial Rural (ITR). TaxapZono Mercado de Terras-o ITR 0 ITRfoi criado atravCs do Estatuto da Terra, em 1964, como objetivo de auxiliar as politicas pfiblicas de desconcentraqgo da terra. A cobranqa do lTR poderia promover uma reduqBo dr6stica do us0 especulativo da terra, promovendo uma maior eficigncia na distribuiqgo de terras. No entanto, por tr6s dessa proposiqBo existem hip6teses muito fortes que n2o se aplicam ao cas0 brasileiro e restringem a implementas80 do imposto. Esta subseq8o inicia-se com um breve retrospect0 da experigncia brasileira com o ITR, que tentar6 demonstrar suas deficigncias. Emseguida, s50 apresentadosalguns aspectoste6ricos importantes da taxaq8o de terras. Desde a sua criaqgo, foram feitas duas grandes reformulaqBes no esquema de cobranqa do ITR, em 1979e 1996, que tentaram mitigar alguns problemas operacionais detectados. Inicialmente, a cobranqa do ITR tornou-se responsabilidade do INCRA (Instituto Nacional de Colonizaqgo e ReformaAgrBria). A partir de uma aliquota b6sica de 0,2% sobre o valor da terra nua, eram aplicados coeficientes relacionados com a dimensgo, localizaqiio, condiqdes sociais e produtividade dos estabelecimentos. E, dadas as faixas de variaqgo de cada coeficiente, a aliquota variava de 0,24% a 3,456%. Entretanto, verificou-se que os objetivos almejados no desenho do imposto estavam longe de ser alcanqados. 0 ITR nunca chegou a constituir uma boa fonte de receita e tampouco 222 conseguiu promover as mudanqas desejadas no meio rural. 0 valor do imposto era excessivamente baixo e apresentava incoergncias ao tributar mais pesadamente o miniftindio, por n8o discriminar o contribuinte segundo a categoria dos im6veis - minifiindio, empresa rural e latifiindio. E, alCm dos parsmetros estabelecidos pel0 INCRA ngo se adequarem B realidade brasileira, o problema de evasiio 6 era grave, com um sistema prechrio de atualizaqb do valor da terra nua [Oliveira e Costa(1979)l. Figura 3.1 ArrecadaCPo do ITR (1972-1991) - 120.00 , I I 100.00 80.00 60.00 40.00 20.00 8 Fonte:Oliveira (1993) I 0 quadro estabelecido nos anos 70 motivou B primeira reformulaqiio importantena legislaqgo do lTR, em 1979. Modificaqaes significativas recafram sobre o artigo 49 do Estatuto da Terra, segundo o qual, "as normas gerais para a fixaqgo do ITR passam a obedecer a critirios de progressividade e regressividade, levando-se em conta os seguintes fatores: o valor da terra nua; a kea do im6vel rural; o grau de utilizaqiio da terra na exploraqao agricola, pecuhia e florestal; o grau de eficiCncia obtido nas diferentes exploraqbes; a kea total, no Pais, do conjunto de imdveis rurais de um mesmo propriethio; a classificaqgo das terras e suas formas de us0 e rentabilidade". A aliquota tornou-se funqBo do grau de utilizaqgo da terra e do grau de eficiCncia da exploraq50, com uma variaqiio entre 0,2% e 3,5% para propriedades acima de 100 m6dulos fiscais". Os dados apresentados por Oliveira (1993) apontam uma frustraqilo com a arrecadaqgo. Os nfveis da arrecadaqiio se elevaram nos anos subsequentes B mudanqa mas retornaram, em 1983, aos niveis anteriores, como mostra a figura 3.1. E, mesmo em 1990, o nivel arrecadado corresponde B insignificante quantia de US$ 20,30 por im6vel rural. A carga tributfiria correspondeu a 25% de umsalhrio m'nimolano emjaneiro de 1992. Segundo a Secretaria de ComunicaqBo de Govemo da PresidCncia da Rep~blica,o percentual do VTN declarado em relaqao ao preqo real da terra na dtcada de 80 variava de 20% para as propriedades com menos de 10 ha a 1,2% para as grandes propriedades com mais de 10mil ha. A Area declarada aproveithvel era muito menor que a real, com os maiores propriethrios de- clarando algo em torno de 50% e os menores 94%. E a declaraqiio da produtividade era ainda mais irreal, com casos, aceitos pel0 INCRA, em que a produtividade era mais de dez vezes superior ao valor esperado calculado pel0 IBGE. 9RA definiGBo de m6dulo fiscal do rnunicipio considera os seguintes fatores: tipo de exploraqIo predomi- nante (hortifrutigranjeira, cultura permanente, cultura ternporiria, pecuiria e florestal); produtividade por cultura e de urnconceit0 de agricultura familiar. 223 0.07 0.40 0.80 1.40 2.00 0.10 0.60 1.30 2.30 3.30 0.15 0.85 1.90 3.30 4.70 0.30 1.60 3.40 6.00 8.60 0.45 3.00 6.40 12.00 20.00 Fonte: Lei9.393, de 19 de dezembro de 1996. Emrespostaaos problemas detectados, foi feita uma reformulaqlo em dezembro de 1996 que, dentre outras modificaq6es, determinou: (i)o aumento da aliquota dos imdveis grandes e improdutivos; (ii) simplificaq2o das faixas de cobranqa de 12 para 6; (iii) da diferenciaqlo fim regional das aliquotas; (iv) o valor declarado pel0 proprietArio, para efeito do pagamento do ITR, sera consideradoem cas0 de desapropriaqlo. As aliquotas diferenciam-se apenas pel0 grau de utilizaqlo e pela Area total do imdvel, de acordo com a tabela 3.1. Pode-se verificar que hA uma acentuada progressividade no tamanho da propriedade e regressividade no grau de utilizaqlo, modificada de forma que os im6veis produtivosforam privilegiados. Figura 3.2 -Arrecadaqfo do ITR (1992-1999) 0.25 250 18 0.20 200 v) u 0 IO a 0.15 150 5 a t b3 8 0.10 [r 100 0.05 50 0.00 0 1992 1993 1994 1995 1996 1997 1998 1999 1 ;' ArrecadaCBo Total do ITR -% ArrecadaCBo Administrada pela SRF] Fonte: Secretaria da ReceitaFederal Reydon et al. (2000) salienta a descontinuidade presente nas aliquotas adotadas, observando que um imdvel com 50,l ha e um grau de utilizaqlo de 80,0% pode pagar um montante de impost0 13 vezes maior que um imdvel de 50,O ha com grau de utilizaqiio igual a 80,1%. E a soluqiio apontadapor alguns autores C o us0 de redutores, como ocorre no Impost0 de Renda. 224 AlCm disto, Reydon et al. (2000) mostram que, apesar dos aperfeiqoamentos administrativos e legais, as expectativas geradas em tomo da reformulaqiio niio se confirmaram. E as principais razaes se associam B dificuldade de avaliaqiio do valor da terra nua e da imprecisiio do conceit0 de Area utilizada. A figura 3.2 mostra, de um lado, a melhoria obtida com a reformulaqiio de 1996 e, por outro, o baixo grau de arrecadaqiio. Segundo cAlculos de Oliveira (1993), a receita potencial do ITR iria variar entre 1,4 e 2,8 bilhaes de ddlares por ano, cas0 fossem utilizadas aliquotas entre 0.5% e 1.0%. Apesar dos cAlculos niio considerarem o efeito da aplicaqiio efetiva destas aliquotas sobre as decisaes dos propriethrios de terra, a magnitude das estimativas deixam claro o espaqo existente para o aumento da arrecadaqiioe a baixa efetividade do mecanismo de taxaqiio utilizado. Aspectos Tedricos Henry George (1839-1897) foi o primeiro a estabelecer uma racionalidade econ6mica para a taxaqiio de terras, em Progress and Poverty, publicado em 1879. 0 autor atribuiu o desemprego e os baixos salirios a uma escassez artificial de terras e ao mau funcionamento do mercado. Essa escassez artificial seria o resultado de uma distribuiqiio desigual das terras p6blicas e de atividades especulativas. Nesse contexto, George propae a utilizaqiio do imposto sobre a propriedade da terra para dinamizar o mercado de terras, sendo capaz de induzir ao pleno us0 do solo, sem distorcer os incentivos marginais. Arnott e Stiglitz (1979) analisam a generalidade da proposiqiio que ficou conhecida como teorema de Henry George, tomando-se uma referencia clhssica nessa direqiio. Outros autores tambCm salientam as vantagens inerentes ao us0 do imposto sobre a terra como fonte de arrecadaqiio [Deininger (1998) e Skinner (1991b)l. 0 imposto sobre terra niio distorce a alocaqiio de recursos e constitui um dos poucos exemplos de imposto lump-sum em termos agregados, o qual poderia garantir um nivel m'nimo de arrecadaqgo, pois a oferta de terra C inelfistica. AlCm disso, o tamanho dos estabelecimentos C observado - principalmente em regiaes onde o direito de propriedade t bem estabelecido, existem informaqijes acessiveis e confiiveis sobre o tamanho das propriedades. Skinner (1991a) estabelece que a perda de capital resultante da aplicaqiio do imposto C transitdria, afetando apenas os atuais proprietArios de terra. Quando os agentes tCm acesso a outros ativos, uma condiqiio de niio-arbitragem garante que o imposto seja completamente absorvido por uma reduqiio no preqo da terra, impedindo que a perda de capital seja duradoura ou estendida aos demais agentes da economia. 0 artigo de Hoff (1991), de outro modo, restringe a utilizaqiio do imposto sobre a terra. A autora argumenta que em um ambiente de incerteza como a agricultura, em que os produtores siio avessos ao risco, o us0 exclusivo do imposto sobre a terra promove uma alocaqiio ineficiente do risco. 0 valor a ser pago niio depende da produsiio, mantendo-se constante independente da ocorr6ncia de uma quebra de safra. No modelo de Hoff (1991), uma combina@io que utilize tambtm o imposto sobre o produto revela-se Pareto-superior. A composi@io dtima de imposto sobre o produto e imposto sobre a terra 6 determinada pel0 dilema entre distorqiio (introduzida pel0 imposto sobre o produto) e compartilhamento de risco. 0 us0 de um esquemamisto de cobransa de imposto sobre a terra e imposto sobre o produto tambCm justifica-se num contexto de informagilo assimktrica, em que os agentes operam com terra ociosa. Assunqiio e Moreira (2000),utilizando um modelo em que o governo maximiza a receita tributAria e tenta combater o us0 especulativo da terra, mostram que o us0 do imposto sobre o produto niio seria necessirio apenas se o governo observasse o grau de utilizaqiio da terra com perfeiqiio ou se niio houvesse terra ociosa em equilibrio. 225 Os resultados acima mostram que a proposiq5o de Henry George sobre as qualidades do imposto sobre a terra requer que os mercados sejam completos e perfeitos e que n5o haja assimetria de inforrnaqb. 0 imposto sobre o produto, apesar de distorcivo, deve ser considerado para a obtenqgo de um compartilhamento eficiente de risco [Hoff (1991)l e de um esquemas de tributaqiio implementavel sob informaqiio assimCtrica [AssunqBo e Moreira (2000)l. Mesmo ap6s a melhoria de uma sCrie de problemas operacionais, o ITR ainda continua pouco efetivo. A experihcia brasileira revela a incapacidade do govemo em aplicar corretamente um esquema de taxaq5o e com isso reduzir os altos graus de evas5o e sub-tributaqiio. Os dados fornecidos apontam ainda para o fato de que esta incapacidade ainda C mais cr8nica as para grandes propriedades. A soluq80 apontada pela literatura te6rica de taxaqgo indica que a utilizaqiio do imposto sobre o produto (ICMS) constitui uma forma de lidar com o problema da assimetria de informaqgo e do compartilhamento de risco, devendo ser considerada no desenho do ITR. 3.2- Programade ReformaAgrhria 0 resultado de que a distribui@o de terras no BrasilC ineficiente, excessivamenteconcentrada, n5o apenas justifica uma intervenqb governamental como tambCm sustenta a adoq5o de politicas de redistribuis5o de terras. E a reforma agrhria com base em desapropriaqbes tem sido o instrumento mais utilizado pel0 govern0 brasileiro nessa direqao, principalmente nos dltimos anos. Abaixo, ser5o apresentadas algumas consideraqbes te6ricas acerca do processo de distribuiqiio de terras, contrapondo, de umlado, a necessidadede que a politica de distribuiq5o de terras seja recorrente e, de outro, possiveis efeitos colaterais sobre o investimento agricola. Em seguida, sera apresentadaa evoluq5o do programa de reforma agriria brasileiro. Aspectos Tedricos 0 resultado de ineficihcia na distribuiqgo de terra apresentado na seq5o II, um contexto em estitico, pode ser estendido tambCm para um modelo dinhico, onde a distribuiqiio dos ativos da economia C endogenamente determinada pel0 processo de acumulaq5o de riquezas na presensa de imperfeiqBes no mercado de crCdito. Esse C um resultado que ficou conhecido a partir dos artigos de Galor e Zeira (1993) e Banerjee e Newman (1993) que consideram a distribuiqgo de riqueza. 0 argument0bisico, alCmde ummercado de crCdito imperfeito, dependeda existzncia de uma n5o-linearidade na produq8o que permite, apenas aos agentes mais ricos (grandes proprietaries de terra), o acesso a uma tecnologia mais lucrativa. Dessaforma, torna-se possivel a ocorrCncia de equilibrios mdltiplos, em que a posiq5o inicial de cada individuo na distribuiqiio de riquezas (distribuiq5o de terra) e os parsmetros tecnol6gicos afetam sua posiq5o no longo-prazo. Nesse contexto, uma redistribuiq5o de riquezas (de terras) teria um efeito permanente sobre a distribuiqiio de longo-prazo da economia. Aghion e Bolton (1997) mostram que, mesmo quando a acumulaq5o de riquezas da classe mais rica afeta positivamente os mais pobres, pode haver espaqo para politicas redistributivas decorrente de problemas de incentivo. Entretanto, essas redistribuiqbes tem apenas um efeito transit6rio e, para aumentar o nivel de eficiencia da economia no longo-prazo, 6 necessaria a adoqiio de politicas sistematicas de redistribuiqgo capazes de mudar a distribuiqiio de riqueza no longo-prazo. Dessa forma, diante do resultado de ineficihcia da concentraq5o fundiiria no Brasil, os argumentos mencionados acima reforqam a necessidade de politicas redistributivas 226 sistemhticas. E o instrumento de reforma agrhria com base em desapropriaGdesC a forma mais direta de intervenqiio e B qual o governo brasileiro vem despendendo mais recursos. Por outro lado, as desapropriaqdespodem distorcer os incentivos a investir na medida em que compromete a estabilidade do direito de propriedade. A experisncia hist6rica de reformas agrhrias redistributivas indica que, em paises (landlord states) onde as terras j h siio cultivadas pelos futuros beneficihrios e tudo que C necessiirio C uma reatribuiqiio de dos direitos de propriedade, os programas tem tido maior sucesso. Nesses casos, ganhos de eficiencia significativos vem sendo observados e os requisitos administrativos necessfirios reforma siio m'nimos. Contudo, em paises caracterizadospor relaqdes de trabalho assalariadas(haciendas)o process0 C bem mais complicado, uma vez que exige uma restruturaqiio do sistema produtivo [Deininger e Feder (1998)l. 0 sucessodosprogramasde reformaagrhriaemlandlord states decorre damaiorcapacidadede seleqiio dos beneficihrios e do fato de que a redistribuiqiio de terras implica em uma maior fixaqiio dos direitos de propriedade, o que constitui umincentivo ao investimento. Nos sistemas de haciendas, entretanto, a reforma agrkia corr6i o direito de propriedade alCm de enfrentar problemas com a seleqiio dos benefici5rios. Portanto, apesar de que a existsncia de um programa sistemdtico de reforma agrhria pode promover ganhos de eficiencia, existem certas 1imitaqBes em sua implementa@o que devem ser consideradas. E mais, como sugerem Deininger e Feder (1998), a habilidade de umgovemo em transferir grandes quantidades de terra niio constitui uma condiqiio suficiente para o sucesso do programa. Experihia Brasileira 0 programa de reforma agrhria brasileiro sofreu, a partir de 1995, uma aceleraqiio nothvel. Enquanto nos trinta anos anteriores, desde a ediqiio do Estatuto da Terra em 1964, o programa havia assentadoapenas 218 mil famlias em 360 projetos de assentamentos, entre 1995 a 2000, o ndmero de fam'lias cresceu para 482 mil, assentadas em 3736 projetos. A tabela 3.2 apresenta algumas estatisticas sobre o balanqo da reforma agrhria no periodo de 1995 a 2000. Os ndmeros revelam a magnitude do programa. Durante o periodo, foram obtidos um total de 18 milhaes de hectares, onde foram assentadas mais de 2.4 milhdes de pessoas. A queda no valor total do im6vel por familia mostra, de um lado, a queda nos preqos das terras caracterizada na se@o anterior, mas C tambdm o resultado do esforqo dos procuradores do INCRA na contestaqiio dos chlculos e nas aqdes rescis6rias das indenizaSdes das terras desapropriadas. Outra informaqiio importante C a significativa queda das invasdes ocorridas no ano de 2000. Tabela 3.2 Balancoda Reforma - 62.044 433 637 4.452 4.395 16.385,04 14.614,59 397 502 446 1 455 I 226 Fonte:INCRA - Balaqo da Reforma Agriui zoo0 Nota: * inclui o Bancoda Terra. 0 procedimento de desapropriagiio tem gerado onerosos dispsndios para o governo devido a contradigdes na legisla~50vigente. Se, de um lado, permite que estabelecimentos ociosos sejam 227 desapropriados, por outro, determina que a indenizaqgo seja calculada com base no valor de mercado, tanto das terras nuas quanto das benfeitorias. Para a teoria econ6mica isso constitui uma contradiqgo em termos pois, para adquirir terras ao preqo de mercado o governo, assim como qualquer outro agente, n20 deveria precisar de fazer nenhum esforqo legal, uma vez que estaria sendo pago umpreqo ao qual o propriettirio estaria disposto a vender. A legislaq20, ao atribuir ao governo um poder de punir grandes latifihdios improdutivos e contraditoriamente exigir que essa puniq2o n5o seja efetiva, acaba por gerar distorqbes que s20 refletidas em processosjudiciais miliontirios. Recentemente, o MinistCrio do Desenvolvimento AgrArio (1999a) publicou o "Livro Branco das Superindenizaqbes" que mostra a situaS2o de alguns processos de desapropriaqb ainda em curso. Para dar um idCia do valor dessas superindenizaqbes, basta dizer que em apenas quatro processos judiciais em anfilise, a Uni2o pode ser condenada a pagar R$ 1,7 bilhbes, o que representa todo o orqamento da reforma agrhria para o ano de 1999. Esse impasse faz com que o custo inicial das desapropriaqties seja aumentado 5 vezes em mddia para o Brasil. Na regib Sudeste, esse fator multiplicativo chega a 14,64. A aq2o dos procuradores do INCRA tem provocado uma reduq2o significativa das indenizaqbes. Enquanto no ano de 1997 foram gastos R$ 780 milhbes, no ano de 2000 o valor foi de R$ 55,7 milhbes. AlCm dos custos com as desapropriaqties, o sucesso das politicas de redistribuiqiiode terras no Brasil depende da sustentabilidade das familias assentadas. Na medida em que as fam'lias beneficiadas tornam-se auto-suficientes, capazes de gerar excedentes produtivos que financiam o seu consumo e o investimento necessfirio 2 plena utilizaqgo de seu potencial, o programa de reforma agrhria promove, de fato, uma melhoriana eficiencia no us0 dos recursos produtivos. 228 Tabela 3.3 FatoresPotencializadores dos Assentamentos - os sistemas de produqiio voltados ao mercado; a produqiio de umbom nivel de subsistCnciafamiliar; o us0 do crkdito; o us0 da miio-de-obra familiar pela maior irea aproveitivel e intensidade do sistema; o crescimento econ8mico menos diferenciado entre as fam'lias nologias pelos assentados; ais com os trCs niveis de Fonte: Bittencourtet al. (1999) Entre os meses de maio e julho de 1998, foi realizada uma pesquisa, resultado de um convCnio de cooperaqiio tCcnica entre o INCFL4e a FAO, cujo principal objetivo foi identificar os fatores que tCm potencializado e restringido o desenvolvimento dos projetos de assentamentos. Para isso, foi feita uma pesquisa de campo em 10 projetos de assentamentosconsiderados de maior desenvolvimento e 10 projetos de menor desenvolvimento. A pesquisa abrangeu os estados da Bahia, Ceari, Goiis, Maranhiio, Minas Gerais, Pari, Paranh, RondGnia, Santa Catarina e S5o Paulo. As tabelas 3.3 e 3.4 apresentamas principais conclusaes desse estudo. 229 Tabela 3.4-Fatores Restritivos dos Assentamentos 0 a produgiio para a subsisthcia familiar; 0 a obtengiio de rendamonet8ria; 0 a eficicia do crCdito; 0 mecanizagiioe o us0 de algumas tecnologias; 0 o consumo de tiguahumano e animal em alguns assentamentos; a miio-de-obra familiar, obrigando a buscarem alternativas de 0 a constiincia da assistenciaticnica; homogCneo entre os assentados, ampliando as diferengas intemas; 0 o acesso a mGquinas, implementos e instalaqdes atravCs de um us0 0 o aumento da produglo e da produtividade; 0 as relagdes com os trCs niveis de govemo; 0 o acesso h infra-estmturasocial e produtiva; Fonte: Bittencourtet ai. (1999) As conclusdes das tabelas 3.3 e 3.4 evidenciam a importiincia das condigdes de formagiio dos projetos. Basicamente, o que C necesskio para o $xito de uma reforma agrBria com desapropriagdes C a transformag50 das condiqdes de restringiam o exercicio da potencialidade produtiva da agricultura familiar. Naqueles assentamentosonde essas transformasdes fizeram- se presentes, os objetivos propostos foram atingidos. 0 crCdito rural, como foi visto, 6 um dos fatores importantes do programa de reforma agrgria. Em 1985, o Conselho MonetBrio Nacional criou o Programa de CrCdito Especial para Reforma AgrBria (Procera) com o objetivo de aumentar a produqiio e a produtividade agricola dos 230 assentados e, assim, permitir a sua emancipaqiio. Mas, segundo Rezende (1999), o programa est6 longe de atingir seus objetivos de tornar as familias independentes da tutela do govemo. 0 autor conclui que a hipdtese mais provivel C de que o crCdito tenha causado apenas uma melhoria artificial no consumo. Durante o period0 de alta inflaqiio, os emprkstimos continhamum subsidio muito elevado, sob a forma de um rebate que cobrava apenas 50% da correqiio monetiria e zero de juros. Usando o exemplo de Rezende (1999), uma inflaq8o de 20% ao mCs faz com que um emprCstimo de R$ 7.650,00, com sete anos de prazo e dois de carCncia, seja pago com apenas R$ 243,32 em tennos reais. A tiltima parcela seria equivalente a R$O,85. E assim o mecanismo de emprkstimo niio contCmincentivos ao investimento, uma vez que niio repassariscos para o agricultor. IS 0,8 18 0,781 0,811 0,800 0,764 0,75 1 0,716 0,708 0,848 0,802 A tabela 3.5 ilustra o efeito das politicas voltadas B distribuiqiio de terras brasileira. Os dados refletem o impacto dos projetos de assentamento, Banco da Terra, a exclusiio das terras p6blicas e as ireas canceladaspel0 combate B grilagem. Pode-se notar uma mudanqaimportante no coeficiente de Gini nacional, liderado principalmente pela desconcentraq2o do Norte e Centro-Oeste. Esses resultados fizeram com que o governo passasse de 5" pais americano com maior concentraqiio de terra para 12". 3.3- Bancoda Terra 0 Banco da Terra foi criado pela Lei Complementar no. 93 de 4 de fevereiro de 1998, sendo regulamentado pel0 Decreto no. 3.475 de 19 de maio de 2000. 0 programa C a expansso do projeto piloto, denominado Ckdula du Terra, implantado em 1997 nos estados do Maranhiio, Ceari, Pernambuco, Bahia e norte de Minas Gerais. Criado para atender 15 mil fam'lias no prazo de 3 anos, o CCdula da Terra superou as expectativas, atingindo essa meta na metade do tempo. A seguir seriio apresentadas as principais caracteristicas do programa, indicando was vantagens em relaqiio ao programa de reforma agriria com base em desapropriaqaes e as principais dificuldades detectadas nos primeiros anos de operaqiio. 0 Banco da Terra atribui aos prdprios beneficiirios, organizados em associa@es ou em cooperativas, a seleqiio, a negociaqiio do preqo da terra e a forma de produqiio nos imdveis adquiridos. Para promover maior agilidade e adequaqiio Bs especificidades regionais, as operaqaes siio descentralizadas, delegadas aos Estados. Os beneficiirios do programa siio os trabalhadores rurais que comprovem um m'nimo de 5 anos de experisncia na atividade rural, preferencialmente assalariados rurais, parceiros, arrendatdrios e proprietkios de imdveis cuja Area niio ultrapasse a dimensiio de propriedade familiar e possua renda insuficiente para o sustento de sua familia. 0governo, atravCs do programa, repassaate'R$40milpor familia paraocusteio daterra, infra- estrutura e assistsncia tCcnica, com prazo de amortizaqiio de at6 20 anos, com 3 anos de carhcia. Para a produqiio, os agricultores podem utilizar os recursos do Pronaf-Planta Brasil, que serd analisado na prdxima subseqzo do artigo. Segundo o Ministe'rio de Politica Fundidria e 231 da Agricultura Familiar (1999b), o govern0 brasileiro, em parceria com o Banco Mundial ir6 aplicar US$2 bilhdes no Banco da Terra nos cinco anos que sucedem B sua regulamentaqb. Segundo Buainain et al. (1999), as caracten'sticas do CCdula da Terra (que ainda se mantCmno Banco da Terra) produziriamuma estrutura eficiente e sustenthvel pelas seguintes razdes: seleqgo de ativos de melhor qualidade - como a terra C comprada e paga pelos pr6prios beneficitirios, hh um incentivo para a seleqzo de Areas com potencial produtivo suficiente para a geraqiio da renda necesskia B quitaqgo da divida. E as associaqdes poderiam reduzir os erros de avaliaqiio, promovendo uma maior compatibilidade entre as caracteristicas da terra e as aptiddes dos beneficihrios; melhor seleqiio dos beneficihrios - como a responsabilidade do empristimo C coletiva, apenas os individuos com capital humano, poupanqaprCvia e conhecimentos adequados ao aproveitamento das oportunidades se selecionariam para participar do programa; eficihcia alocativa e produtiva - a formaqgo dos grupos permite um melhor acesso aos mercados, diversificaqiio de risco e gera incentivos cruzados de monitoramento do esforqo empregado naproduqiio. A tabela 3.6 ilustra o efeito dos incentivos criados sobre o custo de obtenqiio de terras para o programa. 0 aumento de demanda por terras causado pelos subsidios do Banco da Terra poderia determinar um aumento no custo das operaqdes realizadas. Entretanto, os dados mostram que a pressso induzida pel0 programa sobre as negociaqdes de compra das terras mais do que compensamesse aumento de demanda e determinam uma reduqBo significativa no preqo das Areas adquiridas. Eesse efeito, aliado aos custosjudiciais das desapropriaqdes mencionados anteriormente, fazem com que o custo por hectare do programa CCdula da Terra seja bem mais barato que o custo por hectare obtido pelo processo de desapropriaqgo. A questso fundamental por tr6s das vantagens do programa enunciadas acima decorrem da necessidadede que o contrato de emprCstimo seja exeqiiivel, com responsabilidade coletiva. Os incentivos que stio criados pel0 Banco da Terra dependemde forma essencialda necessidadede quitaqgo dos emprkstimos. Navarro (1998), por outro lado, aponta alguns pontos que podem comprometer a efichcia do programa para a alteraqiio da estrutura fundihria. A magnitude dos emprCstimos restringe as operaqdes Bs pequenas e mCdia propriedades. Os grandes latiftindios s20 excluidos do programa, o que reforqa a idCia de que o Banco da Terra constitui um programa complementar ao processode reforma agrhria com desapropriaq6es. 0 autor julga que os agricultores dificilmente pagarso os financiamentos, se mantidas as condiq6es atuais. 0 argument0 consiste na falta de condiq6es sociais de cooperaq8o entre os beneficihrios. A tradiqso de cooperaqiio entre os agricultores mais pobres, no Brasil, e 232 incipiente, restrita a algumas regiaes como o Nordeste, e quase sempre relacionada ao mer0 acesso a programas assistenciaisfinanciados pel0 governo. Dessa forma, o Banco da Terra revela-se umprograma promissor na democratizaqgo do acesso 2 terra, complementar ao programa de reforma agrhria com desapropriaqaes, cujo sucesso est6 atrelado 2 superaqgo de obst6culos importantes. AlCm da consolidaqgo do capital social mencionado por Navarro( 1998) e da exeqiiibilidade do contrato de empre'stimo, d necess6rio tambCm que o prego da terra reflita apenas as possibilidades futuras de ganhos produtivos. Na medida em que a taxag2o ngo seja capaz de inibir a utilizaq2o de terras para fins especulativos, o pagamento dos emprdstimos do Banco da Terra estargo condicionados aos subsidios oferecidos. E, ao final, os beneficiirios teriio incentivo a vender mas propriedades, uma vez que valeriio mais do que o valor presente do fluxos futuros esperados com a atividade agricola. 3.4- ProgramaNacionalde Fortalecimentoda Agricultura Familiar (PRONAF) 0 PRONAF foi criado em 1995 como uma linha de crCdito destinada 2 produggo agricola explorada em regime familiar, incluindo os assentados pel0 programa de reforma agrkia. A partir desse momento, ocorre uma ampliaqiTodas linhas de crddito e um aumento dos recursos destinados ao programa. Para a safra de 1999/2000 foram destinados R$ 3,4 bilhaes disponibilizados em nove linhas de cre'dito, abrangendo 3792 municipios brasileiros. A tabela 3.7 mostra a evoluggo do ndmero de linhas de crddito assim como o aumento do volume de recursos e o aumento dos beneficidrios. AlCm do crddito subsidiado (PRONAF - CrCdito), o govemo ainda transfere recursos orgament6rios para o desenvolvimento rural de municipios selecionados (PRONAF - Infra- Estrutura) e para a capacitaq5o e profissionalizaq8o dos agricultores familiares (PRONAF - Capacitaqb). E em 1999, foi criado o programa Parceria e Mercado, com o objetivo de promover a integraqlo dos produtores familiares nas cadeias comerciais. Fonte: Ministho da Politica Fundiiiriae do Desenvolvimento Agrtirio (1999~) Nota: (*) refere-se apenas aos primeiros6 mesesde financiamento. Os municipios candidatos 2s a@es de infra-estrutura e serviqos do PRONAF devem atender, simultaneamente, aos seguintes crite'rios tkcnicos: (i)proporqiio de estabelecimentos com at6 200 hectares maior que a me'dia estadual; (ii)taxa de urbanizaqiio menor que a verificada no Estado; e (iii) valor da produqiio agricola por pessoaaplicada na agricultura menor que a mkdia estadual. A linha de aq5o da capacitaqiio e profissionalizaqiio tem como objetivo proporcionar 233 aos agricultores familiares e suas organizaqbes: (i)conhecimentos necessfirios 2 elaboraqlo de planos municipais de Desenvolvimento Rural; (ii)conhecimentos, habilidades e tecnologia indispensaveis aos processos de produqZio, beneficiamento e comercializaqlo; e (iii) interdmbio e difuslo das experiencias coerentes com as necessidades da agricultura familiar [Silva (1999)l. A tabela 3.8 apresenta os a evoluqlo dessas duas linhas de a q b do PRONAF. Na avaliaqBo de Silva (1999), que analisou o programa no period0 1995/1998, o desempenho do crkdito rural do PRONAF foi positivo em termos do volume de recursos alocados e aplicados. 0 desempenho tambCm foi devido B diminuiqBo dos encargos financeiros incidentes sobre o crCdito. 0 autor ainda salienta que inicialmente havia uma forte concentraqlo das aplicaqaes nos estados da regilo Sul, que detinha 78,9% do total dos recursos em 1996. E esse movimento est6 sendo revertido, de modo que o Nordeste, que detinha 6% dos recursos em 1996, em 1998, concentra 37,3%. 54.299 1 41.597 I 915 I 136.645 1 ! 71.600 I 31.761 f 1.018 151.693 4 Fonte: Ministkrio da Politica Fundiiria e do DesenvolvimentoAgriirio (1999~) Outra caracteristica apontada por Silva (1999) C o viCs do programa na direqiio dos produtores familiares mais capitalizados, integrados Bs cadeias ago-industriais, tipicamente voltados ao cultivo do fumo, milho e soja. Dessa forma, para alcanqar os objetivos do programa, 6 necessaria a incorporaqlo daqueles agricultores que se encontram fora dos esquemas de integraqiio agro-industrial e pertencem 2s regibes mais afastadasdo pais. Portanto, o PRONAF vem cumprindo um importante papel no fortalecimento de pequenos produtores que, como salientado na tabela 1.3, sgo mais eficientes na exploraqiio de suas keas. Mesmo que o programa ainda enfrente algumas restriqbes em sua abrangencia, constitui um instrumento fundamental para o aprimoramento da eficiCncia agricola brasileira na medida que viabiliza a realizaqgo de todo o potencialprodutivo dos agricultores familiares. 4- Lit$es e Propostas Os dados apresentados nas seqbes anteriores foram organizados em tomo da distribuiqlo de terras, que constitui o principal objeto de analise do artigo. A seqlo Imostrou que a distribuiqlo de terras C injusta, na medida em que impbe condiqbes de vida no campo bastanteinferiores 2s da cidade e C explorada de maneira bastante assimdtrica. Pequenos produtores, ocupando um terqo das terras e 78% dos trabalhadores rurais, convivem com uma agricultura patronal cuja exploraq5o da terra ocorre de forma mais extensiva. Na seqgo 11, os resultados apontam que a distribuiqlo de terras C ineficiente, devido a 3 razbes. 0 alto grau de urbanizaqgoe a industrializaqlo brasileirageraram uma desigualdadeno acesso 2s oportunidades que dificilmente pode constituir o alvo de politicas de desenvolvimento. A especializaqlo produtiva pode ser parte inerente ao desenvolvimento econbmico, o que descartariaessa condiqlo como parte integrante das politicas pliblicas. 234 A outra raziio C a existencia de imperfeiqdes de mercado que determinam uma demanda por terra para fins niio-produtivos. 0 us0 da propriedade da terra como colateral ou como um ativo que serve como seguro contra instabilidades macroeconbmicas produz uma elevaqiio artificial em seu preqo, que deixa de refletir apenas o valor esperado dos fluxos futuros dos retomos da atividade agricola. A terceira condiqiio necessaria para a ineficiencia C o mal funcionamento do mercado de arrendamentos. Na literatura econ6mica, esse problema est6 sempre associado B assimetria de informaqiio que remete ao estabelecimento de contratos de parceria ao invCs de arrendamento. Caso houvesse um mercado de arrendamento perfeito, a propriedade da terra seria completamente irrelevante do ponto de vista de eficikncia da atividade agn'cola, uma vez que o acesso B terra seria garantido por esse mercado. E, diante de uma distribuiqiio de terras que C injusta e ineficiente, a seqgo 111apresentou as politicas empreendidas pel0 govemo. 0 Imposto Territorial Rural C o primeiro instrumento descrito. A taxaq5o de terras tem como objetivo niio apenas a obtenq5o de recursos que possam ser revertidos em politicas de desenvolvimento mas, principalmente, a inibiqiio do us0 de terras para especulaqiio. Como foi demonstrado na seqiio 11esse C um componente importante da ineficiencia que poderia ser mitigado pela adoqb de um mecanismo de taxaqiio efetivo, que reduzaa parcela do preqo da terra que provCm de seu us0para fins niio-produtivos. A experiencia brasileira com o ITR demonstra que o imposto est6 longe de ser efetivo. Apesar dos problemas operacionais levantados por Reydon et al. (2000), a literatura te6rica demonstra restriqdes estruturais quanto B aplicaqiio do imposto. Assunqb e Moreira (2000) provam que o ITR C incapaz de implementar um esquema efetivo de taxasgo em um context0 como o brasileiro, onde h6 terra ociosa em equilibrio e o governo niio observa os parametros de produtividade. A soluq50 seria adotar um esquema de taxaqiio misto, em que a informaqiio proveniente do imposto sobre produto (ICMS - Imposto sobre Circulaqiio de Mercadorias e Serviqos) fosse utilizada no cfilculo do ITR. Dessaforma, diante de um sistema de taxaqiio de terras improdutivas ineficaz, o preqo da terra, refletindo seu valor especulativo, impede o acesso dos pequenos produtores e toma os efeitos das politicas de redistribuiq5o (Reforma Agr6ria e Banco da Terra) transit6rios. Uma vez que as propriedades mantenham um valor nBo-produtivo, os beneficiarios niio tem condiqdes de det&- las sem o subsidio direto do governo. Portanto, para combater a ineficiencia da distribuiqiio de terras e tornar as politicas de redistribuiqiio sustentaveis a longo prazo, C necess6ria uma reformulaqtio no esquema de cobransa do lTR, vinculando-o ao ICMS. Em seguida, a seqiio 111 apresenta os programas de redistribuiq5o de terras. 0 Programa Nacional de Reforma Agr6ria vem concentrando o maior volume de recursos destinados ao desenvolvimento rural nos dltimos anos. As informaqdes apresentadasevidenciam a dimensiio e a abrangencia do programa, que entre 1995 e 2000, assentou mais de 2 milhdes de pessoas. 0 programa tem se mostrado bastante efetivo em alcanqar o pdblico desejado, entretanto, sua implementaqBotem encontrado algumas dificuldades. Os custos judiciais tem elevado em cinco vezes em mCdia o custo das desapropriaqdes. E, alCm disso, o mecanismo de seleqiio dos beneficiarios e das 6reas destinadas aos assentamentos tem apresentado alguns problemas, como mostra Bittencourt et al. (1999). Em alguns assentamentos, as condiqdes que restringiam a produqiio dos agricultores familiares n50 foram revertidas pela estrutura montada. 0 Banco da Terra, ao contrario do programa de reforma agraria combaseem desapropriaqdes, destina-se a oferecer crCdito para a compra de pequenas propriedades. E, devido a essa caracten'stica, constitui um programa complementar B reforma agrhia tradicional. 0 programa, apesar de incipiente, demonstrou was virtudes na seleqiio dos beneficikios e, principalmente, na reduqiio do custo das terras obtidas. U m fator importantepara o funcionamento do Banco da 235 Terra C a chamada consolidaqBo do capital social salientada por Navarro (1998). Como o programa requer que os beneficihrios estejam organizados em associaqdes ou cooperativas, que tCm uma responsabilidade conjunta sobre os projetos individuais, o seu desempenho est6 atrelado B capacidadede interaqgo social dos grupos formados. 0 incentivo ?i deassociaqdesecooperativas,induzidopel0BancodaTerra,constitui formaqgo uma Area promissora de atuaqgo do governo. A base do argument0 de eficiCncia na distribuiqgo de terras desenvolvido no artigo estabelece uma diferenqa entre a produtividade dos grandes latifdndios em relaqgo aos agricultores familiares. E assim, qualquer aumento na produtividade dos pequenos produtores em relaggo aos grandes proprietkios C capaz de ampliar os ganhos de eficiencia obtidos pelas politicas de redistribuiqiio. Dentre as restriqdes que s50 impostas aos pequenos produtores, estgo aquelas relacionadas a indivisibilidades tecnol6gicas e ao acesso a mercado de insumos, beneficiamento e comercializaqgo dos produtos decorrentes da pequena escala de produqgo. A formaqlo de grupos de produtores possibilita uma elevaqiio na escalaprodutiva e, comisso, uma melhoriano acesso a tecnologias mais produtivas, assistencia tCcnica, mhquinas e equipamentos, e outras facilidades que n9o estariam disponiveis aos produtores individualmente. AlCm disso, a consolidaqiio desses grupos possibilita a a@o de programas de microcrCdito e outros programas baseados em responsabilidade coletiva, como o Banco da Terra. 0 PRONAF, tambCm descrito na se@o ID, tem apresentado resultados satisfatdrios no fortalecimento da agricultura familiar. Apesar de que alguns grupos ainda ngo foram atingidos pel0 programa, os resultados demonstram o esforqo empreendido pel0 governo nas diversas linhas de crCdito e nas aqdes de capacitaqgo e infra-estrutura. 0 PRONAF, ao aprimorar a competitividade dos agricultores familiares, passa a ter impactos significativos sobre a eficiencia agn'cola e sobre os resultados dos programas de redistribuiqgo. Por fim, apesar de que a ineficigncia da distribuiqb de terras decorre do us0 de terras para fins ngo-produtivos e de imperfeiqdes no mercado de arrendamento, o governo tem focalizado was politicas em outras 6reas, que sgo relevantes mas podem ter seus efeitos limitados pela presenqa dessas condiqdes. Como foi visto, o atual mecanismo de cobranqa do lTR n2o est6 sendo capaz de inibir a manutensgo de terra ociosa, por parte dos grandes propriethrios. Para alcanqar esse objetivo, o esquemade cobranqa deve ser reformulado. 0 malfuncionamento do mercado de arrendamento de terras deveria constituir maisumfoco de atuaqiio das politicas pdblicas voltadas ao meio rural. Mesmo em um ambiente em que a terra C fortemente utilizada para especulaqHo, o funcionamento do mercado de arrendamento garante o acesso 2 terra. Como mencionado anteriormente na seqtio 11, entretanto, esse mercado tem certas limitaqdes decorrentes da assimetria de informaqgo entre proprietkios e arrendatirios. Mas, como no B r a d esse C ummercado pouco explorado -pouco mais de 3% da area total dos estabelecimentos agn'colas encontra-se sob contrato de arrendamento ou parceria - muito espaqo para melhoria com a reduqgo de barreiras institucionais e incentivos contrarios. E os efeitos desses aprimoramentos sobre a eficiencia sHo bastante claros na estrutura te6rica apresentadana se@o 11. Concluindo, apesar de que a distribuiqiio de terras no Brasil C injusta e ineficiente, o governo tem conseguido alterar sensivelmente a estrutura fundiaria - o coeficiente de Gini da distribuiqiio de terras caiu de 0.848 em 1992 para 0.802 em 2000. Entretanto, a travessia que a agricultura brasileira iniciou nos filtimos anos C longa e envolve caminhos que ainda n2o foram explorados pelas politicas pdblicas. Com base nos resultados e argumentos apresentados, o artigo sugere que existem pel0 menos tr&s trilhas que deveriam ser experimentadas: a reformulaqHo do esquema de cobranqa do ITR, o incentivo 2 formaqkio de associaqdes de produtores e cooperativas, e o aprimoramento do mercado de arrendamento de terras. 236 Bibliografia Aghion, P. e P. Bolton (1997) "A theory of trickle-down growth and development", Review of Economic Studies,64: 151-172. Alesina, A. e D.Rodrik (1994) "Distributive politics and economic growth", Quarterly Journal of Economics, 108: 465-90. Arnott, R. J. e J. E. Stiglitz (1979) "Aggregate land rents, expenditure on public goods, and op- timal city size", Quarterly Journal of Economics, 93(4): 471-500. Assunsgo, J. J. e H. Moreira (2000) "lTR sem mentiras: um comentfirio sobre a taxas50 de te- rras com informasgo assimktrica", PUC-Rio: Departamento de Economia, mimeo. Avila, A. F. D. e R. Evenson (1998) "Total factor productivity growth in Brazilian agriculture and the role of agricultural research", Economia Aplicada, 2(2): 317-356. Barros, R. P.; R. Mendonsa; P. P. Deliberalli e C. Lopes (2000) "Impactos da distribuicao da terra sobre a eficiencia agricola e a pobreza no nordeste", Seminfirio Desajios da Pobreza Rural no Brasil, Rio de Janeiro: Rio Othon Palace. Banerjee, A. V. e A. F. Newman (1993) "Occupational Choice and the Process of Develop- ment", Journal of Political Economy, lOl(2): 274-298. Bittencourt, G. A.; D. S. B. Castilhos; V. Bianchini e H. B. C. Silva (1999) "Principais fatores que afetam o desenvolvimento dos assentamentosde reforma agr6ria no Brasil", Brasilia, INCRAFAO: Projeto UTF/BRA/036/BRA. Brandgo, A. S. P. e G. C. Rezende (1992) "Credit subsidies, inflation and the land market in Brazil: a theoretical and empirical analysis", World Bank, mimeo. Buainain, A. M.; J. M. Silveira e E. Te6filo (1999) "Reforma agr&ia, desenvolvimento e parti- cipaqb: uma discussb das transformagtjes necessfirias e possiveis", Brasilia, NEAD, mimeo. Cardoso, F. H. (1975) "Dos Govemos Militares a Prudente-Campos Sales" in B. Fausto (ed.) 0 B r a d Republicano, Siio Paulo: Difel. Carter, M. R. e D. Mesbah (1993) "Can land market reformmitigate the exclusionary aspects of rapid agroexport growth?", World Development, 21(7): 1085-1100. Deininger, K. e G. Feder (1998) "Land institutions and land markets", World Bank: Policy Re- searchWorking Papers (forthcoming in Handbook of Agricultural Economics). Deininger, K. e P. Olinto (2000) "Asset distribution, inequality, and growth", World Bunk Working Paper #2375, mimeo. Eswaran, M.e A. Kotwal (1985) "A theory of contractual structure in agriculture", 75(3): 352- 367. Feldstein, M. (1980) "Inflation, portfolio choice and the prices of land and corporate stock", American Journal of Agricultural Economics, 62: 910-916. Galor, 0. e J. Zeira (1993) "Income distribution and macroeconomics", Review of Economic Studies,60: 35-52. Goldin, I. G. C. Rezende (1993) "A agricultura brasileira na dkcada de 80: crescimento numa e economia em crise", Rio de Janeiro: Se`rie ZPEA - no. 138. Ghatak, M. e P. Pandey (2000) "Contract choice in agriculture with joint moral hazard in effort and risk`', Journal of Development Economics, 63(2): 303-326. 237 Guanziroli, C. E. e S. E. C. S. Cardim (orgs.) (2000) "Novo retrato da agricultura familiar: o Brasil redescoberto", Projeto de Cooperagtio Tkcnica INCMFAO, mimeo. Hoff, K. (1991) "Land taxes, output taxes, and sharecropping: was Henry George right?", WorldBank Economic Review,5( 1): 93-111. Hoff, K.; A. Bravermane J. E. Stiglitz (eds) (1993) The economics of rural organization: the- ory, practice and policy. New York: OxfordUniversityPress. MinistCrioda PoliticaFundihriae do DesenvolvimentoAgrhrio (1999a) 0 livro branco das su- perindenizag6es: como darfim a essa indhtria, Brasilia:Assessoria de ComunicagBoSo- cial. MinistCrioda Politica Fundihriae do DesenvolvimentoAgrsrio (1999b) Banco da Terra,Brasi- lia. MinistCrio da Politica Fundihriae do DesenvolvimentoAgrsrio (1999~)Ofituro nasce da te- rra: balango da reforma agra`ria e da agriculturafamiliar, Brasilia,mimeo. Mirrlees, J. A. (1971) "An exploration in the theory of optimum income taxation", Review of Economic Studies,38: 175-208. Mirrlees, J. A. (1986) "The theory of optimal taxation". In: Handbook of Mathematical Eco- nomics, vol. 3, K.J. Arrow e M.D. Intriligator, eds., Amsterdam: North-Holland. Navarro, Z. (1998) "0 projeto-piloto Ckdula du Terra - comentirio sobre as condi@es sociais e politico-institucionaisde seu desenvolvimento recente", mimeo. Oliveira, J. T. e I.N. Costa (1979) 0 Impost0 Territorial Rural - avaliagtio econdmica, SBo Paulo: IPE/USP, Relat6riode Pesquisano. 2. Oliveira, J. T. (1983) "0 imposto sobre a propriedade territorial rural 1964-1992", Estudos Econdmicos, 23: 209-224. Otsuka, K. H. Chuma e Y. Hayami (1992) "Land and labor contracts in agrarian economies: theories andfacts", Journal of Economic Literature, 30: 1965-2018. Otsuka, K.e Y. Hayami (1988) "Theories of share tenancy: a critical survey", Economic Devel- opment and Cultural Change, 37: 31-68. Person, T. e G. Tabellini (1994) "Is inequality harmful to growth?', American Economic Re- view, 84(3): 600-621. Reydon, B. P. e L.A. Plata (2000) "EvoluqBo recente do preqo da terra rural no Brasil e os im- pactos do Programada CCdulada Terra", Brasilia,M A D , mimeo. Reydon, B. P.; A. R. Romeiroe L.A. Plata (2000) "Aspectos da questgo agrhria brasileira: lip 6es hluz do mercado de terras", Brasilia, INCRA/FAO: ProjetoUTF/BRA/036/BRA. Rezende, G. C. (1999) "Programa de crCdito especial parareformaagrhria (Procera): institucio- nalidade, subsidio e efichcia", Brasilia,INCRA/FAO: ProjetoUTF/BRA/036/BRA. Romeiro, A. R.; B. P. Reydone L.A. Plata(2000) "Impact0 do lTR nos presos da terra na con- centraGB0fundisria", Brasilia,INCRAIFAO: ProjetoUTF/BRA/036/BRA. Sayad, J. (1992) "EspeculaGBo em terras rurais, efeitos sobre a produqBo agricola e o novo ITR', Pesquisa e Planejamento EconGmico, 12(1): 87-108. Secretaria de ComunicaqBo de Govern0 da Presidencia da RepBblica. Reforma agra`ria: com- promisso de todos. Brasilia, extraido em 10/06/2000 de www.planalto.gov.br/secom/colecao/refagr8.htm. Skinner, J. (1991a) "If agricultural land taxation i s so efficient, why i s it so rarely used?", World Bank Economic Review, 5( 1): 113-133. 238 Silva, E. R. A. (1999) "Programa nacional de fortalecimentoda agriculturafamiliar - relat6rio tCcnico das a@es desenvolvidasno period0 1995/1998", Brasilia: PEA, Text0 para Dis- cussclo no. 664. Skinner, J. (1991b) "Prospects for agricultural land taxation in developing countries", World Bank Economic Review,5(3): 493-511. Stiglitz, J. (1974) "Incentives and Risk-sharingin Sharecropping", Review of Economic Studies, 41: 219-255. Stiglitz, V. L. e A. Weiss (1981) "Credit rationing in markets with imperfect information", American Economic Review, 71(3):393-410. Stock, J. H. e M. W. Watson (1988) "Testing for Common Trends", Journal of the American Statistical Association, 83:1097-1107. 239