American Political Science Review (2018)112.4,1096-1103 doi:10.1017/S0003055418000308 American Political Science Association 2018 Letter Ethnoracial Homogeneity and Public Outcomes:The (Non)effects of Diversity ALEXANDER KUSTOV Princeton University GIULIANA PARDELLI Princeton University ow does ethnoracial demography relate to public goods provision?Many studies find support for the hypothesis that diversity is related to inefficient outcomes by comparing diverse and ho- mogeneous communities.We distinguish between homogeneity of dominant and disadvantaged groups and argue that it is often impossible to identify the effects of diversity due to its collinearity with the share of disadvantaged groups.To disentangle the effects of these variables,we study new data from Brazilian municipalities.While it is possible to interpret the prima facie negative correlation between diversity and public goods as supportive of the prominent"deficit"hypothesis,a closer analysis reveals that,in fact,more homogeneous Afro-descendant communities have lower provision.While we cannot rule out that diversity is consequential in other contexts,our results cast doubt on the reliability ofprevious findings related to the benefits of local ethnoracial homogeneity for public outcomes. INTRODUCTION ticular group shares.We thus argue that,to properly identify the relationship of ethnic diversity and public ow does public goods provision'relate to eth- outcomes,one needs to compare diverse communities 4号元 noracial demography?Political scientists and to homogeneous communities of all groups rather than economists seemed to have reached a consen of a single(usually dominant)group in society,which sus regarding the existence of a robust association be- is nonetheless impossible in many previously studied & tween diversity and a variety of negative social out- contexts.To overcome this limitation,we focus on the comes (i.e.,"diversity deficit").Despite the scarcity of empirically relevant-yet largely overlooked-case of support for a causal link,the sheer number of studies Brazil,which allows us to distinguish between homo- showing diversity to harm provision sufficed to con- geneous local populations composed of either domi- vince the most skeptical of readers.More recently,how- nant or disadvantaged groups.2 When the appropriate ever,these earlier findings have been challenged both group share measures are taken into account,results empirically and theoretically. show that diversity has no discernible effect on public This paper contributes to this ongoing debate by demonstrating that the previously uncovered effects of goods provision. In what follows,we first discuss the limitations of diversity can often be confounded with those of par- previous tests of the diversity hypothesis and empha- size the distinction between the use of group share and Alexander Kustov is a PhD Candidate,Department of Poli- diversity measures (e.g.,fractionalization).To tackle tics.Princeton University.001 Fisher Hall.Princeton.NJ 08544 these issues.we make the case for the analysis of munic- (akustov@princeton.edu). ipal outcomes in the racially diverse and highly decen- Giuliana Pardelli is a PhD Candidate,Department of Poli- tralized case of Brazil.We then show that,when we use tics,Princeton University,001 Fisher Hall,Princeton,NJ 08544 the model specifications adopted in previous studies, (pardelli@princeton.edu). The authors'names appear in alphabetical order.An earlier ver. diversity seems to be negatively correlated with pub- sion of the paper was presented at the 2016 annual meeting of the lic goods,even after controlling for a variety of con- American Political Science Association.We would like to thank founding factors.While this result can be seen as sup- our colleagues,editors,and anonymous reviewers who have read porting the standing hypothesis,a closer examination and commented on previous drafts of this article.We are espe- of the evidence reveals that diversity is not detrimental cially grateful to Samuel Diaz,Mark Kayser,and Ronald Ingle- hart for their helpful suggestions,and Joana Naritomi for kindly per se,but only insofar as it reflects an increase in the sharing her data with us For their useful comments on the pre- share of the disadvantaged group in the local popula- vious versions of our larger project on ethnic cleavages and pub tion.Thus,after re-examining the data and including lic goods provision,we would also like to thank Rafaela Dancy group share measures,we find that,in fact,more homo- gier,Kosuke Imai,Tali Mendelberg,Grigore Pop-Eleches,Edward Telles,Andreas Wimmer,and Deborah Yashar.All errors and omis- geneous Afro-descendant municipalities have worse sions are the sole responsibility of the authors.Replication files are available at the American Political Science Review Dataverse: https://doi.org/10.7910/DVN/AY32JZ. 2 We use the term "disadvantaged"merely to emphasize that eth- Received:August 7 2017;revised:February 19,2018;accepted:May nic groups that are relatively deprived along a given dimension 18,2018.First published online:June 19,2018 (Horowitz 1985)-and conventionally referred to as "minorities"- may constitute demographic majorties.Since social,economic,and I We follow the literature and use "public goods provision"as a gen- political disparities between groups tend to be strongly correlated eral term for government-provided public services such as education and hardly dissociable in many contexts(Stewart 2005).including health care,and infrastructure,even when they do not fit the strict that of Brazil (Bailey 2009),we are agnostic about which particular economic description (i.e.,nonexcludable and nonrivalrous goods) dimension of disadvantage is more consequential. 1096
American Political Science Review (2018) 112, 4, 1096–1103 doi:10.1017/S0003055418000308 © American Political Science Association 2018 Letter Ethnoracial Homogeneity and Public Outcomes: The (Non)effects of Diversity ALEXANDER KUSTOV Princeton University GIULIANA PARDELLI Princeton University How does ethnoracial demography relate to public goods provision? Many studies find support for the hypothesis that diversity is related to inefficient outcomes by comparing diverse and homogeneous communities. We distinguish between homogeneity of dominant and disadvantaged groups and argue that it is often impossible to identify the effects of diversity due to its collinearity with the share of disadvantaged groups. To disentangle the effects of these variables, we study new data from Brazilian municipalities. While it is possible to interpret the prima facie negative correlation between diversity and public goods as supportive of the prominent “deficit” hypothesis, a closer analysis reveals that, in fact, more homogeneous Afro-descendant communities have lower provision. While we cannot rule out that diversity is consequential in other contexts, our results cast doubt on the reliability of previous findings related to the benefits of local ethnoracial homogeneity for public outcomes. INTRODUCTION How does public goods provision1 relate to ethnoracial demography? Political scientists and economists seemed to have reached a consensus regarding the existence of a robust association between diversity and a variety of negative social outcomes (i.e., “diversity deficit”). Despite the scarcity of support for a causal link, the sheer number of studies showing diversity to harm provision sufficed to convince the most skeptical of readers.More recently, however, these earlier findings have been challenged both empirically and theoretically. This paper contributes to this ongoing debate by demonstrating that the previously uncovered effects of diversity can often be confounded with those of parAlexander Kustov is a PhD Candidate, Department of Politics, Princeton University, 001 Fisher Hall, Princeton, NJ 08544 (akustov@princeton.edu). Giuliana Pardelli is a PhD Candidate, Department of Politics, Princeton University, 001 Fisher Hall, Princeton, NJ 08544 (pardelli@princeton.edu). The authors’ names appear in alphabetical order. An earlier version of the paper was presented at the 2016 annual meeting of the American Political Science Association. We would like to thank our colleagues, editors, and anonymous reviewers who have read and commented on previous drafts of this article. We are especially grateful to Samuel Diaz, Mark Kayser, and Ronald Inglehart for their helpful suggestions, and Joana Naritomi for kindly sharing her data with us. For their useful comments on the previous versions of our larger project on ethnic cleavages and public goods provision, we would also like to thank Rafaela Dancygier, Kosuke Imai, Tali Mendelberg, Grigore Pop-Eleches, Edward Telles, Andreas Wimmer, and Deborah Yashar. All errors and omissions are the sole responsibility of the authors. Replication files are available at the American Political Science Review Dataverse: https://doi.org/10.7910/DVN/AY32JZ. Received: August 7, 2017; revised: February 19, 2018; accepted: May 18, 2018. First published online: June 19, 2018. 1 We follow the literature and use “public goods provision” as a general term for government-provided public services such as education, health care, and infrastructure, even when they do not fit the strict economic description (i.e., nonexcludable and nonrivalrous goods). ticular group shares. We thus argue that, to properly identify the relationship of ethnic diversity and public outcomes, one needs to compare diverse communities to homogeneous communities of all groups rather than of a single (usually dominant) group in society, which is nonetheless impossible in many previously studied contexts. To overcome this limitation, we focus on the empirically relevant—yet largely overlooked—case of Brazil, which allows us to distinguish between homogeneous local populations composed of either dominant or disadvantaged groups.2 When the appropriate group share measures are taken into account, results show that diversity has no discernible effect on public goods provision. In what follows, we first discuss the limitations of previous tests of the diversity hypothesis and emphasize the distinction between the use of group share and diversity measures (e.g., fractionalization). To tackle these issues,we make the case for the analysis of municipal outcomes in the racially diverse and highly decentralized case of Brazil.We then show that, when we use the model specifications adopted in previous studies, diversity seems to be negatively correlated with public goods, even after controlling for a variety of confounding factors. While this result can be seen as supporting the standing hypothesis, a closer examination of the evidence reveals that diversity is not detrimental per se, but only insofar as it reflects an increase in the share of the disadvantaged group in the local population. Thus, after re-examining the data and including group share measures, we find that, in fact, more homogeneous Afro-descendant municipalities have worse 2 We use the term “disadvantaged” merely to emphasize that ethnic groups that are relatively deprived along a given dimension (Horowitz 1985)—and conventionally referred to as “minorities”— may constitute demographic majorities. Since social, economic, and political disparities between groups tend to be strongly correlated and hardly dissociable in many contexts (Stewart 2005), including that of Brazil (Bailey 2009), we are agnostic about which particular dimension of disadvantage is more consequential. 1096 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000308
Ethnoracial Homogeneity and Public Outcomes public goods provision than more diverse communities fit from the well-being of a fellow group member and and than homogeneous white majority municipalities. attach lower (or even negative)utility to the welfare of Overall,this paper challenges the "diversity deficit" the out-group(Alesina and Glaeser 2004).Although hypothesis by showing that previous subnational anal- this channel helps to explain why more diverse com- yses have often relied on contexts with"truncated' munities contribute less to the public welfare,it fails to population distributions where disadvantaged groups clarify why homogeneous localities may achieve even never reach a local demographic majority (i.e.,where poorer outcomes.Likewise,mechanisms such as pref- ethnic homogeneity is only defined for one group).In erence homogeneity,expanded technical capabilities. this sense,our results draw attention to the limited and facilitated social sanctions (e.g..see Habyarimana applicability of some of the mechanisms proposed in et al.2007)elucidate the improved ability of more ho- the literature that link diversity to negative social out- mogeneous localities to work collectively.Yet,if groups comes.In particular,we highlight that the failure to dif- are not interchangeable and homogeneous communi- ferentiate between diversity and relevant group shares ties diverge in a systematic way,it might be the case that may cast doubt on the reliability of previous findings these mechanisms do not operate in the same manner related to the "benefits"of ethnoracial homogeneity. across groups. The classic US study by Alesina et al.(1999)ac- ETHNORACIAL DEMOGRAPHY AND PUBLIC knowledges the theoretically-relevant distinction be- GOODS tween racial fractionalization and group shares.How- ever,the US has only a small number of white-minority The "diversity deficit"hypothesis has been investi- localities and,among them,few are racially homo- gated and confirmed across a wide variety of regions geneous (i.e..exhibit low fractionalization levels,see and settings (for a review,see Stichnoth and Van der Figure 1).As a result,it may be empirically difficult, Straeten 2013).However.some of the seminal studies or even unfeasible,to distinguish between the effects in this literature have been criticized for neglecting the of these two variables in this or similar contexts.Simi- 4r元 heterogeneous effects of diversity across various public larly,Schaeffer (2013)shows that,in Europe,most com- goods and for failing to address omitted variable bias peting indices of ethnic diversity are indistinguishable concerns (Gisselquist 2014;Wimmer 2016).The stan- from the mere percentage of immigrant shares.In fact, dard variable used to measure diversity,the fractional- disadvantaged ethnic groups rarely constitute local de- ization index,has also sparked considerable criticism mographic majorities in most democratic,developed (e.g.,see Abascal and Baldassarri 2015).Most impor- countries.Since diversity and group share measures tant,as a summary statistic,it treats groups as equiva- move together,in such contexts,their effects can be lent and fails to indicate which ones are represented confounded. in what proportions in the population (Vigdor 2002; Rushton 2008). EMPIRICAL STRATEGY However,given divergent histories of conflict and migration,there are strong reasons to believe that Brazil is known for being one of the most racially di- ethnic groups are rarely interchangeable (e.g.,see verse and economically unequal democracies in the Horowitz 1985;Sidanius and Pratto 2001).In fact world.Despite this fact,the influence of ethnic demog- between-group disparities tend to be rather ubiquitous, raphy on public goods provision has not yet been in- strikingly persistent and often multidimensional(Tilly vestigated within the country's territory.4 We contend 1999;Stewart 2005).This is important because the over- that the study of Brazilian municipalities can greatly lap of ethnicity and individual socioeconomic charac. contribute to our understanding of the link between teristics may produce an apparent negative association ethnoracial demography and social outcomes for sev- between diversity and social outcomes even if it is,in eral reasons fact,a result of individual and contextual indicators of First,municipalities in Brazil provide a large num- well-being (e.g.,Abascal and Baldassarri 2015).Due ber of comparable cases that reflect consistent politi- to such“compositional effects,”for instance,“major cal jurisdictions,share the same electoral rules,and ex- ity black and minority white"communities may sys hibit wide variation in the dependent variables of in- tematically underperform "majority white and minor- terest.Second,and related,the country's high level of ity black"communities in terms of public outcomes,de- political decentralization implies that the responsibil- spite having the same level of diversity. ity for providing public goods is in the hands of mu- eys Nonetheless.the most commonly used mechanisms nicipal governments.This,in turn,guarantees that our in the literature to elucidate how diversity affects social outcomes are tightly linked to political decisions at the outcomes also assume that ethnic groups are analogous local level rather than at other levels of government.5 and behave in the same manner.According to the"in- group bias"mechanism,for instance,individuals bene- 4 This is particularly surprising given the large amount of studies on the determinants of public expenditures and the vast literature on 3As a measure of diversity,the Herfindahl-Hirschman fractionaliza racial relations in the country (e.g,see Telles 2006).A number of studies have,however,included Brazil as a case in their cross-national tion index indicates the probability that two randomly chosen indi- analyses on the effects of ethnic diversity (e.g.,see La Porta et al.1999: viduals in a community belong to different groups (Alesina et al. Alesina et al.2003;Baldwin and Huber 2010). 1999)F1-where is the proportion of group i in a s To further minimize nonmunicipal influences,we focus on local- locality. level outcomes that are under exclusive municipal responsibility in 1097
Ethnoracial Homogeneity and Public Outcomes public goods provision than more diverse communities and than homogeneous white majority municipalities. Overall, this paper challenges the “diversity deficit” hypothesis by showing that previous subnational analyses have often relied on contexts with “truncated” population distributions where disadvantaged groups never reach a local demographic majority (i.e., where ethnic homogeneity is only defined for one group). In this sense, our results draw attention to the limited applicability of some of the mechanisms proposed in the literature that link diversity to negative social outcomes. In particular, we highlight that the failure to differentiate between diversity and relevant group shares may cast doubt on the reliability of previous findings related to the “benefits” of ethnoracial homogeneity. ETHNORACIAL DEMOGRAPHY AND PUBLIC GOODS The “diversity deficit” hypothesis has been investigated and confirmed across a wide variety of regions and settings (for a review, see Stichnoth and Van der Straeten 2013). However, some of the seminal studies in this literature have been criticized for neglecting the heterogeneous effects of diversity across various public goods and for failing to address omitted variable bias concerns (Gisselquist 2014; Wimmer 2016). The standard variable used to measure diversity, the fractionalization index,3 has also sparked considerable criticism (e.g., see Abascal and Baldassarri 2015). Most important, as a summary statistic, it treats groups as equivalent and fails to indicate which ones are represented in what proportions in the population (Vigdor 2002; Rushton 2008). However, given divergent histories of conflict and migration, there are strong reasons to believe that ethnic groups are rarely interchangeable (e.g., see Horowitz 1985; Sidanius and Pratto 2001). In fact, between-group disparities tend to be rather ubiquitous, strikingly persistent and often multidimensional (Tilly 1999; Stewart 2005).This is important because the overlap of ethnicity and individual socioeconomic characteristics may produce an apparent negative association between diversity and social outcomes even if it is, in fact, a result of individual and contextual indicators of well-being (e.g., Abascal and Baldassarri 2015). Due to such “compositional effects,” for instance, “majority black and minority white” communities may systematically underperform “majority white and minority black” communities in terms of public outcomes, despite having the same level of diversity. Nonetheless, the most commonly used mechanisms in the literature to elucidate how diversity affects social outcomes also assume that ethnic groups are analogous and behave in the same manner. According to the “ingroup bias” mechanism, for instance, individuals bene- 3 As a measure of diversity, the Herfindahl-Hirschman fractionalization index indicates the probability that two randomly chosen individuals in a community belong to different groups (Alesina et al. 1999): F = 1 − N i=1 π2 i , where πi is the proportion of group i in a locality. fit from the well-being of a fellow group member and attach lower (or even negative) utility to the welfare of the out-group (Alesina and Glaeser 2004). Although this channel helps to explain why more diverse communities contribute less to the public welfare, it fails to clarify why homogeneous localities may achieve even poorer outcomes. Likewise, mechanisms such as preference homogeneity, expanded technical capabilities, and facilitated social sanctions (e.g., see Habyarimana et al. 2007) elucidate the improved ability of more homogeneous localities to work collectively. Yet, if groups are not interchangeable and homogeneous communities diverge in a systematic way,it might be the case that these mechanisms do not operate in the same manner across groups. The classic US study by Alesina et al. (1999) acknowledges the theoretically-relevant distinction between racial fractionalization and group shares. However, the US has only a small number of white-minority localities and, among them, few are racially homogeneous (i.e., exhibit low fractionalization levels, see Figure 1). As a result, it may be empirically difficult, or even unfeasible, to distinguish between the effects of these two variables in this or similar contexts. Similarly, Schaeffer (2013) shows that,in Europe,most competing indices of ethnic diversity are indistinguishable from the mere percentage of immigrant shares. In fact, disadvantaged ethnic groups rarely constitute local demographic majorities in most democratic, developed countries. Since diversity and group share measures move together, in such contexts, their effects can be confounded. EMPIRICAL STRATEGY Brazil is known for being one of the most racially diverse and economically unequal democracies in the world. Despite this fact, the influence of ethnic demography on public goods provision has not yet been investigated within the country’s territory.4 We contend that the study of Brazilian municipalities can greatly contribute to our understanding of the link between ethnoracial demography and social outcomes for several reasons. First, municipalities in Brazil provide a large number of comparable cases that reflect consistent political jurisdictions, share the same electoral rules, and exhibit wide variation in the dependent variables of interest. Second, and related, the country’s high level of political decentralization implies that the responsibility for providing public goods is in the hands of municipal governments. This, in turn, guarantees that our outcomes are tightly linked to political decisions at the local level rather than at other levels of government.5 4 This is particularly surprising given the large amount of studies on the determinants of public expenditures and the vast literature on racial relations in the country (e.g., see Telles 2006). A number of studies have, however,included Brazil as a case in their cross-national analyses on the effects of ethnic diversity (e.g., see La Porta et al.1999; Alesina et al. 2003; Baldwin and Huber 2010). 5 To further minimize nonmunicipal influences, we focus on locallevel outcomes that are under exclusive municipal responsibility in 1097 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000308
Alexander Kustov and Giuliana Pardelli FIGURE 1.The Distribution of Racial Demography across US Localities. Cities Counties Metro areas 00 100 100 r=-0.86 r=-0.95 -0.98 0.75 0.75 0.75 0.50 0.50 0.50 025 0.25- 025 0.00 0.00 0.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 100 Fractionalization Fractionalization Fractionalization Each dot represents local racial demography in terms of fractionalization or group shares(whites).The graph is based on the data from Alesina et al.(1999). Finally,and most important,Brazil offers enough varia- Data tion in the local predominance of racial groups to allow for a clear empirical differentiation between this vari- We use a new purpose-built dataset of 5,505 Brazilian able and diversity.The country has a near equal pro- municipalities(2010),including a variety of racial de- & portion of African and European descendants(50.74% mography,public goods,and economic geography vari- negros and 4773%brancos),and almost as many ma- ables(for more details,see Appendix).Individual-level jority white as majority black municipalities-which census data are used to construct the indices of racial may display the same level of diversity despite having fractionalization and group shares at the smallest po- a rather different population composition(Figure 2). litically relevant administrative unit (municipalities).s Our model specification builds on the classic US Additionally,we examine a range of dependent vari- study of Alesina et al.(1999)and its subsequent repli- ables to identify the (potentially)diverging effects of cation and extension by Gisselquist(2014).We regress racial divisions on different types and aspects of service a set of public outcomes related to local service pro- provision.These variables include the total amount of vision on different racial demography'measures and public resources allocated to social spending,disag- control for the most relevant confounders identified in gregated spending indicators,and two different mea- the literature.In particular,our analysis differentiates sures of public goods quality.Our covariates incorpo- between the three most relevant "dimensions of dis- rate a set of other municipal characteristics that influ- advantage"recognized in the case of Brazil:race,class. ence the capacity of local governments to provide pub- and geographic location.By controlling for the average lic services,such as size of the locality,age,education, income,proportion of poor population,regional loca- urbanization rate,local GDP,interpersonal inequality tion,and geographic characteristics of municipalities, (GINI),poverty rate,as well as geography (Naritomi we thus distinguish between the effects of these differ- et al.2012)(for summary statistics,see Table A1). ent local features on provision,but also minimize the concern that group shares or diversity may be merely ANALYSIS AND RESULTS proxying for other types of group disadvantage. Our analysis is divided into two steps.First,we repli- cate the model used in the seminal US study of Alesina et al.(1999)using municipal-level data from Brazil in 2010.Results from this estimation.shown in Table 1. eys Brazil (municipal schools,hospitals,and so on).Additionally,to take suggest that the relationship between racial fraction- potential state interventions into account,we include state fixed ef- alization and public goods provision in Brazil is very fects in our regression analyses.Finally,although federal interference similar to the one observed in the US.More specif- in local affairs occasionally takes place,there is no evidence that it ically,higher diversity seems to be related to higher is systematically tied to the racial composition of municipalities and therefore should not affect our results. overall government expenditure,but lower education 6 This commonly used classification encompasses both Brown(par dos,43.13%)and Black (pretos,761%)Census categories.Other cat- egories include Asian (amarelos,109%)and Indigenous(indigenas, 8 Our diversity measure considers each one of the census categories 0.43%)populations. as a separate group,but our results are also robust to the use of Race has been shown to be the most salient ethnic cleavage in the an alternative fractionalization measure based on a unified Afro- country (for a detailed comparison,see Lieberman and Singh 2012). descendant category composed of pardos and pretos (not shown). 1098
Alexander Kustov and Giuliana Pardelli FIGURE 1. The Distribution of Racial Demography across US Localities. r = −0.86 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 Fractionalization Whites, share Cities r = −0.95 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 Fractionalization Counties r = −0.98 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 Fractionalization Metro areas Each dot represents local racial demography in terms of fractionalization or group shares (whites). The graph is based on the data from Alesina et al. (1999). Finally, and most important, Brazil offers enough variation in the local predominance of racial groups to allow for a clear empirical differentiation between this variable and diversity. The country has a near equal proportion of African and European descendants (50.74% negros6 and 47.73% brancos), and almost as many majority white as majority black municipalities—which may display the same level of diversity despite having a rather different population composition (Figure 2). Our model specification builds on the classic US study of Alesina et al. (1999) and its subsequent replication and extension by Gisselquist (2014). We regress a set of public outcomes related to local service provision on different racial demography7 measures and control for the most relevant confounders identified in the literature. In particular, our analysis differentiates between the three most relevant “dimensions of disadvantage” recognized in the case of Brazil: race, class, and geographic location. By controlling for the average income, proportion of poor population, regional location, and geographic characteristics of municipalities, we thus distinguish between the effects of these different local features on provision, but also minimize the concern that group shares or diversity may be merely proxying for other types of group disadvantage. Brazil (municipal schools, hospitals, and so on). Additionally, to take potential state interventions into account, we include state fixed effects in our regression analyses. Finally, although federal interference in local affairs occasionally takes place, there is no evidence that it is systematically tied to the racial composition of municipalities and therefore should not affect our results. 6 This commonly used classification encompasses both Brown (pardos, 43.13%) and Black (pretos, 7.61%) Census categories. Other categories include Asian (amarelos, 1.09%) and Indigenous (indígenas, 0.43%) populations. 7 Race has been shown to be the most salient ethnic cleavage in the country (for a detailed comparison, see Lieberman and Singh 2012). Data We use a new purpose-built dataset of 5,505 Brazilian municipalities (2010), including a variety of racial demography, public goods, and economic geography variables (for more details, see Appendix). Individual-level census data are used to construct the indices of racial fractionalization and group shares at the smallest politically relevant administrative unit (municipalities).8 Additionally, we examine a range of dependent variables to identify the (potentially) diverging effects of racial divisions on different types and aspects of service provision. These variables include the total amount of public resources allocated to social spending, disaggregated spending indicators, and two different measures of public goods quality. Our covariates incorporate a set of other municipal characteristics that influence the capacity of local governments to provide public services, such as size of the locality, age, education, urbanization rate, local GDP, interpersonal inequality (GINI), poverty rate, as well as geography (Naritomi et al. 2012) (for summary statistics, see Table A1). ANALYSIS AND RESULTS Our analysis is divided into two steps. First, we replicate the model used in the seminal US study of Alesina et al. (1999) using municipal-level data from Brazil in 2010. Results from this estimation, shown in Table 1, suggest that the relationship between racial fractionalization and public goods provision in Brazil is very similar to the one observed in the US. More specifically, higher diversity seems to be related to higher overall government expenditure, but lower education 8 Our diversity measure considers each one of the census categories as a separate group, but our results are also robust to the use of an alternative fractionalization measure based on a unified Afrodescendant category composed of pardos and pretos (not shown). 1098 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000308
Ethnoracial Homogeneity and Public Outcomes FIGURE 2.The Distribution of Racial Demography across Brazilian Municipalities. 1.00 r=-0.60 0.75 0.50 0.25 0.00 0.00 0.25 0.50 0.75 1.00 'asn Fractionalization Each dot represents local racial demography in terms of fractionalization or group shares(whites).The graph is based on Brazil's 2010 Census. spending.Additionally,we find a strong negative as- between the "white share"11 variable and the various sociation between fractionalization and the quality of provision measures remains consistent across both health care and education across Brazilian municipali- samples,and mostly positive and significant with ties.Overall,this is precisely the pattern we would ex- respect to the different dependent variables.12 pect to see where diversity is associated with the un- Finally,explaining public outcomes may require tak- derprovision of public goods. ing into account the uneven distribution of groups The same(and even stronger)relationship,however, across the country's territory (Naritomi et al.2012). can be observed using the white group share as a Some groups may be overrepresented in areas with un- measure of ethnic demography (see Table A4).To favorable geographic characteristics,which may in turn better understand these findings,in the second portion hinder service provision.As a result,the relationship of the analysis we divide our sample into majority between racial demography and public goods provi- white and minority white municipalities (Table 2)and sion may itself be confounded by economic geography. re-examine the effects of fractionalization and group As Table A5 indicates,however,the significant rela- shares.10 Results from these estimations show that tionship between group shares and public goods pro- the diversity coefficient remains negative only in the vision largely withstands the inclusion of geographic models using the first sample of municipalities-that is, controls.13 those where the majority of the population is classified as white according to the Census.In the sample of minority white localities,however,the diversity 11 To further understand the role played by different racial groups. variable has no effect.Conversely,the relationship Table A7 looks at each group's effect separately and confirms that more homogeneous Afro-descendant communities have poorer pro- the large number of outcomes tested and samples apply the Bonferroni-Holm p-value adjustment for 15 different com. parisons in the case of fractionalization (Tables 1 and 2)and white The positive relationship between diversity and health care spend- shares (Tables A4 and 2)to check whether some associations may be ing,for which there is no compelling theoretical explanation,is also statistically significant by chance.Our results remain unchanged. observed in the case of the US. 13 The inclusion of geographic covariates does reduce the magnitude 10 For the summary statistics of each sample,see Tables A2 and A3. of effects in some of the models,but changes are not systematic.The As these tables indicate,majority white municipalities are on average role of geography itself appears to be modest and,sometimes,am- better-off compared to minority white municipalities,illustrating the biguous (for details on the coefficients of geographic covariates,see relevance of including socioeconomic controls in our analysis. Table A6).This does not,however,imply that its effects should be 1099
Ethnoracial Homogeneity and Public Outcomes FIGURE 2. The Distribution of Racial Demography across Brazilian Municipalities. r = −0.60 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 Fractionalization Whites, share Each dot represents local racial demography in terms of fractionalization or group shares (whites). The graph is based on Brazil’s 2010 Census. spending.9 Additionally, we find a strong negative association between fractionalization and the quality of health care and education across Brazilian municipalities. Overall, this is precisely the pattern we would expect to see where diversity is associated with the underprovision of public goods. The same (and even stronger) relationship, however, can be observed using the white group share as a measure of ethnic demography (see Table A4). To better understand these findings, in the second portion of the analysis we divide our sample into majority white and minority white municipalities (Table 2) and re-examine the effects of fractionalization and group shares.10 Results from these estimations show that the diversity coefficient remains negative only in the models using the first sample of municipalities—that is, those where the majority of the population is classified as white according to the Census. In the sample of minority white localities, however, the diversity variable has no effect. Conversely, the relationship 9 The positive relationship between diversity and health care spending, for which there is no compelling theoretical explanation, is also observed in the case of the US. 10 For the summary statistics of each sample, see Tables A2 and A3. As these tables indicate,majority white municipalities are on average better-off compared to minority white municipalities, illustrating the relevance of including socioeconomic controls in our analysis. between the “white share”11 variable and the various provision measures remains consistent across both samples, and mostly positive and significant with respect to the different dependent variables.12 Finally, explaining public outcomes may require taking into account the uneven distribution of groups across the country’s territory (Naritomi et al. 2012). Some groups may be overrepresented in areas with unfavorable geographic characteristics, which may in turn hinder service provision. As a result, the relationship between racial demography and public goods provision may itself be confounded by economic geography. As Table A5 indicates, however, the significant relationship between group shares and public goods provision largely withstands the inclusion of geographic controls.13 11 To further understand the role played by different racial groups, Table A7 looks at each group’s effect separately and confirms that more homogeneous Afro-descendant communities have poorer provision. 12 Given the large number of outcomes tested and samples used, we apply the Bonferroni-Holm p-value adjustment for 15 different comparisons in the case of fractionalization (Tables 1 and 2) and white shares (Tables A4 and 2) to check whether some associations may be statistically significant by chance. Our results remain unchanged. 13 The inclusion of geographic covariates does reduce the magnitude of effects in some of the models, but changes are not systematic. The role of geography itself appears to be modest and, sometimes, ambiguous (for details on the coefficients of geographic covariates, see Table A6). This does not, however, imply that its effects should be 1099 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000308
Alexander Kustov and Giuliana Pardelli TABLE 1.Racial Diversity and Public Goods Provision Total spending Educ.share Heal.share Educ.quality Heal.quality (1) (2) (3) (4) (5) Fractionalization 0.285* -0.072** 0.082* -1.080* -0.761** (0.050) (0.016) (0.016) (0.154) (0.146) Income PC,log 0.821** -0.044** 0.033* 1.844* 0.147 (0.045) (0.014) (0.014) (0.141) 0.133) Population,log -0.287* 0.030* 0.029* -0.215* -0.208* (0.005) (0.002) (0.002) (0.016) (0.015) Pop.over 65,share -3.147* -0.393* 0.217体 -1.458* -2.597** (0.284) (0.090) (0.089) (0.885) 0.837) Pop.under 18,share 0.012 0.267* -0.262* -4.442* -3.294* (0.174) (0.055) (0.055) (0.541) (0.512) GINI -1.123** -0.022 -0.042 -3.643* -0.433 (0.134) (0.043) (0.042) (0.421) (0.398) Years of schooling 0.013 0.007* 0.006* 0.207* 0.050* (0.005) (0.002) (0.002) (0.015) 0.015) Area,log 0.011* -0.0002 0.002 0.006 0.057** (0.004) (0.001) (0.001) (0.014) (0.013) Urban,share -0.062* -0.030* 0.014* -0.226* -1.030** (0.026) (0.008) (0.008) (0.081) (0.077) Poor,share 1.461* 0.149** 0.084* 2.246* 0.548 (0.140) (0.045) (0.044) (0.438) (0.414) Constant 4.873** 0.441* -0.229* -1.033 9.133* (0.283) (0.090) (0.089) (0.885) (0.837) State FE Yes Yes Yes Yes Yes Observations 5.150 5,146 5,149 5.503 5.503 Adjusted R2 0.596 0.560 0.238 0.789 0.384 All specifications include "state fixed effects"based on 26 Brazilian states. For variable descriptions,see Appendix. The standard errors are given in parentheses,+p<0.1;*p<0.05;**p<0.01:***p<0.001. Together these findings illustrate that our initial re- which the correlation between fractionalization and sults on the negative effects of diversity are mislead- the group share measure is minimized.14 Within this ing.In fact,more diverse communities outperform artificially restricted sample,fractionalization does not homogeneous nonwhite localities in terms of service robustly relate to any provision measure after control- 685:50190 provision-and are thus found to have poorer outcomes ling for group shares.At the same time,as before,mu- only when compared to homogeneous white munici- nicipalities with a greater proportion of white popula- palities.In other words,racial fractionalization is detri- tion exhibit consistently better public goods regardless mental to the provision of public goods only to the of fractionalization levels. extent that it reflects an increase in the nonwhite population share.That is,when we restrict our analysis to the sample of majority nonwhite localities-where DISCUSSION diversity's increase represents a higher proportion of In our analysis of Brazilian municipalities,we find that white population-fractionalization ceases to be asso- the prima facie negative relationship between diversity ciated with worse outcomes. and public goods provision stems from the fact that These findings seem to suggest that diversity may have heterogeneous effects in different contexts.Be- higher levels of fractionalization reflect a larger propor- fore we can make this statement,however,we have tion of disadvantaged ethnic groups in the local popu- lation.Our case and data allow us to measure the effect to consider that the very reason why fractionalization of diversity in localities where either dominant or dis- is associated with public goods outcomes in'majority white'but not in'minority white'municipalities may be advantaged groups constitute a demographic majority. due to its higher correlation with white group shares Yet,this may not always be feasible in other settings. In fact,in cases where fractionalization is almost in- in the former subsample (-0.98 versus 0.58).To ex- amine the independent effect of diversity on provi- distinguishable from group share measures,diversity's sion,we thus restrict our analysis to the interval within 4 This produces a selection of observations with fractionalization ignored;rather,it suggests that a full understanding of geography's levels between 0.35 and 0.7 (see Figure 1 and Table A8).We would nuanced infuences requires more detailed examination. like to thank an anonymous reviewer for suggesting this analysis. 1100
Alexander Kustov and Giuliana Pardelli TABLE 1. Racial Diversity and Public Goods Provision Total spending Educ. share Heal. share Educ. quality Heal. quality (1) (2) (3) (4) (5) Fractionalization 0.285∗∗∗ − 0.072∗∗∗ 0.082∗∗∗ − 1.080∗∗∗ − 0.761∗∗∗ (0.050) (0.016) (0.016) (0.154) (0.146) Income PC, log 0.821∗∗∗ − 0.044∗∗∗ 0.033∗∗ 1.844∗∗∗ 0.147 (0.045) (0.014) (0.014) (0.141) (0.133) Population, log − 0.287∗∗∗ 0.030∗∗∗ 0.029∗∗∗ − 0.215∗∗∗ − 0.208∗∗∗ (0.005) (0.002) (0.002) (0.016) (0.015) Pop. over 65, share − 3.147∗∗∗ − 0.393∗∗∗ 0.217∗∗ − 1.458∗ − 2.597∗∗∗ (0.284) (0.090) (0.089) (0.885) (0.837) Pop. under 18, share 0.012 0.267∗∗∗ − 0.262∗∗∗ − 4.442∗∗∗ − 3.294∗∗∗ (0.174) (0.055) (0.055) (0.541) (0.512) GINI − 1.123∗∗∗ − 0.022 − 0.042 − 3.643∗∗∗ − 0.433 (0.134) (0.043) (0.042) (0.421) (0.398) Years of schooling 0.013∗∗ − 0.007∗∗∗ 0.006∗∗∗ 0.207∗∗∗ 0.050∗∗∗ (0.005) (0.002) (0.002) (0.015) (0.015) Area, log 0.011∗∗ − 0.0002 0.002 0.006 − 0.057∗∗∗ (0.004) (0.001) (0.001) (0.014) (0.013) Urban, share − 0.062∗∗ − 0.030∗∗∗ 0.014∗ − 0.226∗∗∗ − 1.030∗∗∗ (0.026) (0.008) (0.008) (0.081) (0.077) Poor, share 1.461∗∗∗ 0.149∗∗∗ 0.084∗ 2.246∗∗∗ − 0.548 (0.140) (0.045) (0.044) (0.438) (0.414) Constant 4.873∗∗∗ 0.441∗∗∗ − 0.229∗∗∗ − 1.033 9.133∗∗∗ (0.283) (0.090) (0.089) (0.885) (0.837) State FE Yes Yes Yes Yes Yes Observations 5,150 5,146 5,149 5,503 5,503 Adjusted R2 0.596 0.560 0.238 0.789 0.384 All specifications include “state fixed effects” based on 26 Brazilian states. For variable descriptions, see Appendix. The standard errors are given in parentheses, +p < 0.1; ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001. Together these findings illustrate that our initial results on the negative effects of diversity are misleading. In fact, more diverse communities outperform homogeneous nonwhite localities in terms of service provision–and are thus found to have poorer outcomes only when compared to homogeneous white municipalities. In other words, racial fractionalization is detrimental to the provision of public goods only to the extent that it reflects an increase in the nonwhite population share. That is, when we restrict our analysis to the sample of majority nonwhite localities—where diversity’s increase represents a higher proportion of white population—fractionalization ceases to be associated with worse outcomes. These findings seem to suggest that diversity may have heterogeneous effects in different contexts. Before we can make this statement, however, we have to consider that the very reason why fractionalization is associated with public goods outcomes in ‘majority white’ but not in ‘minority white’ municipalities may be due to its higher correlation with white group shares in the former subsample (−0.98 versus 0.58). To examine the independent effect of diversity on provision, we thus restrict our analysis to the interval within ignored; rather, it suggests that a full understanding of geography’s nuanced influences requires more detailed examination. which the correlation between fractionalization and the group share measure is minimized.14 Within this artificially restricted sample, fractionalization does not robustly relate to any provision measure after controlling for group shares. At the same time, as before, municipalities with a greater proportion of white population exhibit consistently better public goods regardless of fractionalization levels. DISCUSSION In our analysis of Brazilian municipalities, we find that the prima facie negative relationship between diversity and public goods provision stems from the fact that higher levels of fractionalization reflect a larger proportion of disadvantaged ethnic groups in the local population. Our case and data allow us to measure the effect of diversity in localities where either dominant or disadvantaged groups constitute a demographic majority. Yet, this may not always be feasible in other settings. In fact, in cases where fractionalization is almost indistinguishable from group share measures, diversity’s 14 This produces a selection of observations with fractionalization levels between 0.35 and 0.7 (see Figure 1 and Table A8). We would like to thank an anonymous reviewer for suggesting this analysis. 1100 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:56:49, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000308