Abstract
A considerable amount of recent work in political science and economics builds from the hypothesis that ethnic heterogeneity leads to poor provision of public goods, a key component of poor governance. Much of this work cites Alesina, Baqir and Easterly as providing empirical proof. This paper argues that the findings of this article have been significantly overstated. Through a simple re-analysis of the data, it shows that ethnic diversity does not straightforwardly undermine public goods provision. Rather, at least in these data, the relationship is mixed for different public goods: ethnic diversity is related to lower provision of some public goods and to higher provision of others. In some cases, there is no clear relationship. The differences between the findings presented here and those of the original article are arguably subtle, but are worth noting because of Alesina, et al.'s important contribution to the literature.
Introduction
A considerable amount of recent work in political science and economics builds from the hypothesis that high levels of ethnic diversity lead to poor governance, and in particular to the ‘under-provision’ of public goods. While this hypothesis has gained particular attention in recent work, it also has roots in classic theories of democratic governance. John Stuart Mill (1991, 428), for instance, declared that ‘free institutions are next to impossible in a country made up of different nationalities.’
The literature suggests that ethnic division has a negative effect on democratic governance through several different mechanisms. One key mechanism developed in the political economy literature posits that, because ethnic groups have heterogeneous preferences, ethnically diverse societies will support lower contributions than ethnically homogenous societies to the provision of ‘productive’ public goods, such as schools, health services and roads. The most cited formalization and testing of this argument is Alesina, Baqir, and Easterly (1999)'s article, ‘Public goods and ethnic divisions’ (hereafter, ABE). ABE tests and finds support for its model using data on urban counties, cities and metropolitan areas in the USA. Banerjee, Iyer, and Somanathan (2005, 639) note that it addresses ‘one of the most powerful hypotheses in political economy’, ‘the notion that social divisions undermine economic progress, not just in extremis, as in the case of a civil war, but also in more normal times.’
Although ABE is based on analysis of public spending in the USA, it has become an important article in the literature on ethnic politics more generally and is referenced in work on many regions. It has been cited in at least 360 articles, including over 50 published in 2011 (see, e.g., Miguel 2004; Keefer and Khemani 2005; Kimenyi 2006; Habyarimana et al. 2007; Baldwin and Huber 2010).1 Much of this work cites ABE as proof of the negative relationship between ethnic heterogeneity and public goods provision, or even between ethnic heterogeneity and good governance more broadly, and proceeds to analyse other topics that build upon this finding. Habyarimana et al. (2007, 709), for instance, takes the finding that ‘ethnic diversity undermines public goods provision’ as its starting point and seeks to explain why, using an experimental approach to test three mechanisms about the individual behaviour behind this relationship.
Through a relatively simple re-analysis of the data presented in ABE, this article shows that its central finding is not as robust as citations of it would suggest: the data do not support the straightforward claim that ethnic diversity undermines the provision of public goods. Rather, the relationship between ethnic diversity and public goods provision is mixed for different goods: it is sometimes positive, sometimes negative, and sometimes insignificant. Analysis of other outcome measures is also consistent with an ambiguous relationship between ethnic divisions and poor governance.
More specifically, using ABE's own data and similar methods, the relationship between ethnic fractionalization and lower spending on public goods is shown to be significantly weaker when simple omitted variables biases are corrected and other outcomes are considered. The relationship between ethnic fractionalization and the share of total spending on public goods like welfare and health is statistically insignificant, while the relationship between ethnic fractionalization and the share of spending on police is positive and statistically significant in both the original ABE analysis and in this analysis. The strongest support for the ABE hypothesis is in analysis of the share of spending on education and on roads, which continues to have a negative relationship with ethnic fractionalization. However, higher ethnic fractionalization is also related in a statistically significant way to higher levels of educational expenditure per child (i.e. the amount spent on education relative to other goods may be lower, but the absolute amount is higher). Further, some of the clearest relationships in the data are between higher ethnic fractionalization and higher levels of local government expenditure, intergovernmental revenue and local taxation – which are not well explained by the ABE model.
The differences between the analysis and findings presented in this article and those in ABE's original article are arguably subtle. However, they are worth revisiting because of ABE's importance to the literature on ethnic politics and governance. Taking ABE's findings as a starting point, a significant amount of research has focused, like Habyarimana et al. (2007, 709), ‘not on whether ethnic diversity undermines public goods provision, but on why.’ This article suggests that the foundations of this line of argument are shakier than claimed. A few other articles are sometimes cited along with ABE to support this claim, but all provide only incomplete ‘tests’: Miguel (2004), in particular, shows that higher ethnic diversity is associated with lower funding for primary education in Kenya – i.e. only one public good, not public goods in general, a finding that is broadly consistent with those of this article. Easterly and Levine (1997) find that ethnic diversity in sub-Saharan Africa helps to explain poor governance, measured in terms of low schooling, political instability, underdeveloped financial systems, distorted foreign exchange markets, government deficits and weak infrastructure. More recently, consistent with Easterly and Levine's findings, Patsiurko, Campbell, and Hall (2012) find an inverse relationship between ethnic fractionalization and economic performance in Organisation for Economic Cooperation and Development (OECD) countries, but neither focuses on testing the relationship between ethnic diversity and public goods provision directly. The analysis presented here suggests that testing this hypothesis remains an open area for research. It also suggests that researchers interested in ‘why’ questions should focus on why ethnic diversity undermines public goods provision in particular areas such as education, rather than on public goods provision in general.
This article first reviews ABE's model and key findings. It then presents an empirical critique of these findings based on a re-analysis of the data used in the original article. It concludes with a discussion of the relationship between ethnic diversity and patronage, an important claim that is advanced, but not directly tested, by ABE.
Public goods and ethnic divisions
The basic argument in ABE that ‘when individuals have different preferences, they want to pull fewer resources together for public projects’, builds explicitly on Alesina and Spolaore (1997) and on other work in political economy (Alesina, Baqir, and Easterly 1999, 1243). As Easterly and Levine (1997, 1215–1216) summarize, this argument has important implications for economic growth: it means that in ethnically diverse situations ‘less of the public good is chosen by society’, which although ‘socially optimal’ is ‘costly for output and growth compared with a homogenous society’. In ABE, heterogeneity of preferences across ethnic groups means that ethnically diverse societies will be less supportive of contributions to the provision of public goods than more homogeneous societies, and thus that more government spending will go towards ‘private’ goods (which they equate with patronage spending). ABE is important and well cited in part because it presents a clear model along these lines, but even more so because it presents empirical tests of the model's predictions across a range of public goods.
The model considers a jurisdiction (a city, county or state) with a stable population, in which there is no in- or out-migration, and decisions are made by majority rule (Alesina, Baqir, and Easterly 1999, 1247–1254). It assumes that the average individual's preferences regarding the type of public good provided will differ by ethnic group.
The average individual's utility is given by 
, where α is between 0 and 1; g is the public good; li is the distance between individual i's preferred type of public good and the public good provided; and c is private consumption.2 Private consumption c is equal to exogenous pre-tax income (y) minus the lump-sum tax (t), which is identical for everyone (c = y–t). The public budget constraint implies that g = t. For tractability, the model further assumes that individuals first vote on the size of the public good (i.e. the amount of taxation) and then on the type of public good.
The article highlights three key results. First, building on the median voter theorem (Downs 1957), for any positive amount of spending g, the type of public good chosen will be the one preferred by the median voter (proposition 1). Second, the size of the public good provided in equilibrium will be given by 
, where
is the ‘median distance from the type most preferred by the median voter’ or the ‘median distance from the median’, an indicator of the polarization of preferences (proposition 2) (Alesina, Baqir, and Easterly 1999, 1249). The median distance from the median (
) will be small in an un-polarized society and large in a polarized society, where there are homogeneous within-group but distinct cross-group preferences. Further, the size of the public good in equilibrium will be decreasing in
(corollary to proposition 2).
Finally, it develops a key prediction with regard to the share of government spending on public goods. This is the key relationship explored in the empirical analysis: by adding an additional assumption that some government spending (g) can be targeted to specific groups, g is decomposed into two parts, spending on (targeted) patronage (g1) and spending on a non-excludable public good (g2), which cannot be well targeted to specific groups. Although interest group politics, ABE argues, may lead to an increase in patronage spending (g1), higher levels of polarization will be tied (as predicted by their basic model) to less spending on public goods (g2). Thus, as polarization increases, the share of spending on non-excludable public goods should decline relative to total spending. ABE also builds on a related literature to predict that fiscal discipline should be more problematic in ethnically fractionalized jurisdictions (Alesina, Baqir, and Easterly 1999, 1254).
Empirical support
ABE tests the predictions of the model using data on US local government – cities, metropolitan areas and urban counties with populations of at least 25,000. These data are drawn from the County and City Data Book 1994 (CCDB) and City and County Plus (US Census Bureau 1994; CCP 1994). ABE considers a range of dependent variables: the share of local government spending on health, education, police, fire protection, roads, welfare, and sewerage and trash pickup, and spending on roads per capita; intergovernmental revenue; taxes per capita; surplus per capita; and expenditure per capita. Data on government finances are for 1992 for county areas and metropolitan areas, and for 1990–91 for cities. All other data are for 1990, unless otherwise noted.
The key independent variable of interest, ethnic polarization, is proxied with a measure of racial fractionalization (ETHNIC), which is also used in other indices of ethno-linguistic fractionalization (ELF) (see Taylor and Hudson 1972; Montalvo and Reynal-Querol 2005). It measures the probability that two individuals selected at random will be from different ethnic groups. ETHNIC is calculated as 1−σ (Racei)2, where Racei is the share of the population that self-identified as ‘white’, ‘black’, ‘Asian and Pacific Islander’, ‘American Indian’ or ‘other’ in the 1990 US census. Regressions also control for other socio-economic factors that might influence fiscal priorities: income per capita; log of population; percentage of BA graduates; income inequality measured as the mean-to-median income ratio; and the fraction of the population over 65 years of age.
Analysis is conducted separately for cities, metropolitan areas and urban counties. Ordinary least squares (OLS) regressions are presented for the city sample and two-stage least squares for the metropolitan area and county samples, but it is noted that OLS results for the latter two samples are similar (Alesina, Baqir, and Easterly 1999, 1261). The two-stage least squares analysis instruments for ETHNIC and income per capita using 1979–80 values.
The analysis presented in ABE suggests strong support for the model's key prediction: ETHNIC is negatively associated (in a statistically significant way) with the share of spending on roads, education, welfare, and sewage and trash pickup, as well as with spending per capita on roads.3 It is also negatively associated with fire protection, although the result is not statistically significant. Two results are inconsistent with the model's predictions: spending on police is positively related to ethnic fractionalization in all samples, as is spending on health in the metropolitan and county areas samples.4 The article further finds that ethnic fractionalization tends to be positively associated with spending overall, which appears to be financed through higher debt and intergovernmental revenue, rather than through local taxes. It concludes that these results are ‘broadly consistent with political economy theories in which heterogeneous and polarized societies will value public goods less, patronage more, and will be collectively careless about fiscal discipline’ (Alesina, Baqir, and Easterly 1999, 1274).
Omitted variables
The revised analysis presented here considers three types of omitted variables, in other words, factors that should (arguably) have been included in ABE's empirical analysis but were not. The first are state effects. These are especially important because a variety of laws, regulations and other factors differ systematically across US states. Murray, Evans, and Schwab (1998), for instance, map variation in education-finance reform across states between 1971 and 1996. Looking at variation in spending, they find that most variation (two-thirds versus one-third) is between states rather than within states (Murray, Evans, and Schwab 1998, 808). Ladd (2005, 144) notes that ‘measuring fiscal condition, especially that of local governments, is complicated by the observation that expenditure responsibilities often differ among jurisdictions of the same type. This problem is particularly severe if one compares fiscal conditions across large cities in different states.’ Because both ethnic demography and the fiscal responsibilities of local governments vary across states, it is possible that the main ABE results for ETHNIC picked up some of the effects of variation in state-level policies and institutions.
The second type of omitted variables relate to community characteristics. Non-urban communities with low population densities, for instance, may benefit less from economies of scale in the provision of certain public goods. In an area with a highly dispersed population, more miles of roads may be needed to serve a community of the same size living in close proximity, just as more schools and health clinics may be needed to provide service to the same number of individuals at a reasonable distance from their homes. Likewise, urban communities with highly concentrated populations may have (regardless of ethnic diversity) higher demands for some public goods such as policing. The ABE analysis is limited to cities, counties and metropolitan areas with populations greater than 25,000, but there is still considerable variation of this sort across communities.
The third type of omitted variables relate to additional controls relevant to spending on some public goods. For analysis of education spending in particular, the fraction of the population of school age is a key missing variable.5 A community with few children to educate should naturally spend less on schools. It is clear that the authors chose to hold constant the same indicators across all areas of analysis for reasons of consistency, but support for the model would be stronger if ETHNIC were still significant when such additional controls were added.
Another key concern with the empirical analysis is over the way in which ethnic polarization is measured. Racial divisions, while clearly salient in the USA in many contexts, are not the only ethnic divisions, or even necessarily the most salient ethnic divisions in US subnational politics. Historically, ethnic divisions within the ‘white’ category, such as between Irish Americans and Italian Americans, have been more salient in many US cities. In many jurisdictions, political issues such as bilingual education and immigration reform divide Hispanic and non- Hispanic populations, immigrant and non-immigrant populations, and those who speak English versus another language at home, much more than they divide the racial communities captured in the US census and in ABE's ETHNIC measure. Further, use of the Herfindahl-Hirschman Index is only one way of capturing the degree of ethnic division in a population, and indeed may not be the most appropriate one for assessing the causal mechanism posited in this model (Herfindahl 1950; Hirschman 1945). To partially account for these concerns, several different ethnic variables are discussed below. The broader issue of measuring ethnic diversity is explored more fully in other work (e.g. Roeder 2001; Alesina et al. 2003; Posner 2004; Campos and Kuzeyev 2007).
It should be noted that ABE includes discussion of state dummy variables and population density among the additional variables considered in its sensitivity analysis (Alesina, Baqir, and Easterly 1999, 1267–1274). However, it reports results only for the share of spending on roads and on education (see Table VII, Alesina, Baqir, and Easterly 1999, 1270). These particular results are broadly consistent with the findings discussed here, but the findings on other variables suggest additional important relationships in the data that are not discussed in ABE.
Revised analysis
The revised analysis presented here uses the ABE replication data sets for counties, cities and metropolitan areas and supplements these data with additional information principally from the CCDB.6 It analyses all three samples, but focuses on the county areas sample, which includes the largest number of observations and for which the most additional information was available.
In order to correct for omitted variables biases in the original analysis, additional control variables were added using data either directly from the CCDB or the CCDB and City and County Plus as reported in the ABE replication data sets: population density and state dummy variables, for instance, are available for all three samples in the original ABE data sets. For the county areas sample, new controls were added for land area and the share of the population employed in farming (to assess community type) and for the share of the population of school age. A number of new ethnic variables were also considered, including the share of the population that identify as Hispanic, the share of the population that speak a language other than English at home, and the share of the population that is foreign born, along with analogous fractionalization measures. For the metropolitan areas sample, new controls were included for land area and for the fraction of the population employed in manufacturing (in lieu of the fraction employed in farming).7 Two unique measures of ethnic segregation used in Cutler and Glaeser (1997) and Cutler, Glaeser, and Vigdor (1999b) were also considered: an Index of Racial Dissimilarity (DISM), measuring the extent to which African Americans have contact with non-African Americans based on the percentage of people who would need to move across census tracts to get an even racial division across the entire metropolitan statistical area; and an Index of Racial Isolation (ISOL), measuring the extent to which African Americans have contact with non-African Americans based on an adjusted percentage of African American residents in their census tract (Cutler, Glaeser, and Vigdor 1999a). These indicators were available only for metropolitan areas. Although they provide useful new measures of ethnic division, there is some difficulty in their interpretation. A high ISOL value, for instance, may be due to significant disparities in the economic status of African Americans and others, or a high degree of racial tension and discrimination in housing markets. It might also be a function of how the community developed and which cities are newer (see Glaeser and Vigdor 2001).
Additional indicators of the outcomes of interest – public goods provision and local finances – were also considered. New variables were constructed and added in the county sample, for instance, for spending on education per school-aged child, spending on police per capita, federal employment as a share of total employment, and state and local employment as a share of total employment. An indicator for the fraction of students enrolled in private schools was also considered in order to investigate whether county residents were substituting private education for public education, as the ABE model would suggest. Analysis was also done using a measure of violent crime in 1991 from the ABE data sets.
For simplicity, the ABE analysis was first replicated using OLS regressions with robust standard errors for each of the dependent variables analysed in ABE for each of the three samples. Largely because ABE reports two-stage least squares results for the county and metropolitan areas samples, the results of the revised analysis are slightly different, but the key findings are the same.8 Next, state effects were controlled for by using fixed-effects regressions for states in all three samples for all outcomes. The analysis first included all original control variables (log of population, income per capita, mean-to-median income, fraction of the population over 65, and percentage of university graduates), and then considered additional controls.
Because the exact same indicators were not available for county areas, metropolitan areas and cities, the additional controls added varied across the three samples. Analysis was most complete for county areas, which included the following additional control variables: state-fixed effects; land area; population density; and fraction employed in farming. The school-aged population was also controlled for in analysis of education spending. The metropolitan areas regressions included the following additional control variables: state-fixed effects; land area; population density; and the fraction of the labour force employed in manufacturing. The cities regressions included additional controls for state effects and population density only. Although some potentially relevant variables thus are still omitted in the metropolitan areas and cities analysis, analysis of the county areas sample suggests that state effects were in fact the most significant omitted variables with respect to our variable of interest: in general, results on ETHNIC do not change significantly with the addition of the other new variables, but they do change in magnitude and statistical significance with the addition of controls for state variation. Finally, analysis was done using the various ethnic indicators.
provides an example of various specifications tested, showing results for the dependent variable of spending on roads per capita in the replicated ABE and revised models. and summarize results on the key variable of interest, ETHNIC, in all regressions. follows ABE's Table IV and reports results for shares of government spending on all public goods, as well as spending on roads per capita (Alesina, Baqir, and Easterly 1999, 1262). reports results for other measures and fiscal aggregates.
Table 1. Sample results for spending on roads per capita in the ABE and revised models, county areas sample.
Table 2. Summary of results on ETHNIC in regressions for shares of government spending on public goods.
Table 3. Coefficient on ETHNIC in regressions for fiscal aggregates.
The revised analysis shows clearly mixed support for the key predictions of the ABE model. As shows, when state-fixed effects and other omitted variables are added, ETHNIC has a statistically significant negative relationship only with the shares of public spending in three areas: education (in the county areas sample only, but not in the metropolitan areas sample); sewage and sanitation (tested in the city sample only); and roads overall, but not per capita. As in the original ABE analysis, it also has a positive and statistically significant relationship with the share of spending on police (in the revised analysis, in the county areas and cities samples only). The original ABE analysis also shows a positive and statistically significant relationship between ETHNIC and the share of spending on health in the metropolitan and county areas samples, but it is not statistically significant in the revised analysis. In the metropolitan areas sample, none of the coefficients on ETHNIC are statistically significant when controls for state effects are added. This is possibly due to the smaller sample size.
Findings regarding the share of spending on education, roads, and sewage and sanitation lend some support to ABE's hypothesis. Results for the share of spending on education in particular appear consistent with the story told in ABE, and in fact the magnitude of the coefficient on ETHNIC in the county areas sample increases with the addition of omitted variables when compared with results from the replicated ABE model. Intuitively, this result seems plausible as education is an area of public spending where the mechanism described in ABE is most straightforward. Additional support is suggested by analysis of private school enrolment. The ABE model suggests that high ethnic fractionalization should be related to a higher percentage of enrolment in private schools as individuals should prefer private over public provision of education, and fixed-effects analysis shows a positive relationship between ETHNIC and the dependent variable of private school enrolment as a percentage of total school enrolment.9
Taken together, these results also raise some key questions. For one, if ethnic polarization and divergent preferences are indeed driving the findings on education, it is puzzling that there is not a similar relationship with other public goods such as welfare spending. Luttmer (2001) shows that individuals are more likely to be supportive of welfare spending when it goes to members of their own ethnic group, rather than other groups, suggesting that a similar process should be at work here. One possibility for further research is that US county and metropolitan areas may have less discretion over public spending on welfare, which may be more tied to intergovernmental transfers, and thus cannot change spending levels on welfare easily in response to local pressures.
In addition, as shows, ethnic fractionalization has a positive and statistically significant relationship with higher levels of educational expenditure per child. In other words, higher ethnic fractionalization is not necessarily related to lower provision of the public good of education in absolute terms, although it is related to a lower share of total spending on education. (Total expenditure is discussed further below.)
The intuition behind the empirical results on the share of public spending on sewage and sanitation and on roads is also not clear. Why should members of one ethnic group lose out if members of another ethnic group consume this public good? It is possible that if there is a high degree of ethnic segregation in residence patterns, heterogeneous preferences across ethnic groups might be due to the location of sewage facilities or of roads. However, if ethnic segregation were driving this result, we should also see a significant negative coefficient on ETHNIC for other spatially distributed public goods such as fire protection and hospitals. Analysis of levels of spending is also instructive here: when additional controls are added, there is no statistically significant relationship between ethnic fractionalization and spending on roads per capita in any of the three samples.
Other ethnic indicators
One explanation for some of these results is that ETHNIC, that is racial fractionalization, does not capture the most politically relevant cleavages – either because other, non-racial ethnic divisions are more salient in these jurisdictions, or because the ethnic fractionalization index does not capture the salient aspects of racial polarization. In particular, as suggested above, we might expect indicators of the size of the Hispanic, immigrant and ‘non-English’-speaking populations to be more relevant in the analysis of public spending on education. In addition, it is possible that these other ethnic divisions are additional omitted variables in the analysis described in and .
In order to address these possibilities, analysis was conducted using the various non-racial measures of ethnic fractionalization in the county areas sample and the two new measures of racial segregation in the metropolitan areas sample. For each dependent variable, fixed-effects regressions were conducted first using all controls and all ethnic variables, and then including ETHNIC, all other controls, and each separate ethnic indicator in turn.
In the county areas sample, there is some evidence that Hispanic identity and language spoken at home may be related to the share of spending on public goods and to public finance outcomes. For instance, fractionalization measures for Hispanic identity and language spoken at home (both highly correlated with each other) have an additional negative, statistically significant relationship with the share of spending on roads, while the fractionalization measure for language spoken at home has an additional positive, statistically significant relationship with the share of spending on police. Both measures are also positively related in a statistically significant way with the share of spending on welfare, and negatively related with spending on roads per capita. However, it seems likely that regional factors in particular are driving some of these results as the share of the population that identifies as Hispanic tends to be higher in the US South and West (e.g. Texas, New Mexico and California). If we analyse spending on roads per capita by state, for instance, Hispanic ‘fractionalization’ has a negative and statistically significant relationship in Texas and California, but not in New York, Florida, Ohio and Pennsylvania.
In the metropolitan areas sample, there is some evidence that the ISOL (but not the DISM) has additional explanatory power beyond ETHNIC in the analysis of spending on roads per capita (positive relationship), and on the share of spending on education (negative) and on police (positive). Results also suggest that measures like ISOL could better capture ethnic polarization than simple fractionalization values as the inclusion of ISOL changes the significance of the results on ETHNIC in several regressions.
Other fiscal aggregates
In the revised analysis as compared to the original model, ETHNIC has a higher coefficient and is statistically significant in the regressions for total spending per capita in all three samples, and for intergovernmental revenue in both cities and metropolitan areas. It is positively related to local government taxes per capita in all samples, statistically significant at the 95 per cent level in the county areas and cities samples, and at the 90 per cent level in the metropolitan areas sample. In the county areas sample, it also has a positive and statistically significant relationship with educational expenditure per child, spending on police per capita, and federal and state and local government employment as a percentage of the total labour force.
Looking at spending per capita, a one-unit increase in ETHNIC corresponds to roughly $1,018 more in spending per capita in county areas, $477 in cities and $1,152 in metropolitan areas (based on the fixed-effects analysis with all controls). These figures are notable in that average spending per capita is $1,850 in county areas, $876 in cities and $2,023 in metropolitan areas. Thus, although a one-unit increase in the value of ETHNIC represents an enormous demographic difference, from complete ethnic homogeneity to complete ethnic heterogeneity, even smaller increases correspond to significant increases in spending when compared to average values.
Contrary to ABE's findings, the revised analysis suggests that higher spending is financed both through higher intergovernmental revenue and higher local taxes. In analysis of local taxes per capita, there is a notable change in the estimated effect of ETHNIC when state effects are added. In county areas, for instance, the coefficient on ETHNIC is $339, suggesting that in an area with an ETHNIC value of 0.5 versus 0.25, local government taxes are predicted to be roughly $85 higher per capita. This is a sizeable amount considering that average local taxes per capita are $648 in county areas.
These results also raise questions about the ABE finding that higher government spending in ethnically diverse communities is financed largely through higher budget deficits – arguably another measure of poor governance. The analysis here suggests that local government deficits before intergovernmental transfers are higher with higher ETHNIC values, but that after transfers, deficits in more ethnically diverse US areas are not significantly different from those in more homogeneous regions. Further, a sizeable amount of increased spending appears to go to productive public goods like education and police (see Ajilore and Smith 2011).10
Conclusion and extensions
This paper revisits the much-cited claim that there is a negative relationship between ethnic heterogeneity and public goods provision by re-analysing the data presented in one of the articles most cited as providing empirical evidence of this claim. Analysis shows that support is in fact mixed, largely depending on the public good. The differences between the findings presented here and those of the original article are arguably subtle, but are worth highlighting because of the importance of this argument to the literature.
This analysis raises several issues for further research. First, the relationship between ethnic diversity and the under-provision of public goods is only a hypothesis, not a robust finding, and remains an open question for research. A strong argument could be made that the predictions of the ABE model should hold in other ethnically diverse countries, even if support in the USA is not very strong. Ethnic division in the USA is arguably of a different order today than ethnic polarization in countries where ethnic parties are the norm, ethnic civil war is likely, and democratic institutions that might mediate such conflict are much weaker. Ideally, we would be able to move this literature forward substantially by testing the predictions of the ABE model using analogous data on local government spending in sub- Saharan Africa, for instance, but this has not been possible because reliable disaggregated data of this type has not been available. Several studies have conducted related analyses, suggesting the plausibility of the ABE model in Africa (e.g. Miguel 2004; Kimenyi 2006), but these studies offer only partial empirical ‘tests’.
Second, neither ABE's analysis nor the analysis presented here fully addresses the important corollary to the central prediction in ABE's discussion: that higher ethnic fractionalization also should be related to a higher share of spending on private or ‘patronage’ goods, such as group-targeted jobs or other benefits, as opposed to spending on ‘productive’ public goods. ABE's analysis focuses on explaining the share of spending on various ‘productive’ public goods (roads, education, health, etc.); it does not directly measure patronage spending, nor directly test the relationship between ethnic heterogeneity and spending on patronage goods. The revised analysis presented here suggests that in the USA, higher levels of ethnic fractionalization are related to larger governments with higher taxes and higher intergovernmental revenue, but it is not clear what this additional money is funding. Further empirical analysis and additional data are needed to explore this relationship fully.
A related question concerns the causality of the relationship itself. ABE hypothesizes that ethnic division leads to patronage politics: because of ethnic divisions, voters favour private goods over public goods, and governments will target spending to particular groups to the exclusion of others. Other studies highlight the link from patronage politics to ethnic politics (i.e. the endogeneity of ethnic divisions to patronage politics): Wantchekon (2003, 403), for instance, finds that the results of his study of the 2001 Beninese elections ‘further develop and expand the conventional wisdom in African politics by establishing that…clientelist appeals reinforce ethnic voting (not the other way round).’ Fearon (1999, 18–21) predicts that ‘if we can order governments or political systems by the amount of political pork they make available, then we would predict ethnic political coalitions to be more common in governments where pork is more prevalent and accessible to politicians.’ Chandra (2004, 1) argues that voters in ‘patronage-democracies’ ‘choose between parties by conducting ethnic headcounts rather than by comparing policy platforms or ideological positions.’ In short, the literature suggests that ethnic divisions and patronage politics are indeed correlated, but unravelling causality is a much harder issue, especially with quantitative methods alone.
A final question concerns good governance more generally and its relationship to spending on ‘private’ or ‘excludable’ goods. The clear assumption in ABE is that government spending on such ‘patronage’ goods is a negative outcome. ABE suggests that ethnic divisions are especially problematic for governance because they lead simultaneously to the under-provision of public goods and the over-provision of patronage goods. Wantchekon (2003) and others similarly favour more policy-focused electoral promises over ‘clientistic’ appeals. But is all spending targeted to particular groups pernicious? Surely it is easy to criticize the most unsophisticated forms of targeted spending, such as the awarding of jobs or contracts to unqualified members of one's ethnic group, but what about other forms of targeted spending? Targeting transfers to particular disadvantaged ethnic constituencies, for instance, might be done on the basis of supporting key principles of good governance, such as equity. The degree to which elected representatives have a duty first to serve the constituencies that elected them (e.g. paying particular attention to jobs and investment in their own constituencies) is also by no means a settled public debate.
| ABE model (OLS, robust standard errors) | Revised model (fixed effects for states) | |||
|---|---|---|---|---|
| 1 | 2 | 3 | 4 | |
| Ethnic fractionalization | −85.441* (12.11) | −62.853* (11.53) | −9.987 (11.70) | −6.257 (11.79) |
| Income per capita | −0.000 (0.00) | 0.000 (0.00) | 0.001 (0.00) | |
| Log of population | −8.298* (2.21) | −1.054 (2.37) | −11.239* (1.67) | −8.140* (1.80) |
| Percentage BA graduates | 187.570* (36.10) | 215.779* (34.51) | 36.834 (28.79) | 47.164 (28.83) |
| Mean-to-median income ratio | −100.980* (23.61) | −85.118* (21.88) | −35.753* (17.23) | −30.135 (17.50) |
| Fraction of population >65 | 262.997* (52.37) | 263.880* (48.27) | 203.919* (40.46) | 202.697* (40.15) |
| Population density | 1.664* (0.59) | −0.626 (0.73) | ||
| Area | 0.006* (0.00) | 0.002* (0.00) | ||
| Farm population | 751.979* (87.54) | 271.356* (51.91) | ||
| Constant | 281.933* (38.01) | 138.780* (40.72) | 235.417* (26.68) | 175.956* (29.65) |
| Number of observations | 1386 | 1386 | 1386 | 1386 |
| Number of groups | 49 | 49 | ||
| R2 | 0.1307 | 0.2304 | ||
| R2 overall | 0.0740 | 0.1608 | ||
| R2 within | 0.0853 | 0.1077 | ||
| R2 between | 0.0100 | 0.1513 | ||
Note: Standard errors are in parentheses.
* 95% confidence
| Revised analysis (fixed effects) | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| ABE model using OLS regression with robust standard errors | State-fixed effects only | All controls | |||||||
| ETHNIC coefficient | n | R2 | ETHNIC coefficient | n (groups) | Adj. R2 | ETHNIC coefficient | n (group) | Adj. R2 | |
| County areas | |||||||||
| Education | −0.089* (0.02) | 1,386 | 0.15 | −0.102* (0.03) | 1,386 (49) | 0.14 | −0.138* (0.03) | 1,386 (49) | 0.16 |
| Roads | −0.052* (0.01) | 1,386 | 0.21 | −0.027* (0.01) | 1,386 (49) | 0.20 | −0.026* (0.01) | 1,386 (49) | 0.28 |
| Welfare | −0.039* (0 .01) | 1,386 | 0.06 | 0.009 (0.01) | 1,386 (49) | 0.01 | 0.005 (0.01) | 1,386 (49) | 0.00 |
| Police | 0.035* (0.00) | 1,386 | 0.22 | 0.020* (0.00) | 1,386 (49) | 0.19 | 0.019* (0.00) | 1,386 (49) | 0.19 |
| Health | 0.119* (0.02) | 1,386 | 0.07 | 0.006 (0.02) | 1,386 (49) | 0.03 | 0.004 (0.03) | 1,386 (49) | 0.03 |
| Roads (per capita) | −85.441* (12.11) | 1,386 | 0.13 | −9.987 (11.70) | 1,386 (49) | 0.07 | −6.257 (11.79) | 1,386 (49) | 0.16 |
| Metropolitan areas | |||||||||
| Education | −0.172* (0.04) | 305 | 0.20 | −0.079 (0.05) | 274 (48) | 0.17 | −0.046 (0.05) | 242 (41) | 0.18 |
| Roads | −0.050* (0.01) | 305 | 0.28 | −0.029* (0.01) | 274 (48) | 0.24 | −0.025 (0.01) | 242 (41) | 0.23 |
| Welfare | −0.012 (0.02) | 305 | 0.04 | 0.019 (0.01) | 274 (48) | 0.02 | 0.007 (0.01) | 242 (41) | 0.02 |
| Police | 0.022* (0.01) | 305 | 0.25 | 0.014 (0.01) | 274 (48) | 0.23 | 0.013 (0.01) | 242 (41) | 0.28 |
| Health | 0.217* (0.06) | 305 | 0.12 | 0.071 (0.07) | 274 (48) | 0.06 | 0.027 (0.08) | 242 (41) | 0.03 |
| Roads (per capita) | −86.545* (21.13) | 305 | 0.21 | −28.900 (24.68) | 274 (48) | 0.16 | −16.562 (27.73) | 242 (41) | 0.17 |
| Cities | |||||||||
| Roads | −0.082* (0.01) | 1,020 | 0.13 | −0.064* (0.02) | 1,020 (51) | 0.13 | −0.061* (0.02) | 1,020 (51) | 0.14 |
| Police | 0.096* (0.01) | 1,020 | 0.10 | 0.052* (0.01) | 1,020 (51) | 0.09 | 0.049* (0.01) | 1,020 (51) | 0.09 |
| Fire protection | −0.004 (0.01) | 1,020 | 0.05 | 0.005 (0.01) | 1,020 (51) | 0.05 | 0.007 (0.01) | 1,020 (51) | 0.05 |
| Sewage and sanitation | −0.076* (0.02) | 1,020 | 0.09 | −0.080* (0.02) | 1,020 (51) | 0.08 | −0.078* (0.02) | 1,020 (51) | 0.09 |
| Roads (per capita) | −36.76* (9.94) | 1,020 | 0.08 | −18.477 (12.57) | 1,020 (51) | 0.07 | −14.480 (12.60) | 1,020 (51) | 0.09 |
Note: Standard errors are in parentheses.
* 95% confidence interval
| New analysis (fixed effects) | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| ABE model using OLS regression with robust standard errors | State-fixed effects only | All new controls | |||||||
| ETHNIC coefficient | n | R2 | ETHNIC coefficient | n (groups) | R2 overall | ETHNIC coefficient | n (groups) | R2 overall | |
| County areas | |||||||||
| Total spending (per capita) | 493.091* (113.146) | 1,386 | 0.17 | 1070.900* (114.32) | 1,386 (49) | 0.11 | 1017.564* (116.10) | 1,386 (49) | 0.12 |
| Local government taxes (per capita) | −1.098 (44.32) | 1,386 | 0.46 | 364.097* (46.25) | 1,386 (49) | 0.41 | 339.000* (46.76) | 1,386 (49) | 0.42 |
| Intergovernmental revenue | 254.488* (61.02) | 1,385 | 0.06 | 445.395** (49.98) | 1,385 (49) | 0.05 | 345.927* (48.97) | 1,385 (49) | 0.13 |
| Local government surplus before transfers | −214.165* (65.77) | 1,385 | 0.04 | −436.056* (59.28) | 1,385 (49) | 0.03 | −407.467* (60.74) | 1,385 (49) | 0.05 |
| Local government surplus after transfers | 40.322 (44.81) | 1,385 | 0.01 | 9.340 (55.14) | 1,385 (49) | 0.00 | −61.540 (56.00) | 1,385 (49) | 0.04 |
| Federal employment (share of total employment) | 0.054* (0.01) | 1,400 | 0.10 | 0.084* (0.02) | 1,400 (49) | 0.09 | 0.091* (0.02) | 1,400 (49) | 0.09 |
| State and local government employment (share of total) | 0.033* (0.01) | 1,400 | 0.38 | 0.061* (0.01) | 1,400 (49) | 0.38 | 0.052* (0.01) | 1,400 (49) | 0.38 |
| Expenditure on education (per child) | −183.638 (226.29) | 1,386 | 0.23 | 801.399* (234.47) | 1,386 (49) | 0.13 | 915.726* (242.71) | 1,386 (49) | 0.13 |
| Expenditure on police (per capita) | 84.145* (7.80) | 1,386 | 0.40 | 90.599* (7.40) | 1,386 (49) | 0.40 | 81.630* (7.41) | 1,386 (49) | 0.38 |
| Metropolitan areas | |||||||||
| Total spending (per capita) | 746.934* (288.20) | 305 | 0.24 | 1215.535* (286.29) | 274 (48) | 0.18 | 1152.302* (306.36) | 242 (41) | 0.20 |
| Local government taxes (per capita) | −119.428 (94.66) | 305 | 0.54 | 206.820* (103.65) | 274 (48) | 0.50 | 196.811 (110.16) | 242 (41) | 0.48 |
| Intergovernmental revenue | 338.679* (153.58) | 304 | 0.11 | 601.903* (111.08) | 273 (47) | 0.06 | 597.337* (116.85) | 241 (40) | 0.04 |
| Local government surplus before transfers | −278.837 (149.00) | 304 | 0.09 | −568.565* (126.43) | 273 (47) | 0.04 | −593.598* (140.53) | 241 (40) | 0.02 |
| Local government surplus after transfers | 59.842 (79.36) | 304 | 0.05 | 33.339 (98.59) | 273 (47) | 0.04 | 3.740 (107.44) | 241 (40) | 0.06 |
| Cities | |||||||||
| Total spending (per capita) | 290.555* (103.49) | 1,020 | 0.13 | 496.959* (101.56) | 1,020 (51) | 0.12 | 477.211* (102.10) | 1,020 (51) | 0.12 |
| Local government taxes (per capita) | 148.565* (53.40) | 1,020 | 0.18 | 247.706* (45.10) | 1,020 (51) | 0.17 | 236.147* (45.29) | 1,020 (51) | 0.17 |
| Intergovernmental revenue | 53.752 (44.32) | 1,020 | 0.08 | 115.873* (39.97) | 1,020 (51) | 0.06 | 94.554* (39.78) | 1,020 (51) | 0.09 |
| Local government surplus before transfers | −129.000* (53.53) | 1,020 | 0.05 | −109.502* (54.74) | 1,020 (51) | 0.04 | −93.025 (54.91) | 1,020 (51) | 0.06 |
| Local government surplus after transfers | −75.249** (27.72) | 1,020 | 0.01 | 6.372 38.81 | 1,020 (51) | 0.01 | 1.529 (39.05) | 1,020 (51) | 0.01 |
Note: Standard errors are in parentheses.
* 95% confidence interval
Acknowledgements
Special thanks for comments on earlier versions of this paper go to Nikolaos Biziouras, Jiyoon Kim, Maria Koinova, Chappell Lawson, Neophytos Loizides, Omar McDoom, Jonathan Rodden, Jim Snyder and several anonymous reviewers. Jacob Vigdor kindly provided access to data sets. This article was written while the author was a Visiting Fellow in the Department of Government at the London School of Economics and Political Science (LSE), and the LSE's support is acknowledged with thanks.
Notes
1. Based on a search in the Thomson Reuters Web of Science on 5 November 2012.
2. Population size is normalized at 1, so that g represents the per capita and aggregate size of the public good.
3. Spending on roads is analysed in all three samples, while spending on welfare and education is only studied in the county and metropolitan areas samples, and spending on sewage and trash pickup and on fire protection only in the cities sample.
4. Appendix Table 1 summarizes key results. (The Appendix is available upon request.)
5. Including the share of the population over 65 controls for this only partially.
6. The Alesina, Baqir, and Easterly (1999) data sets are available from http://www.worldbank.org
7. These were based on US census data as reported in Cutler, Glaeser, and Vigdor (1999b).
8. The Appendix provides a summary of results. With regard to spending on public goods, the only difference between the ABE results and the replicated ABE analysis is that the relationship between ETHNIC and spending on welfare in the metropolitan areas sample is negative but not statistically significant in the replicated analysis. There are several similar differences in the results on other public finance outcomes, such as total expenditure and local taxes.
9. See Appendix Table 7.
10. Higher degrees of racial fractionalization are associated in the data with more spending on police, both per capita and as a share of total spending. The data also suggest higher levels of violent crime in more diverse areas. See Appendix Table 8.