Africa s Growth Tragedy, 20 Years On

Africa s Growth Tragedy, 20 Years On Gwen-Jiro Clochard * and Guillaume Hollard January 30, 2018 Abstract In their influential analysis of the determinants of growth in Sub-Saharan Africa, Easterly and Levine (1997) singled out the effect of ethnic divisions, revealing a diversity burden. Twenty years later, we here propose to revisit this diversity burden in the light of the substantial progress that has been made in data availability, econometric analysis and the understanding of the channels via which ethnic fractionalization may affect public-good provision. We use a much richer and larger dataset, instrumental-variable (rather than OLS) estimation to avoid potential biases in coefficient estimation, and test different assumptions regarding the channels via which ethnic diversity may influence public-good provision. We here instrument current ethnic diversity with a pre-colonial measure of the fractionalization index. We confirm the diversity burden hypothesis for eight out of the ten public goods considered, finding a substantial causal negative impact of ethnic diversity on public-good provision. However, the size of this effect varies greatly from *Ecole polytechnique - gwen-jiro.clochard@polytechnique.edu. Ecole polytechnique and CNRS - guillaume.hollard@polytechnique.edu. 1

one public good to another, and the opposite result is found for two public goods: ethnic diversity here positively affects access to the electricity grid and cellphone coverage. Our results therefore call for a better understanding of the channels that link ethnic diversity to public-good provision. The available models allow us to identify three public-good characteristics that produce heterogeneity in the regression coefficients: returns to scale, fixed costs and the complexity of managing the good. The comparison of the theoretical and empirical results leads us to suggest that returns to scale are the main driver of the relationship. Overall, our results suggest that, under some conditions, the coexistence of ethnic groups can positively affect the provision of public goods. In particular, when returns to scale are high, ethnic diversity appear to play a positive role. 2

1 Introduction In their influential analysis of the determinants of growth in Sub-Saharan Africa, Easterly and Levine (1997) singled out the effect of ethnic divisions. They found that ceteris paribus GDP per capita growth in totally heterogeneous societies is two percent per year lower than that in homogeneous societies. Subsequent work by Collier (2000) and Easterly (2001) has also supported the existence of this "diversity burden". In particular, more heterogeneous societies are found to provide fewer public goods. In La Porta et al. (1999), Alesina et al. (2003) and Banerjee et al. (2007), for example, more heterogeneous societies have worse infrastructure, and greater illiteracy and infant-mortality rates. Twenty years after the pioneering work of Easterly and Levine (1997), we here propose to revisit the idea of a diversity burden in the light of the substantial progress that has been made in data availability, econometric analysis and the understanding of the channels via which ethnic fractionalization may affect public-good provision. Our aim here is to re-examine the link between ethnic diversity and public-good access in the following ways. We use Afrobarometer survey data covering over 50,000 individuals in 36 African countries. This enables us to run regressions at the individual level while including country fixed effects in the regressions. In contrast, Easterly and Levine ran their analysis at the country level, with only three observations per country (each observation corresponding to average GDP growth over a decade). As in previous work, ethnic diversity is measured using the standard ethno- 3

linguistic fractionalization index. However the data now available allows us to calculate this at the regional level (a region being the first administrative level within a given country), rather than at the country level used in previous work. We can thus account for the large observed differences in public goods within countries. We use instrumental-variable estimation, rather than OLS, to establish a causal link between ethnic heterogeneity and the provision of public goods. This allows us to avoid potential problems such as measurement error, reverse causality and omitted-variable biases. We here propose the use of the pre-colonial measure of ethnic fractionalization in the areas around the main administrative centers to produce exogenous variation in ethnic diversity. Our assumption is that these cities were populated by the surrounding population, who were sometimes forced to move. Regions that were more heterogeneous thus produced greater ethnic diversity. We estimate the effect of ethnic diversity on the access to ten different public goods: the electricity grid, cellphone service (which in many African countries is provided publicly), clean water, sewage treatment, post offices, schools, police stations, health centers, transportation services and paved roads. We thus add public goods that were not available or were far less developed 20 years ago, like cellphone networks and the electricity grid We compare a number of different theoretical channels through which ethnic diversity affects public-good production, and in particular focus on the differences 4

between public goods. For eight out of ten of the public goods considered, we confirm the "diversity burden" hypothesis of Easterly and Levine (1997), since there is a substantial negiative causal impact of ethnic diversity on public-good provision. As an example, completely heterogeneous societies have 25 percent less access, ceteris paribus, to schools. This is comparable to the figure of 17 percent in Alesina et al. (1999). However, the size of the effect varies greatly from one public good to the other. Furthermore, for some important public goods the opposite result is found: ethnic diversity positively affects access to the electricity grid and cellphone coverage. Overall, our results then call for a better understanding of the channels that link ethnic diversity to public-good provision. The existing literature identifies three public-good characteristics that could lie behind the heterogeneity in the regression coefficients: returns to scale, fixed costs and the complexity of managing the good. The comparison of the theoretical and empirical results leads us to suggest that returns to scale are the main driver of the relationship. Despite the fact that (on average) the "diversity burden" argument of Easterly and Levine (1997) seems to hold, our results suggest that, under some conditions, the coexistence of ethnic groups can positively affect the provision of public goods. In particular, when returns to scale are high, ethnic diversity appears to play a positive role. The remainder of the paper is organized as follows. In Section 2, we present the data and the variables of interest in our regressions. The instrument and the estimation framework appear in Section 3, and the results in Section 4. The mechanisms that might explain the causal effects of ethnic diversity are then investigated in Section 5. 5

Last, Section 6 concludes. 2 Data and variables We use two different data sets to test the causal impact of ethnic diversity on access to public goods: the Afrobarometer (2015) and the geographical data on soil quality, access to river streams and ethnic homeland territories calculated by Nunn and Wantchekon (2011) and Michalopoulos and Papaioannou (2013). The Afrobarometer data comes from nationally-representative samples of primary sampling units (PSUs) selected with a probability proportional to population size (with a minimum size of 1200). We use data from 36 African countries: Algeria, Benin, Botswana, Burkina Faso, Burundi, Cameroon, Cape Verde, Côte d Ivoire, Egypt, Gabon, Ghana, Guinea, Kenya, Lesotho, Liberia, Madagascar, Malawi, Mali, Mauritius, Morocco, Mozambique, Namibia, Niger, Nigeria, Sao Tome and Principe, Senegal, Sierra Leone, South Africa, Sudan, Swaziland, Tanzania, Togo, Tunisia, Uganda, Zambia and Zimbabwe. The surveys were conducted face-to-face with respondents in their language of choice. Public goods The influential paper by Alesina et al. (1999) investigated the relationship between ethnic diversity and the inputs to public goods, i.e. public spending. Other papers (Alesina et al. (2003), Hollard and Sene (2015)) have considered the analogous influence on the output of the good, once produced (the literacy rate, school quality, the infant-mortality rate etc.). In this paper, we are situated somewhat in the middle as we look at the influence of fractionalization on access to public goods. Access is 6

a necessary but not sufficient condition for individuals to enjoy the good: having an electricity pylon close to one s home does not imply that one is connected to the grid; having a school nearby does not mean that it is of good quality, or even that one can afford to send one s children there (Hollard and Sene, 2015); health centers can be of poor quality if the doctors never show up (Banerjee and Duflo, 2006). We use the sixth round of the Afrobarometer, which contains data on individual characteristics, including whether the individual has access to a number of local public goods, and information about ethnicity. The variables of interest in this paper appear in Table 1. We want to evaluate the impact of ethnic diversity on a set of public goods: the electricity grid, cell-phone service (in many countries these two are publicly provided), piped-water systems, sewage-treatment mechanisms, post offices, schools, police stations, health centers, a form of paid transport and paved roads. The controls we introduce include age, education, participation in a religious group, urbanization, wealth, perceived corruption and gender. As will be seen in the results below, and as expected, urbanization and wealth are strongly correlated with access to public goods: they are therefore essential controls to include in the regression. We also calculate data on wealth and corruption, based on a set of related questions. 1 The Ethno-linguistic Fractionalization Index As is common in the literature, we measure ethnic diversity using the Ethno-linguistic Fractionalization Index (EFI - sometimes referred to as the ELF in the literature), which is the probability that two randomlyselected individuals belong to different ethnic groups (see Equation 1, where s i repre- 1 The wealth questions include possession of a television, radio or mobile phone. The corruption questions ask individuals about their beliefs regarding the involvement of public officials in corruption 7

sents the share of group i in the population). In this index, the larger the coefficient, the more ethnically diverse is the population. EF I = 1 i s 2 i (1) Table 1: Summary of variables Variable N Mean Std. Dev. Public goods Presence in the area Electricity grid 53919 0.65 0.48 Piped water 53759 0.60 0.49 Sewage treatment 53376 0.30 0.46 Cellphone service 53887 0.94 0.25 Post office 53291 0.25 0.43 School 53719 0.88 0.33 Police 53347 0.37 0.48 Clinic 53495 0.60 0.49 Transport system 53762 0.83 0.38 Road 53935 0.55 0.50 Controls Urbanized area 53335 0.41 0.49 Age 53641 37.27 14.54 Education 53780 3.47 2.23 Religion 52962 0.94 0.24 Gender 53935 0.50 0.50 Wealth 53634 0 0.74 Corruption 34471 0 0.86 Regional controls Surface (x1000km 2 ) 45505 29.89 46.07 Population (mill. hab) 45505 1.65 2.06 Suitability for agriculture 45505 0.44 0.23 Institutions 45505 1.69 0.84 Presence of rivers 45505 0.16 0.18 Ethnic Fractionalization Index EFI 456 0.45 0.3 Historic EFI 473 0.55 0.28 8

Using data from the Afrobarometer, we construct the fractionalization index at the regional level (there are 500 regions in the data set). In the data, the EF I has a mean of 0.45 and a standard deviation of 0.30, reflecting the great variability across the continent (see Figure 1). Figure 1: Distribution of the Ethno-Linguistic Fractionalization Index The second data base is derived from the work of Murdock (1967), who compiled a large amount of ethnographic data covering over 1,000 societies worldwide. In particular the data set contains information on the characteristics of ethnic groups and tribes in Africa prior to colonial times. The paper by Nunn and Wantchekon (2011), which considers the impact of the slave trade on social capital in Africa, provides additional information on the location of ethnic homelands, which will be of use for the construction of our instrument. 9

3 Using an IV approach to ensure causality The main contribution of this paper is to explore the causal impact of fractionalization on public goods using a large data set covering 36 countries. Many contributions here are based on simple correlations, which may produce misleading conclusions due to reverse causality, measurement error and omitted variables. The empirical technique of instrument variables (IV) allows us to solve both of these problems, provided that our instrument is relevant, i.e. a variable that affects public-good provision only through its impact on ethnic diversity. Identifying such an instrument is not easy, as the criteria that the instrument must satisfy (exogeneity and relevance) are challenging. One solution in the literature is to use information about history or geography as instruments. For instance,? use colonial settlers death as an instrument for institutional capacity, Nunn and Wantchekon (2011) consider distance to the coast to instrument the number of slaves taken from a country, and Atkin (2013) use altitude, rainfall and temperature to instrument income from agriculture. The instrument we propose here is a precolonial measure of the Ethnic Fractionalization Index, calculated using the historic homelands of ethnic groups. 3.1 Presentation of the instrument The instrument we construct is based on geographical data, compiled using data from Nunn and Wantchekon (2011) and Michalopoulos and Papaioannou (2016) (see Figure 2). The arguments regarding the exogeneity of the instrument are presented in the next section. The hypothesis behind our instrument is that if a city is located in an area with little 10

ethnic diversity, it is likely that population later will be less ethnically diverse than in cities that were located in more-diverse areas. For a given region, we draw a circle from the administrative center. 2 Within this circle, we calculate the area corresponding to the historical homeland of each ethnic group, which we use as a proxy for the number of people from this ethnic group within the circle (see Figure 3). 3 We then calculate the historic ethnic fractionalization index (historicef I) using the formula in Equation 1. This variable, which we use as an instrument, has a mean of 0.55 and a standard deviation of 0.28, which is quite similar to the values for EF I (see Figure 4). 2 We use a radius of 300 km (around 185 miles) for our estimations. As a robustness test, we checked alternative values ranging from 200 to 500 km, and found no significant differences. 3 We do this as we do not have direct information on population density in precolonial times. 11

Figure 2: The map of ethno-linguistic homelands (Source: Nunn and Wantchekon (2011)) Figure 3: The creation of the instrument (Source: authors calculations based on data from Nunn and Wantchekon (2011)) Figure 4: The distribution of the difference between current and historic EF I 12

3.2 Relevance According to ethnologists, African populations before the colonial period were characterized by low density, with few cities, and little ethnic diversity (Boserup, 1985). It was only during the colonial period that cities started to emerge, created by colonizers to extract resources. As they required labor, colonizers forced the local populations to move to the cities. In particular, for simplicity, they took populations who lived close to the newly-created cities. The idea on which our instrument is based is what we call an "aspiration effect": if the city was located in an area where the population was ethnically diverse in precolonial times, the current population is also more likely to be ethnically diverse than if the city was located in an area which was more homogeneous. 3.3 Conditional exogeneity There are three potentially significant caveats in the empirical estimation of the impact of ethnic diversity on the provision of public goods. The first is reverse causality: individuals may migrate to given areas due to the presence of certain public goods. This migration in turn affects the ethnic fractionalization of the area, either increasing or reducing EFI, depending on the incoming populations. If there is increasing diversity, the reverse impact is positive and the OLS coefficient is upwardly biased. As our instrument is based on precolonial data, when there were very few public goods (at least of the type of public goods considered here), our instrument enables us to avoid any issues of reverse causality. 13

The second caveat is the risk of an omitted variable that influences both ethnic diversity and the provision of public goods. The instrument we use is based on precolonial data, so that we can eliminate omitted variables that are more recent (including colonial treatments and post-colonial policies). However, there exist some factors that may have influenced migration before colonization, and continue to affect the provision of public goods today. Ethnic institutions existed before colonization, and these institutions have a significant influence on current economic outcomes, in particular public goods (see Gennaioli and Rainer (2007), Michalopoulos and Papaioannou (2013)). Tastes for institutions could produce population movements, mixing and merging ethnic groups. Controlling for pre-colonial institutions is therefore crucial for our exclusion restriction. We do this using the institutional data in Murdock (1967). Land characteristics could also influence both migration and the provision of public goods. We here control for two factors that could affect migration in the same way, but impact the provision of public goods differently. The first of these is land quality. Fertile land can encourage greater population movement, thus affecting ethnic diversity (as was the case for institutions, this could either increase or reduce the EF I). Greater fertility also produces more revenue, thus increasing the capacity to supply public goods. We control for fertility with a suitability index, which is the product of two components reflecting the climatic and soil suitability for cultivation (Source: Michalopoulos (2012), based on the Atlas of the Biosphere). This index is calculated after the year 2000, which could introduce error if soil suitability has changed since precolonial times; in particular this will bias our 14

instrument if this change is correlated with the other variables of interest. The second factor is access to water (rivers and oceans, as calculated by Michalopoulos (2012)). Since we control for land suitability, we assume that the impact of access to rivers will only affect ethnic fractionalization and the provision of public goods through access to fishing resources and transport (and not via the better soil that is found closer to coasts and rivers). Access to fisheries will impact our two variables of interest in the same way as land fertility by increasing the ability to supply public goods. On the other hand, access to transport could increase the demand for infrastructure (especially for transport, such as road or port facilities). The third caveat is measurement error. We use the same ethnic fractionalization index, which means that the errors in the calculation of this index will still be present, even in IV estimation. These errors are mainly twofold (see Posner (2004)): the Herfindahl measure is uni-dimensional, and therefore does not capture the many possible dimensions of diversity (language, religion, caste etc.) and the lack of information on the depth of the divisions between ethnic groups. Apart from these errors embedded in the fractionalization index, other errors may arise during the calculation of the index. The first refers to population size. It is easier to create a public good within the boundaries of an ethnic group if the population is larger, so that the impact of ethnic diversity would differ between two regions with the same EFI if they are of different size. We use information on population in 2000 at the regional level to control for region size. A measure of population in pre-colonial times might be preferable in terms of precision, but the lack of relevant data prevents the use of such a measure. The second error refers to region size. This could bias the 15

estimation in the same way as population. We therefore control for this at the regional level. Once we have controlled for these regional factors, our hypothesis is that having a more heterogeneous population in pre-colonial times in a given location is a random variable. We therefore use our pre-colonial measure of ethnic fractonalization as an instrument for current ethnic diversity. 3.4 Regression framework We wish to estimate the causal impact of ethno-linguistic fractionalization on a set of public goods. The Equation we estimate is the following for individual i in region j and country k: Y ijk = β 0 + β 1 EF I jk + γ 1 X ijk + γ 2 X jk + δ k + ɛ i (2) Here Y ijk is a public-good dummy for the presence of a school, clinic, paved road, etc. The vector X ijk represents a set of controls at the individual level, while X jk is the set of regional controls. The δ parameter represents the country fixed effects, including country characteristics such as national institutions (which according to Collier (2000) are key determinants of the impact of heterogeneity) and factor endowments (which could increase the correlations with ethnic diversity: Easterly and Levine (2003)). In the case of endogenous ethnic diversity, a simple ordinary least squares (OLS) estimation will not be consistent. We therefore estimate a Two-Stage Least Squares Instrumental Variable (2SLS-IV) model with the following equations: 16

Y ijk = β 0 + β 1 ˆ EF I jk + γ 1 X ijk + γ 2 X jk + δ k + ɛ i ˆ EF I jk = π 0 + π 1 historic_ef I jk + ω 1 X ijk + ω 2 X jk + γ k + µ i (3) As robustness checks, we added the controls one by one to see if this had an effect on the impact of ethnic diversity; we also varied the radius of the instrument. 4 4 Results 4.1 OLS results We first estimate Equation 2 without any controls: the results appear in Table 2. The coefficients here are not negative for all the goods, in contrast with the previous results in the literature. Table 2 also shows that there is no single pattern linking ethnic division and access to public goods, as the coefficients are significantly different from one good to the other. These differences in coefficients could reflect differences in the impact of omitted variables. As a test, we estimate Equation 2 via OLS, but this time with controls for other variables and country fixed effects. The results (in the Appendix) continue to show that there is not a unique pattern, with some coefficients being, as suggested in the literature, negative (piped water systems, schools and health centers), and others positive (such as sewage-treatment facilities), which is not a common result in the 4 The results from this robustness check are not shown in the paper, but are available upon request. Overall, this change did not affect the sign of the results. 17

Table 2: The results of OLS regressions without controls Dependent Electricity Cellphone Piped-water Sewage Post variable grid service system treatment office EFI -0.059*** -0.005 0.019** -0.064*** 0.017** (0.007) (0.004) (0.007) (0.008) (0.007) Constant 0.683*** 0.939*** 0.294*** 0.634*** 0.238*** (0.004) (0.002) (0.004) (0.004) (0.004) Schools Police Clinic Transport Road EFI 0.016** 0.037*** 0.035*** 0.036*** 0.072*** (0.005) (0.007) (0.008) (0.006) (0.008) Constant 0.867*** 0.352*** 0.581*** 0.811*** 0.511*** (0.003) (0.004) (0.004) (0.003) (0.005) * p<0.10, ** p<0.05, *** p<0.01. literature, and still others insignificant. The different patterns in these coefficients could reflect different levels of reverse causality or omitted variables which affect only some of the public goods. This is why we now turn to causality. 4.2 2SLS - IV results Our objective here is to estimate the causal link between ethnic fractionalization and access to public goods. We first check the relevance of our instrument. To do so we run an OLS regression on our instrument, controlling for all the controls we use in the regression (in order to avoid biases: see Angrist and Pischke (2008, Chapter 4)): this is the first stage of our 2SLS-IV regression (Equation 3). The results in Table 3 show that our instrument is relevant, as the Rule of Thumb (F _Stat > 10, see?) is respected. With our relevant instrument, we now estimate the 2SLS-IV model (Equations 3). As discussed above, this enables us to interpret the coefficient on ethnic fractionaliza- 18

Table 3: The first stage of the 2SLS-IV regression. Dependent variable: EFI Variable Coefficient (Std. Err.) Historic EFI 0.232 (0.009) Urbanized area 0.051 (0.002) Age 0.000 (0.000) Education 0.000 (0.001) Wealth -0.002 (0.002) Perceived corruption -0.002 (0.001) Religious -0.010 (0.005) Gender 0.001 (0.002) Region population 0.000 (0.000) Region size 0.000 (0.000) Suitability for agriculture 0.145 (0.007) Presence of rivers 0.158 (0.006) Intercept 0.302 (0.009) N 30129 R 2 0.081 F (42,30086) 219.668 Significance levels : : 10% : 5% : 1% tion in the regression causally. The IV-regression results appear in Tables 4 and 5. The comparisons between the results of the OLS and IV regressions are shown in Table 4.3. For the majority of public goods, our results suggest that the adverse impact of ethnic diversity found previously in the literature continues to hold. 19

Table 4: 2SLS-IV results I Dependent Electricity Cellphone Piped-water Sewage Post variable grid system treatment office EFI 0.168** 0.093* -0.224** -0.515*** -0.028 (0.078) (0.054) (0.078) (0.091) (0.079) Urbanized area 0.365*** 0.065*** 0.438*** 0.405*** 0.295*** (0.006) (0.004) (0.006) (0.007) (0.006) Age -0.000 0.000** -0.000 0.000 0.000** (0.000) Education 0.019*** 0.004*** 0.009*** 0.011*** 0.007*** (0.001) (0.001) (0.001) (0.001) (0.001) Wealth 0.103*** 0.022*** 0.039*** 0.064*** 0.027*** (0.003) (0.002) (0.003) (0.004) (0.003) Perceived corruption 0.010*** 0.002 0.008*** 0.013*** 0.002 (0.002) (0.002) (0.002) (0.003) (0.002) Religious 0.044*** 0.019** 0.005 0.046*** 0.052*** (0.009) (0.006) (0.009) (0.011) (0.009) Gender -0.032*** -0.006** -0.018*** -0.019*** -0.012** (0.004) (0.003) (0.004) (0.005) (0.004) Region population 0.000*** 0.000*** -0.000** -0.000-0.000** (0.000) Region size -0.000*** -0.000* -0.000 0.000** 0.000*** (0.000) Suitability for agriculture -0.112*** -0.062*** -0.069*** 0.065** -0.139*** (0.019) (0.013) (0.019) (0.022) (0.020) Presence of rivers 0.063*** -0.073*** 0.102*** 0.200*** -0.064*** (0.017) (0.012) (0.017) (0.020) (0.017) Constant 0.325*** 0.852*** 0.220*** 0.546*** 0.090** (0.037) (0.025) (0.037) (0.043) (0.037) * p<0.10, ** p<0.05, *** p<0.01. 20

Table 5: 2SLS-IV results II Dependent variable Schools Police Clinic Transport Road EFI -0.252*** -0.241** -0.850*** -0.164** -0.431*** (0.070) (0.095) (0.107) (0.078) (0.093) Urbanized area 0.046*** 0.346*** 0.235*** 0.195*** 0.326*** (0.005) (0.007) (0.008) (0.006) (0.007) Age -0.000-0.000* -0.000-0.000 0.000 (0.000) Education 0.005*** 0.008*** 0.011*** 0.007*** 0.009*** (0.001) (0.001) (0.002) (0.001) (0.001) Wealth 0.015*** 0.047*** 0.046*** 0.043*** 0.040*** (0.003) (0.004) (0.005) (0.003) (0.004) Perceived corruption -0.002-0.000-0.001 0.001 0.008** (0.002) (0.003) (0.003) (0.002) (0.003) Religious 0.017** 0.043*** 0.044*** 0.024** 0.021* (0.008) (0.011) (0.012) (0.009) (0.011) Gender -0.004-0.017*** -0.018** -0.011** -0.014** (0.004) (0.005) (0.006) (0.004) (0.005) Region population -0.000-0.000*** -0.000*** -0.000-0.000* (0.000) Region size 0.000** 0.000*** 0.000*** -0.000 0.000*** (0.000) Suitability for agriculture 0.066*** -0.058** 0.072** 0.011-0.119*** (0.017) (0.024) (0.026) (0.019) (0.023) Presence of rivers 0.037** -0.017 0.086*** 0.103*** 0.130*** (0.015) (0.021) (0.023) (0.017) (0.020) Constant 0.923*** 0.318*** 0.832*** 0.769*** 0.573*** (0.033) (0.045) (0.050) (0.037) (0.044) * p<0.10, ** p<0.05, *** p<0.01. 21

4.3 Heterogeneity The causality results (Tables 4 and 5) suggest that the results in the papers on the link between ethnic diversity and access to public goods were almost right: for eight of the ten public goods considered here there is a negative causal impact of ethnic diversity. However, for the other two goods (Table 4), namely the development of the electricity grid and the cellphone service, the link is causally positive. This is in contradiction with what is found in the literature, which has suggested that no matter the public good, the impact of ethnic diversity is negative. Having controlled for country fixed effects, which capture national characteristics such as democracy or the enforcement of property rights, the argument of Collier (2000) that the impact of ethnic heterogeneity should be diluted when controlling for such variables does not seem to hold. In terms of size, ceteris paribus a completely heterogeneous country is predicted to have 20% fewer schools than a fully homogeneous country. This figure is similar to that of 17% in Alesina et al. (1999) for education spending. The estimate for roads, which was consistent with the figure in Alesina et al. (1999) in the OLS regressions (around 7%: see the Appendix), is much larger in the causal analysis (43%). There could thus be a reverse-causality effect for this good: this would be positive (areas which have more roads tend to attract more heterogeneous populations), producing an upward bias in the OLS estimate that brings it closer to zero. The regression results here show that there are differences between public goods, which in turn produce differential impacts of ethnic diversity. The next section discusses the mechanisms that might explain these differences. 22

Table 6: EFI regressions: OLS without controls, OLS with controls and IV OLS without controls OLS with controls 2SLS-IV Electricity grid -0.059*** -0.034** 0.168** (0.007) (0.011) (0.078) Cellphone service -0.005-0.035*** 0.093* (0.004) (0.007) (0.054) Sewage treatment 0.019** 0.127*** -0.224** (0.007) (0.011) (0.078) Piped water -0.064*** -0.063*** -0.515*** (0.008) (0.012) (0.091) Post offices 0.017** 0.036** -0.028 (0.007) (0.011) (0.079) Schools 0.016** -0.078*** -0.252*** (0.005) (0.010) (0.070) Police stations 0.037*** 0.083*** -0.241** (0.007) (0.013) (0.095) Health centers 0.035*** -0.065*** -0.850*** (0.008) (0.014) (0.107) Transport service 0.036*** -0.070*** -0.164** (0.006) (0.011) (0.078) Roads 0.072*** -0.034** -0.431*** (0.008) (0.013) (0.093) * p<0.10, ** p<0.05, *** p<0.01. 5 Mechanisms to explain the differences between public goods: A comparison We here try to understand the differences in the effect of ethnic fractionalization on the provision of different public goods. A comparison paper (Clochard and Hollard, forthcoming) compares a number of the models that have appeared in the literature following Easterly and Levine (1997), which have shaped our understanding of the channels through which ethnic diversity affects public-good provision. This section summarizes the findings in this companion paper. 23

Three public-good parameters have been identified in the literature. For each of these, we calculate in Table 7 the expected ranking of the coefficient in the ethnicfractionalization regressions. The first parameter is the returns to scale of the good. In the models of Alesina et al. (1999) and Alesina and La Ferrara (2000), the optimal contribution to the public good is directly dependent on returns to scale (g = (α(1 ˆl m )) 1 1 α ). The second parameter is the complexity of the good, from Banerjee and Duflo (2006). The underlying idea is that communities will have difficulty in monitoring complex goods. Regarding the impact of ethnic diversity, a more heterogeneous society will have greater difficulty in agreeing on efficient monitoring processes for complex goods. We here proxy complexity by the level of postsecondary education required to manage the public good. The third and final characteristic is the fixed cost. If large investments are required to create the public good (for example, the creation of a dam), individuals need to have more trust in the participation of other members. Ethnic diversity, in this model (as developed by Hollard and Sene (2015)), reduces trust between individuals, and thus reduces the individuals contributions. Results Table 7 is constructed using the coefficients found in the literature. These parameters enable us to calculate the expected ranking of the EFI coefficients in the regressions from the previous section. 5 Based on the rankings in Table 7, we can calculate, for each model, the standard error of the difference between the prediction 5 The ranking is from the highest to the lowest coefficient. 24

and the estimation result. These standard errors are shown in Table 8. We consider that the lowest standard error corresponds to the best fit, and thus the best model. Table 7: Predictions of the different models Public Characteristic Expected sign Expected Good of the EFI coefficient Ranking Returns to scale α Electricity grid 1.14 >0 1 Post offices 1.25 >0 2 Schools 1 0 4/5 Sewage treatment 1.34 >0 3 Water sanitation 1 0 4/5 Police 0.8 <0 6 Clinics 0.9 <0 7 Complexity (proxied by the level of post-secondary education needed to produce and manage the public good) Education Electricity grid 5 <0 6 Post offices 0 <0 1/2 Schools 4 <0 5 Sewage treatment 0 <0 1/2 Water sanitation 2 <0 3/4 Police 2 <0 3/4 Clinics 8 <0 7 Fixed costs Cost ranking Electricity grid 1 <0 7 Post offices 7 <0 1 Schools 6 <0 2 Sewage treatment 2 <0 6 Water sanitation 3 <0 5 Police 5 <0 3 Clinics 4 <0 4 The results show that the best model fit comes from returns to scale. The regression results (based on only seven public goods) are significant at the 1% level for the returns-to-scale ranking of the estimation results. However, for post offices and 25

Model Table 8: Comparison of models and estimation results Standard error of the comparison of rankings Returns to scale 1.00 Complexity 2.65 Fixed costs 3.16 sewage-treatment facilities, despite being in the right position in the ranking, the sign of the coefficient is not as expected. This means that returns to scale are the most likely solution to explain the errors. Despite having the largest standard deviation in the estimation result (the largest error) when controlling for returns to scale, the coefficient for fixed costs is significant at the 10% level. As such, a significant part of the variance that is not explained by returns to scale is captured by fixed costs. 6 Conclusion A large stream of research, deriving from the influential work of Easterly and Levine (1997) and Alesina et al. (1999), has found an adverse impact of ethnic heterogeneity on economic outcomes, and in particular on public goods. More recently, however, a number of papers have expressed doubts about this correlation, and found only small (or even positive) effects of ethnic diversity (Gisselquist et al. (2016), Hollard and Sene (2015)). This paper aims to reassess these findings by providing, via IV estimation, an analysis of the causal impact. In addition, we here use more complete and reliable datasets, which produce more reliable results. 26

The estimation results are overall in line with the "diversity burden". However, the impact of ethnic diversity on public goods depends on the type of public good, both in terms of magnitude and sign. The existing models in the literature suggests three channels through which ethnic diversity affects public-good provision: returns to scale, fixed costs and the complexity of the management of the good. The comparison of the predictions from these models and the estimation results suggests that the best fit comes from returns to scale, although fixed costs also appear to play a role. Overall, the results from both theory and the empirical estimations show that the central result of Easterly and Levine (1997) and Alesina et al. (1999) does not seem to hold unconditionally. For eight out of the ten public goods in this paper, the "diversity debit" holds. However, in part by not focussing on the consequences of the returns to scale in public goods, these contributions missed differences between them, which can produce a positive relationship between fractionalization and public-good provision. Under some conditions, it is then possible to make people collaborate in publicgood production. Further research here should therefore focus on policies that increase individuals ability to collaborate, appealing to the more global notion of social capital. For instance, Attanasio et al. (2015) show that mandatory meetings were effective in enhancing social capital by encouraging communication between members. In Africa in particular, there exists one specific field in which ethnic lines appear to blur: football. In addition to having a direct impact on social ties (see Seippel (2006)), major sporting events may be occasions for creating meetings (for instance during public screenings of the game), which were found by Attanasio et al. (2015) to enhance communication 27

between communities. 28

References Alberto Alesina and Eliana La Ferrara. Participation in heterogeneous communities. The quarterly journal of economics, 115(3):847 904, 2000. Alberto Alesina, Reza Baqir, and William Easterly. Public goods and ethnic divisions. The Quarterly Journal of Economics, 114(4):1243 1284, 1999. Alberto Alesina, Arnaud Devleeschauwer, William Easterly, Sergio Kurlat, and Romain Wacziarg. Fractionalization. Journal of Economic growth, 8(2):155 194, 2003. Joshua D Angrist and Jörn-Steffen Pischke. Mostly harmless econometrics: An empiricist s companion. Princeton university press, 2008. David Atkin. Trade, tastes, and nutrition in india. The American Economic Review, 103(5):1629 1663, 2013. Orazio Attanasio, Sandra Polania-Reyes, and Luca Pellerano. Building social capital: Conditional cash transfers and cooperation. Journal of Economic Behavior & Organization, 118:22 39, 2015. Abhijit Banerjee and Esther Duflo. Addressing absence. The Journal of Economic Perspectives, 20(1):117 132, 2006. Abhijit Banerjee, Lakshmi Iyer, and Rohini Somanathan. Public action for public goods. Handbook of development economics, 4:3117 3154, 2007. Ester Boserup. Economic and demographic interrelationships in sub-saharan africa. Population and Development Review, pages 383 397, 1985. 29

Gwen-Jiro Clochard and Guillaume Hollard. Ethnic fractionalization and the provision of public goods: They were almost right. forthcoming. Paul Collier. Ethnicity, politics and economic performance. Economics & Politics, 12 (3):225 245, 2000. William Easterly. Can institutions resolve ethnic conflict? Economic Development and Cultural Change, 49(4):687 706, 2001. William Easterly and Ross Levine. Africa s growth tragedy: policies and ethnic divisions. The Quarterly Journal of Economics, pages 1203 1250, 1997. William Easterly and Ross Levine. Tropics, germs, and crops: how endowments influence economic development. Journal of monetary economics, 50(1):3 39, 2003. Nicola Gennaioli and Ilia Rainer. The modern impact of precolonial centralization in africa. Journal of Economic Growth, 12(3):185, 2007. Rachel M Gisselquist, Stefan Leiderer, and Miguel Niño-Zarazúa. Ethnic heterogeneity and public goods provision in zambia: Evidence of a subnational diversity dividend. World Development, 78:308 323, 2016. Guillaume Hollard and Omar Sene. What drives quality of schools in africa? disentangling social capital and ethnic divisions. 2015. Rafael La Porta, Florencio Lopez-de Silanes, Andrei Shleifer, and Robert Vishny. The quality of government. Journal of Law, Economics, and organization, 15(1):222 279, 1999. 30

Stelios Michalopoulos. The origins of ethnolinguistic diversity. The American economic review, 102(4):1508 1539, 2012. Stelios Michalopoulos and Elias Papaioannou. Pre-colonial ethnic institutions and contemporary african development. Econometrica, 81(1):113 152, 2013. Stelios Michalopoulos and Elias Papaioannou. The long-run effects of the scramble for africa. The American Economic Review, 106(7):1802 1848, 2016. George Peter Murdock. Ethnographic atlas. 1967. Nathan Nunn and Leonard Wantchekon. The slave trade and the origins of mistrust in africa. The American Economic Review, 101(7):3221 3252, 2011. Daniel N Posner. Measuring ethnic fractionalization in africa. American journal of political science, 48(4):849 863, 2004. Ørnulf Seippel. Sport and social capital. Acta sociologica, 49(2):169 183, 2006. 31

Annex 1: Regression results for each public good Table 9: Regression results for the Electricity grid OLS without controls OLS with controls IV EFI -0.059*** -0.034** 0.168** (0.007) (0.011) (0.078) Urbanized area 0.375*** 0.365*** (0.005) (0.006) Age -0.000-0.000 Education 0.019*** 0.019*** (0.001) (0.001) Wealth 0.103*** 0.103*** (0.003) (0.003) Perceived corruption 0.010*** 0.010*** (0.002) (0.002) Religious 0.042*** 0.044*** (0.009) (0.009) Gender -0.032*** -0.032*** (0.004) (0.004) Region population 0.000*** 0.000*** Region size -0.000*** -0.000*** Suitability for agriculture -0.077*** -0.112*** (0.014) (0.019) Presence of rivers 0.093*** 0.063*** (0.012) (0.017) Constant 0.683*** 0.413*** 0.325*** (0.004) (0.014) (0.037) * p<0.10, ** p<0.05, *** p<0.01. 32

Table 10: Regression results for cellphone service OLS without controls OLS with controls IV EFI -0.005-0.035*** 0.093* (0.004) (0.007) (0.054) Urbanized area 0.071*** 0.065*** (0.003) (0.004) Age 0.000** 0.000** Education 0.004*** 0.004*** (0.001) (0.001) Wealth 0.022*** 0.022*** (0.002) (0.002) Perceived corruption 0.002 0.002 (0.002) (0.002) Religious 0.018** 0.019** (0.006) (0.006) Gender -0.006** -0.006** (0.003) (0.003) Region population 0.000*** 0.000*** Region size -0.000-0.000* Suitability for agriculture -0.039*** -0.062*** (0.010) (0.013) Presence of rivers -0.052*** -0.073*** (0.008) (0.012) Constant 0.939*** 0.907*** 0.852*** (0.002) (0.010) (0.025) * p<0.10, ** p<0.05, *** p<0.01. 33

Table 11: Regression results for sewage-treatment systems OLS without controls OLS with controls IV EFI 0.019** 0.127*** -0.224** (0.007) (0.011) (0.078) Urbanized area 0.422*** 0.438*** (0.004) (0.006) Age 0.000-0.000 Education 0.009*** 0.009*** (0.001) (0.001) Wealth 0.039*** 0.039*** (0.003) (0.003) Perceived corruption 0.009*** 0.008*** (0.002) (0.002) Religious 0.009 0.005 (0.009) (0.009) Gender -0.018*** -0.018*** (0.004) (0.004) Region population 0.000* -0.000** Region size -0.000*** -0.000 Suitability for agriculture -0.129*** -0.069*** (0.014) (0.019) Presence of rivers 0.050*** 0.102*** (0.012) (0.017) Constant 0.294*** 0.065*** 0.220*** (0.004) (0.014) (0.037) * p<0.10, ** p<0.05, *** p<0.01. 34

Table 12: Regression results for piped-water systems OLS without controls OLS with controls IV EFI -0.064*** -0.063*** -0.515*** (0.008) (0.012) (0.091) Urbanized area 0.383*** 0.405*** (0.005) (0.007) Age 0.000** 0.000 Education 0.011*** 0.011*** (0.001) (0.001) Wealth 0.065*** 0.064*** (0.004) (0.004) Perceived corruption 0.014*** 0.013*** (0.003) (0.003) Religious 0.053*** 0.046*** (0.010) (0.011) Gender -0.019*** -0.019*** (0.005) (0.005) Region population 0.000** -0.000 Region size 0.000 0.000** Suitability for agriculture -0.012 0.065** (0.016) (0.022) Presence of rivers 0.134*** 0.200*** (0.014) (0.020) Constant 0.634*** 0.346*** 0.546*** (0.004) (0.016) (0.043) * p<0.10, ** p<0.05, *** p<0.01. 35

Table 13: Regression results for Post Offices OLS without controls OLS with controls IV EFI 0.017** 0.036** -0.028 (0.007) (0.011) (0.079) Urbanized area 0.292*** 0.295*** (0.005) (0.006) Age 0.000** 0.000** Education 0.007*** 0.007*** (0.001) (0.001) Wealth 0.027*** 0.027*** (0.003) (0.003) Perceived corruption 0.002 0.002 (0.002) (0.002) Religious 0.053*** 0.052*** (0.009) (0.009) Gender -0.012** -0.012** (0.004) (0.004) Region population -0.000** -0.000** Region size 0.000*** 0.000*** Suitability for agriculture -0.150*** -0.139*** (0.014) (0.020) Presence of rivers -0.072*** -0.064*** (0.013) (0.017) Constant 0.238*** 0.062*** 0.090** (0.004) (0.014) (0.037) * p<0.10, ** p<0.05, *** p<0.01. 36

Table 14: Regression results for Schools OLS without controls OLS with controls IV EFI 0.016** -0.078*** -0.252*** (0.005) (0.010) (0.070) Urbanized area 0.038*** 0.046*** (0.004) (0.005) Age -0.000-0.000 Education 0.005*** 0.005*** (0.001) (0.001) Wealth 0.015*** 0.015*** (0.003) (0.003) Perceived corruption -0.001-0.002 (0.002) (0.002) Religious 0.020** 0.017** (0.008) (0.008) Gender -0.004-0.004 (0.004) (0.004) Region population 0.000* -0.000 Region size 0.000* 0.000** Suitability for agriculture 0.037** 0.066*** (0.012) (0.017) Presence of rivers 0.013 0.037** (0.011) (0.015) Constant 0.867*** 0.845*** 0.923*** (0.003) (0.013) (0.033) * p<0.10, ** p<0.05, *** p<0.01. 37

Table 15: Regression results for Police Stations OLS without controls OLS with controls IV EFI 0.037*** 0.083*** -0.241** (0.007) (0.013) (0.095) Urbanized area 0.331*** 0.346*** (0.006) (0.007) Age -0.000-0.000* Education 0.008*** 0.008*** (0.001) (0.001) Wealth 0.047*** 0.047*** (0.004) (0.004) Perceived corruption 0.001-0.000 (0.003) (0.003) Religious 0.048*** 0.043*** (0.011) (0.011) Gender -0.017*** -0.017*** (0.005) (0.005) Region population -0.000-0.000*** Region size 0.000*** 0.000*** Suitability for agriculture -0.114*** -0.058** (0.017) (0.024) Presence of rivers -0.064*** -0.017 (0.015) (0.021) Constant 0.352*** 0.175*** 0.318*** (0.004) (0.017) (0.045) * p<0.10, ** p<0.05, *** p<0.01. 38

Table 16: Regression results for Health Centers OLS without controls OLS with controls IV EFI 0.035*** -0.065*** -0.850*** (0.008) (0.014) (0.107) Urbanized area 0.197*** 0.235*** (0.006) (0.008) Age 0.000-0.000 Education 0.011*** 0.011*** (0.001) (0.002) Wealth 0.048*** 0.046*** (0.004) (0.005) Perceived corruption 0.001-0.001 (0.003) (0.003) Religious 0.054*** 0.044*** (0.012) (0.012) Gender -0.019*** -0.018** (0.005) (0.006) Region population -0.000* -0.000*** Region size 0.000** 0.000*** Suitability for agriculture -0.062*** 0.072** (0.018) (0.026) Presence of rivers -0.029* 0.086*** (0.016) (0.023) Constant 0.581*** 0.486*** 0.832*** (0.004) (0.018) (0.050) * p<0.10, ** p<0.05, *** p<0.01. 39

Table 17: Regression results for Transport systems OLS without controls OLS with controls IV EFI 0.036*** -0.070*** -0.164** (0.006) (0.011) (0.078) Urbanized area 0.190*** 0.195*** (0.005) (0.006) Age -0.000-0.000 Education 0.007*** 0.007*** (0.001) (0.001) Wealth 0.044*** 0.043*** (0.003) (0.003) Perceived corruption 0.001 0.001 (0.002) (0.002) Religious 0.025** 0.024** (0.009) (0.009) Gender -0.011** -0.011** (0.004) (0.004) Region population -0.000-0.000 Region size -0.000-0.000 Suitability for agriculture -0.005 0.011 (0.014) (0.019) Presence of rivers 0.092*** 0.103*** (0.012) (0.017) Constant 0.811*** 0.727*** 0.769*** (0.003) (0.014) (0.037) * p<0.10, ** p<0.05, *** p<0.01. 40

Table 18: Regression results for Roads OLS without controls OLS with controls IV EFI 0.072*** -0.034** -0.431*** (0.008) (0.013) (0.093) Urbanized area 0.307*** 0.326*** (0.005) (0.007) Age 0.000 0.000 Education 0.009*** 0.009*** (0.001) (0.001) Wealth 0.041*** 0.040*** (0.004) (0.004) Perceived corruption 0.009** 0.008** (0.003) (0.003) Religious 0.026** 0.021* (0.011) (0.011) Gender -0.015** -0.014** (0.005) (0.005) Region population 0.000* -0.000* Region size 0.000 0.000*** Suitability for agriculture -0.188*** -0.119*** (0.016) (0.023) Presence of rivers 0.073*** 0.130*** (0.014) (0.020) Constant 0.511*** 0.398*** 0.573*** (0.005) (0.017) (0.044) * p<0.10, ** p<0.05, *** p<0.01. 41

Annex 2: IV regressions with different controls for the schools regression 42

Table 19: Different controls for schools (1) (2) (3) (4) (5) (6) (7) (8) (9) EFI -1.119*** -0.854*** -0.859*** -0.804*** -0.758*** -0.233*** -0.209*** -0.209*** -0.252*** (0.221) (0.177) (0.182) (0.182) (0.180) (0.062) (0.062) (0.062) (0.070) Urbanized 0.090*** 0.090*** 0.079*** 0.072*** 0.043*** 0.040*** 0.040*** 0.046*** area (0.010) (0.010) (0.010) (0.010) (0.005) (0.005) (0.005) (0.005) Age -0.000** -0.000-0.000-0.000-0.000-0.000-0.000 (0.000) Education 0.007*** 0.005*** 0.005*** 0.005*** 0.005*** 0.005*** (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Wealth 0.018*** 0.015*** 0.015*** 0.015*** 0.015*** (0.003) (0.003) (0.003) (0.003) (0.003) Perceived -0.001-0.001-0.001-0.002 corruption (0.002) (0.002) (0.002) (0.002) Religious 0.020** 0.020** 0.017** (0.008) (0.008) (0.008) Gender -0.004-0.004 (0.003) (0.004) Region -0.000 population (0.000) Region size 0.000** (0.000) Suitability for 0.066*** agriculture (0.017) Presence 0.037** of rivers (0.015) Constant 1.455*** 1.279*** 1.293*** 1.239*** 1.229*** 0.969*** 0.938*** 0.939*** 0.923*** (0.115) (0.088) (0.093) (0.092) (0.091) (0.032) (0.033) (0.033) (0.033) * p<0.10, ** p<0.05, *** p<0.01. 43