## Abstract

The relationship between ethnic heterogeneity and economic growth is complex. Empirical research working with cross-country data finds a negative, or statistically insignificant, relationship. However, analysis at the city level finds a positive effect of diversity on wages and productivity. Generally there is a trade-off between the economic benefits of diversity and the costs of heterogeneity. Using cells of fixed size, we find that the relationship between diversity and growth is positive for small geographical areas. In the case of Africa, we argue that the explanation is the increase in trade at the boundaries between ethnic groups due to ethnic specialization.

## I. Introduction

THE issue of the effect of ethnic diversity on economic development has generated a large body of literature. Initially the literature analyzed the issue by running cross-country regressions. For example, using cross-country differences in ethnic diversity, Easterly and Levine (1997) show that Africa's low level of economic development is associated with its high degree of ethnic heterogeneity. Alesina et al. (2003) and Alesina and La Ferrara (2005), also using cross-country data, similarly show a consistent negative effect of ethnic fractionalization on growth. Further research has qualified the conditions for this negative relationship.1 By contrast, research based on data from small geographical areas, such as cities, frequently finds a positive effect of diversity on wages and productivity.2

Is ethnic diversity good or bad for economic growth? The literature has often emphasized the trade-off between the benefits of diversity and the costs of heterogeneity. On the one hand, ethnic diversity can be beneficial by enhancing productivity through innovation, skill complementarities, increased creativity, trade, and product variety. On the other hand, diversity can generate an inefficient provision of public goods, ethnically biased policies, and conflict or disagreement over common public goods and policies. All of these theories generally imply that there is a size at which benefits and costs are equalized, implying that on a smaller scale, we should find a positive effect of ethnic diversity, and on a larger scale we should find a negative effect. However, theoretical models on the effects of ethnic diversity on economic development are mostly agnostic about the scale of analysis. The literature has thus generally found that diversity seems to be negative, or irrelevant, for development at high levels of aggregation but positive at low levels of geographical aggregation, such as the city level. Detailed geographical data are, in fact, becoming increasingly popular in the analysis of ethnicity.3

In this paper, we argue that the answer to the question of the nature of the relationship between ethnic heterogeneity and growth is, in fact, different depending on the size of the unit of analysis. From a theoretical perspective, if we assume that ethnic specialization increases the variety of goods in the presence of ethnic diversity, it is simple to write down a growth model that depicts a positive relationship between growth and ethnic diversity. In fact, the growth rate of most product-variety models increases with the size of the economy, as measured by total labor supply,4 which is positively correlated with diversity. However, as we increase the size of the unit and therefore the degree of diversity, the cost of heterogeneity increases. There is an optimal level of diversity determined by the trade-off between the benefits and the costs of ethnic heterogeneity. The larger the benefits from skill variety in production, the larger the size, while the higher the effect of heterogeneity on the unwillingness to share public goods, the smaller the size. Obviously the specific mechanism that explains the benefits and costs when ethnic heterogeneity increases may be different depending on the level of development and the sectoral structure of the economy.

One mechanism, which we explore in this paper, is based on the increase in trade due to the production specialization or service complementarities of different ethnic groups. The issue of the impact of spatial ethnic heterogeneity on intranational trade is an underdeveloped topic of research. Aker et al. (2014) argue, using information from Niger, that transaction costs are higher for trade between regions with different ethnicities versus trade in homogeneous areas. It is well known that trust is higher among people of the same ethnic group and that this consequently reduces transaction costs. However, ethnic specialization in production generates a motive for trade that is absent within homogeneous groups. There are, in fact, many mitigating mechanisms that can reduce the cost of transactions with other ethnic groups. Several papers have analyzed mechanisms that can support trade among agents that belong to different groups—for instance, Glaeser (2005), Greif (1993, 2000, 2006), and Jha (2013).

At a very high level of resolution of the grid, there is obviously no possibility of finding measures of output, value added, or even wages for most countries. We consequently take advantage of luminosity data to proxy for local economic activity. Recent research has shown that light density at night is a good proxy of economic activity.5 We find that at the highest degree of resolution, there is a positive association between ethnic heterogeneity and economic growth. Finding this correlation at the country level does not, however, resolve the issue of endogeneity caused by the possibility that other unobserved characteristics can drive the association via, for instance, institutional differences. Using artificially constructed cells mitigates this concern. Taking advantage of arbitrarily drawn borders for arbitrary levels of aggregation allows for the control using fixed effects and mitigates the concern of endogeneity of the contemporaneous boundaries of countries.

In addition, we show that the results are robust to a large number of potential issues. First, we show that the results are unaffected by the use of a large number of controls to account for within-country variation, such as geography, climate, soil quality, and proximity to lakes, or political capitals. Second, we run 100 different regressions, randomly changing the initial location of the point that defines, together with the level of resolution, the exact location of the area covered by each cell. The results are robust to the location of the origin of the grid that defines the cells. We also show that the positive effect of heterogeneity on economic growth is not only due to the contribution of urban areas and is not simply capturing an agglomeration effect. Finally, we find that reducing the degree of resolution of the grid decreases the association between ethnic diversity and development, to the point of finding no association between heterogeneity and development.6

These findings are likely to derive from different mechanisms, depending on the sectoral structure of the economy or its level of development. In particular, we consider the case of Africa. In order to understand why regional development is faster along ethnic borders, we propose a mechanism related to trade. We find that areas that have more ethnic diversity also have a higher proportion of markets. Previous research (Michalopoulos, 2012) has shown that ethnic groups in Africa typically specialize in different agricultural products and therefore have incentives to exchange goods. Ethnic groups that are geographically close to one another and can potentially monitor each other may therefore have a large volume of trade despite their different ethnic origins. In order to provide evidence of this mechanism, we show that local markets in Africa are located close to ethnic borders, supporting this interpretation.

The structure of the paper is as follows. Section II describes the data. Section III presents the basic results and discusses some exercises that show the robustness of the results. Section IV proposes a mechanism for the case of Africa and presents supporting evidence. Section V concludes.

## II. Data

The previous section argues that the effect of ethnic heterogeneity on growth depends on the balance between the positive effects of diversity (e.g., skill complementarities, propensity to innovate, specialization and trading) and its the negative effects (e.g., problems in the provision of public goods, potential conflict). The empirical literature generally finds that the negative effect of diversity prevails over the positive effect at the country level. From a theoretical perspective, it is clear that for small areas, the positive effect of diversity should generally dominate the negative effect, while the opposite should be true for large geographical areas. In this paper, we therefore investigate the effect of ethnic diversity on development using small geographical units to investigate the sign of that relationship.

Our units of observation are grid-country cells that generate country-level data at very high resolution. We construct grids and calculate the value of the explanatory variables and the outcome for each of these cells. We also check the robustness of the results by changing the origin of the grid that generates the country cells. Table 1 presents the descriptive statistics of the variables used in the empirical analysis. The basic variables for the specification are measures of local growth and ethnic diversity. We also describe the control variables included in the regressions.

Table 1.

Summary Statistics

ObservationsMeanSDQ50Q90
Growth 21,514 1.509 2.800 0.623 6.266
Log night light 1992 pc 21,514 $-$5.312 4.159 $-$4.332 $-$0.645
Ethnic Fractionalization 21,514 0.113 0.191 0.000 0.467
Distance to Coastline (km) 21,514 144.219 181.713 72.026 387.069
Terrain Ruggedness Index, 100 m 21,514 9,869.915 3,338.446 10,704.371 11,778.844
Coastline Border 21,514 0.450 0.498 0.000 1.000
Average temperature from 1961–1980 21,514 13.380 11.836 15.190 26.867
Average precipitation from 1961–1980 21,514 79.630 73.480 53.872 187.556
Log. Population Density 21,514 1.204 2.871 1.507 4.665
Share Mining 21,514 0.000 0.011 0.000 0.000
% Fertile Soil 21,514 27.683 36.213 2.778 96.528
Distance to River (km) 21,514 13.652 14.940 13.334 29.817
Distance to Lake (km) 21,514 8.429 54.004 2.778 12.423
Distance to Equatorial Line 21,514 24.844 28.699 28.000 60.161
Log. OBS area 21,514 7.324 3.585 9.351 9.968
Border (yes$=$1) 21,514 0.271 0.444 0.000 1.000
Ecological Fractionalization 21,514 0.220 0.292 0.020 0.610
Pathogen Stress Index 21,514 11.48 3.63 11 16
ObservationsMeanSDQ50Q90
Growth 21,514 1.509 2.800 0.623 6.266
Log night light 1992 pc 21,514 $-$5.312 4.159 $-$4.332 $-$0.645
Ethnic Fractionalization 21,514 0.113 0.191 0.000 0.467
Distance to Coastline (km) 21,514 144.219 181.713 72.026 387.069
Terrain Ruggedness Index, 100 m 21,514 9,869.915 3,338.446 10,704.371 11,778.844
Coastline Border 21,514 0.450 0.498 0.000 1.000
Average temperature from 1961–1980 21,514 13.380 11.836 15.190 26.867
Average precipitation from 1961–1980 21,514 79.630 73.480 53.872 187.556
Log. Population Density 21,514 1.204 2.871 1.507 4.665
Share Mining 21,514 0.000 0.011 0.000 0.000
% Fertile Soil 21,514 27.683 36.213 2.778 96.528
Distance to River (km) 21,514 13.652 14.940 13.334 29.817
Distance to Lake (km) 21,514 8.429 54.004 2.778 12.423
Distance to Equatorial Line 21,514 24.844 28.699 28.000 60.161
Log. OBS area 21,514 7.324 3.585 9.351 9.968
Border (yes$=$1) 21,514 0.271 0.444 0.000 1.000
Ecological Fractionalization 21,514 0.220 0.292 0.020 0.610
Pathogen Stress Index 21,514 11.48 3.63 11 16

Own elaboration from several sources.

### A. Local Growth

To measure growth in each cell, we need information on economic development. At high levels of resolution, it is difficult to find estimations of GDP and, certainly, many areas of the world do not have information on geocoded high-resolution measures of economic development. It has, however, become increasingly common to use satellite night light density as a proxy for local economic activity when working with small geographical areas. Satellite night light density data are available from the National Oceanic and Atmospheric Administration and have been used recently by scholars such as Henderson, Storeygard, and Weil (2012), Michalopoulos (2012), Michalopoulos and Papaioannou (2013, 2014), and Alesina, Michalopoulos et al. (2016). There is also a series of papers that specifically corroborate a high within-country correlation between GDP and light density at night. Chen and Nordhaus (2011) find that luminosity has informational value for countries, regions, and areas with poor-quality or missing data. They also argue that night light has a large estimated optimal weight in the estimation of growth rates in countries with low-quality statistical systems, following the A to D classification of the Penn World Tables (PWT). In particular, the authors show that the weight is, in these cases, larger than in the estimation of the level of GDP per capita. The importance of night light, as measured by its weight, in the estimation of growth is always higher in low-GDP density countries than those of high-GDP density, for any level of quality of the statistical system.7 More recently, Pinkovskiy and Sala-i-Martin (2016) have used nighttime lights to show that national accounts are an excellent proxy for actual income, while survey means have very little, if any, informative content to estimate true income. They show that growth rates of GDP per capita are very highly correlated with the growth of night light per capita, while the growth rate of survey means is very weakly correlated with the growth of night light per capita. Along similar lines, Jean et al. (2016) use satellite images and machine learning techniques to predict poverty at small scales. In their application, they use daytime satellite photos to capture details of the landscape (e.g., metal roof, water) that they correlate, using neural networks, with satellite night lights as a proxy for economic activity.8

All data were obtained from the National Geophysical Data Center, specifically the Earth Observation Group (EOG) reference to version 4 of the Defense Meteorological Satellite Program-Operational Linescan System (DMSP-OLS) Night-time Lights Time Series. The data are divided by year, from 1992 to 2013, and by six satellites from F10 to F18. From the three available image types, we use the stable light version, which is quantitized into 63 levels of light intensity. We have information on the nighttime light and the total nighttime light density by pixel from 1992 to 2010. Population data come from the Gridded Population of the World. For each cell, we constructed measures of luminosity per capita. Our basic dependent variable is the per capita growth between 1992 and 2010.

### B. Spatial Ethnic Diversity

We use data from GREG (georeferencing of ethnic groups) for the geospatial location of ethnic groups (Weidman, Rod, & Cederman, 2010). Relying on maps and data drawn from the classical Soviet Atlas Narodov Mira (AnM), the GREG data set employs geographic information systems (GIS) to represent group territories as polygons. The full GREG data set has global coverage and consists of 929 groups.9

For each country cell, we construct two types of measures of diversity. For the first measure, we use the percentage of territory that the homeland of the ethnic group covers in a particular cell. The second measure uses the percentage of the population living in the homeland of the ethnic group in a particular cell. We use the traditional fractionalization measure (Herfindhal index). Since data on population living in the specific homeland of a cell–country unit can only be computed from 1990 on, we use this second measure as a robustness check.

To capture ethnic diversity, we also use the Ethnolingustic Fractionalization Index (ELF). In particular, the index takes the form
$FRAC=1-∑i=1Nπi2=∑i=1Nπi(1-πi),$
(1)
where $π$ is the proportion of people who belong to ethnic group $i$. The broad popularity of the ELF index is based on its intuitive appeal: the index captures the probability that two randomly selected individuals from a given area will not belong to the same ethnolinguistic group.

### C. Control Variables

The regressions include a long list of control variables. Among the geographic controls, we consider the distance to the coastline, the distance to closest river or lake, and the ruggedness index. The climate controls include average temperature and average precipitation from 1961 to 1980. Additionally we include population density, the log of the area, the share of mining and fertile soil, the distance to the equator, ecological diversity, and the degree of pathogen stress. Ecological diversity captures biogeographic diversity based on eco-regions, which are defined as relatively large units of land or water containing a distinct assemblage of natural communities sharing a large majority of species, dynamics, and environmental conditions. We construct an index of ecological fractionalization as in Fenske (2014). Pathogen stress measures the extent of disease prevalence during precolonial times using a general measure of pathogen that includes the presence and intensity of seven pathogens.10

## III. Basic Results

The basic specification is
$lnyijt-lnyij0=αj+βlnyij0+γFRACij+∑γkzkij+εij,$
where $i$ and $j$ refer to a cell and a country respectively, and $yijt$ and $yij0$ are night light per capita in 2010 and 1992, respectively.11 FRAC is the level of ethnic fractionalization at each country cell. Using this arbitrary geographical area, we minimize concerns over the possible endogeneity of the political boundaries highlighted in the cross-sectional empirical literature. We also include controls for geographic and climate variables. To control for other factors that are country specific, we include country fixed effects. This is another advantage with respect to cross-country regressions. In fact, if we increase the size of the country cells, we ultimately reach the size of each country.12 We should notice that cells that include areas of two or more countries are divided in as many cells as countries. In this way, we make sure that the limiting case of the expansion of the size of the cells is a particular country. This is also the reason that we describe the elements of the grid as country cells: each cell, or part of a standard size cell, belongs only to one country. Table 2 shows the results. Column 1 presents the estimators for the specification that only includes country fixed effects and robust standard error. The estimation shows a positive relationship between ethnic fractionalization and growth controlling for the initial level of development.13 Column 2 presents the same basic regression using clustered standard errors at the country level. The standard errors almost double, but the estimators are still statistically significant and, in the case of ethnic diversity, positive.
Table 2.

Ethnic Diversity and Growth

Dependent Variable: Growth
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.370$***$ −0.370$***$ −0.370$***$ −0.367$***$ −0.369$***$ −0.362$***$ −0.365$***$ −0.364$***$
[0.007] [0.022] [0.021] [0.021] [0.021] [0.021] [0.021] [0.021]
(0.013))
Ethnic Fractionalization 0.617$***$ 0.617$**$ 0.696$***$ 0.709$***$ 0.700$***$ 0.568$**$ 0.608$**$ 0.620$**$
[0.100] [0.287] [0.265] [0.251] [0.247] [0.266] [0.258] [0.240]
(0.146)
Geographic Variables No No Yes Yes Yes Yes Yes Yes
Climate Variables No No Yes Yes Yes Yes Yes Yes
Population Density No No No Yes Yes Yes Yes Yes
Share Mining and Fertile Soil No No No No Yes Yes Yes Yes
Distance to River and Lake No No Yes Yes Yes Yes Yes Yes
Log. Area Obs No No No No No Yes Yes Yes
Country and Coastline Border No No No No No No Yes Yes
Dist. Equatorial Line No No No No No No Yes Yes
Ecological Diversity No No No No No No No Yes
Pathogen stress No No No No No No No Yes
Observations 21,514 21,514 21,514 21,514 21,514 21,514 21,514 21,514
$R$-squared 0.294 0.294 0.298 0.299 0.299 0.301 0.304 0.306
Dependent Variable: Growth
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.370$***$ −0.370$***$ −0.370$***$ −0.367$***$ −0.369$***$ −0.362$***$ −0.365$***$ −0.364$***$
[0.007] [0.022] [0.021] [0.021] [0.021] [0.021] [0.021] [0.021]
(0.013))
Ethnic Fractionalization 0.617$***$ 0.617$**$ 0.696$***$ 0.709$***$ 0.700$***$ 0.568$**$ 0.608$**$ 0.620$**$
[0.100] [0.287] [0.265] [0.251] [0.247] [0.266] [0.258] [0.240]
(0.146)
Geographic Variables No No Yes Yes Yes Yes Yes Yes
Climate Variables No No Yes Yes Yes Yes Yes Yes
Population Density No No No Yes Yes Yes Yes Yes
Share Mining and Fertile Soil No No No No Yes Yes Yes Yes
Distance to River and Lake No No Yes Yes Yes Yes Yes Yes
Log. Area Obs No No No No No Yes Yes Yes
Country and Coastline Border No No No No No No Yes Yes
Dist. Equatorial Line No No No No No No Yes Yes
Ecological Diversity No No No No No No No Yes
Pathogen stress No No No No No No No Yes
Observations 21,514 21,514 21,514 21,514 21,514 21,514 21,514 21,514
$R$-squared 0.294 0.294 0.298 0.299 0.299 0.301 0.304 0.306

In column 1 robust standard error in brackets. In columns 2 to 7, Robust standard error clustered at country level are reported in brackets. In column 8, Conley standard errors in parentheses (Spatial correlation kernel cutoff $=$ 200 km). Significant at $*$10%, $**$5%, and $***$1%. Country fixed effects are included. Geographic variables include distance to coastline, distance to closest river or lake, and the ruggedness index. Climate controls: Average temperature from 1961 to 1980 and average precipitation from 1961 to 1980.

### A. Controlling for Observable Variables

The basic results are robust to adding a large number of geographical and climate controls. Column 3 of table 2 adds the basic geographic (distance to the coastline, distance to the closest river or lake, and ruggedness) and climate controls (average temperature and precipitation). The effect of ethnic diversity is positive and a bit higher than in the previous specification. Adding population density (column 4) and the share of mining/fertile soil (column 5) does not affect the estimate. It is interesting to note that including population density does not affect the size of the coefficient on the effect of ethnic diversity on growth.14 When using cross-country data, the high correlation between ethnic diversity and population density has an important effect on the estimation of the parameter associated with ethnic diversity. However, at the cell level, we do not observe that effect. The borders of ethnic homelands do not concentrate areas with higher or lower proportion of cities, or population density, than other areas. Many big cities, with high population density, are not located at these borders.

The results are also robust to including many other observable variables, like the inclusion of the area of the cells (column 6); the results are also robust to the inclusion of a dummy for country or coastal borders and distance to the equator (column 7). Finally, column 8 includes ecological diversity and pathogen stress as controls.15 The results are basically unaffected. Inferential conclusions are not altered by using Conley's approach to correct the standard errors for spatial correlation as reflected in the braces of column 8. These results indicate a positive relationship between local ethnic diversity and local growth at a very high degree of geographical resolution. Using the results of the last column, we find an increase in the degree of ethnic heterogeneity of 2 standard deviations implies an annual increase of output per capita of 1.1 percentage points.16

### B. Urban Agglomeration and Migration

The positive relationship between ethnic diversity and local growth could be capturing an agglomeration effect related to the presence of large cities in the cell and therefore have a level of relationship different from the country cells that we use as the basic unit of observation. The empirical literature showing a positive impact of ethnic diversity on growth refers mostly to cities. Moreover, related to this “urban premium,” cities have higher productivity and higher wages than other areas. Is it only the urban premium that drives our results? The results of table 3 indicate that it is not only agglomeration effects that support the previous findings. In column 1 we include a dummy for the national capital, while in column 2, we also add dummies for provincial capitals. The effect of diversity is still present after controlling for capital cities. Column 3 adds dummies for urban agglomerations, considering as such urban areas with more than 500,000 inhabitants. In all the cases, the basic results of table 2 are maintained: ethnic diversity has a positive and significant effect on growth.

Table 3.

Ethnic Diversity and Growth: The Role of Agglomeration

Dependent Variable: Growth
Drop all-urban centerDrop 10% richestDrop 20% richestDrop 10% most denselyDrop 20% most densely
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.363$***$ −0.368$***$ −0.367$***$ −0.358$***$ −0.364$***$ −0.372$***$ −0.369$***$ −0.358$***$
[0.021] [0.021] [0.021] [0.021] [0.015] [0.016] [0.022] [0.022]
Ethnic Fractionalization 0.619$**$ 0.611$**$ 0.609$**$ 0.620$**$ 0.575$**$ 0.746$***$ 0.628$**$ 0.523$*$
[0.240] [0.238] [0.237] [0.282] [0.251] [0.163] [0.265] [0.275]
Nat. Capital (yes$=$1) −0.344$***$ −0.575$***$ −0.536$***$  −0.571$***$ −0.605$***$ −0.681$***$ −1.312$***$
[0.091] [0.090] [0.078]  [0.074] [0.076] [0.190] [0.274]
Prov. Capital (yes$=$1)  0.416$***$ 0.451$***$  0.465$***$ 0.472$***$ 0.845$***$ 1.246$***$
[0.128] [0.132]  [0.128] [0.124] [0.169] [0.213]
Urb. Agglom (yes$=$1)   $-$0.150  −0.337$**$ −0.620$***$
[0.144]  [0.131] [0.117]
Controls from table 2  Yes Yes Yes Yes Yes Yes Yes Yes
Observations 21,514 21,514 21,514 18,174 19,151 16,997 19,148 16,997
$R$-squared 0.306 0.307 0.307 0.298 0.295 0.314 0.303 0.294
Dependent Variable: Growth
Drop all-urban centerDrop 10% richestDrop 20% richestDrop 10% most denselyDrop 20% most densely
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.363$***$ −0.368$***$ −0.367$***$ −0.358$***$ −0.364$***$ −0.372$***$ −0.369$***$ −0.358$***$
[0.021] [0.021] [0.021] [0.021] [0.015] [0.016] [0.022] [0.022]
Ethnic Fractionalization 0.619$**$ 0.611$**$ 0.609$**$ 0.620$**$ 0.575$**$ 0.746$***$ 0.628$**$ 0.523$*$
[0.240] [0.238] [0.237] [0.282] [0.251] [0.163] [0.265] [0.275]
Nat. Capital (yes$=$1) −0.344$***$ −0.575$***$ −0.536$***$  −0.571$***$ −0.605$***$ −0.681$***$ −1.312$***$
[0.091] [0.090] [0.078]  [0.074] [0.076] [0.190] [0.274]
Prov. Capital (yes$=$1)  0.416$***$ 0.451$***$  0.465$***$ 0.472$***$ 0.845$***$ 1.246$***$
[0.128] [0.132]  [0.128] [0.124] [0.169] [0.213]
Urb. Agglom (yes$=$1)   $-$0.150  −0.337$**$ −0.620$***$
[0.144]  [0.131] [0.117]
Controls from table 2  Yes Yes Yes Yes Yes Yes Yes Yes
Observations 21,514 21,514 21,514 18,174 19,151 16,997 19,148 16,997
$R$-squared 0.306 0.307 0.307 0.298 0.295 0.314 0.303 0.294

Robust standard errors clustered at country level are reported in brackets. Significant at $*$10%, $**$5%, and $***$1%. Country fixed effects are included. Controls from table 2 include distance to coastline, Ruggedness Index, Average temperature 1961–1980, average precipitation 1961–1980, Population Density, Area, Share Mining, % Fertile Soil, Distance to River, Distance to Lake and Border (yes $=$ 1), Country and Coastline Border, Dist. Equatorial Line, Ecological Diversity, and Pathogen stress.

To continue investigating the issue of the influence of urban agglomeration, columns 4 to 8 of table 3 restrict the sample to different subsets of country cells without urban centers. Column 4 excludes the cells that contain an urban center.17 Columns 5 and 6 exclude the richest areas (upper 10% and 20%, respectively), mostly associated with the presence of urban areas. The most densely populated cells are also associated with the presence of urban metropolitan areas. For this reason, columns 7 and 8 present the estimation, dropping from the sample the cells with the highest population density (upper 10% and 20%, respectively). Overall, the results of table 3 indicate that the relationship between ethnic diversity and local growth is not driven only by the agglomeration effect associated with the presence of urban areas.

This exercise also addresses a potential measurement problem associated with the ethnic distribution of the population living in urban centers and, in particular, in capitals. Michalopoulos and Papaionnaou (2014) claim that under the assumption that in a given urban center, the respective indigenous group is relatively more populous than recent migrant groups, this should not be an important concern. We have shown in this section that our basic results are robust to the elimination of urban centers from the sample.

Finally, there is the issue of potential simultaneity of ethnic heterogeneity and growth: high-growth areas may attract more diverse populations than stagnant ones.18 In order to address the potential impact of postcolonial migration to prosperous countries and subsequent increases in ethnic diversity in those areas, we follow the strategy of Ashraf and Galor (2013).19 We perform the analysis restricting the sample to specific sets of countries depending on their attractiveness to migrants or the migratory distance from East Africa: countries that do not belong to the OECD and therefore are less attractive to migrants; non-neo-European countries (excluding the United States, Canada, Australia, and New Zealand), non-Latin American countries; non-sub-Saharan African countries, and the complementary of all the previous samples. Montalvo and Reynal-Querol (2017a) show that the effect of diversity on growth remains statistically significant in all these restricted samples, with the parameter estimate moving between 0.7 and 0.8.

### C. Some Additional Analyses of Robustness

We have run many other robustness analyses. In table 4, we check the robustness of the results to the use of alternative measures of ethnic diversity. In column 1, we find that the results are robust to the use of ethnic fractionalization calculated as the percentage of population living in the ethnic homeland. The empirical findings are basically unaffected if we use other sources of ethnic diversity such as Ethnologue (column 2). We also calculate fractionalization using ancestral ethnic homelands. To construct this variable, we use Murdock's data in the analysis of the African case (columns 3).20 This is reassuring because calculating ELF using ancestral ethnic homelands mitigates the concerns about the endogeneity of migration to local growth given that ancestral ethnic homelands do not reflect migrations.

Table 4.

Ethnic Diversity and Growth: Additional Robustness Checks

Dependent Variable: Growth
Without Outliers FRAC
(1)(2)(3)(4)
Log night light 1992 −0.363$***$ −0.378$***$ −0.385$***$ −0.355$***$
[0.021] [0.022] [0.036] [0.022]
Ethnic Fractionalization POP 0.646$***$
[0.224]
Ethnic Fractionalization (Ethnologue)  0.548$***$
[0.162]
Ethnic Fractionalization 1800 (Murdock)   0.893$***$
[0.259]
Ethnic Fractionalization    0.681$**$
[0.341]
Controls from table 2  Yes Yes Yes Yes
Observations 21,481 19,822 3,654 19,148
$R$-squared 0.306 0.314 0.270 0.307
Dependent Variable: Growth
Without Outliers FRAC
(1)(2)(3)(4)
Log night light 1992 −0.363$***$ −0.378$***$ −0.385$***$ −0.355$***$
[0.021] [0.022] [0.036] [0.022]
Ethnic Fractionalization POP 0.646$***$
[0.224]
Ethnic Fractionalization (Ethnologue)  0.548$***$
[0.162]
Ethnic Fractionalization 1800 (Murdock)   0.893$***$
[0.259]
Ethnic Fractionalization    0.681$**$
[0.341]
Controls from table 2  Yes Yes Yes Yes
Observations 21,481 19,822 3,654 19,148
$R$-squared 0.306 0.314 0.270 0.307

Robust standard errors clustered at country level are reported in brackets. Significant at $*$10%, $**$5%, and $***$1%. Country fixed effects are included. Controls from table 2 include distance to coastline, Ruggedness Index, Average temperature 1961–1980, average precipitation 1961–1980, Population Density, Area, Share Mining, % Fertile Soil, Distance to River, Distance to Lake and Border (yes$=$1), Country and Coastline Border,Dist. Equatorial Line, Ecological Diversity, Pathogen stress.

In column 4, we show that results are robust to dropping the outliers.21

We have also analyzed the sensitivity of the results to the grid generating coordinates. Therefore, we consider the possibility that the results shown in previous sections are the outcome of a specific initial point generating the grid. For this purpose, we produce 100 grids with random initial coordinates. More specifically, we take our initial coordinates (longitude $-$180; latitude $-$89) and add to both a random number generated by two uniform distributions.22 The results show that all the parameter estimates are statistically significant no matter the initial coordinates of the grid. In addition, the estimates move mostly in a close range between 0.7 and 0.9.23

## IV. The Case of Africa

Why is diversity good for units of small size? In the first section, we discussed several mechanisms that could explain the reduction of the influence of diversity on development as the size of the relevant unit of observation increases. The trade-off between the positive effect of diversity on the intensity of innovation and creation of knowledge and the negative effect of reduction of social capital depend on the size of the area. We argued that the specific mechanisms that explain the basic findings of this paper may depend on the characteristics of the group of countries being analyzed. In this section, we suggest a mechanism to explain the positive relationship between ethnic fractionalization and growth in Africa based on the possibility that ethnic diversity can increase trade when observed at high resolution. Assuming that members of different ethnic groups have less trust in one another than in members of the same group, trade across ethnic groups implies the need to monitor and be able to retaliate in cases of nonfulfillment of contract conditions. Trade across ethnic groups therefore requires proximity. As it is not possible to find data on trade across ethnic groups and thus provide some evidence of the likelihood of this mechanism, we rely on an indirect argument. As Michalopoulous (2012) has shown, ethnic groups tend to specialize. Using our data, we similarly find evidence of ethnic specialization. Moreover, high variability in the proportion of crops in an area is associated with high growth. Therefore, one mechanism that can explain the positive effect of ethnic diversity on development is the fact that the specialization in production of different ethnic groups provides larger opportunities for welfare improvement through trading with other ethnic groups. This implies that if this effect is larger than the transaction cost associated with lower levels of trust or communication issues, we should find local markets at the ethnic borders.

A theoretical explanation for these effects can be derived from a variation of the trade game with social norms presented in Rohner, Thoening, and Zilibotti (2013). The salience of social norms is heterogeneous across individuals and groups and determines the psychological benefit derived from agents by playing cooperatively. This benefit is assumed to be group specific but exogenous. However, it seems reasonable to assume that this psychological benefit depends on the proximity of ethnic groups; it is more likely that nearby groups share some social norms and have less prejudice than do those far away from one another. The contact theory that Allport (1954) proposed, a well-established idea in social psychology, suggests that contact between members of different groups can work to reduce prejudice and intergroup conflict. Desmet et al. (2019) find, in agreement with contact theory, that local learning reduces the antagonism felt toward other ethnic groups. The findings of Robinson (2017) are also consistent with contact theory: local diversity increases interethnic trust. It is important to get along with your neighbors but less important to follow the social norms of individuals with whom little interaction is foreseen. Therefore, the boundaries between ethnic groups should attract trade.

To investigate this mechanism, we have chosen the case of Africa, as this is the region of the world where most of the research on the issue of ethnic diversity is concentrated. Africa is a particularly interesting case for analysis when dealing with the relationship between ethnic diversity and growth as the latter is the most ethnically diverse region of the world. Before analyzing the location of markets in Africa, we first show that the general results on the relationship between ethnicity and growth also hold for Africa.

In table 5, we run the main regression for the African continent, and the relationship between diversity and growth remains positive at the same level of resolution as for the original results for the whole world. Columns 1 to 6 show that the basic result is unaffected by the inclusion of an increasing set of observables. Within Africa, areas around the equator seem to be the poorest and most diverse. For this reason, we check carefully the role of the equator in the basic relationship. In column 7, we include a dummy that has value 1 if the cell is on or within an area of +10/−10 degrees from the equatorial line and 0 otherwise. Column 8 reports the results of the regression using a dummy that takes value 1 if the cell is within +5/−5 degrees from the equator. The results are robust to the inclusion of these controls for the equator. Therefore, we also observe in Africa that more ethnically diverse areas are also those that grow faster.

Table 5.

Ethnic Diversity and Growth: The Case of Africa

Dependent Variable: Growth
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.382$***$ −0.378$***$ −0.379$***$ −0.380$***$ −0.380$***$ −0.382$***$ −0.381$***$ −0.382$***$
[0.034] [0.033] [0.032] [0.033] [0.035] [0.036] [0.035] [0.036]
Ethnic Fractionalization 0.912$**$ 0.801$**$ 0.773$**$ 0.743$**$ 0.751$**$ 0.748$**$ 0.940$**$ 0.808$**$
[0.374] [0.369] [0.372] [0.377] [0.380] [0.378] [0.462] [0.401]
$+/-$ 10 degree from Equatorial Line       $-$0.548
[0.397]
$+/-$ 10 degree from Equatorial Line $×$ FRACgregallv2       $-$0.534
[0.866]
$+/-$ 5 degree from Equatorial Line        0.101
[0.344]
$+/-$ 5 degree from Equatorial Line $×$ FRACgregallv2        $-$0.301
[0.964]
Geographic Variables Yes Yes Yes Yes Yes Yes Yes Yes
Climate Variables Yes Yes Yes Yes Yes Yes Yes Yes
Population Density No Yes Yes Yes Yes Yes Yes Yes
Share Mining and Fertile Soil No No Yes Yes Yes Yes Yes Yes
Distance to River and Lake No No No Yes Yes Yes Yes Yes
Log. Area Obs No No No No Yes Yes Yes Yes
Country and Coastline Border No No No No Yes Yes Yes Yes
Dist. Equatorial Line No No No No Yes Yes Yes Yes
Ecological Diversity No No No No No Yes Yes Yes
Pathogen stress No No No No No Yes Yes Yes
Observations 3,713 3,713 3,713 3,713 3,713 3,713 3,713 3,713
$R$-squared 0.259 0.264 0.265 0.267 0.268 0.269 0.271 0.269
Dependent Variable: Growth
(1)(2)(3)(4)(5)(6)(7)(8)
Log night light 1992 −0.382$***$ −0.378$***$ −0.379$***$ −0.380$***$ −0.380$***$ −0.382$***$ −0.381$***$ −0.382$***$
[0.034] [0.033] [0.032] [0.033] [0.035] [0.036] [0.035] [0.036]
Ethnic Fractionalization 0.912$**$ 0.801$**$ 0.773$**$ 0.743$**$ 0.751$**$ 0.748$**$ 0.940$**$ 0.808$**$
[0.374] [0.369] [0.372] [0.377] [0.380] [0.378] [0.462] [0.401]
$+/-$ 10 degree from Equatorial Line       $-$0.548
[0.397]
$+/-$ 10 degree from Equatorial Line $×$ FRACgregallv2       $-$0.534
[0.866]
$+/-$ 5 degree from Equatorial Line        0.101
[0.344]
$+/-$ 5 degree from Equatorial Line $×$ FRACgregallv2        $-$0.301
[0.964]
Geographic Variables Yes Yes Yes Yes Yes Yes Yes Yes
Climate Variables Yes Yes Yes Yes Yes Yes Yes Yes
Population Density No Yes Yes Yes Yes Yes Yes Yes
Share Mining and Fertile Soil No No Yes Yes Yes Yes Yes Yes
Distance to River and Lake No No No Yes Yes Yes Yes Yes
Log. Area Obs No No No No Yes Yes Yes Yes
Country and Coastline Border No No No No Yes Yes Yes Yes
Dist. Equatorial Line No No No No Yes Yes Yes Yes
Ecological Diversity No No No No No Yes Yes Yes
Pathogen stress No No No No No Yes Yes Yes
Observations 3,713 3,713 3,713 3,713 3,713 3,713 3,713 3,713
$R$-squared 0.259 0.264 0.265 0.267 0.268 0.269 0.271 0.269

Robust standard errors clustered at country level are reported in brackets. Significant at $*$10%, $**$5%, and $***$1%. Country fixed effects are included. Controls from table 2 include distance to coastline, Ruggedness Index, Average temperature 1961–1980, average precipitation 1961–1980, Population Density, Area, Share Mining, % Fertile Soil, Distance to River, Distance to Lake and Border (yes$=$1), Country and Coastline Border, Dist. Equatorial Line, Ecological Diversity, Pathogen stress.

### A. Interethnic Trade and Markets in Africa

There are a number of indications that interethnic trade is at the origin of many African local markets. Most of this trade took place at the boundaries of ethnic homelands. The evidence draws mostly on work by geographers, anthropologists, and historians who have studied the origin of traditional markets in Africa before the arrival of Europeans. For example, Hodder (1965) provides a voluminous body of evidence to support the view that external trading contacts were critical for the genesis of markets in Africa. Hodder (1965) remarks that “the analysis of Yorubaland, for example, indicated that traditional markets were often located at junction zones, areas in which products of each area could be easily exchanged. Also markets were found at the junction of different people: Ketu market, for instance, was regarded as an important link between Yoruba and Dahomey peoples; Iperu market was a contact point between Egba and Ijebu groups of the Yoruba; mamu market was a traditionally frontier market between Ijebu and Ibadan Yoruba” (99). Hodder (1965) adds that

There is plenty of additional evidence of the interethnic origin of markets in Africa. Meillassoux (1965) analyzes the case of the Guro land in Ivory Coast, where “markets among the Guro of central Ivory Coast tend to be localized at the contact area between complementary zones,” supporting the conclusion that “markets are primarily induced by external exchange of complementary products with an alien population. When such a situation occurs, the markets tend to be localized at the contact area between complementary zones. Hence, they can help to indicate the limits of substantive economic areas” (297–298). Vansina (1965) draws an almost identical conclusion with regard to the traditional markets of the Kuba peoples of present-day Zaire.

Roberts (1970) highlights that precolonial commerce in the interior of Tanzania was an activity involving different peoples, including Nyamwezi, Sumbwa, Gogo, Taturu, Sukuma, Vinza, and Sagara, who exchanged complementary products which circulated within and between regional trade networks.

Yet another compelling example is that of the Abyssinian market town covered by Messing (1965): “There is relatively little exchange of any kind outside the extended-family and rural hamlets except for that taking place on daily and weekly markets. On certain seasonal occasions, over 1,000 persons may gather in and about the weekly market at Gondar. Money is used as both medium of exchange and standard of value. The market is closely related to the division of labor which is caste-like in its ethnic specialization of occupations, such as smiting, pottery-making and tanning” (387).

### B. Empirical Results

Assuming that trade at the local level usually takes place in markets, we look for data on the latter. It is difficult to find the location of markets in Africa since there are potentially thousands of very small markets where farmers may sell their products. Therefore, we need some criteria to identify relevant markets that are exogenous to our methodology. Porteous (2019) identifies the location of 223 regionally important hub markets in Africa.24 Around 60% of markets are located in cities of more than 100,000 inhabitants and 40% in smaller villages.

We overlap Porteous's map with our own map of the spatial distribution of ethnic groups in Africa in figure 1. It is easy to see that many of the markets are very close to ethnic borders. In fact, the average distance to the closest ethnic border among all the markets is just 27 km, which seems to indicate that trading markets are located close to ethnic borders.25 In order to show how far the actual distribution of markets is with respect to a random geographic distribution, we run a simulation with 500 random samples of 223 locations in Africa (equal to the number of markets taken from Porteous, 2019).26 We consider continental sub-Saharan Africa, the area similarly covered by Porteous (2019), and use the Haversine formula to estimate the distance of each simulated market to the closest ethnic border. Finally, we take the average distance to the closest ethnic border for each of the 500 simulations. The results show that the average distance of 27 km is at the 1% of the distribution.27 This indicates that markets are much closer to ethnic borders than randomly generated locations.
Figure 1.
Market Location and Ethnic Borders
Figure 1.
Market Location and Ethnic Borders
Figure 2 provides3 additional evidence for the concentration of trading along the ethnic borders. It shows the average ethnic diversity index for the actual location of the markets and the 500 simulations of markets' distribution in sub-Saharan Africa. For the placebo analysis, we randomly generate the location of 223 “virtual” markets, or the actual number of markets identified by Porteous (2019). In each simulation we use a buffer of 50 km around each point to calculate their index of ethnic diversity. Figure 2 shows that the ethnic diversity of the actual markets is in the tail of the distribution of heteregeneity indices of the virtual markets.
Figure 2.
Market Location and Ethnic Borders
Figure 2.
Market Location and Ethnic Borders
Figure 3 runs a similar exercise to figure 2 but relating market location and growth. It shows that the growth rate around the markets is much higher than the growth rate around the virtual markets.
Figure 3.
Market Location and Ethnic Borders
Figure 3.
Market Location and Ethnic Borders

Previous evidence shows that areas that have more ethnic diversity also have more markets. An explanation is based on the specialization of ethnic groups. The geographic proximity of ethnic groups may also increase trade if they are highly specialized in the production of specific agricultural products or services. While the level of trust among different ethnic groups may, in general, be smaller than intragroup trust, the fact that they are geographically close to one another can facilitate monitoring and therefore counterbalance the potential lack of trust. Jha (2013) shows that medieval Hindus and Muslims could provide complementary services and a mechanism to share gain from trade, which increased tolerance between these two groups. The development of these practices into formal institutions generated inertia in the degree of ethnic tolerance.28 The location of local markets in Africa seems to support this interpretation.

## V. Conclusion

The relationship between ethnic heterogeneity and development is complicated. Empirical research working with cross-country data finds a negative, or null, relationship. However, research at the city level usually finds a positive relationship between diversity and wages and/or productivity. In this paper, we find that small areas tend to generate a positive relationship. We argue that an explanation of the positive relationship between diversity and growth in Africa consistent with the data is the increase in trade at the boundaries between ethnic groups due to ethnic specialization.

## Notes

1

Collier (2001), Easterly (2001), Alesina, Spolaore, and Wacziarg (2000), Montalvo and Reynal-Querol (2005), or Gören (2014).

2

Ottaviano and Peri (2005, 2006), Alesina, Harnoss, and Rapoport (2016), and Sparber (2010) study the case of the United States. Nathan (2011) analyzes the case of the United Kingdom, while Suedekum, Wolf, and Blien (2014) discuss the results of Germany, and Bakens, Mulder, and Nijkamp (2013) the example of the Netherlands.

3

Michalopoulos and Papaioannou (2013, 2014), Alesina, Michalopoulos, and Papaioannou (2016), and Michalopoulos (2012). Several papers have recently used the boundaries between ethnic groups to generate quasi-experiments like Michalopoulos and Papaioannou (2014, 2016).

4

Aghion and Howitt (2009).

5

See, for example, Chen and Northaus (2011) or Pinkowskiy and Sala-i-Martin (2016).

6

Montalvo and Reynal-Querol (2017a).

7

The cross-validation analysis in Michalopoulos and Papaioannou (2013) shows that light density at night is highly correlated with a wealth index across households in four large African countries.

8

Night light intensity has also being used to measure inequality at low levels of geographical aggregation. See, for example, Alesina, Michalopoulos et al. (2016) and Montalvo and Reynal-Querol (2017b).

9

Desmet, Ortuo-Ortin, and Wacziarg (2012) use a linguistic tree to calculate measures of diversity at different levels of aggregation. They argue that while deep cleavages are relevant for conflict, more superficial cleavages are relevant for economic growth. We tend to agree with this conclusion, although given the computational challenges of the exercises in this paper, an analysis of that incorporates the degree of ethnic cleavages must necessarily be left for a future project.

10

For a full description of the variables included in the empirical analysis, see the online appendix.

11

The measure of night light per capita is generally adopted as the good proxy for GDP per capita at high resolution. Pinkovskiy and Sala-i-Martin (2016) similarly use night light per capita in all of their baseline regressions.

12

Montalvo and Reynal-Querol (2017a) perform a systematic analysis of the effect of the size of geographical units on the relationship between ethnic diversity and growth. They find a positive relationship for small geographical areas and no effect for large areas and countries.

13

In the regressions of table 2, the speed of convergence implied by the coefficient of the log of the initial level of development ranges from 1.6% to 2.4%, values similar to those typically found for the speed of convergence across regions or countries. This result indicates that night light density generates similar results to the those found with other indicators of economic development.

14

We are indebted to Roman Wacziarg for suggesting this exercise.

15

For the description of these variables, see the online data description.

16

To relate this to GDP per capita, we note that Pinkovskiy and Sala-i-Martin (2016) cannot reject the null hypothesis that the weight of log GDP per capita is 1 in the optimal light night-based proxy of true income. The same result holds for the subsample of Africa.

17

See the online appendix.

18

Ashraf and Galor (2013) argue that the direction of the potential endogeneity bias is ambiguous a priori since wealthier societies can have advanced military technology to minimize invasions of foreigners.

19

Previous versions of the paper addressed this issue using two instrumental variables. The results confirmed the findings discussed above.

20

The results are also unaffected if we measure diversity as the number of ethnic groups.

21

The results are also robust to the use of the level of night light per capita instead of the growth rate. Including regressions by continents do not alter the results.

22

See Montalvo and Reynal-Querol (2017a) for a detailed description of the process of generating these 100 grids by using random initial coordinates.

23

We have also investigated if the results are heterogeneous depending on country characteristics (political institutions, level of development of the country, and country-wide ethnic diversity) showing that the effect of ethnic heterogeneity at this level of local analysis does not depend robustly on the quality of country's political institutions. See Montalvo and Reynal-Querol (2017a).

24

Porteous (2019) considers 230 markets, but there are 223 placed in different locations. We should notice that, obviously, there are many small markets not included in the list.

25

Montalvo and Reynal-Querol (2017a) zoom in one of figure 1's typical markets (the Gambela market), to show how ethnic groups and ethnic borders are distributed around it.

26

We thank Stelios Michalopoulos for this suggestion.

27

See Montalvo and Reynal-Querol (2017a).

28

Jha (2013) also finds that medieval ports, despite being more ethnically diverse, were less prone to conflicts between ethnic groups.

## REFERENCES

Aghion
,
Philippe
, and
Peter W.
Howitt
,
The Economics of Growth
(
Cambridge, MA
:
MIT Press
,
2009
).
Aker
,
Jenny C.
,
Michael
Klein
,
Stephen A.
O'Connell
, and
Muzhe
Yang
, “
Journal of Development Economics
107
(
2014
),
1
16
.
Alesina
,
Alberto
,
Arnaud
Devleeschauwer
,
William
Easterly
,
Sergio
Kurlat
, and
Romain
Wacziarg
, “
Fractionalization,
Journal of Economic Growth
8
(
2003
),
155
194
.
Alesina
,
Alberto
,
Johann
Harnoss
, and
Hillel
Rapoport
, “
Birthplace Diversity and Economic Prosperity,
Journal of Economic Growth
21
(
2016
),
101
138
.
Alesina
,
Alberto
, and
Eliana La
Ferrara
, “
Ethnic Diversity and Economic Performance,
Journal of Economic Literature
43
(
2005
),
762
800
.
Alesina
,
Alberto
,
Stelios
Michalopoulos
, and
Elias
Papaioannou
, “
Ethnic Inequality,
Journal of Political Economy
124
(
2016
),
428
488
.
Alesina
,
Alberto
,
Enrico
Spolaore
, and
Romain
Wacziarg
, “
Economic Integration and Political Disintegration,
American Economic Review
90
(
2000
),
1276
96
.
Allport
,
Gordon Willard
,
Kenneth
Clark
, and
Thomas
Pettigrew
,
The Nature of Prejudice
, (
:
,
1954
).
Ashraf
,
Quamrul
, and
Oded
Galor
, “
The Out of Africa Hypothesis, Human Genetic Diversity and Comparative Development,
American Economic Review
103
(
2013
),
1
46
.
Bakens
,
Jessie
,
Peter
Mulder
, and
Peter
Nijkamp
, “
Economic Impacts of Cultural Diversity in the Netherlands: Productivity, Utility, and Sorting,
Journal of Regional Science
53
(
2013
),
8
36
.
Chen
,
Xi
, and
William D.
Nordhaus
, “
Using Luminosity Data as a Proxy for Economic Statistics,
Proceedings of the National Academy of Sciences
108
(
2011
),
8589
8594
.
Collier
,
Paul
, “
Ethnic Diversity: An Economic Analysis of Its Implications,
Economic Policy
32
(
2001
),
129
166
.
Desmet
,
Klaus
,
Joseph Flavian
Gomes
, and
Ignacio
Ortuño-Ortín
, “
The Geography of Linguistic Diversity and the Provision of Public Goods
,”
Journal of Development Economics
143
:
1
(
2019
), 102384.
Desmet
,
Klaus
,
Ignacio
Ortuño-Ortín
, and
Romain
Wacziarg
, “
The Political Economy of Linguistic Cleavages,
Journal of Development Economics
97
(
2012
),
322
338
.
Easterly
,
William
, “
Can Institutions Resolve Ethnic Conflict?
Economic Development and Cultural Change
49
(
2001
),
687
706
.
Easterly
,
William
, and
Ross
Levine
, “
Africa's Growth Tragedy: Policies and Ethnic Divisions
,”
Quarterly Journal of Economics
112
(
4
) (
1997
),
1203
1250
.
Fenske
,
James
, “
Ecology, Trade and States in Pre-Colonial Africa,
Journal of the European Economic Association
12
(
2014
),
612
640
.
Glaeser
,
Edward L.
, “
The Political Economy of Hatred,
Quarterly Journal of Economics
120
(
2005
),
45
86
.
Gören
,
Erkan
, “
How Ethnic Diversity Affects Economic Growth,
World Development
59
(
2014
),
275
297
.
Greif
,
Avner
, “
Contract Enforceability and Economic Institutions in Early Trade: The Maghribi Traders' Coalition,
American Economic Review
83
(
1993
),
525
548
.
Greif
,
Avner
The Fundamental Problem of Exchange: A Research Agenda in Historical Institutional Analysis,
European Review of Economic History
4
(
2000
),
251
284
.
Greif
,
Avner
The Birth of Impersonal Exchange: The Community Responsibility System and Impartial Justice,
Journal of Economic Perspectives
20
(
2006
),
221
236
.
Henderson
,
J. Vernon
,
Storeygard
, and
David N.
Weil
, “
Measuring Economic Growth from Outer Space,
American Economic Review
102
(
2012
),
994
1028
.
Hodder
,
Bramwell William
, “
Some Comments on the Origin of Traditional Markets in Africa South of the Sahara,
Transactions of the Institute of British Geographers
36
(
1965
),
97
105
.
Jha
,
Saumitra
, “
Trade, Institutions and Ethnic Tolerance: Evidence from South Asia,
American Political Science Review
107
(
2013
),
806
832
.
Jean
,
Neal
,
Burke
Marshall
,
Xie
Michael
,
Davis W.
Matthew
,
David B.
Lobell
, and
Ermon
Stefano
, “
Combining Satellite Imagery and Machine Learning to Predict Poverty
,”
Science
353
:
6301
(
2016
),
790
794
.
Meillassoux
,
Claude
, “
Social and Economic Factors Affecting Markets in Guro Land
,” in
P.
Bohannan
, and
G.
Dalton
, eds.,
Markets in Africa
(
Evanston, IL
:
Northwestern University Press
,
1965
).
Messing
,
Simon D.
, “
The Abyssinian Market Town
,” in
P.
Bohannan
and
G.
Dalton
, eds.
Markets in Africa
(
Evanston, IL
:
Northwestern University Press
,
1965
).
Michalopoulos
,
Stelios
, “
The Origins of Ethnolinguistic Diversity,
American Economic Review
102
(
2012
),
1508
1539
.
Michalopoulos
,
Stelios
, and
Elias
Papaioannou
, “
Pre-Colonial Ethnic Institutions and Contemporary African Development,
Econometrica
81
(
2013
),
113
152
.
Michalopoulos
,
Stelios
, and
Elias
Papaioannou
National Institutions and Subnational Development in Africa
,”
Quarterly Journal of Economics
129
:
1
(
2014
),
151
213
.
Michalopoulos
,
Stelios
, and
Elias
Papaioannou
The Long Run Effects of the Scramble for Africa,
American Economic Review
106
(
2016
),
1802
1848
.
Montalvo
,
José G.
, and
Marta
Reynal-Querol
, “
Ethnic Diversity and Economic Development,
Journal of Development Economics
76
(
2005
),
293
323
.
Montalvo
,
José G.
, and
Marta
Reynal-Querol
Ethnic Diversity and Growth: Revisiting the Evidence
,”
Barcelona GSE working paper
992
(
2017a
).
Montalvo
,
José G.
, and
Marta
Reynal-Querol
Earthquakes and Inequality
,” mimeo (
2017b
).
Nathan
,
Max
, “
The Long Term Impact of Migration in British Cities: Diversity, Wages, Employment and Prices
,”
SERC discussion paper
67
(
2011
).
Ottaviano
,
Gianmarco I. P
, and
Giovanni
Peri
, “
Cities and Cultures,
Journal of Urban Economics
58
(
2005
),
304
307
.
Ottaviano
,
Gianmarco I. P
, and
Giovanni
Peri
The Economic Value of Cultural Diversity: Evidence from US Cities,
Journal of Economic Geography
6
(
2006
),
9
44
.
Pinkovskiy
,
Maxim
, and
Xavier
Sala-i-Martin
, “
Lights, Camera . . . Income! Illuminating the National Account-Household Surveys Debate,
Quarterly Journal of Economics
131
(
2016
),
579
631
.
Porteous
,
Obie
, “
High Trade Costs and Their Consequences: An Estimated Dynamic Model of African Agricultural Storage and Trade
,”
American Economic Journal: Applied Economics
11
(
2019
),
327
366
Roberts
,
Andrew
, “
,” in
R.
Gray
. and
D.
Birmingham
, eds.,
(
London: Oxford University Press
,
1970
).
Robinson
,
Amanda Lea
, “
Ethnic Diversity, Segregation and Ethnocentric Trust in Africa
,”
British Journal of Political Science
33
(
2017
),
607
620
Rohner
,
Dominic
,
Mathias
Thoenig
, and
Fabrizio
Zilibotti
, “
War Signals: A Theory of Trade, Trust and Conflict,
Review of Economic Studies
80
(
2013
),
1114
1147
.
Sparber
,
, “
Racial Diversity and Macroeconomic Productivity across US States and Cities
,”
Regional Studies
44
:
1
(
2010
),
71
85
.
Suedekum
,
Jens
,
Katja
Wolf
, and
Uwe
Blien
, “
Cultural Diversity and Local Labor Markets,
Regional Studies
40
(
2014
),
173
191
.
Vansina
,
Jan
, “
Trades and Markets among the Kuba
,” in
P.
Bohannan
and
G.
Dalton
, eds.
Markets in Africa
(
Evanston, IL
:
Northwestern University Press
,
1965
).
Weidmann
,
Nils B.
,
Jan Ketil
Rd
, and
Lars–Erik
Cederman
, “
Representing Ethnic Groups in Space: A New Dataset,
Journal of Peace Research
47
(
2010
),
491
499
.

## Author notes

We thank Daron Acemoglu, Alberto Alesina, Tim Besley, Klaus Desmet, Ruben Durante, Oded Galor, Saumitra Jha, Massimo Morelli, Gerard Padro, Elias Pa- paioannou, Luigi Pascali, Felipe Valencia, Hans-Joachim Voth, and, especially, Antonio Ciccone, Stelios Michalopoulos, and Roman Wacziarg for their helpful comments. We also thank the participants in the Bocconi Workshop on Conflict, the Meeting of the European Public Choice Society, the EPCS of Groningen, the Berlin Workshop on Advances on the Political Economy of Conflict and Redistribution III, the Invited Session on Political Economy of Development at the EEA Meetings of 2015 in Mannheim, the Yale University Workshop on the Political Economy of Social Conflict, and the NBER Summer Institute. We have also benefited from the comments of participants in seminars at Stanford GSB, Oxford, LSE, Warwick, Bonn, Stockholm, UPF and CREI. Financial support from the European Research Council, the Spanish National Science Foundation, the Barcelona GSE Research Network, and the government of Catalonia is gratefully acknowledged.

A supplemental appendix is available online at https://doi.org/10.1162/rest_a_00901.