ethnic inequality - accueilpiketty.pse.ens.fr/files/alesinaetal2016.pdf · fort. consistent with...

All

Ethnic Inequality

Alberto Alesina

Harvard University, Innocenzo Gasparini Institute for Economic Research, Centrefor Economic Policy Research, and National Bureau of Economic Research

Stelios Michalopoulos

Harvard University, Brown University, Centre for Economic Policy Research,and National Bureau of Economic Research

Elias Papaioannou

London Business School, Centre for Economic Policy Research,and National Bureau of Economic Research

WementreseaDubeEnricticipaBritisNew Y

Electro[ Journa© 2016

use su

This study explores the consequences and origins of between-ethnicityinequality for a large sample of countries. First, combining satellite im-ages of nighttime luminosity with the homelands of ethnolinguisticgroups, we construct measures of ethnic inequality. Second, we uncovera strong inverse association between ethnic inequality and contempo-rary development above and beyond its relationship with cross-regionand cross–administrative unit inequality. Third, we establish that differ-ences in geographic endowments across ethnic homelands explain a siz-able fraction of the variation in economic disparities across groups.Fourth, we show that inequality in geographic endowments across eth-nic homelands is a negative correlate of development.

428

thank Harald Uhlig ðthe editorÞ and two anonymous referees for excellent com-s. We would like to thank Nathan Fleming and Sebastian Hohmann for superlativerch assistance. For valuable suggestions we also thank Christian Dippel, Oeindrila, Sebastian Hohmann, Michele Lenza, Nathan Nunn, Debraj Ray, Andrei Shleifer,o Spolaore, Pierre Yared, Romain Wacziarg, David Weil, Ivo Welch, and seminar par-nts at Dartmouth, the Athens University of Economics and Business, University ofh Columbia, Brown, Centre de Recerca en Economia Internacional, Oxford, Bocconi,ork University, Michigan, Paris School of Economics, Institute for International Eco-

nically published March 4, 2016l of Political Economy, 2016, vol. 124, no. 2]by The University of Chicago. All rights reserved. 0022-3808/2016/12402-0004$10.00

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AMbject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago.edu/t-and-c).

ethnic inequality 429

I. Introduction

Ethnic diversity has costs and benefits. On the one hand, diversity in skills,education, and endowments can enhance productivity by promoting inno-vation. On the other hand, diversity is often associated with poor and eth-nically targeted policies, inefficient provision of public goods, and ethnic-based hatred and conflict. In fact, a large literature finds a negative impactof ethnolinguistic fragmentation on various aspects of economic perfor-mance, with the possible exception of wealthy economies ðsee Alesinaand La Ferrara [2005] for a reviewÞ. Income inequality may also haveboth positive and negative effects on development. On the negative side,a higher degree of income inequality may lead to conflict and crime, pre-vent the poor from acquiring education, or lead to expropriation and loftytaxation discouraging investment. On the positive side, income inequalitymay spur innovation and entrepreneurship by motivating individuals andby providing the necessary pools of capital for capital-intensive modes ofproduction. Further complicating the relationship between the two, apositive correlation between inequality and development may reflect Si-mon Kuznetz’s ð1955Þ conjecture that industrialization translates intohigher levels of inequality at the early stages of development, while at laterstages, the association becomes negative. Given the theoretical ambigui-ties ðand data issuesÞ, perhaps it comes as no surprise that it has been veryhard to detect empirically a robust association between inequality and de-velopment ðsee Benabou [2005] and Galor [2011] for surveysÞ.This paper puts forward and tests an alternative conjecture that focuses

on the intersection of ethnic diversity and inequality. Our thesis is thatwhat matter most for comparative development are economic differencesbetween ethnic groups coexisting in the same country rather than the de-gree of fractionalization per se or income inequality conventionally mea-sured ði.e., independent of ethnicityÞ.1The first contribution of this study is to provide measures of within-

country differences in well-being across ethnic groups, defined as “ethnicinequality.”Toovercome the sparsity of incomedata along ethnic lines and

1 Stewart ð2002Þ and Chua ð2003Þ are early precedents. Providing case study evidence,they argue that horizontal inequalities across ethnic/religious/racial groups are importantfeatures of underdeveloped and conflict-prone societies. Yet to the best of our knowledge,there have been very few systematic empirical works, if any, that directly examine this con-jecture. We discuss parallel studies that touch on this issue below.

nomic Studies, Columbia, Centre of Planning and Economic Research, Warwick, LondonSchool of Economics, Nottingham, the National Bureau of Economic Research SummerInstitute Meetings in Political Economy, the Centre for Economic Policy Research Devel-opment Economics Workshop, the conference “How Long Is the Shadow of History?The Long-Term Persistence of Economic Outcomes” at University of California, Los Ange-les, and the Nemmers Conference in the Political Economy of Growth and Development atNorthwestern University. Papaioannou greatly acknowledges financial support from theLondon Business School Research and Materials Development Fund grant. All errorsare our own responsibility. Data are provided as supplementary material online.

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AMAll use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago.edu/t-and-c).

430 journal of political economy

All

in order to construct country-level indicators of ethnic inequality for thelargest possible set of states, we combine ethnographic and linguisticmaps on the location of groups with satellite images of light density atnight, which are available at a fine grid. Recent studies have shown thatluminosity is a strong proxy of development ðe.g.,Henderson, Storeygard,and Weil 2012Þ. The cross–ethnic group inequality index is weakly corre-lated with the commonly employed—and notoriously poorly measured—income inequality measures at the country level and is modestly corre-lated with ethnic fractionalization. To isolate the cross-ethnic componentof inequality from the overall regional inequality, we also construct prox-ies of spatial inequality and measures capturing regional differences inwell-being across first- and second-level administrative units.Second, we document a strong negative association between ethnic in-

equality and real GDP per capita across countries. This correlation holdseven when we control for the overall degree of spatial inequality and in-equality across administrative regions. The latter is also inversely relatedto a country’s economic performance, a novel finding in itself. We alsouncover that the negative correlation between ethnolinguistic fragmen-tation and development weakens considerably ðand becomes statisticallyindistinguishable from zeroÞ when we account for ethnic inequality; thissuggests that it is the unequal concentration of wealth across ethnic linesthat correlates with development rather than diversity per se.Third, in an effort to shed light on the roots of ethnic inequality, we

explore its geographic underpinnings. In particular, motivated by recentwork showing that linguistic groups tend to reside in distinct land endow-ments ðsee Michalopoulos 2012Þ, we construct Gini coefficients reflect-ing differences in various geographic attributes across ethnic homelandsand show that the latter is a strong predictor of ethnic inequality. On thecontrary, there is no link between contemporary ethnic inequality andoften-used historical variables capturing the type of colonization and le-gal origin, among others.Fourth, we show that contemporary development at the country level

is also inversely related to inequality in geographic endowments acrossethnic homelands. Yet, once we condition on between-group income in-equality, differences in geographic endowments are no longer a significantcorrelate of underdevelopment. These results suggest that geographic dif-ferences across ethnic homelands influence comparative developmentmostly via shaping economic inequality across groups.Mechanisms and related works.—Income disparities along ethnic lines

are likely to lead to political inequality based on ethnic affiliation, in-crease between-group animosity, and lead to discriminatory policies ofone ðor moreÞ group against the others. In line with this idea, in recentwork Huber and Suryanarayan ð2013Þ document that party ethnificationin India is more pronounced in states with a high degree of inequality

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AM use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago.edu/t-and-c).


across subcastes.2 Furthermore, differences in preferences along bothethnic and income lines may lead to inadequate public goods provision,as groups’ ideal allocations of resources will be quite distant. Baldwinand Huber ð2010Þ provide empirical evidence linking between-group in-equality to the underprovision of public goods for 46 democracies. InAlesina, Michalopoulos, and Papaioannou ð2014Þ, we show that thereis a strong inverse link between ethnic inequality and public goods within18 sub-Saharan African countries ðand that this effect partly stems frompolitical inequality and ethnic-based discriminationÞ.3 Ethnic inequalitymay also impede institutional development and the consolidation of de-mocracy ðRobinson 2001Þ. In line with this conjecture, Kyriacou ð2013Þexploits survey data from 29 developing countries and shows that socio-economic ethnic group inequalities reduce government quality.Chua ð2003Þ presents case study evidence arguing that the presence of

an economically dominant ethnicminoritymay lower support for democ-racy and free-market institutions, as the majority of the population usu-ally feels that the benefits of capitalism go to just a handful of ethnicgroups. She discusses, among others, the influence of Chineseminoritiesin the Philippines, Indonesia, Malaysia, and other eastern Asian coun-tries; the dominant role of ðsmallÞ Lebanese communities in western Af-rica; and the similarly strong influence of Indian societies in eastern Af-rica. Other examples include the IðgÞbo in Nigeria and the Kikuyu inKenya. Finally, to the extent that ethnic inequality implies that well-beingdepends on one’s ethnic identity, then it is more likely to generate envyand perceptions that the system is “unfair” and reduce interpersonaltrust, more so than the conventionally measured economic inequality,since the latter can be more easily thought of as the result of ability or ef-fort. Consistent with the view that ethnic inequality is detrimental to theformation of social ties across groups, Tesei ð2014Þ finds that greater ra-cial inequality across US metropolitan areas is associated with low levelsof social capital.

2 Ethnic inequality may impede development by spurring civil conflict ðHorowitz 1985Þ.However, Esteban and Ray ð2011Þ show that the effect of ethnic inequality on conflict isambiguous, as it also depends on within-group inequality. Recent works in political scienceprovide opposing results. Cederman, Weidman, and Gleditch ð2011Þ combine proxies oflocal economic activity from the G-Econ database with ethnolinguistic maps to constructan index of ethnic inequality for a subset of “politically relevant ethnic groups” ðas definedby the Ethnic Power Relations DatasetÞ and then show that in highly unequal countries,both rich and poor groups fight more often than those groups whose wealth is closer tothe country average. However, in parallel work, Huber and Mayoral ð2013Þ find no link be-tween inequality across ethnic lines and conflict.

3 Similarly, Deshpande ð2000Þ and Anderson ð2011Þ focus on income inequality acrosscastes in India and associate between-caste inequality to public goods provision. See alsoLoury ð2002Þ for an overview of works studying the evolution of racial inequality in theUnited States and its implications.



All

Organization.—The paper is organized as follows. In Section II, we de-scribe the construction of the ethnic ðand regionalÞ inequality measuresand present summary statistics and the basic correlations. In Section III,we report the results of our analysis associating income per capita withethnic inequality across 173 countries. Besides reporting various sensitiv-ity checks, we also examine the link between development and inequalityacross administrative regions. In Section IV, we explore the geographicorigins of contemporary differences in economic performance acrossgroups. In Section V, we report estimates associating contemporary devel-opment with inequality in geographic endowments across ethnic home-lands. In Section VI, we summarize our findings and discuss avenues forfuture research.

II. Data

To construct proxies of ethnic inequality for the largest set of countries,we combine information from ethnographic-linguistic maps on the loca-tion of groups with satellite images of light density at night that are avail-able at a fine grid. In this section, we discuss the construction of the cross-country measures reflecting inequality in development ðas captured byluminosity per capitaÞ across ethnic homelands within 173 countries.We also describe in detail the construction of the other measures of spa-tial inequality and discuss the main patterns.

A. Ethnic Inequality Measures

1. Location of Ethnic Groups

We identify the location of ethnic groups employing two data sets/maps.4

First, we use the Geo-Referencing of Ethnic Groups ðGREGÞ, which is thedigitized version of the Soviet Atlas Narodov Mira ðWeidmann, Rod, andCederman 2010Þ. GREG portrays the homelands of 928 ethnic groupsaround the world. The information pertains to the early 1960s, so formany countries, in Africa in particular and to a lesser extent in SoutheastAsia, it corresponds to the time of independence.5 The data set uses thepolitical boundaries of 1964 to allocate groups to different countries. Wethus project the ethnic homelands to the political boundaries of the 2000Digital Chart of the World; this results in 2,129 ethnic homelands within

4 Note that across all units of analysis in the construction of the respective indexes, weexclude polygons of less than 1 square kilometer ðkm2Þ to minimize measurement error inthe drawing of the underlying maps.

5 The original Atlas Narodov Mira consists of 57 maps. The original sources are ð1Þ eth-nographic and geographic maps assembled by the Institute of Ethnography at the USSRAcademy of Sciences, ð2Þ population census data, and ð3Þ ethnographic publications ofgovernment agencies at the time.



contemporary countries. Most areas ð1,637Þ are coded as pertaining to asingle group, whereas in the remaining 492 homelands, there can be upto three overlapping groups. For example, in northeastern India over anarea of 4,380 km2, the Assamese, the Oriyas, and the Santals overlap. Theluminosity of a region wheremultiple groups reside contributes to the av-erage luminosity of each group. The size of ethnic homelands varies con-siderably. The smallest polygon occupies an area of 1.09 km2 ðFrench inMonacoÞ, and the largest extends over 7,335,476 km2 ðAmerican Englishin the United StatesÞ. The median ðmeanÞ group size is approximately4,200 ð61,000Þ km2. The median ðmeanÞ country in our sample has eightð11.5Þ ethnicities, with the most diverse being Indonesia with 95 groups.Our second source is the fifteenth edition of the Ethnologue ðGordon

2005Þ, which maps 7,581 language-country groups worldwide in themid/late 1990s, using the political boundaries of 2000. In spite of thecomprehensive linguistic mapping, Ethnologue’s coverage of the Amer-icas and Australia is rather limited while for others ði.e., Africa and AsiaÞ,it is very detailed. Each polygon delineates a traditional linguistic region;populations away from their homelands ðin cities, refugee campsÞ arenot mapped. Groups of unknown location, as well as widespread and ex-tinct languages, are not mapped either; the only exception is the Englishin the United States. Ethnologue also records areas where languages over-lap. Ethnologue provides a more refined linguistic aggregation comparedto the GREG. As a result, the median ðmeanÞ homeland extends to 726ð12,676Þ km2. The smallest language is the Domari in Israel, which covers1.18 km2, and the largest group is the English in the United States, cov-ering 7,330,520 km2. The median ðmeanÞ country has nine ð42.3Þgroups, with Papua New Guinea being the most diverse with 809 linguis-tic groups.GREG attempts to map major immigrant groups whereas Ethnologue

generally does not. This is important for countries in the New World.For example, in Argentina, GREG reports 16 groups, among them Ger-mans, Italians, and Chileans, whereas Ethnologue reports 20 purely indig-enous groups ðe.g., the Toba and the QuechuaÞ. For Canada, Ethnologuelists 77mostly indigenous groups, such as the Blackfoot and the Chipewyan,with only English and French being nonindigenous; in contrast, GREGlists 23 groups featuring many nonindigenous ones, such as Swedes, Rus-sians, Norwegians, and Germans. Hence, the two ethnolinguistic map-pings capture different cleavages, at least in some continents. Thoughwe have performed various sensitivity checks, for our benchmark resultswe are including all groups without attempting to make a distinction asto which cleavage is more salient.6

6 A thorough exploration of ethnic inequality across different linguistic cleavages is rel-egated to the online supplementary appendix.



All

It is important to note that the underlying maps do include regionswhere groups overlap, and we take that into account in our measure,as we show below. However, both maps do not capture relatively recentwithin-country migrations toward the urban centers, for example. Thereason is that the original sources attempt to trace the historical home-land of each group. Hence, actual ethnic mixing is likely higher thanwhat the ethnographic maps reflect. This will induce measurement errorto our proxies of ethnic inequality. Nevertheless, under the assumptionthat in a given urban center the respective indigenous group is relativelymore populous than recent migrant ones, assigning the observed lumi-nosity per capita to this group is not entirely ad hoc. Moreover, there is alarge literature documenting that migrant workers channel systemati-cally a fraction of their earnings back to their homelands. This would im-ply that although we do not observe migrant workers in our data set tothe extent that they send remittances to their families and influencetheir livelihoods, this will be reflected in the luminosity per capita ofthe ancestral homelands, which we directly measure. Moreover, to atleast partially account for this issue, we have constructed all inequalitymeasures also excluding the regions where capitals fall.

2. Data on Luminosity and Population

Comparable data on income per capita at the ethnicity level are scarce.Hence, following Henderson et al. ð2012Þ and subsequent studies ðe.g.,Chen and Nordhaus 2011; Michalopoulos and Papaioannou 2013, 2014;Pinkovskiy 2013; Hodler and Raschky 2014; Pinkovskiy and Sala-i-Martin2014Þ, we use satellite image data on light density at night as a proxy.These—and other works—show that luminosity is a strong correlate ofdevelopment at various levels of aggregation ðcountries, regions, ethnichomelandsÞ. The luminosity data come from the Defense Meteorologi-cal Satellite Program’s Operational Linescan System that reports imagesof the earth at night ðfrom 20:30 till 22:00Þ. The six-bit number that rangesfrom 0 to 63 is available approximately at every square kilometer since1992.To construct luminosity at the desired level of aggregation, we average

all observations falling within the boundaries of an ethnic group andthen divide by the population of each area using data from the GriddedPopulation of the World, which reports georeferenced pixel-level popu-lation estimates for 1990 and 2000.7

7 The data are constructed using subnational censuses and other population surveys atvarious levels ðcity, neighborhood, regionÞ. For details, see http://sedac.ciesin.columbia.edu/data/collection/gpw-v. In the online supplementary appendix, we present varioussensitivity checks that examine the role of measurement error in both the population es-timates and the luminosity data.



3. New Ethnic Inequality Measures

We proxy the level of economic development in ethnic homeland i withmean luminosity per capita, yi; and we then construct an ethnic Gini co-efficient for each country that reflects inequality across ethnolinguisticregions. Specifically, the Gini coefficient for a country’s population con-sisting of n groups with values of luminosity per capita for the historicalhomeland of group i, yi, where i 5 1 to n are indexed in nondecreasingorder ðyi � yi11Þ, is calculated as follows:

G 51

n

�n 1 12 2

∑ni51ðn 1 12 iÞyi

∑ni51 yi

�:

The ethnic Gini index captures differences in mean income—as cap-tured by luminosity per capita at the ethnic homeland—across groups.For each of the two different ethnic-linguistic maps ðAtlas Narodov Miraand EthnologueÞ, we construct Gini coefficients for the maximum sampleof countries using cross–ethnic homeland data in 1992, 2000, and 2012.For each mapping we construct three ethnic Ginis. First, for our baselineestimates we use information from all groups. Second, we construct theGini coefficient dropping the capital cities. This allows us to accountboth for extreme values in luminosity and also for population mixing,which is naturally higher in capitals. Third, we compile measures exclud-ing small ethnicities, defined as those representing less than 1 percent ofthe 2000 population in a country.8

B. Measures of Spatial Inequality

Since we use ethnic homelands ðrather than individual-levelÞdata tomea-sure between-group inequality, the ethnic inequality measures also re-flect regional disparities in income or public goods provision that maynot be related to ethnicity per se. To isolate the between-ethnicity compo-nent of inequality from the regional one, we also construct Gini coeffi-cients reflecting ðiÞ the overall degree of spatial inequality and ðiiÞ in-equality across ðfirst- and second-levelÞ administrative units for eachcountry. Moreover, in an attempt to assess the accuracy of the underlyinggroups’ mappings, we perturbed the original homelands and compiledGini coefficients based on these altered ethnic homelands.

8 For example, in Kenya the Atlas Narodov Mira ðthe EthnologueÞ maps 18 ð53Þ ethnicðlinguisticÞ groups. Yet eight ethnic ð37 linguisticÞ areas host less than 1 percent of Kenya’spopulation as of 2000. So we construct Gini coefficients ðiÞ using all ethnic groups ð18 and53Þ, ðiiÞ dropping ethnic regions where the capital ðNairobiÞ falls ð17 and 52Þ, and ðiiiÞ us-ing the 11 ethnic and 16 linguistic groups, respectively, whose populations exceed 1 per-cent of Kenya’s population.



All

1. Overall Spatial Inequality Index

Our baseline index reflecting the overall degree of spatial inequality isbased on aggregating ðvia the Gini coefficient formulaÞ luminosity percapita across roughly equally sized pixels in each country. We first gener-ate a global grid of 2.5� 2.5 decimal degrees ðthat extends from2180 to180 degrees longitude and from 75 degrees latitude to 265 degreeslatitudeÞ. Second, we intersect the resulting grid with the 2000 DigitalChart of the World, which portrays contemporary national borders. Thisresults in 4,865 pixels across the globe falling within country bound-aries. The median ðmeanÞ box extends to roughly 22,500 ð27,500Þ km2.Note that boxes intersected by the coastline and national boundariesare smaller. Third, for each box we compute luminosity per capita in1992, 2000, and 2012. Fourth, we aggregate the data at the country levelestimating a Gini coefficient that captures the overall degree of spatial in-equality. The cross-countrymean ðmedianÞ number of pixels used for theestimation of the spatial Ginis is 24.9 ð8Þ, so these Ginis are quite compa-rable to the ethnic inequality measures.

2. Inequality across Administrative Regions

We also compiled inequality measures across administrative units, usingdata from the GADM Global Administrative Areas database on theboundaries of administrative regions. Following a procedure similar tothe derivation of the ethnic inequality and the overall spatial inequalityindexes, we construct measures reflecting inequality ðin lights per cap-itaÞ across first-level and second-level administrative units. In our sampleof 173 countries, the median ðmeanÞ number of first-level administrativeunits is 13 ð17Þ. A median ðmeanÞ first-level administrative unit spansroughly 7,197 ð44,050Þ km2, which is somewhat larger than the mediansize ð4,578Þ for groups in the Atlas Narodov Mira. The cross-country me-dian ðmeanÞ size of second-level administrative units is 110 ð301Þ km2.So, these units are much smaller than the Ethnologue or the Atlas NarodovMira homelands.

3. Inequality across Perturbed Ethnic Regions

We have also created ethnic Gini coefficients using perturbed ethnic re-gions. Using as inputs the centroids of ethnic-linguistic homelands, wegenerate Thiessen polygons that have the unique property that each poly-gon contains only one input point and that any location within a polygonis closer to its associated point than to a point within any other polygon.Thiessen polygons have the exact same centroids as the actual linguisticand ethnic homelands in the Ethnologue and GREG databases, respec-



tively, the key difference being that the actual homelands have idiosyn-cratic shapes.9 We then construct a spatial Gini coefficient that reflects in-equality in lights per capita across these sets of Thiessen polygons. Themean size of the Thiessen polygons based on the Ethnologue ðGREGÞ data-base is 11,862 ð58,784Þ km2, very similar to themean size of homelands inthe Ethnologue ðGREGÞ—12,676 ð61,213Þ km2.Comment.—All three proxies of the spatial inequality also reflect in-

equality across ethnic homelands, since ðiÞ there is clearly some degreeof measurement error on the exact boundaries of ethnic regions, andðiiÞ populationmixing is likely higher than the one we observe in the data.Moreover, in several countries, administrative boundaries follow ethniclines, while in the case of large groups, the spatial Gini coefficients mayalso ðpartiallyÞ capture within–ethnic group inequality.10

C. Example

Figures 1 and 2 provide an illustration of the construction of the ethnicinequality measures for Afghanistan. The Atlas Narodov Mira ðGREGÞmaps 31 ethnicities ðfig. 1AÞ, whereas the Ethnologue reports 39 languagesðfig. 2AÞ. According to GREG, the Afghans ðPashtunsÞ are the largestgroup residing in the southern and central-southern regions. This groupmakes up 51 percent of the population in 2000. The second-largestgroup are the Tajik people, who compose 22 percent of the populationand are located in the northeastern regions and in scattered pockets inthe western part of the country. There are eight territories in whichgroups overlap.Figures 1B and 2B portray the distribution of lights per capita for each

group, with lighter colors indicating more brightly lit homelands. Thecenter of the country, where the Hazara-Berberi reside, is poor; the sameapplies to the eastern provinces, where the Nuristani, the Pamir Tajiks,the Pashai, and the Kyrgyz groups are located. Luminosity is higher inthe Pashtun/Pathans homelands and to some lesser extent in the Tajikregions. Second, using lights per capita across all homelands, we esti-

9 Note that there are very few instances in which the number of Thiessen polygons is notidentical to the number of the underlying groups ðe.g., there is a difference of one groupfor six out of the 173 countries in the EthnologueÞ. This difference is due to the fact that ahandful of border/coastal groups have such a peculiar shape that their centroid falls out ofthe country’s boundaries they belong to. Hence, since Thiessen polygons are based on thecentroids of the ethnic-linguistic groups that fall within the country, those groups whosecentroids fall outside are not taken into account. Note that this has virtually no effecton the results since when we focus on the countries where the number of Thiessen poly-gons is identical to the number of groups, the pattern is the same.

10 In principle, one could generate within-group inequality measures using the finerstructure of the luminosity data. However, within-group mobility and risk-sharing issuesmake a luminosity-based, within-group inequality index less satisfactory.


FIG. 1.—Illustration of the construction of the ethnic inequality measures for Afghani-stan. The Atlas Narodov Mira ðGREGÞ maps 31 ethnicities ðpanel AÞ; the Ethnologue reports39 languages ðpanel BÞ.


All

mate the Gini coefficient in 1992, in 2000, and in 2012. In 2000 the Ginicoefficient estimated from GREG ðEthnologueÞ is 0.95 ð0.90Þ.11 We also es-timated the ethnic inequality measures excluding the ethnic homelandwhere the capital, Kabul, falls. In this case the ethnic Ginis are similarð0.95 with GREG and 0.91 with EthnologueÞ. For robustness, we also esti-mated Gini coefficients of ethnic inequality excluding groups constitut-ing less than 1 percent of the country’s population. In this case the Ginicoefficient with the GREG mapping is based on just four groups, whilethe Ethnologue -based Gini is based on seven ethnic homelands.

11 Since the Gridded Population of the World reports zero population for some ethnicareas, the Gini index with the GREGmapping is based on 27 ethnic observations, while theGini coefficient with the Ethnologue mapping is based on 39 linguistic groups ðno gaps inthis caseÞ.



Figure 3 illustrates the construction of the overall spatial inequality.When we divide the globe into boxes of 2.5 � 2.5 decimal degree boxes,we get 24 areas in Afghanistan. The estimated Gini index capturing theoverall degree of spatial inequality in Afghanistan is 0.73. For consistencywe also estimated the overall spatial inequality ðGiniÞ index excluding thepixel where the capital city falls or those boxes where less than 1 percent

FIG. 1 (Continued )



All

of the country’s population lived in 2000. Figures 4A and 4B illustrate theconstruction of inequality measures across administrative regions usingboth the first-level and second-level units. There are 32 provinces ðve-layatÞ that constitute the first-level administrative units, and there are328 second-level administrative units ðwuleswaliÞ. After estimating aver-

FIG. 2.—Distribution of lights per capita for each group based on GREG and Ethnologuemapping, with lighter colors indicating more brightly lit areas.



age luminosity per capita for each unit, we construct Gini indexes captur-ing inequality in development across administrative regions. Again, weconstruct these inequality measures using all regions, dropping the cap-ital, and also excluding those units with less than 1 percent of total pop-ulation. In our example, the first-level administrative unit Gini index is0.76 and the second-level administrative unit Gini coefficient is 0.93. Fig-ures 5A and 5B illustrate the derivation of the perturbed ethnic home-lands Gini index for Afghanistan based on the Atlas Narodov Mira andthe Ethnologue, respectively. There are 31 and 39 Thiessen polygons,as many as the number of ethnic and linguistic groups. The centroidsof the Thiessen polygons are identical to the ones of the actual home-lands, the only difference being that the actual homelands have ratherpeculiar shapes.

D. Descriptive Evidence

1. Ethnic Inequality around the World

Table 1 reports summary statistics for the baseline ethnic inequality mea-sures and the proxies of the overall degree of spatial inequality and re-gional inequality across administrative units. The average and medianvalues of the ethnic Gini coefficients are quite similar with both map-pings in each year ðaround 0.42–0.49 in 2000Þ. The average ðmedianÞ valueof the overall spatial Gini coefficient in 2000 is similar, 0.42 ð0.43Þ. The

FIG. 3.—Construction of the overall spatial inequality. When we divide the globe intoboxes of 2.5 � 2.5 decimal degree boxes, we get 24 areas in Afghanistan.



All

Gini coefficients based on administrative regions are, on average, smallerwhen estimated across first-level units ðmean 0.37Þ and larger when esti-mated at the finer second level ðmean 0.57Þ. Moreover, regional inequal-ity seems to be slightly trending downward, as all Gini coefficients aresmaller in 2012 ðand in 2000Þ. This may be driven by the expansion of

FIG. 4.—Construction of inequality measures across administrative regions using boththe first-level and second-level units.



electrification ðand regional convergenceÞ in many underdeveloped anddeveloping countries ðmostly in Africa and South AsiaÞ.Figures 6A and 6B illustrate the global distribution of ethnic inequality

with the GREG and Ethnologue mappings, respectively. Sub-Saharan Af-rica and East and South Asia host the most ethnically unequal countries.For example, with the Ethnologue mapping, the mean ðmedianÞ of the

FIG. 5.—Perturbed ethnic homelands for Afghanistan based on the Atlas Narodov Miraðpanel AÞ and the Ethnologue ðpanel BÞ.


TABLE1

Summar

ySt

atistics:C

ross-Countr

yIneq

ual

ityMea

sure

s

Observations

Mean

Stan

dard

Deviation

Minim

um

25th

Percentile

Med

ian

75th

Percentile

Maxim

um

A.Atlas

Narodov

MiraðG

REGÞ

Number

ofethnic

homelan

ds

173

10.994

13.579

13

812

94Ethnic

Giniin

2012

173

.406

.248

.00

.21

.44

.57

.96

Ethnic

Giniin

2000

173

.424

.260

.00

.20

.48

.63

.97

Ethnic

Giniin

1992

173

.473

.280

.00

.27

.54

.70

.96

B.Ethnologue

Number

oflingu

istichomelan

ds

173

38.619

94.029

13

835

753

Ethnolingu

isticGiniin

2012

173

.439

.325

.00

.13

.44

.74

.98

Ethnolingu

isticGiniin

2000

173

.446

.333

.00

.12

.49

.77

.98

Ethnolingu

isticGiniin

1992

173

.487

.346

.00

.14

.54

.80

.99

C.Pixels

Number

ofpixels

173

19.572

36.838

1.00

4.00

8.00

2029

2Sp

atialGiniin

2012

173

.411

.270

.00

.18

.42

.60

.98

SpatialGiniin

2000

173

.421

.269

.00

.18

.43

.64

.97

SpatialGiniin

1992

173

.463

.271

.00

.21

.49

.70

.96

444


TABLE1(C

ontinued)

Observations

Mean

Stan

dard

Deviation

Minim

um

25th

Percentile

Med

ian

75th

Percentile

Maxim

um

D.First-Level

AdministrativeUnits

Nfirst-levelad

ministrativeunits

173

16.873

14.962

1.00

8.00

12.00

2088

Administrativeunitðfirst-levelÞ

Giniin

2012

173

.353

.202

.00

.20

.30

.45

.93


Giniin

2000

173

.368

.215

.00

.21

.31

.48

.94


Giniin

1992

173

.422

.241

.00

.21

.38

.59

.94

E.Se

cond-Level

AdministrativeUnits

Nseco

nd-le

velad

ministrativeunits

135

290.10

462

6.28

511

.00

48.00

99.00

248

5,47

8Administrativeunitðse

cond-le

velÞ

Giniin

2012

135

.562

.237

.14

.35

.56

.78

.95

Administrativeunitðse

cond-le

velÞ

Giniin

2000

135

.572

.250

.15

.36

.55

.83

.96

Administrativeunitðse

cond-le

velÞ

Giniin

1992

135

.639

.249

.13

.43

.68

.87

.98

Note

.—Thetable

reportssummarystatistics

forthemainethnic

ineq

uality,overallspatialineq

uality,an

dad

ministrativeunitineq

ualitymeasuresem

-ployed

inthecross-countryan

alysis.Se

ction

IIan

dtheDataAppen

dix

give

details

on

theco

nstruction

oftheethnic

ineq

ualitymeasuresðG

ini

coefficien

tsÞ.AllGinicoefficien

tsareestimated

usingluminosityper

capitaacrossethnichomelan

dsðin

pan

elAÞ,acrosslingu

istichomelan

dsðin

pan

elBÞ,

pixels/boxe

sofrough

lythesamesize

ðinpan

elCÞ,first-levelad

ministrativeunitsðin

pan

elDÞ,an

dseco

nd-le

velad

ministrativeunitsðin

pan

elEÞ.

445


FIG. 6.—Global distribution of ethnic inequality with the GREG and the Ethnologuemap-ping ðpanels A and BÞ, world distribution of the overall degree of spatial inequality and re-gional inequality across first-level administrative units ðpanels C andDÞ, and global distribu-tion of ethnic inequality partialing out the effect of the overall spatial inequality ðpanels Eand F Þ.


All

baseline ethnic inequality index for sub-Saharan African countries is0.63 ð0.728Þ, while for South and East Asian countries the correspondingmean ðand medianÞ value of the ethnic Gini index is 0.59 ð0.69Þ.12 In

12 Specifically, ethnic inequality is particularly high across South Asia ðin total sevencountries, namely, Afghanistan, Bangladesh, Bhutan, India, Nepal, Pakistan, and Sri


FIG. 6 ðContinued Þ


contrast, western Europe is the region with the lowest level of ethnic in-equality ðmean and median values of ethnic Gini around 0.24Þ. Accord-ing to the Atlas Narodov Mira, the five most ethnically unequal countriesare Sudan, Afghanistan, Mongolia, Zambia, and Central African Repub-

LankaÞ. The mean and median Gini index is 0.635 on the basis of the Ethnologue and 0.55when we use the GREG. Ethnic inequality is also high in the 21 countries of the East Asiaand Pacific region, but only when we use the Ethnologue, where the mean is 0.58.



All

lic, with an average Gini coefficient in luminosity across ethnic home-lands of 0.91. According to the Ethnologue’s more detailed mapping oflanguage groups, the countries with the highest cross–ethnic group in-equality ðwhere Gini exceeds 0.95Þ are Democratic Republic of Congo,Papua New Guinea, Sudan, Ethiopia, and Chad.




Figures 6C and 6D plot the world distribution of the overall degree ofspatial inequality and regional inequality across first-level administrativeunits, respectively. As is evident, spatial and regional inequality is muchhigher in Asia and Africa as compared to western Europe and LatinAmerica. The countries with the highest overall spatial inequality accord-ing to the measure based on the 2.5� 2.5 decimal degree boxes are Rus-sia, Mongolia, Sudan, Peru, and Egypt; in all these countries the spatialGini coefficient exceeds 0.90. The countries with the highest regionalinequality across first-level administrative units are Libya, Chad, andGuinea ðGini around 0.90Þ. We should stress that in some countriesfirst-level administrative units cover large territories ðin terms of bothpopulation and land areaÞ. Hence, inequality measured across theseunits may not adequately capture existing regional inequalities. To partlyaccount for this, we have also constructed Gini coefficients using second-level administrative regions that in many countries are numerous. How-ever, an important caveat to keep in mind throughout the analysis is thatin several countries regional inequalities and, more importantly, ethnicdisparities in income may occur at much finer levels of aggregation ðe.g.,neighborhoodsÞ than what our ancestral-ethnic homeland approach allowsfor.13

Appendix table 1 reports the correlation structure of the ethnic Ginicoefficients between the two global maps at different points in time. Acouple of interesting patterns emerge. First, the correlation of the Ginicoefficients across the two alternative mappings is strong but not over-whelming. The correlation with the baseline measures that use all ethnicareas is around .75, but when we drop small groups or capitals, the cor-relation falls to .65. In line with our discussion above, these correlationssuggest that the two maps capture somewhat different aspects of ethnic-linguistic cleavages. Second, in the 20-year period for which luminositydata are available ð1992–2012Þ, ethnic inequality appears very persistent,as the correlations of the Gini coefficients over time exceed .90. Giventhe high inertia, in our empirical analysis below we will exploit cross-country variation. Third, not surprisingly, the correlation between eth-nic inequality and the Gini coefficient capturing the overall degree of

13 A case exemplifying this situation is that of South Africa, a country with sizable incomedifferences between ethnic groups. Since segregation, after the fall of apartheid, occurs ata much finer level than in the ancestral homelands, our data cannot capture this phenom-enon. South Africa looks also quite equal when inequality is measured across first-level ad-ministrative units ð0.22Þ. This is very similar to the ethnic Ginis, which are 0.20 with GREGand 0.28 with Ethnologue. However, regional inequality in South Africa is significantlyhigher when estimated across second-level administrative units ðGini index is 0.40, verysimilar to the global mean and median valuesÞ.



All

spatial inequality and regional inequality across ðfirst-levelÞ administra-tive units is positive, but again far from perfect. In particular, the corre-lation of the ethnic Gini with the overall spatial Gini ðbased on artificialboxesÞ ranges between .55 and .70, while the correlation of the ethnicGini coefficients with the administrative unit Ginis is lower, around .50.Since we are primarily interested in documenting the explanatory

power of ethnic inequality beyond the overall spatial inequality in mostspecifications, we control for the latter. Figures 6E and 6F portray theglobal distribution of ethnic inequality partialing out the effect of theoverall spatial inequality.

2. Basic Correlations

Ethnic diversity.—Appendix table 2, panel A, reports the correlation struc-ture between the various ethnic inequality and spatial inequalitymeasuresand the widely used ethnolinguistic fragmentation indexes. We observe apositive correlation between ethnic inequality and linguistic-ethnic frac-tionalization ð.35–.45Þ. Figures 7A and 7B provide a graphical illustra-tion of the association between the two proxies of ethnic inequality andthe ethnic and linguistic fragmentation measures of Alesina et al. ð2003Þand Desmet, Ortuño-Ortín, and Wacziarg ð2012Þ, respectively. The corre-lation between ethnic inequality and the segregation measures compiledby Alesina and Zhuravskaya ð2011Þ is also positive ð.20–.45Þ. Ethnic in-equality tends to go in tandem with segregation. This is reasonable sinceeconomic differences between groups are more likely to persist whengroups are also geographically separated. We also examine the associa-tion between ethnic inequality and spatial inequality with the ethnic po-larization indicators of Montalvo and Reynal-Querol ð2005Þ, failing todetect a systematic association. These patterns suggest that the ethnicinequality measure captures a dimension distinct from already-proposedaspects of a country’s ethnic composition.Income inequality.—We then examined the association between ethnic

inequality and income inequality, as reflected in the standard Gini coef-ficient ðapp. table 2, panel BÞ. The income Gini coefficient is taken fromEasterly ð2007Þ, who using survey and census data compiled by the UN’sWorld Institute for Development Economics Research constructs ad-justed cross-country Gini coefficients for more than a hundred coun-tries over the period 1965–2000. Figures 8A and 8B illustrate this asso-ciation using the GREG and the Ethnologue mapping, respectively. Thecorrelation between ethnic inequality and economic inequality is mod-erate, around .25–.30. Yet this correlation weakens considerably and be-comes statistically insignificant once we simply condition on continen-tal constants.


FIG. 7.—Graphical illustration of the association between the two proxies of ethnic in-equality and the ethnic and linguistic fragmentation measures of Alesina et al. ð2003Þ andDesmet et al. ð2012Þ, respectively.



All

III. Ethnic Inequality and Development

A. Baseline Estimates

In table 2 we report cross-country ordinary least squares ðOLSÞ estimates,relating the log of per capita GDP in 2000 with ethnic inequality. In panel

FIG. 8.—Association between income inequality and ethnic inequality using the Ethnologueand GREG mapping of groups’ homelands.



A we use the ethnic inequality measure based on the Atlas Narodov Miramapping, while in panel B we use the measures derived from Ethnologue’smapping. In all specifications we include region-specific constants ðfol-lowing the World Bank’s classificationÞ to account for continental differ-ences in ethnic inequality and comparative economic development.The coefficient of the ethnic inequality index in column 1 is nega-

tive and significant at the 1 percent level. Figures 9A–9D illustrate theunconditional and the conditional on regional fixed effects association.Specification 2 also reveals a negative association between economic de-velopment and the overall degree of spatial inequality, as reflected in theGini coefficient based on pixels of 2.5 � 2.5 degrees. This suggests thatunderdevelopment goes in tandem with regional inequalities. In col-umn 3 we include both the ethnic inequality Gini index and the spatialGini coefficient. The estimate on the ethnic inequality Gini is stablewith both the GREG and the Ethnologuemappings. In contrast, the coeffi-cient on the overall spatial inequality measure drops considerably andbecomes statistically indistinguishable from zero in both models. Thissuggests that the ethnic component of spatial inequality is the relativelystronger negative correlate of development.In column 4we associate the log of per capita GDPwith the log number

of ethnic-linguistic groups. In line with previous works, income per capitais significantly lower in countries with many ethnic ðpanel AÞ and linguis-tic ðpanel BÞ groups; yet the estimates in column 5, where we jointlyinclude in the empirical model the proxies of ethnic inequality andfractionalization, show that it is income differences along ethnic linesrather than ethnolinguistic heterogeneity per se that correlate with un-derdevelopment. The results are similar when we jointly include in thespecification the ethnic Gini index, the overall spatial inequality mea-sure, and the fractionalizationmeasure in column 6. Although, as a resultof the small number of observations and multicollinearity ðsee app. ta-ble 1Þ, these results should be interpreted with caution, only the ethnicinequality measure enters with a statistically significant estimate.In columns 7 and 8 we examine whether the significantly negative as-

sociation between ethnic inequality and income per capita is driven byan unequal clustering of population across ethnic homelands or bythe skewness in the size of ethnic homelands; to do so we construct Ginicoefficients of population and land area that capture inequality in thesize of ethnic homelands. The ethnic inequality Gini index retains itseconomic and statistical significance, while both the population andthe homeland size ethnic Ginis enter with estimates statistically indistin-guishable from zero. This suggests that the association between ethnicinequality and underdevelopment is not driven by inequality in the sizeof ethnic homelands captured by either the population of each group orthe area of each homeland. In column 9 we also control for a country’ssize including in the empirical model the log of population in 2000 and


TABLE2

BaselineEstim

ates:E

thnic

Ineq

ual

ityan

dEco

nomic

Dev

elopm

entin

2000

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

A.Atlas

Narodov

MiraðG

REGÞ

Ethnic

ineq

ualityðG

iniÞ

21.39

11***

21.39

00***

2.951

8**

2.927

6*21.34

49***

21.10

32**

21.11

72**

ð.258

8Þ

ð.341

6Þð.3

953Þ

ð.484

5Þ

ð.494

3Þ

ð.518

8Þð.5

492Þ

Spatialineq

ualityðG

iniÞ

2.997

3***

2.001

52.031

52.004

6.010

42.559

2ð.2

774Þ

ð.351

0Þð.3

568Þ

ð.353

9Þ

ð.359

1Þð.4

749Þ

Lognumber

ofethnicities

2.313

6***

2.142

92.143

32.217

4*2.186

3ð.0

612Þ

ð.090

8Þ

ð.091

7Þ

ð.127

7Þð.1

440Þ

Ethnicineq

ualityin

populationðG

iniÞ

.651

71.15

541.08

58ð1.150

0Þ

ð1.155

4Þð1.254

6ÞEthnic

ineq

ualityin

size

ðarea;GiniÞ

2.793

32.794

92.827

6ð1.173

2Þ

ð1.106

0Þð1.229

7ÞLoglandarea

.144

2ð.0

879Þ

Logpopulationðin

2000

Þ2.136

8ð.0

829Þ

Adjusted

R2

.654

.623

.652

.646

.657

.655

.65

.656

.662

Observations

173

173

173

173

173

173

173

173

173

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

B.Ethnologue

Ethnic

ineq

ualityðG

iniÞ

21.16

03***

21.07

45***

21.16

88***

21.06

36**

21.34

52***

21.23

68***

21.71

86***

ð.232

8Þ

ð.265

2Þð.3

587Þ

ð.424

1Þ

ð.346

9Þ

ð.437

7Þð.4

549Þ

Spatialineq

ualityðG

ini;pixelsÞ

2.997

3***

2.154

92.156

42.176

02.199

22.535

1ð.2

774Þ

ð.302

1Þð.3

136Þ

ð.301

0Þ

ð.316

5Þð.4

235Þ

454


TABLE2(C

ontinued)

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

Lognumber

oflangu

ages

2.192

1***

.002

12.002

52.037

8.072

6ð.0

466Þ

ð.068

8Þ

ð.071

1Þ

ð.082

6Þ

ð.096

7Þ

Ethnicineq

ualityin

populationðG

iniÞ

.801

2.846

0.845

3ð.9

324Þ

ð.938

7Þ

ð.914

4Þ

Ethnic

ineq

ualityin

size

ðarea;

GiniÞ

2.489

82.457

42.295

8ð.8

948Þ

ð.904

5Þ

ð.885

3Þ

Loglandarea

.136

6*ð.0

808Þ

Logpopulation

2.210

5**

ð.084

0Þ

Adjusted

R2

.654

.623

.652

.632

.652

.65

.651

.649

.665

Observations

173

173

173

173

173

173

173

173

173

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Note

.—Thetable

reportscross-countryOLSestimates.Thedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.Theethnic

Ginico

efficien

tsreflectineq

ualityin

ligh

tsper

capitaacrossethnic-lingu

istichomelan

ds.In

pan

elA,weuse

thedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞtoag-

greg

ateligh

tsper

capitaacrossethnic

homelan

ds.In

pan

elB,w

euse

thedigitized

versionoftheEthnologuedatab

aseto

aggreg

ateligh

tsper

capitaacross

lingu

istichomelan

ds.Theoverallspatialineq

ualityindex

ðGinicoefficien

tÞcapturesthedegreeofspatialineq

ualityacross2.5�

2.5decim

aldegreeboxes/

pixelsin

each

countryðboxe

sintersectedbynational

boundariesareofsm

allersizeÞ.Se

ctionIIgivesdetailsontheco

nstructionoftheethnic

ineq

uality

andspatialineq

ualityðG

iniÞindex

es.Thelognumber

ofethnicitiesin

cols.4–

6,8,

and9den

otesthelogarithm

ofthenumber

ofethnic

andlingu

istic

groupsin

each

countryacco

rdingto

theGREGðin

pan

elAÞa

ndtheEthnologueðin

pan

elBÞ.C

olumns7,8,an

d9includeas

controlsaGiniindex

capturing

ineq

ualityin

populationacross

ethnic

ðlingu

isticÞ

homelan

dsan

daGiniindex

capturingineq

ualityin

landarea

across

ethnic

ðlingu

isticÞ

homelan

ds.

Column9includes

thelogofaco

untry’slandarea

andthelogofpopulationin

2000

.Allspecificationsincluderegional

fixe

deffectsðco

nstan

tsnot

reported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.Robust

ðheterosked

asticity-adjusted

Þstandarderrors

arereported

inparen

theses

belowtheestimates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.

***Statisticallysign

ificantat

the1percentlevel.

455


FIG. 9.—Unconditional and conditional on regional fixed effects association betweenethnic inequality and GDP per capita across countries.


All

log of land area, as ethnic heterogeneity, ethnic inequality, and the over-all degree of spatial inequality are likely to be increasing in size. Doingso has little effect on our results. Ethnic inequality remains a systematiccorrelate of underdevelopment.



The estimate on the ethnic inequality index with the Atlas NarodovMira mapping in panel A ðcol. 9Þ implies that a reduction in the ethnicGini coefficient by 0.25 ðone standard deviation, from the level of Nige-ria, where the ethnic Gini is 0.76, to the level of Namibia, where the eth-nic Gini is 0.53Þ is associated with a 28 percent ð0.25 log pointsÞ increasein per capita GDP ðthese countries have very similar overall spatial Ginis




All

of around 0.8Þ. The standardized beta coefficient of the ethnic inequal-ity index is around 0.20–0.30, quite similar to that of the works on therole of institutions on development ðe.g., Acemoglu, Johnson, and Rob-inson 2001Þ.Other aspects of the ethnic composition.—In table 3 we investigate whether

other dimensions of the distribution of the population across groups, re-lated to fractionalization, polarization, and genetic diversity, rather thanincome inequality across ethnic lines, influence comparative develop-ment. In columns 1 and 6 we augment the specification with the Alesinaet al. ð2003Þ and Desmet et al. ð2012Þ ethnic and linguistic fractionaliza-tion measures, respectively. Doing so has no effect on the coefficient onethnic inequality, which retains its economic and statistical significance.Moreover, the fractionalization indicators enter with unstable and statis-tically insignificant estimates, suggesting that it is differences in well-being across ethnic lines that explain underdevelopment rather thanfragmentation per se.14 In columns 2 and 7 we experiment with Fearon’sð2003Þ cultural fragmentation index that adjusts the fractionalization in-dex for linguistic distances among ethnic groups. Cultural fractionaliza-tion enters with a statistically insignificant estimate, while the ethnic in-equality Gini index retains its economic and statistical significance.Motivated by recent works highlighting the importance of polarization

ðMontalvo and Reynal-Querol 2005; Esteban, Mayoral, and Ray 2012Þ, incolumns 3 and 8 we condition on an index of ethnic polarization. Ethnicinequality correlates strongly with development, while the polarizationmeasures enter with insignificant estimates.15

Building on the recent work of Alesina and Zhuravskaya ð2011Þ show-ing that countries with a high degree of ethnolinguistic segregation tendto have low-quality national institutions and inefficient bureaucracies, incolumns 4 and 9 we include in the specifications their measures of eth-nic and linguistic segregation, respectively. The sample falls consider-ably, as these measures are available for approximately 90 countries.While there is some evidence that ethnic segregation is a feature of un-derdevelopment, the coefficient on the ethnic inequality proxy contin-ues to be quite stable and significant at standard confidence levels.In columns 5 and 10 we condition on a proxy of within-country ge-

netic diversity, based on migratory distance of each country’s capital from

14 When we do not include the ethnic inequality Ginis, the ethnic and linguistic frag-mentation measures enter with negative and significant ðat the 10 percent to 5 percentlevelÞ estimates ðapproximately 20.55Þ.

15 The same applies if we use alternative measures of ethnic-linguistic polarization. Over-all polarization is significantly related to civil conflict but not to income per capita. We alsoestimated specifications including both the polarization and the fractionalization indica-tors; in all perturbations the coefficient on ethnic inequality retains its statistical and eco-nomic significance.



Ethiopia. Since Ashraf and Galor ð2013Þ argue that the effect of geneticdiversity on development is nonlinear, we enter the latter in a quadraticfashion ðthough this has no effect on our resultsÞ. In all permutationsthe ethnic inequality proxy enters with a stable ðaround 21Þ and highlysignificant estimate.

Overall the results in table 3 show that the strong negative associa-tion between ethnic inequality and income across countries is not medi-ated by differences in the societies’ ethnic or genetic composition.16

Alternative measures of ethnic inequality and geographic controls.—In table 4we augment the main specification with an array of geographic traits andexperiment with alternative measures of ethnic inequality. In columns 1and 4 we use the baseline ethnic inequality measures based on all home-lands. In columns 2 and 5 we use ethnic Ginis that exclude from the es-timation regions where capitals fall. Note that the sample drops as inthese models we do not consider monoethnic and monolinguistic coun-tries. In columns 3 and 6 we introduce ethnic Ginis that exclude groupswith less than 1 percent of a country’s population. Note that, a priori,there is no reason to exclude small groups, since ethnic hatred may bedirected to minorities that, nevertheless, control a significant portionof the economy ðChua 2003Þ. Moreover, when these groups are dropped,the sample of ethnic homelands used to estimate the ethnic Ginis dropsconsiderably.17 To avoid concerns of self-selecting the conditioning set,we follow the baseline specification of Nunn and Puga ð2012Þ and in-clude ðon top of log population and log land areaÞ an index of terrainruggedness, distance to the coast, an index of gemquality, the percentageof each country with fertile soil, and the percentage of tropical land ðtheData Appendix gives variable definitionsÞ. To isolate the role of ethnicinequality on development from regional inequalities and ethnic frag-mentation, in all specifications we control for the overall degree of spatialinequality in lights per capita and ethnic-linguistic fractionalization.The negative correlation between ethnic inequality and income per

capita remains strong. This applies to all proxies of ethnic inequality.While compared to the unconditional specifications the estimate on eth-nic Gini declines somewhat, it retains significance at standard confidencelevels. Thus, while still an unobserved or omitted countrywide factor may

16 We also experimented with the newly constructed index of birthplace diversity ofAlesina, Harnoss, and Rapoport ð2016Þ, again finding that the link between ethnic inequal-ity and underdevelopment is robust.

17 On average, the number of ethnic ðlinguisticÞ groups per country falls from 11 ð39Þ to4.2 ð7Þ. While the median number of groups per country across the 173 countries is eightðwith both GREG and EthnologueÞ, when we drop groups consisting of less than one of acountry’s population, the medians fall to 3 ðGREGÞ and 4 ðEthnologueÞ. In contrast tothe ethnic inequality measures, the spatial Gini and the administrative unit Gini coeffi-cients do not get affected much when we drop small ðin terms of populationÞ pixels andadministrative regions.


TABLE3

Eth

nic

Ineq

ual

ityan

dEco

nomic

Dev

elopm

entEth

nic:Ineq

ual

ityan

dOth

erFe

ature

sofEth

nolinguisticFr

agmen

tation,P

ola

riza

tion,a

ndDiver

sity

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

ð10Þ

Ethnic

ineq

uality

ðGiniÞ

21.39

11***

21.61

16***

21.39

87***

21.43

69***

21.36

84***

21.07

16***

2.962

0***

21.08

93***

21.04

77*

21.01

08***

ð.351

1Þ

ð.387

6Þð.3

478Þ

ð.527

4Þð.3

551Þ

ð.293

2Þ

ð.294

2Þ

ð.261

9Þ

ð.580

8Þ

ð.290

1Þ

Ethnic-lingu

istic

fragmen

tation

.005

52.006

1ð.3

549Þ

ð.294

3Þ

Culturalfragmen

ta-

tion

2.372

1.010

8ð.3

486Þ

ð.363

7Þ

Ethnolingu

istic

polarization

.444

2.621

6ð1.006

1Þð1.000

4ÞEthnic-lingu

istic

segreg

ation

21.32

94*

2.189

0ð.7

294Þ

ð.897

4Þ

Gen

etic

diversity

183.03

60**

195.31

81**

ð84.51

87Þ

ð81.16

72Þ

Gen

etic

diversity

2213

0.16

18**

214

0.53

93**

460

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AMAll use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago.

ð60.25

10Þ

ð58.36

10Þ

edu/t-and-c).

TABLE3(C

ontinued)

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

ð10Þ

Spatialineq

uality

ðGiniÞ

2.002

1.209

62.033

5.208

6.061

52.155

72.169

22.187

8.165

12.101

5ð.3

541Þ

ð.371

7Þð.3

540Þ

ð.414

4Þ

ð.340

7Þ

ð.304

6Þ

ð.308

6Þ

ð.307

2Þð.4

348Þ

ð.303

8ÞAdjusted

R2

.650

.684

.646

.731

.678

.650

.669

.646

.674

.675

Observations

173

150

172

9615

717

315

017

292

157

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Note

.—Thetable

reportscross-countryOLSestimates.T

hedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.TheethnicGinico

efficien

tsreflectineq

ualityin

ligh

tsper


ds,based

onthedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞinco

ls.1

–5an

dbased

on

theEthnologuein

cols.6–10

.Theoverallspatialineq

ualityindex

ðGinicoefficien

tÞcapturesthedeg

reeofspatialineq

ualityacross2.5�2.5decim

aldeg

ree

boxe

s/pixelsin

each

countryðboxe


boundariesan

dtheco

astlineareofsm

allersizeÞ.Se


nstruction

oftheethnicineq

ualityan

dspatialineq

ualityðG

iniÞindex

es.Inco

lumns1an

d6,

weco

ntrolforethnican

dlingu

isticfragmen

tationusingindicatorsthat

reflectthelike

lihoodthat

tworandomlych

osenindividualsin

oneco

untrywillnotbemem

bersofthesamegroupðth

eethnic

fragmen

tationindex

inco

l.1co

mes

from

Alesinaet

al.[20

03]an

dthelingu

isticfragmen

tationindex

inco

l.6co

mes

from

Desmet

etal.[20

12]Þ.

Inco

ls.2

and7,

weco

ntrolfor

culturalðlingu

isticÞ

fragmen

tationusingan

index

ðfrom

Fearon20

03Þthat

acco

untsforlingu

isticdistancesam

onggroups.In

cols.3

and8,

weco

ntrolfor

ethnic

polarization,usingtheMontalvoan

dReynal-Q

uerolð200

5Þindex

.In

cols.4an

d9,

weco

ntrolforethnic

andlingu

isticsegreg

ation,respectively,

usingthemeasuresofAlesinaan

dZhuravskaya

ð201

1Þ.In

cols.5an

d10

,weco

ntrolforthege

netic

diversity

index

ofAshrafan

dGalorð201

3Þ.Allspec-

ificationsincluderegionalfixe

deffectsðco

nstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.R

obuststan

dard

errors

arereported

inparen

theses

belowtheestimates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.


ificantat

the1percentlevel.

461


TABLE4

Eth

nic

Ineq

ual

ityan

dEco

nomic

Dev

elopm

entin

2000

:Additional

Contr

ols

andAlt

ernat

iveMea

sure

sofEth

nic

Ineq

ual

ity

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

AllEthnic

Areas

ð1Þ

Excluding

Cap

itals

ð2Þ

Excluding

SmallGroups

ð3Þ

AllEthnic

Areas

ð4Þ

Excluding

Cap

itals

ð5Þ

Excluding

SmallGroups

ð6Þ

Ethnic

ineq

ualityðG

iniÞ

2.917

2***

2.636

7*2.983

5**

2.769

2***

2.626

2**

2.935

6**

ð.328

7Þ

ð.326

6Þ

ð.450

3Þ

ð.290

2Þð.2

909Þ

ð.395

5ÞSp

atialineq

ualityðG

iniÞ

2.569

821.20

35***

21.12

54*

2.695

2*21.33

66***

21.04

95*

ð.368

6Þ

ð.370

0Þ

ð.623

7Þ

ð.382

5Þð.3

535Þ

ð.610

5ÞEthnic-lingu

isticfragmen

tation

.140

9.318

5.379

3.107

52.059

7.141

1ð.3

145Þ

ð.300

9Þ

ð.344

4Þ

ð.267

1Þð.2

696Þ

ð.252

5ÞAdjusted

R2

.724

.760

.736

.723

.759

.734

Observations

173

155

173

173

147

173

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Simple

controls

Yes

Yes

Yes

Yes

Yes

Yes

Geo

grap

hic

controls

Yes

Yes

Yes

Yes

Yes

Yes

Note

.—Thetable


hedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.Theethnic

Ginico

efficien

tsreflectineq

ualityin

ligh

tsper

capitaacross

ethnic

homelan

ds,based

onthedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞinco

ls.1–

3an

donthe

Ethnologuein

cols.4–6.

Theoverallspatialineq

ualityindex

ðGinico

efficien

tÞcapturesthedeg

reeofspatialineq

ualityacross

2.5�

2.5decim

aldeg

ree

boxe

s/pixelsin

each

countryðboxe


boundariesan

dtheco

astlineareofsm

allersizeÞ.Fortheco


and

thespatialineq

ualitymeasuresðG

inicoefficien

tsÞinco

ls.1

and4,

weuse

alle

thnicðlingu

isticÞ

homelan

dsðan

dpixelsÞ;

inco

ls.2

and5,

weex

cludeethnic

areasðan

dpixelsÞ

wherecapital

cities

fall;in

cols.3an

d6,

weex

cludepolygo

nsðlingu

istic,

ethnic,boxe

sÞwithless

than

1percentofaco

untry’spopu-

lation.S

ectionIIgivesdetailsontheco

nstructionoftheethnicineq

ualityan

dspatialineq

ualityðG

iniÞindex

es.Inallspecificationsweco

ntrolforethnic-

lingu

isticfragmen

tationusingindicators

reflectingthelike

lihoodthat

tworandomlych

osenindividualsin

oneco

untrywillnotbemem

bersofthesame

groupðth

eethnic

fragmen

tationindex

inco

ls.1–

3co

mes

from

Alesinaet

al.[200

3]an

dthelingu

isticfragmen

tationindex

inco

ls.4–6co

mes

from

Desmet

etal.[201

2]Þ.In

allspecificationsweincludeas

controlsloglandarea

andlogpopulationin

2000

ðsimple

setofco

ntrolsÞ,ameasure

ofterrain

rugg

edness,thepercentage

ofeach

countrywithfertilesoil,thepercentage

ofeach

countrywithtropicalclim

ate,averagedistance

tonearestice-free

coast,

andan

index

ofge

m-qualitydiamondex

tractionðgeo

grap

hic

setofco

ntrolsÞ.Allspecificationsincluderegional

fixe

deffectsðco

nstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.Robuststan

darderrors

arereported

inparen

theses

belowtheestimates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.


ificantat

the1percentlevel.

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AMAll use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago
.edu/t-and-c).


jointly affect development and ethnic inequality, the estimates clearlypoint out that the correlation does not reflect ðobservableÞ mean differ-ences in commonly employed geographical characteristics.

B. Inequality across Administrative Units, Ethnic Inequality,and Development

We now examine the relationship between ethnic inequality and com-parative development, accounting for regional disparities across admin-istrative units. In this regard, as described in Section II, we have con-structed Gini coefficients reflecting inequality in lights per capitaacross first- and second-level administrative units. This variable is quiteuseful in many ways. First, as administrative units are well defined, theregional Ginis are easily interpretable. Second, examining the link be-tween spatial inequality across administrative regions and developmentis interesting by itself. A vast literature that goes back at least to the workof Williamson ð1965Þ has studied theoretically and empirically the in-terlinkages between development and spatial ðregionalÞ inequality. ðSeethe reviews of Kanbur and Venables [2008] and Kim [2009] for recentworks.Þ Third, since in some countries ethnic boundaries have formedthe basis for the delineation of administrative units, we can directly testwhether the strong cross-country correlation between inequality acrossethnic homelands and GDP per capita reflects an inverse relationshipbetween inequality across politically defined regions and comparativedevelopment.Table 5 reports the results. Let us start with panel A, where we use Gini

coefficients of regional inequality estimated across first-level administra-tive units. On average, there are 18 first-level administrative units in eachcountry. Examples of first-level unit regions include the German landerð16Þ; the US ð50Þ, Brazilian ð27Þ, and Indian ð35Þ states; the Swiss can-tons ð26Þ; and the Chinese provinces and autonomous regions ð32Þ.The coefficient on the administrative unit Gini index in the uncondi-tional specification ðin col. 1Þ is negative and highly significant ð21.60Þ.This suggests that underdevelopment is characterized by large regionaldifferences in well-being ðor public goods provisionÞ. This is in accordwith our earlier results ðe.g., table 2, col. 2Þ showing a similar pattern whenusing the overall spatial inequality Gini. In columns 2 and 6 we includeboth the administrative unit and the ethnic inequality Ginis ðusing theAtlas Narodov Mira and Ethnologue mappings, respectivelyÞ. Both inequal-ity measures enter with negative and significant estimates ðmagnitudearound21Þ. In columns 3 and 7 we control for ethnic and linguistic frac-tionalization ðusing the Alesina et al. [2003] and Desmet et al. [2012]measures, respectivelyÞ. In line with our previous estimates, once we ac-count for inequalities across ethnic ðand now also across administrativeÞ


TABLE5

Eth

nic

Ineq

ual

ity,

Administr

ativeUnit

Ineq

ual

ity,

andEco

nomic

Dev

elopm

ent

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

Unco

nditional

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

A.Ineq

ualityacross

AdministrativeUnitsðFirst-LevelÞ

Administrativeunitineq

uality

ðGiniÞ

21.67

17***

21.01

39**

21.01

85**

21.50

74***

21.26

51***

21.00

80**

21.01

50**

21.51

77***

21.27

52***

ð.400

4Þð.4

188Þ

ð.423

0Þð.4

096Þ

ð.430

7Þ

ð.444

6Þ

ð.448

6Þ

ð.433

3Þð.4

494Þ

Ethnic

ineq

ualityðG

iniÞ

21.02

52***

21.04

01***

21.04

78***

2.808

7**

2.852

7***

2.828

0***

2.795

2**

2.609

6**

ð.261

1Þ

ð.276

5Þð.3

557Þ

ð.319

7Þ

ð.248

7Þ

ð.285

6Þ

ð.315

8Þð.2

941Þ

Ethnic-lingu

istic

fragmen

tation

.058

42.020

6.133

72.054

82.110

3.023

9ð.3

479Þ

ð.338

7Þ

ð.306

2Þ

ð.278

3Þ

ð.273

6Þð.2

572Þ

Adjusted

R2

.642

.667

.665

.686

.741

.667

.665

.686

.738

Observations

173

173

173

173

173

173

173

173

173

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Simple

controls

No

No

No

Yes

Yes

No

No

Yes

Yes

Geo

grap

hic

controls

No

No

No

No

Yes

No

No

No

Yes

All

464


TABLE5(C

ontinued)

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

Unco

nditional

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

ð9Þ

B.Ineq

ualityacross

AdministrativeUnitsðSecond-LevelÞ

Administrativeunitineq

uality

ðGiniÞ

2.609

5*2.157

92.171

321.08

86***

2.697

6*2.283

42.309

921.19

67***

2.799

5**

ð.339

1Þð.3

421Þ

ð.340

9Þð.3

572Þ

ð.367

5Þ

ð.351

8Þ

ð.353

5Þ

ð.362

7Þð.3

482Þ

Ethnic

ineq

ualityðG

iniÞ

21.10

68***

21.04

95***

21.08

63***

2.788

2**

2.708

6**

2.617

6*2.888

1***

2.687

3**

ð.331

5Þ

ð.334

6Þð.3

812Þ

ð.344

5Þð.2

974Þ

ð.332

3Þ

ð.338

9Þð.3

415Þ

Ethnic-lingu

istic

fragmen

tation

2.223

72.514

2.025

2.182

72.202

7.059

9ð.4

018Þ

ð.377

2Þ

ð.351

4Þ

ð.320

7Þ

ð.293

7Þð.2

557Þ

Adjusted

R2

.677

.698

.696

.726

.773

.692

.690

.723

.773

Observations

135

135

135

135

135

135

135

135

135

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Simple

controls

No

No

No

Yes

Yes

No

No

Yes

Yes

Geo

grap

hic

controls

No

No

No

No

Yes

No

No

No

Yes

Note

.—Thetable

ðboth

pan

elsAan

dBÞr

eportscross-countryOLSestimates.Thedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.The

ethnicGinicoefficien

tsreflectineq

ualityin

ligh

tsper


ds,based

onthedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞin

cols.2–

5an

dontheEthnologuein

cols.6–

9.Thead

ministrativeunitGiniindex

reflects

ineq

ualityin

ligh

tsper

capitaacross

administrativeregions.In

pan

elAweuse

first-leveladministrativeunits.In

pan

elBweuse

seco

nd-le

veladministrativeunits.Se



ineq

ualityan

dspatialineq

ualityðG

iniÞindex

es.In

specifications3–

5an

d7–

9,weco

ntrolforethnic-lingu

isticfragmen

tationusingindicators

reflecting

thelike

lihoodthat

tworandomlych

osenindividualsin

oneco

untrywillnotbemem


eethnicfragmen

tationindex

inco

ls.3

–5

comes

from

Alesinaet

al.[20

03]an

dthelingu

isticfragmen

tationindex

inco

ls.7–9co

mes

from

Desmet

etal.[20

12]Þ.

Specifications4,5,8,an

d9include

asco

ntrolsloglandarea

andlogpopulationin

2000

ðsimple

setofco

ntrolsÞ.Sp

ecifications5an

d10

includeas

controlsameasure

ofterrainrugg

edness,

thepercentage

ofeach

countrywithfertilesoil,thepercentage

ofeach

countrywithtropical

clim

ate,

averagedistance

tonearestice-free

coast,an

dan

index

ofge

m-qualitydiamondex

tractionðgeo

grap

hicsetofco


fixe

deffectsðco

nstan

tsnotreported

Þ.TheData

Appen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.Robuststan

darderrors

arereported

inparen

theses

belowtheestimates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.

***

Statisticallysign

ificantat

the1percentlevel.

465



All

regions, there is no systematic link between ethnolinguistic fragmentationand development. In columns 4, 5, 8, and 9 we control for country sizeðlog population and log land areaÞ and the rich set of geographic features.The results remain intact. Across all permutations, both the ethnic in-equality measure and the Gini index capturing inequality across first-leveladministrative units enter with negative and highly significant coeffi-cients. The “standardized” beta coefficients that summarize in terms ofstandard deviations the change in the outcome variable ðlog of per capitaGDPÞ induced by a one standard deviation change in the independentvariables are comparable for the two inequality measures, around 0.20.Table 5, panel B, reports similar specifications in which administrative-

level inequality is estimated across second-level units. The Global Admin-istrative Areas database does not report second-level administrative unitsfor all countries; hence the sample drops to 135 ðwe mostly lose smallcountries, such as Singapore, Jamaica, and SwazilandÞ. The results aresimilar if we assign to these countries the first-level administrative unitGini coefficients. As the median ðmeanÞ number of such units is 110ð301Þ, the respective Ginis are estimated using a very fine aggregation. Ex-amples include the German ðregierungsbezirkÞ government regions ð40Þ,the French département ð96Þ, and the Brazilian municipalities ð5,503Þ.The coefficient on the administrative region Gini index in column 1 isnegative and significant at the 90 percent level; yet its magnitude is con-siderably smaller than the analogous one with the first-level administra-tive Gini index ð20.61Þ. ðThe implied beta coefficient is 20.10.Þ Thecoefficient on the administrative region Gini drops considerably andloses its statistical significance once we include the ethnic inequalityproxy ðcols. 2 and 5Þ and condition on ethnolinguistic fragmentationðcols. 3 and 6Þ. In contrast, the ethnic inequality measure retains its sta-tistical and economic significance. The coefficient on the ethnic Gini isunaffected when we condition on size and geography ðin cols. 4, 5, 8,and 9Þ.The evidence in table 5 reveals two important findings. First, in a large

cross section of countries there is a clear negative association betweeneconomic performance and regional inequalities across first-level ad-ministrative units. This new ðto the best of our knowledgeÞ finding addsto the literature in urban economics and economic geography that stud-ies the relationship between regional economic disparities and the pro-cess of development.18 Second, and more important given our focus, thestrong cross-country link between ethnic inequality and underdevelop-

18 Note that because of the lack of comparable regional income data across countries,empirical works on spatial inequalities have mostly been country specific, and the few ex-isting comparative studies have relied on small samples ðe.g., Ezcurra 2013; Lessmann2014Þ.



ment does not capture the similarly negative association between GDPper capita and economic differences across politically defined spatialunits.

C. Perturbing Ethnic Homelands

Wenow explore whether the pattern uncovered so far survives a horse racebetween ethnic inequality constructed using the original mappings andethnic inequality based on slightly modified ethnic homelands.19 Showingthat our original ethnic inequality measures dominate the Gini indexbased on perturbed ethnic homelands would suggest that not only arethe centroids of the groups correctly identified in the original maps butthat also the specific boundaries delineated are more precise than theThiessen-based ones. Effectively, this sensitivity check investigates how pre-cisely drawn the groups’ boundaries are in the underlying data sets.Table 6 reports the results of the “horse race” regressions, examining

the link between the log of per capita GDP and ethnic inequality, condi-tional on the perturbed ethnic homelands Gini index. Across all specifi-cations the ethnic Gini index enters with a negative and significant esti-mate that is quite similar ðaround 20.9Þ to the more parsimoniousspecifications in tables 2–4. In contrast, the Gini index based on the per-turbed ethnic areas ðThiessen polygonsÞ enters with an estimate unsta-ble and statistically indistinguishable from zero. It is perhaps instructiveto point out that the perturbed linguistic homelands of Ethnologue seemto have little predictive power on GDP per capita beyond the role of eth-nic inequality based on the Ethnologue homelands themselves, whereasfor the case of GREG the perturbed ethnic inequality index enters witha ðconsistentÞ negative sign and has a moderate magnitude. This patternis in line with the idea that the Ethnologue compared to GREG’s mappingmay have less measurement error since the former draws from a wealthof resources that are up to date and more precisely documented, unlikeGREG, which derives from maps of the 1960s.

D. Further Robustness Checks

We have performed numerous sensitivity checks to investigate the ro-bustness of the strong cross-country association between ethnic inequal-ity and underdevelopment. We report and discuss in detail these robust-

19 As explained in Sec. II, we are creating the modified groups’ homelands generatingtwo alternative sets of Thiessen polygons, one using as input points the centroids of thelinguistic homelands according to the Ethnologue data set and the other using the respec-tive centroids of the Atlas NarodovMira. Thiessen polygons have the exact same centroids asthe actual linguistic and ethnic homelands in the Ethnologue and GREG databases, respec-tively.


TABLE6

Eth

nic

Ineq

ual

ityan

dDev

elopm

entConditioningonPe

rturb

edEth

nic

Homelan

ds

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

Ethnic

ineq

ualityðG

iniÞ

2.951

0*2.956

6*2.865

2*2.927

6**

21.43

89***

21.43

65***

21.24

13***

2.806

3*ð.5

004Þ

ð.507

4Þ

ð.517

4Þ

ð.438

3Þ

ð.407

5Þ

ð.411

5Þ

ð.415

9Þð.4

166Þ

Perturbed

ethnic

ineq

ualityðG

iniÞ

2.537

12.545

72.822

32.264

5.330

2.337

6.025

92.184

5ð.5

258Þ

ð.526

2Þ

ð.546

9Þ

ð.482

9Þ

ð.464

0Þ

ð.471

8Þ

ð.495

0Þð.4

410Þ

Ethnic-lingu

isticfragmen

tation

.044

62.010

7.161

42.021

32.014

7.151

0ð.3

521Þ

ð.349

8Þ

ð.315

9Þ

ð.296

2Þ

ð.296

2Þð.2

656Þ

Adjusted

R2

.654

.652

.662

.721

.653

.650

.657

.718

Observations

173

173

173

173

173

173

173

173

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Simple

controls

No

No

Yes

Yes

No

No

Yes

Yes

Geo

grap

hic

controls

No

No

No

Yes

No

No

No

Yes

Note

.—Thetable


hedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.TheethnicGinico

efficien

tsreflectineq

ualityin

ligh

tsper

capitaacross

ethnic

homelan

ds,based

onthedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞinco

ls.1–

4an

donthe

Ethnologuein

cols.5

–8.

Theperturbed

ethnicineq

ualitymeasuresðG

inico

efficien

tsÞc

apture

thedeg

reeofspatialineq

ualityacrossThiessen

polygo

nsin

each

countrythat

use

asinputpoints

thecentroidsoftheethnic-lingu

istichomelan

dsacco

rdingto

theAtlas

Narodov

Miraðin

cols.1–

4Þan

dto

the

Ethnologueðin

cols.5

–8Þ.T

hiessen

polygo

nshavetheuniquepropertythat

each

polygo

nco

ntainsonlyoneinputpoint,an

dan

ylocationwithin

apolygo

niscloserto

itsassociated

pointthan

toapointofan

yother

polygo

n.In

specifications2–

4an

d6–

8,weco

ntrolforethnic-lingu

isticfragmen

tationusing

indicators

reflectingthelike

lihoodthat

tworandomlych

osenindividualsin

oneco

untrywillnotbemem


eethnic

fragmen

ta-

tionindex

inco

ls.2

–4co

mes

from

Alesinaet

al.[200

3]an

dthelingu

isticfragmen

tationindex

inco

ls.6

–8co

mes

from

Desmet

etal.[201

2]Þ.Sp

ecifica-

tions3,4,7,an

d8includeas

controlsloglandarea

andlogpopulationin

2000

ðsimplesetofco

ntrolsÞ.S

pecifications4an

d8includeas

controlsan

index

ofterrainrugg

edness,thepercentage

ofe

achco

untrywithfertilesoil,thepercentage

ofe

achco

untrywithtropicalclim

ate,averagedistance

tonearestice-

free

coast,an

dan

index

ofge

m-qualitydiamondex

tractionðgeo

grap

hic

setofco


fixe

deffectsðco

nstan

tsnot

reported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.Robust

stan

darderrors

arereported

inparen

theses

belowtheesti-

mates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.


ificantat

the1percentlevel.

This content downloaded from 129.199.207.113 on January 23, 2017 09:29:38 AMAll use subject to University of Chicago Press Terms and Conditions (http://www.journals.uchicago
.edu/t-and-c).


ness checks in the online supplementary appendix. Specifically, we showthat the results are similar when ðiÞ we do not include region fixed ef-fects; ðiiÞ we estimate ethnic Ginis without taking into account observa-tions either from capitals or from small groups; ðiiiÞ we drop from theestimation ðtypically smallÞ countries with just one ethnic or linguisticgroup; ðivÞ we use radiance-calibrated luminosity data to construct allinequality measures ðso as to account for top coding in the lights datathat occurs at the major urban centersÞ; ðvÞ we account for the resolu-tion of population estimates at the grid level that are used to compilethe inequality measures; ðviÞ we use inequality measures ðbased on lightsÞnonstandardized by population and control for inequality in the distribu-tion of population across ethnic areas; ðviiÞwe perform the analysis at var-ious nodes of Ethnologue’s linguistic tree ðthis approach follows Desmetet al. [2012], who show that the impact of ethnic fractionalization ongrowth, public goods, and conflict depends on the level of linguisticaggregationÞ; ðviiiÞ we try accounting for measurement error of the un-derlying mapping of groups estimating two-stage least-squares modelsthat extract the common component of ethnic inequality from bothEthnologue and the Atlas Narodov Mira; ðixÞ on top of the rich set of geo-graphic variables, we also condition on various historical controls; andðxÞ we drop iteratively from the estimation a different continent/regionand focus within each region separately. The regional analysis reveals thatthe development-ethnic inequality nexus is nonexistent for countries inwestern Europe and North America and weak in Latin America. On thecontrary, the association is especially strong within East and South Asiaas well as for countries in the Middle East and North Africa.

IV. On the Origins of Ethnic Inequality

Given the strong correlation between ethnic inequality and underdevel-opment, we have investigated the roots of inequality across ethnic lines.

A. Historical ðColonial Þ Origins
We started by examining the association between ethnic inequality andcommonly used historical correlates of contemporary development.There is little evidence linking contemporary differences in well-beingacross ethnic groups to the legal tradition ðLa Porta et al. 1998Þ; theconditions that European settlers faced at the time of colonization,as captured by settler mortality ðAcemoglu et al. 2001Þ or precolonialpopulation densities ðAcemoglu, Johnson, and Robinson 2002Þ; theshare of Europeans in the population ðHall and Jones 1999; Puttermanand Weil 2010Þ; and border design and state artificiality ðAlesina, East-erly, and Matuszeski 2011Þ. For brevity, we report these results in the


All

online supplementary appendix.20 These insignificant associations sug-gest that the strong negative correlation between ethnic inequality anddevelopment does not reflect the aforementioned aspects of history.

B. Geographic Origins

Motivated by the insight of Michalopoulos ð2012Þ that differences inland endowments gave rise to location-specific human capital, leadingto the formation of ethnolinguistic groups, we investigated whether dif-ferences in geographic and ecological attributes play a role in explainingcontemporary income disparities across ethnic lines. To the extent thatland endowments shape ethnic human capital and affect the diffusionand adoption of technology and innovation ðe.g., Diamond 1997Þ, ethnic-specific inequality in the distribution of geographic features would man-ifest itself in contemporary differences in well-being across groups.21

To construct proxies of geographic inequality, we obtained georefer-enced data on elevation, land suitability for agriculture, distance to thecoast, precipitation, and temperature and calculated for each ethnic areathe mean value.22 We then derived Gini coefficients at the country levelthat reflect group-specific inequality in each of these ðfiveÞ dimensions.We also estimated measures of the overall degree of inequality in geo-graphic endowments, constructing for each of the five geographic traitsspatial Gini coefficients across boxes of 2.5 � 2.5 decimal degrees andacross administrative units.Preliminary evidence.—In table 7 we explore the association between

ethnic inequality ðin lights per capitaÞ and these measures of inequalityin geographic endowments across ethnic homelands. Specifications 1and 5 simply condition on region fixed effects. To isolate the ethnic-specific component, in columns 2 and 6 we include in the empiricalmodel Gini coefficients capturing the overall degree of spatial inequalityacross each of these five traits, while in columns 3 and 7 we include Ginicoefficients of inequality in the same five geographic features acrossfirst-level administrative units. In specifications 4 and 8 we include ascontrols the country averages of each of the five variables. In almostall permutations, all five ethnic Ginis enter with positive estimates; thissuggests that ethnic-specific differences in geo-ecological endowments

20 There is also no association between ethnic inequality and proxies of the inclusivenessof early institutions ðAcemoglu et al. 2008Þ and state history ðBockstette, Chanda, andPutterman 2002Þ.

21 Language differences between groups are likely to exacerbate the limited mobilityacross ethnic homelands induced by the underlying differences in ethnic-specific humancapital.

22 In the previous draft of the paper, we also used information on the share of each ethnicarea covered by water bodies ðlakes, rivers, and other streamsÞ. The results are similar; we omitthis variable because luminosity getsmagnified over water areas because of bleeding-blooming.


http://www.journals.uchicago.edu/action/showLinks?crossref=10.1257%2F002205105774431243


TABLE7

Onth

eOriginsofConte

mpo

rary

Eth

nic

Ineq

ual

ity:

Ineq

ual

ityin

Geo

gra

phic

Endowmen

tsac

ross

Eth

nic

Homelan

dsan

dConte

mpo

rary

Eth

nic

Ineq

ual

ity

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

Lan

dqualityðG

iniÞ

.303

5**

2.007

72.091

8.184

.404

2***

.191

3.427

1**

.393

0***

ð.134

8Þ

ð.219

1Þ

ð.186

4Þð.1

414Þ

ð.113

2Þ

ð.178

4Þð.1

924Þ

ð.123

0Þ

Tem

perature

ðGiniÞ

1.65

129.19

567.37

843.08

7419

.560

0***

47.652

9***

39.702

4***

36.885

9***

ð7.246

2Þð11.16

86Þ

ð10.70

68Þ

ð8.050

7Þð6.860

8Þð12.37

94Þ

ð10.69

88Þ

ð10.11

17Þ

PrecipitationðG

iniÞ

.342

1*.784

5*.796

9**

.291

6.188

4.560

6.376

2.287

8ð.2

034Þ

ð.422

7Þ

ð.344

5Þð.2

259Þ

ð.263

6Þ

ð.486

2Þð.4

119Þ

ð.262

0Þ

Distance

totheco

astðG

iniÞ

.285

2**

.395

4**

.158

9.464

0***

.143

3.181

9.001

2.346

2**

ð.112

3Þ

ð.180

1Þ

ð.163

1Þð.1

458Þ

ð.138

8Þ

ð.165

9Þð.1

994Þ

ð.133

2Þ

ElevationðG

iniÞ

.500

2*.684

4**

.901

2***

.301

9.441

3***

.344

9.639

6***

2.035

6ð.2

674Þ

ð.325

5Þ

ð.306

4Þð.2

772Þ

ð.158

4Þ

ð.223

2Þð.2

013Þ

ð.160

1Þ

Adjusted

R2

.450

.467

.493

.491

.583

.611

.617

.667

Observations

164

164

164

164

164

164

164

164

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Controls

No

Spatial

Admin.unit

Levels

No

Spatial

Admin.unit

Levels

Note

.—Thetable

reportscross-countryOLSestimates,associatingco

ntemporary

ethnic

ineq

ualitywithineq

ualityin

geograp

hic

endowmen

tsacross

ethnic

homelan

ds.Thedep

enden

tvariab

leistheethnic

Ginico

efficien

tthat

reflects

ineq

ualityin

ligh

tsper

capitaacross

ethnic-lingu

istichomelan

ds,

usingthedigitized

versionofAtlas

Narodov

MiraðG

REGÞinco

ls.1–

4an

dEthnologuein

cols.5–

8.Toco

nstruct

theineq

ualitymeasuresin

geograp

hic

en-

dowmen

tsacross

ethnic

homelan

ds,wefirstestimatethedistance

from

thecentroid

ofeach

ethnic

homelan

dto

theclosestseacoast,averageelevation,

precipitation,tem

perature,a

ndlandqualityforagricu

lture

andthen

construct

Ginico

efficien

tscapturingineq

ualityacrossethnichomelan

dsin

each

of

thesege

ograp

hic

featuresforeach

country.In

cols.2an

d6,

weco

ntrolfortheoveralldeg

reeofspatialineq

ualityin

geograp

hic

endowmen

tsusingthe

Ginicoefficien

tofeach

ofthesefeaturesðdistance

totheclosestseacoast,elevation,p

recipitation,tem

perature,andlandqualityforagricu

ltureÞe

stim

ated

across2.5�2.5decim

aldeg

reeboxe

s/pixelsin

each

countryðboxe

sintersectedbynationalboundariesan

dtheco

astlineareofsmallersizeÞ.I

nco

ls.3

and

7,weco

ntrolfortheregionalineq

ualityacrossad

ministrativeunitsin

geograp

hicen

dowmen

tsusingtheGinicoefficien

tofeach

ofthesefeaturesðdistance

totheco

ast,elevation,precipitation,temperature,an

dlandqualityforagricu

ltureÞe

stim

ated

across

first-levelad

ministrativeunitsin

each

country.Col-

umns4an

d8includeas

controlsthemeanvalues

ðforeach

countryÞ

ofdistance

toseacoast,elevation,precipitation,temperature,an

dlandqualityfor

agricu

lture.Allspecificationsincluderegional

fixe

deffectsðco

nstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddata

sources.Robuststan

darderrors

arereported

inparen

theses

belowtheestimates.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.

***

Statisticallysign

ificantat

the1percentlevel.



All

translate into larger disparities in ethnic contemporary development.Depending on the specification details—GREG or Ethnologue mapping,whether we condition on the level of each geographical trait and re-gional inequality in each of the five geographic features—different Ginicoefficients of geographic inequality enter with significant estimates. Forexample, in the specifications using the GREGmapping, the Ginis captur-ing inequality in elevation and proximity to the coast enter with signifi-cant estimates, while in the Ethnologue -based models the Gini indicatorsreflecting inequality in land quality for agriculture and temperature arethe key correlates of ethnic inequality. Moreover, the controls capturinginequality across random pixels or administrative regions all enter withstatistically insignificant estimates ðcoefficients not shownÞ. Thus, whilewe cannot precisely identify which geographic featureðsÞ matters most,the message from table 7 is that differences in geography across ethnic re-gions translate into differences in contemporary ethnic inequality.A composite index of inequality in geographic endowments.—We thus aggre-

gate the five Gini indexes of ethnic inequality in geographic endow-ments via principal components. The use of factor analysis techniquesis appropriate in our context because we have many variables ðGinicoefficientsÞ that aim at capturing a similar concept ðwith some degreeof noiseÞ, namely, inequality in ethnic-specific geographic attributes.Moreover, we are not sure about which aspects of geographic inequalityshould matter the most. Table 8 reports the results of the principal com-ponent analysis. The first principal component explains approximately60 percent of the common variance of the five measures of inequalityin geographic endowments across ethnic homelands and close to 50 per-cent when we estimate Gini coefficients across pixels of ðroughlyÞ thesame size and across first-level administrative units. The second principalcomponent explains around 20 percent of the total variance, while jointlythe other principal components explain a bit less than a fourth of the to-tal variance. All five inequality measures load positively on the first prin-cipal component. This applies to all inequality measures ðacross ethnicand linguistic homelands, administrative regions, and boxesÞ. The eigen-value of the first principal component is greater than two in all permuta-tions ðone being the rule of thumbÞ, while the eigenvalues of the otherprincipal components are close to and less than one. We thus focus onthe first principal component, which, given the significant positive load-ings of all Gini coefficients, we label as “inequality in geographic endow-ments across ethnic homelands.”Inequality in geography across ethnic homelands and ethnic inequality.—In

figures 10A and 10B we plot the baseline index of ethnic inequality ðbasedon lights per capitaÞ against the first principal component of inequalityin ethnic-specific geographic endowments. There is a strong positive cor-relation for both mappings ðaround .55Þ, suggesting that differences in


TABLE8

Principa

lCompo

nen

tAnal

ysis:Ineq

ual

ityin

Geo

gra

phic

Endowmen

ts

Fact

orLoad

ings

Principa

lCompo

nen

tEigen

valu

eVar

iance

Explained

Var

iabl

e1stPrincipal

Componen

t2n

dPrincipal

Componen

t3rdPrincipal

Componen

t4thPrincipal

Componen

t5thPrincipal

Componen

t

A.GiniCoefficien

tGREG:AllGroups

1st

2.98

2.596

Ginilandquality

.460

2.133

2.572

.647

2.159

2nd

.844

.169

Ginitemperature

.482

2.358

.388

.097

.693

3rd

.746

.149

Giniprecipitation

.485

.025

2.463

2.741

.038

4th

.251

.050

Giniseadistance

.297

.918

.143

.145

.167

5th

.177

.035

Ginielevation

.482

2.106

.536

2.058

2.682

B.GiniCoefficien

tEthnologue:

AllGroups

1st

2.94

2.588

Ginilandquality

.432

2.423

.585

.496

2.215

2nd

1.06

2.213

Ginitemperature

.448

2.305

2.688

.319

.362

3rd

.551

.110

Giniprecipitation

.505

2.250

.183

2.781

.201

4th

.260

.052

Giniseadistance

.353

.710

.274

.196

.509

5th

.184

.037

Ginielevation

.484

.403

2.275

2.067

2.724

C.GiniCoefficien

t:Sp

atialIneq

ualityIndex

1st

2.71

1.542

Ginilandquality

.482

2.444

.259

.118

.700

2nd

1.14

9.230

Ginitemperature

.476

2.060

2.678

.519

2.203

3rd

.629

.126

Giniprecipitation

.497

2.350

.360

2.281

2.650

4th

.313

.063

Giniseadistance

.306

.677

.504

.439

2.042

5th

.199

.040

Ginielevation

.448

.468

2.301

2.667

.212

D.GiniCoefficien

t:AdministrativeUnitIneq

uality

1st

2.31

4.463

Ginilandquality

.484

2.414

.428

2.084

.637

2nd

1.27

0.254

Ginitemperature

.482

2.063

2.640

.586

.100

3rd

.771

.154

Giniprecipitation

.574

2.200

.253

2.089

2.748

4th

.403

.081

Giniseadistance

.181

.717

.492

.455

.059

5th

.242

.048

Ginielevation

.415

.520

2.318

2.659

.150

Note

.—Thetable

reportstheresultsoftheprincipal

componen

tan

alysisthat

isbased

onfive

measuresðG

inicoefficien

tsÞr

eflectingineq

ualityin

geo-

grap

hic

endowmen

tsin

distance

totheco



lture

across

ethnic-lingu

istichomelan

ds

ðpan

elsAan

dBÞ,pixelsof2.5�

2.5decim

aldeg

rees

ðinpan

elCÞ,an

dfirst-levelad

ministrativeregionsðin

pan

elDÞ.Column1reportstheeige

nvalueof

each

principal

componen

t,an

dco

l.2givesthepercentage

ofthetotalvarian

ceex

plained

byeach

principal

componen

t.Theother

columnsgive

the

factorload

ings

foreach

ofthefive

principal

componen

ts.



All

geography explain a sizable portion of contemporary differences in devel-opment across ethnic homelands.In table 9 we formally assess the role of ethnic-specific geographic in-

equality, as captured by the composite index of inequality in geographicendowments across ethnic-linguistic homelands, on contemporary ethnicinequality.23 Columns 1 and 5 show that the strong correlation illustratedin the figures is not driven by continentwide differences. In columns 2and 6 we control for the overall degree of spatial inequality in geographicendowments augmenting the specifications with the first principal compo-nent of theGini coefficients in geography using pixels of 2.5� 2.5 decimaldegrees. Likewise, in specifications 3 and 7 we add the first principal com-ponent of the geographic inequality measures across first-level administra-tive regions. This has little effect on the coefficient of the ethnic inequalityin geographic endowments, which retains its economic and statistical sig-nificance ðat the 99 percent levelÞ. Moreover, the two proxies of the overalldegree of spatial inequality in geography enter with small coefficients thatalso have the “wrong sign” and are not always statistically significant. In col-umns 4 and 8 we control for the level effects of geography, augmenting thespecification with the country average values of elevation, precipitation,temperature, distance to the coast, and land suitability for agriculture.The composite index reflecting differences in geographic endowmentsacross ethnic homelands continues entering with a positive and significantcoefficient. The estimate with the Ethnologuemapping ð0.12Þ implies that aone standard deviation increase in the inequality in geography across eth-nic homelands index ð1.74 points, from Zambia to EthiopiaÞ translatesinto a 20 percentage point increase in the ethnic inequality index ðexactlyas the difference in ethnic inequality between Zambia and Ethiopia, some-what more than half a standard deviation; see table 1Þ.In the online supplementary appendix, we show that the link between

ethnic inequality and inequality in geographic endowments across ethnichomelands prevails ðiÞ when we compile cross-country composite indica-tors of inequality in geographic endowments across ethnic lines using aricher set of geographic variables, ðiiÞwhenwe condition on contemporary

23 In this as well as in the subsequent tables, we also report bootstrap standard errors thataccount for the fact that the key independent variable—inequality in geographic endow-ments across ethnic homelands—is a “generated” regressor ðas it is a principal componentcapturing a geography factor; see Wooldridge 2002Þ. Our bootstrap method works as fol-lows. A random sample with replacement is generated from the full sample of countries. Inthis random sample, we extract the first principal component of the five Gini indicatorsthat capture inequality in geography across ethnic lines on elevation, precipitation, tem-perature, distance to coast, and land quality. We then use this principal component ðfromthe random sampleÞ in the regression ðwhere the dependent variable is ethnic inequalityÞ.This process is repeated 10,000 times. Table 9 gives the standard deviation of the coeffi-cient estimates across all ð10,000Þ replications ðfor a similar approach, see the recent studyby Ashraf and Galor [2013]Þ. As can be seen, bootstrap standard errors are very similar tostandard heteroskedasticity-adjusted ðWhiteÞ standard errors.


FIG. 10.—Panels A and B plot the baseline index of ethnic inequality ðbased on lightsper capitaÞ against the first principal component of inequality in ethnic-specific geographicendowments. Panels C and D plot the conditional on regional fixed effects association.


All


476


TABLE9

Onth

eOriginsofConte

mpo

rary

Eth

nic

Ineq

ual

ity:

Ineq

ual

ityin

Geo

gra

phic

Endowmen

tsac

ross

Eth

nic

Homelan

dsan

dConte

mpo

rary

Eth

nic

Ineq

ual

ity

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

Ineq

ualityin

geograp

hyacross

ethnic

homelan

dsðprincipal

componen

tsÞ

.090

2***

.105

3***

.107

7***

.089

8***

.121

0***

.145

8***

.159

5***

.118

1***

ð.010

1Þð.0

195Þ

ð.019

0Þ

ð.010

7Þð.0

124Þ

ð.019

8Þ

ð.018

6Þð.0

123Þ

[.01

02]

[.01

92]

[.01

96]

[.01

08]

[.01

18]

[.02

00]

[.01

92]

[.01

24]

Spatialineq

ualityin

geograp

hyðprincipal

componen

tsÞ

2.019

52.033

2*ð.0

201Þ

ð.019

5Þ

[.02

00]

[.01

984]

Ineq

ualityin

geograp

hyacross

administrativeunitsðprincipal

componen

tsÞ

2.025

12.056

9***

ð.021

0Þ

ð.018

5Þ[.02

14]

[.01

95]

Adjusted

R2

.452

.453

.456

.483

.587

.594

.608

.656

Observations

164

164

164

164

164

164

164

164

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Additional

controls

No

No

No

Levels

No

No

No

Levels

Note

.—Thetable


ntemporary

ethnicineq

ualitywithineq

ualityin

geograp

hicen

dowmen

tsacross

ethnic

homelan

ds.Thedep

enden

tvariab

leistheethnic

Ginico

efficien

tthat

reflectsineq

ualityin

ligh

tsper

capitaacrossethnic-lingu

istichomelan

ds,

usingthedigitized

versionofAtlas

Narodov

MiraðG

REGÞinco

ls.1–

4an

dEthnologuein

cols.5–

8.Theco

mposite

index

ofineq

ualityin

geograp

hic

en-

dowmen

tsisthefirstprincipal

componen

toffive

ineq

ualitymeasuresðG

inico

efficien

tsÞm

easuringineq

ualityacross

ethnic-lingu

istichomelan

dsin

distance

totheco



lture.Themap

pingofethnic

homelan

dsfollowsthedigitized

versionofGREG

inco

ls.1–

4an

dEthnologuein

cols.5–

8.Columns2an

d6includeaco

mposite

index

reflectingtheoveralldeg

reeofspatialineq

uality

inge

ograp

hicen

dowmen

ts.T

heco

mposite

index

aggreg

ates

ðviaprincipalco

mponen

tsÞG

inicoefficien

tsacross2.5�2.5decim

aldeg

reeboxe

s/pixels

ineach

countryðboxe

s/pixelsintersectedbynational

boundariesan

dtheco

astlineareofsm

allersizeÞo

fdistance

totheco

ast,elevation,p

recipitation,

temperature,an


lture.Columns3an

d7includeaco

mposite

index

reflectingregional

ineq

ualityin

geograp

hic

endowmen

tsacross

administrativeunits.Theco

mposite

index

aggreg

ates

ðviaprincipal

componen

tsÞG

inico

efficien

tsacross

first-levelad

ministrativeunitsin

dis-

tance

totheco



lture.Columns4an

d8includeas


ðfor

each

countryÞ

ofdistance

toseacoast,elevation,precipitation,temperature,an



fixe

deffectsðco

nstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.R

obuststan

darderrors

arereported

inparen

-theses

inthefirstrowbelowtheestimates.Intheseco

ndrowbelowtheestimates

wealso

reportin

bracketsbootstrap

stan

darderrorsthat

acco

untforthe

use

ofage

nerated

ðprincipal

componen

tÞregressor.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.

***

Statisticallysign

ificantat

the1percentlevel.

477



All

differences in development across space or administrative unions, andðiiiÞ when we iteratively drop different regions from the estimation.

V. Geographic Inequality and Development

A. Inequality in Geographic Endowments across Ethnic Homelands
and Economic Development
Given the strong positive association between ethnic inequality and in-equality in geographic endowments, it is interesting to examine whethercontemporary development is systematically linked to the unequal dis-tribution of geographic endowments across ethnic homelands. We thusestimated least-squares specifications associating the log of real GDPper capita in 2000 with the composite index of ethnic-specific inequalityin geography ðacross the five geographic dimensionsÞ. While omitted-variables concerns cannot be eliminated, examining the role of inequal-ity in geographic endowments across ethnic homelands on comparativedevelopment assuages concerns that the estimates in tables 2–4 are drivenby reverse causation. Moreover, geographic inequality can be thought ofas an alternative “primitive” measure of economic differences across lin-guistic homelands ðcompared to the ethnic inequality index based onluminosityÞ.Table 10 reports the results. The coefficient on the proxy of ethnic in-

equality in geographic endowments in columns 1 and 5 is negativeðaround 0.13Þ and significant at the 99 percent confidence level. Thissuggests that countries with sizable inequalities in geographic endow-ments across ethnic homelands are, on average, less developed. In col-umns 2 and 6 we condition on the overall degree of inequality in geog-raphy with the spatial Gini index based on boxes, while in columns 3and 7 we control for inequality in the geography across first-level admin-istrative units. This allows examining whether the negative associationbetween development and geographic disparities across ethnic home-lands—revealed in columns 1 and 5—captures the role of overall spatialgeographic inequalities, unrelated to ethnicity. The composite measurescapturing geographic inequalities across space and across administrativeregions enter with estimates statistically indistinguishable from zero ðwhichalso have the opposite signÞ. In contrast, the composite index capturing in-equality in geographic endowments across ethnic homelands retains its sta-tistical and economic significance. These results further show that it is in-equality across ethnic lines ðin geography in this caseÞ rather than acrossspace or administrative regions that correlates with underdevelopment.The same applies when we control for the mean values of the five geo-graphic variables ðin cols. 4 and 8Þ. Themost conservative estimate impliesthat a one standard deviation increase in geographic inequality across


TABLE10

Ineq

ual

ityin

Geo

gra

phic

Endowmen

tsac

ross

Eth

nic

Homelan

dsan

dConte

mpo

rary

Dev

elopm

ent

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

Ineq

ualityin

geograp

hyacross

ethnic

homelan

ds

ðprincipal

componen

tÞ2.134

7***

2.208

7***

2.161

0**

2.099

3***

2.129

3***

2.168

7**

2.132

4**

2.111

9***

ð.038

3Þ

ð.075

4Þð.0

695Þ

ð.036

7Þ

ð.043

2Þ

ð.070

8Þð.0

659Þ

ð.036

4Þ

[.03

87]

[.07

58]

[.07

09]

[.03

89]

[.04

32]

[.07

17]

[.06

75]

[.03

79]

Spatialineq

ualityin

geograp

hyðprincipal

componen

tÞ.095

4.052

7ð.0

854Þ

ð.073

7Þ[.08

55]

[.07

49]

Ineq

ualityin

geograp

hyacross

administrativeunits

ðprincipal

componen

tÞ.037

5.004

5ð.0

903Þ

ð.076

4Þ[.09

07]

[.07

76]

Adjusted

R2

.635

.636

.633

.691

.633

.632

.630

.696

Observations

164

164

164

164

164

164

164

164

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Additional

controls

No

No

No

Levels

No

No

No

Levels

Note

.—Thetablereportscross-countryOLSestimates,associatingco

ntemporary

developmen

twithineq

ualityin

geograp

hicen

dowmen

tsacrossethnic

homelan

ds.Thedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.Theco

mposite

index

ofineq

ualityin

geograp

hic

endowmen

tsisthefirst

principal

componen

toffive

ineq

ualitymeasuresðG

inico

efficien

tsÞm

easuringineq

ualityacross

ethnic-lingu

istichomelan

dsin

distance

totheco

ast,el-

evation,p

recipitation,tem

perature,andlandqualityforagricu

lture.T

hemap

pingofethnichomelan


versionofAtlas

Narodov

Mira

ðGREGÞinco

ls.1–

4an

dofEthnologuein

cols.5–

8.Columns2an

d6includeaco

mposite

index

reflectingtheoveralldeg

reeofspatialineq

ualityin

geo-

grap

hicen

dowmen

ts.T

heco

mposite

index

aggreg

ates

ðviaprincipal

componen

tsÞG

inico

efficien

tsacross2.5�

2.5decim

aldeg

reeboxe

s/pixelsin

each

countryðboxe


boundariesan

dtheco

astlineareofsm

allersizeÞo

fdistance

totheco

ast,elevation,precipitation,temper-

ature,an


lture.Columns3an

d7includeaco

mposite

index

reflectingregional

ineq

ualityin

geograp

hic

endowmen

tsacross

ad-

ministrativeunits.Theco

mposite

index

aggreg

ates

ðviaprincipal

componen

tsÞG

inico

efficien

tsacross

first-levelad

ministrativeunitsin

distance

tothe

coast,elevation,precipitation,temperature,an


lture.Columns4an

d8includeas


ðforeach

countryÞ

ofdistance

totheco

ast,elevation,p

recipitation,temperature,a

ndlandqualityforagricu

lture.A

llspecificationsincluderegional

fixe

deffectsðco

nstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.R

obuststan

darderrorsarereported

inparen

theses

inthefirstrow

belowtheestimates.In

theseco


wealso

report

inbracketsbootstrap

stan

darderrors

that

acco

untfortheuse

ofage

nerated

ðprincipal

componen

tÞregressor.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.


ificantat

the1percentlevel.



All

ethnic homelands ð1.7 pointsÞ decreases income per capita by approxi-mately 25 percent ð0.22 log pointsÞ.In the online supplementary appendix, we show that the negative asso-

ciation between the log income per capita and inequality in geographicendowments across ethnic lines is present also when ðiÞ we drop itera-tively a different region and ðiiÞ we control for contemporary differencesin spatial development or inequality in lights per capita across adminis-trative regions.

B. Geographic Inequality across Ethnic Homelands, Ethnic Inequality,and Economic Development

Given the strong negative correlation between development and ethnicinequality when the latter is proxied by differences in geographic endow-ments ðtable 10Þ or in disparities in luminosity per capita ðtables 2–6Þ, intable 11 we report specifications linking development to both measures.The results reveal that once we condition on contemporary ethnic in-come inequality, differences in geography across groups lose their powerin explaining cross-country variation in development. While some pecu-liar type of measurement error may explain this finding, it indicates thatethnic-specific inequality in geographic endowments relates to contem-porary development primarily via its influence on ethnic inequality.Since geographic inequality across ethnic lines does not seem to exert

an independent influence onGDP once we account for ethnic differencesin well-being, we also estimated two-stage least-squares ð2SLSÞ estimatesassociating geographic inequality across ethnic homelands to ethnic in-equality in lights per capita in the first stage and the component of eth-nic inequality explained by geographic disparities with the log per capitaGDP in 2000 in the second stage. While the 2SLS estimates do not iden-tify causal effects, they account for measurement error in the proxy mea-sure of development ðlights per capitaÞ and also isolate the geography-driven component of ethnic inequality. The 2SLS results ðreported inthe online supplementary appendix, table 23Þ reveal that the part of eth-nic inequality that reflects geographic differences across ethnic home-lands is a significant correlate of development.Remark.—These results should not be interpreted as showing that un-

equal geography across ethnic lines necessarily “causes” ethnic inequalityðand underdevelopmentÞ. It is possible that certain groups for a plethoraof reasons ðe.g., higher early development, superior military technologyÞconquered better-quality territories. In this regard the correlation be-tween inequality in geographic endowments across ethnic lines and eth-nic inequality ðcaptured by lights per capitaÞ indicates the sizable persis-tence of inequality. Hence, one might view an unequal ethnic geographyas amanifestation of deeper ethnic differences. Nevertheless, even in this


TABLE11

Ineq

ual

ityin

Geo

gra

phic

Endowmen

tsac

ross

Eth

nic

Homelan

ds,Eth

nic

Ineq

ual

ity,

andConte

mpo

rary

Dev

elopm

ent

Atl

asNar

odovMiraðG

REGÞ

Eth

nolo

gue

ð1Þ

ð2Þ

ð3Þ

ð4Þ

ð5Þ

ð6Þ

ð7Þ

ð8Þ

Ethnic

ineq

ualityðdevelopmen

tÞ21.29

84***

21.26

73***

21.29

59***

21.06

02***

21.16

03***

21.15

05***

21.22

58***

2.780

5**

ð.368

5Þð.3

601Þ

ð.365

0Þð.3

440Þ

ð.332

6Þð.3

320Þ

ð.339

1Þð.3

294Þ

[.37

32]

[.36

46]

[.37

30]

[.34

91]

[.33

42]

[.33

63]

[.34

44]

[.33

87]

Ineq

ualityin

geograp

hyacross

ethnic

homelan

dsðprincipal

componen

tÞ2.017

72.075

32.021

42.004

1.011

2.001

.063

22.019

7ð.0

495Þ

ð.071

9Þð.0

661Þ

ð.045

5Þð.0

560Þ

ð.075

2Þð.0

737Þ

ð.055

2Þ[.05

02]

[.07

35]

[.06

94]

[.04

70]

[.05

62]

[.07

69]

[.07

70]

[.05

63]

Spatialineq

ualityin

geograp

hyðprincipal

componen

tÞ.070

7.014

5ð.0

803Þ

ð.066

5Þ[.08

12]

[.06

88]

Ineq

ualityin

geograp

hyacrossad

mininstrative

unitsðprincipal

componen

tÞ.005

2.065

3ð.0

804Þ

ð.069

4Þ[.08

22]

[.07

19]

Adjusted

R2

.663

.662

.660

.707

.662

.660

.661

.705

Observations

164

164

164

164

164

164

164

164

Reg

ionfixe

deffects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Additional

controls

No

No

No

Levels

No

No

No

Levels

Note

.—Thetable


ntemporary

developmen

twithethnic

ineq

ualityan

dineq

ualityin

geograp

hic

endow-

men

tsacrossethnic

homelan

ds.Thedep

enden

tvariab

leisthelogofreal

GDPper

capitain

2000

.Theethnic

Ginico

efficien

tsreflectineq

ualityin

ligh

tsper

capitaacross

ethnic

homelan

ds,based

onthedigitized

versionoftheAtlas

Narodov

MiraðG

REGÞinco

ls.1–

4an

dbased

ontheEthnologuein

cols.5–

8.The

composite

index

ofineq

ualityin

geograp

hicen

dowmen

tsisthefirstprincipal

componen

toffive

ineq

ualitymeasuresðG

inico

efficien

tsÞm

easuringineq

uality

across

ethnic-lingu

istichomelan

dsin

distance

totheco



lture.Themap

pingofethnic

homelan


versionofGREGin

cols.1

–4an

dofEthnologuein

cols.5

–8.

Columns2an

d6includeaco

mposite

index

reflectingtheoverall

deg

reeofspatialineq

ualityin

geograp

hicen

dowmen

ts.T

heco

mposite

index

aggreg

ates

ðviaprincipalco

mponen

tsÞG

inicoefficien

tsacross2.5�

2.5decim

aldeg

reeboxe

s/pixelsin

each

countryðboxe


boundariesan

dtheco

astlineareofsm

allersizeÞindistance

totheco

ast,elevation,

precipitation,temperature,an


lture.Columns3an

d7includeaco

mposite

index

reflectingregional

ineq

ualityin

geograp

hic

endow-

men

tsacross

administrativeunits.Theco

mposite

index

aggreg

ates

ðviaprincipal

componen

tsÞG

inico

efficien

tsacross

first-levelad

ministrativeunitsin

dis-

tance

totheco



lture.Columns4an

d8includeas

controls

themeanvalues

ðforeach

countryÞ

ofdistance

totheco




fixe

deffects

ðconstan

tsnotreported

Þ.TheDataAppen

dix

givesdetailedvariab

ledefi

nitionsan

ddatasources.Robust

stan

darderrors

arereported

inparen

theses

inthefirstrowbelowtheestimates.Intheseco


wealso

reportin

bracketsbootstrap

stan

darderrors

that

acco

untfortheuse

ofage

n-

erated

ðprincipal

componen

tÞregressor.

*Statisticallysign

ificantat

the10

percentlevel.

**Statisticallysign

ificantat

the5percentlevel.

***

Statisticallysign

ificantat

the1percentlevel.



All

case, it is the presence of an inherently unequal geography that partiallyallows these primordial ethnic differences to become salient ðotherwisethere would be no “better land” for stronger groups to conquer and everygroup would have the same land endowmentÞ.

VI. Conclusion

This study shows that ethnic differences in economic performance ratherthan the degree of ethnic diversity or the overall level of inequality arenegatively correlated with economic development. While a large litera-ture has examined ðaÞ the interplay between inequality and developmentand ðbÞ the effects of various aspects of the ethnic composition ðsuch asfragmentation, polarization, segregationÞ on economic performance,there is little work studying the linkages between ethnicity, inequality,and cross-country comparative development. This paper is a first effortto fill this gap.First, combining linguistic maps on the spatial distribution of ethnic

groups within countries with satellite images of light density at night, weconstructGini coefficients reflecting inequality inwell-being across ethniclines for a large number of countries. Ethnic inequality is weakly corre-lated with the standard measures of income inequality and modestly cor-related with ethnolinguistic fractionalization, polarization, and segrega-tion. Second, we show that the newly constructed proxy of ethnic inequalityis negatively related to per capita GDP. The association retains its economicand statistical significance when we condition on inequality across admin-istrative units, which is also inversely related to development. Includingin the empirical specification both the ethnic inequality index and thewidely used ethnolinguistic fragmentation indicators, the latter loses sig-nificance, suggesting that it is inequality across ethnic lines that is corre-lated with poor economic performance rather than fractionalization perse. Third, we conduct a preliminary exploration of the roots of contempo-rary differences in well-being across ethnic groups. In this regard, we con-struct indicators of ethnic inequality in various geographic endowmentsand show that contemporary differences in development across ethnichomelands have a significant geographic component. Fourth, we showthat geographic inequality across ethnic lines is also inversely related tocontemporary development and that this correlation seems to operatevia ethnic inequality.Our study calls for future work on both the empirical and the theoret-

ical fronts. One could employ our cross-country data and approach to ex-amine the role of specific policies, such as trade openness and democra-tization, in shaping inequality across ethnic lines ðand even administrativeregionsÞ over time. Furthermore, building on the literature on institu-tions ðe.g., Acemoglu, Johnson, and Robinson 2005Þ and state formation



ðe.g., Besley and Persson 2011Þ, one could explore the role of initial—atindependence—differences in standards of living across ethnic groupson the subsequent path of economic and political modernization. Futureworks should also employ within-country approaches that are more suit-able for identifying the mechanisms at play. For instance, it is of great in-terest to understand the channels via which ethnic differences in incomeshapedevelopment.Does the linkoperate via theprovisionof public goods,via spurring conflict and animosity, or by shaping trust and beliefs? For ex-ample, in ongoing work ðAlesina et al. 2014Þ, we use a plethora of micro-level data fromAfrica to assess the role of between-groupandwithin–ethnicgroup inequality on public goods provision, trust, and civic and politicalparticipation within ðrather than acrossÞ countries. Moreover, given thelarge literature on inequality, fragmentation, and conflict, future workshould explore in detail the role of ethnic-level incomedifferences on con-flict ðas Mitra and Ray [2014] do in the case of Hindu-Muslim conflict inIndiaÞ. Another avenue of future research is to compile between-group in-equality measures over time using detailed data from censuses, surveys, ortax records that are available for some developed countries, in the spirit ofAtkinson, Piketty, and Saez ð2011Þ.

Appendix

Data

Country-Level Data

Income level: Log of real per capitaGDP at purchasing power parity ðchain indexÞin 2000. Source: PennWorld Tables, edition 7 ðHeston, Summers, andAten 2011Þ.

Population: Log population in 2000. Source: Penn World Tables, edition 7ðHeston et al. 2011Þ.

Land area: Log surface area. Source: Nunn and Puga ð2012Þ.Income inequality: Adjusted Gini coefficient index averaged over the period

1965–98. Source: Easterly ð2007Þ, based on the UN World Institute for Develop-ment Economics Research data.

Ethnic-linguistic fractionalization: Index of ethnic-linguistic heterogeneity. Itreflects the probability that two randomly selected individuals belong to differ-ent ethnolinguistic/religious groups. For completeness we use two measures,one from Alesina et al. ð2003Þ, which in turn is based on the CIA Factbook and En-cyclopaedia Britannica, and one from Desmet et al. ð2012Þ, which is based onEthnologue ðlevel 15Þ.

Ethnic-linguistic segregation: Index ranging from zero to one capturing ethnic-linguistic segregation ðclusteringÞ within countries. If each region is composed ofa separate group, then the index is equal to one; this is the case of complete seg-regation. If every region has the same fraction of each group as the country as awhole, the index is equal to zero; this is the case of no segregation. The index isincreasing in the square deviation of regional-level fractions of groups relativeto the national average. The index gives higher weight to the deviation of group



All

composition from the national average in bigger regions than in smaller regions.Source: Alesina and Zhuravskaya ð2011Þ.

Ethnolinguistic polarization: Index of ethnolinguistic polarization that achievesits maximum score when a country consists of two groups of equal size. Source:Montalvo and Reynal-Querol ð2005Þ.

Cultural fragmentation: Index of ethnolinguistic fractionalization that ac-counts for the degree of similarity between linguistic groups using the Ethnologuelinguistic tree. Source: Fearon ð2003Þ.

Genetic diversity: The expected heterozygosity ðgenetic diversityÞ of a country’scontemporary population. The index is based on distances fromEast Africa to theyear 1500 locations of the ancestral populations of the country’s component eth-nic groups in 2000 and on the pairwisemigratory distances among these ancestralpopulations. The source countries of the ancestral populations are identifiedfrom the World Migration Matrix ðPutterman and Weil 2010Þ, and the moderncapital cities of these countries are used to compute the aforementioned migra-tory distances. The measure of genetic diversity is then computed by applyingðiÞ the coefficients obtained from regressing expected heterozygosity on migra-tory distance from East Africa at the ethnic group level, using a worldwide sampleof 53 ethnic groups constituting the Human Genome Diversity Cell Line Panel,compiled by the Human Genome Diversity Project ðHGDPÞ and the Centred’Étude du Polymorphisme Humain ðCEPHÞ; ðiiÞ the coefficients obtained fromregressing pairwise genetic distance on pairwisemigratory distance in a sample of1,378 HGDP-CEPH ethnic group pairs; and ðiiiÞ the ancestry weights represent-ing the fractions of the year 2000 national population that can trace their ances-tral origins to different source countries in the year 1500. Source: Ashraf andGalor ð2013Þ.

Soil quality: Percentage of each country with fertile soil. Source: Nunn andPuga ð2012Þ.

Ruggedness: The terrain ruggedness index quantifies topographic heterogene-ity. The index is the average across all grid cells in the country not covered by water.The units for the terrain ruggedness index correspond to the units used to mea-sure elevation differences. Ruggedness is measured in hundreds of meters of ele-vation difference for grid points 30 arc-seconds ð926meters on the equator or anymeridianÞ apart. Source: Nunn and Puga ð2012Þ.

Tropical: The percentage of the land surface of each country with tropical cli-mate. Source: Nunn and Puga ð2012Þ.

Gem-quality diamond extraction: Carats of gem-quality diamond extractionbetween 1958 and 2000, normalized by land area. Source: Nunn and Puga ð2012Þ.

Common law: Indicator variable that identifies countries that have a commonlaw legal system. Source: La Porta et al. ð1999Þ and Nunn and Puga ð2012Þ.

European descent: The variable, calculated from version 1.1 of the migrationmatrix of Putterman and Weil ð2010Þ, estimates the percentage of the year 2000population in every country that is descended from people who resided in Eu-rope in 1500. Source: Nunn and Puga ð2012Þ.

Settler mortality: Log of mortality rates faced by European colonizers in thelate nineteenth century. Source: Acemoglu et al. ð2001Þ.

Population density before colonization: Log of population density around theyear 1500. Source: Acemoglu et al. ð2002Þ and Nunn and Puga ð2012Þ.



Border straightness index: The 0-1 index reflects how straight—and thus likelyto be nonorganic—national borders are. Source: Alesina et al. ð2011Þ.

Neolithic transition: The logarithm of the number of thousand years elapsedðas of the year 2000Þ since themajority of the population residingwithin a country’smodern national borders began practicing sedentary agriculture as the primarymode of subsistence. This measure, reported by Putterman ð2008Þ, is compiled us-ing a wide variety of both region- and country-specific archaeological studies as wellas more general encyclopedic works on the transition from hunting and gatheringto agriculture during the neolithic revolution. Source: Putterman andWeil ð2010Þand Ashraf and Galor ð2013Þ.

Ethnic partitioning: Percentage of the population of a country that belongs topartitioned ethnic groups. Source: Alesina et al. ð2011Þ.

Regional fixed effects: The region constants correspond to Southeast Asia andthe Pacific, Latin America and the Caribbean, North America, western Europe,eastern Europe and central Asia, the Middle East and northern Africa, and sub-Saharan Africa. The classification follows the World Bank’s World DevelopmentIndicators database.

Group-Level Data

Light density at night per capita: Light density is calculated by averaging luminos-ity observations across pixels that fall within each territory ðethnic-linguistic home-lands, boxes of 2.5 � 2.5 decimal degrees, administrative units, and ThiessenpolygonsÞ and then dividing by population density. Source: National Oceanic andAtmospheric Administration, National Geophysical Data Center ðhttp://ngdc.noaa.gov/eog/dmsp/downloadV4composites.htmlÞ.

Population density: Average number of people per square kilometer for 1990and 2000. Source: Center for International Earth Science Information Network,Columbia University, and Centro Internacional de Agricultura Tropical, 2005.Gridded Population of the World Version 3: Population Density Grids. Palisades,NY: Socioeconomic Data and Applications Center, Columbia University ðhttp://sedac.ciesin.columbia.edu/gpwÞ.

Area: Total area ðin square kilometersÞ of each territory ðethnic-linguistic home-lands, boxes of 2.5 � 2.5 decimal degrees, administrative units, and ThiessenpolygonsÞ.

Elevation: Average elevation above country minimum value in meters. Source:WorldClim—Global Climate Data. Data were originally collected by NASA-JPLSRTM. http://www.worldclim.org/current.

Land suitability for agriculture: Average land quality for cultivation within eachcountry. The index is the product of two components capturing the climatic andsoil suitability for farming. Source: Michalopoulos ð2012Þ; original source: Atlasof the Biosphere ðhttp://www.sage.wisc.edu/iamdata/grid_data_sel.phpÞ.

Distance to the seacoast: The geodesic distance from the centroid of each coun-try to the nearest coastline, measured in 1,000s of kilometers. Source: Global Map-ping International, Colorado Springs, Colorado. Series name: Global MinistryMapping System. Series issue: Version 3.0.

Average annual precipitation: Average annual precipitation ðmillimetersÞ for theapproximate 1950–2000 time frame within the respective territory ðethnic-linguistic



All

homelands, boxes of 2.5 � 2.5 decimal degrees, administrative units, and ThiessenpolygonsÞ. Source: WorldClim—Global Climate Data ðhttp://www.worldclim.org/bioclimÞ.

Average annual temperature: Average annual temperature for the approximate1950–2000 time frame within the respective territory ðethnic-linguistic home-lands, boxes of 2.5 � 2.5 decimal degrees, administrative units, and ThiessenpolygonsÞ. Source: WorldClim—Global Climate Data ðhttp://www.worldclim.org/bioclimÞ.

Precipitation seasonality: Coefficient of variation of annual precipitation forthe approximate 1950–2000 time frame within the respective territory ðethnic-linguistic homelands, boxes of 2.5 � 2.5 decimal degrees, administrative units,and Thiessen polygonsÞ. Source: WorldClim—Global Climate Data ðhttp://www.worldclim.org/bioclimÞ.

Temperature range: Range ðestimated as the difference of the maximum valueof the warmest month minus the minimum value of the coldest monthÞ of an-nual temperature for approximately the period 1950–2000 within the respectiveterritory ðethnic-linguistic homelands, boxes of 2.5 � 2.5 decimal degrees, ad-ministrative units, and Thiessen polygonsÞ. Source: WorldClim—Global ClimateData ðhttp://www.worldclim.org/bioclimÞ.

References

Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2001. “TheColonialOr-igins of Comparative Development: An Empirical Investigation.” A.E.R. 91 ð5Þ:1369–1401.

———. 2002. “Reversal of Fortune: Geography and Institutions in the Making ofthe Modern World Income Distribution.” Q.J.E. 107 ð4Þ: 1231–94.

———. 2005. “Institutions as a Fundamental Cause of Long-Run Growth.” InHandbook of Economic Growth, edited by Philippe Aghion and Steven N. Durlauf,109–39. Amsterdam: North-Holland.

Acemoglu, Daron, Simon Johnson, James A. Robinson, and Pierre Yared. 2008.“Income and Democracy.” A.E.R. 98 ð3Þ: 808–42.

Alesina, Alberto, Arnaud Devleeschauwer, William Easterly, Sergio Kurlat, andRomain Wacziarg. 2003. “Fractionalization.” J. Econ. Growth 8 ð2Þ: 155–94.

Alesina, Alberto, William Easterly, and Janina Matuszeski. 2011. “ArtificialStates.” J. European Econ. Assoc. 9 ð2Þ: 246–77.

Alesina, Alberto, Johann Harnoss, and Hillel Rapoport. 2016. “Birthplace Diver-sity and Economic Prosperity.” J. Econ. Growth, forthcoming.

Alesina, Alberto, and Eliana La Ferrara. 2005. “Ethnic Diversity and EconomicPerformance.” J. Econ. Literature 43:762–800.

Alesina, Alberto, Stelios Michalopoulos, and Elias Papaioannou. 2014. “Inequalityin Africa.” Manuscript, Brown Univ., Harvard Univ., and London Bus. School.

Alesina, Alberto, and Ekaterina Zhuravskaya. 2011. “Segregation and the Qualityof Government in a Cross-Section of Countries.” A.E.R. 101 ð5Þ: 1872–1911.

Anderson, Siwan. 2011. “Caste as an Impediment to Trade.” American Econ. J.:Appl. Econ. 3 ð1Þ: 239–63.

Ashraf, Quamrul, andOdedGalor. 2013. “TheOut of AfricaHypothesis, HumanGe-netic Diversity, and Comparative Economic Development.” A.E.R. 103 ð1Þ: 1–46.

Atkinson, Anthony B., Thomas Piketty, and Emmanuel Saez. 2011. “Top Incomesin the Long Run of History.” J. Econ. Perspectives 49 ð1Þ: 3–71.


http://www.journals.uchicago.edu/action/showLinks?crossref=10.1023%2FA%3A1024471506938


http://www.journals.uchicago.edu/action/showLinks?crossref=10.1111%2Fj.1542-4774.2010.01009.x

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1257%2Fapp.3.1.239

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1257%2Fapp.3.1.239



Baldwin, Kate, and John D. Huber. 2010. “Economic versus Cultural Differences:Forms of Ethnic Diversity and Public Goods Provision.” American Polit. Sci. Rev.104 ð4Þ: 644–62.

Benabou, Roland. 2005. “Inequality, Technology, and the Social Contract.” InHandbook of Economic Growth, vol. 1, edited by Philippe Aghion and StevenDurlauf, 1595–1638. Amsterdam: North-Holland.

Besley, Timothy, and Torsten Persson. 2011. Pillars of Prosperity: The Political Eco-nomics of Development Clusters. Princeton, NJ: Princeton Univ. Press.

Bockstette, Valerie, Areendam Chanda, and Louis Putterman. 2002. “States andMarkets: The Advantage of an Early Start.” J. Econ. Growth 7 ð4Þ: 347–69.

Cederman, Lars-Eric, Nils B. Weidman, and Kristian Skrede Gleditch. 2011.“Horizontal Inequalities and Ethnonationalist Civil War: A Global Compari-son.” American Polit. Sci. Rev. 105 ð1Þ: 478–95.

Chen, Xi, and William D. Nordhaus. 2011. “Using Luminosity Data as a Proxy forEconomic Statistics.” Proc. Nat. Acad. Sci. 108 ð21Þ: 8589–94.

Chua, Amy. 2003. World on Fire: How Exporting Free-Market Democracy Breeds EthnicHatred and Global Instability. London: Heinemann.

Deshpande, Aswini. 2000. “Does Caste Still Define Disparity? A Look at Inequal-ity in Kerala, India.” A.E.R. Papers and Proc. 90 ð2Þ: 322–25.

Desmet, Klaus, Ignacio Ortuño-Ortín, and Romain Wacziarg. 2012. “The Polit-ical Economy of Ethnolinguistic Cleavages.” J. Development Econ. 97 ð2Þ: 322–38.

Diamond, Jared. 1997. Guns, Germs, and Steel: The Fates of Human Societies. NewYork: Norton.

Easterly, William. 2007. “Inequality Does Cause Underdevelopment.” J. Develop-ment Econ. 84 ð2Þ: 755–76.

Esteban, Joan, Laura Mayoral, and Debraj Ray. 2012. “Ethnicity and Conflict: AnEmpirical Study.” A.E.R. 102 ð4Þ: 1310–42.

Esteban, Joan, and Debraj Ray. 2011. “Linking Conflict to Inequality and Polar-ization.” A.E.R. 101 ð4Þ: 1345–74.

Ezcurra, Roberto. 2013. “Does Economic Globalization Affect Regional Inequal-ity? A Cross-Country Analysis.” World Development 52:92–103.

Fearon, James D. 2003. “Ethnic Structure and Cultural Diversity by Country.”J. Econ. Growth 8:195–222.

Galor, Oded. 2011. “Inequality, Human Capital and the Process of Develop-ment.” In Handbook of the Economics of Education, edited by Stephen Machin,Eric A. Hanushek, and Ludger Woessmann. Amesterdam: Elsevier.

Gordon, Raymond G., Jr., ed. 2005. Ethnologue: Languages of the World. 15th ed.Dallas: SIL Internat.

Hall, Robert E., and Charles I. Jones. 1999. “Why Do Some Countries Produce SoMuch More Output per Worker than Others?” Q.J.E. 114 ð1Þ: 83–116.

Henderson, J. Vernon, Adam Storeygard, and David N. Weil. 2012. “MeasuringEconomic Growth from Outer Space.” A.E.R. 102 ð2Þ: 994–1028.

Heston, Alan, Robert Summers, and Bettina Aten. 2011. “Penn World Table Ver-sion 7.0.” Center Internat. Comparisons of Production, Income, and Prices,Univ. Pennsylvania.

Hodler, Roland, and Paul A. Raschky. 2014. “Regional Favoritism.”Q. J.E. 129 ð2Þ:995–1033.

Horowitz, DonaldL. 1985. Ethnic Groups in Conflict. Berkeley: Univ. California Press.Huber, John D., and Laura Mayoral. 2013. “Inequality, Ethnicity and Civil Con-

flict.” Working paper, Dept. Polit. Sci., Columbia Univ.Huber, John D., and Pavithra Suryanarayan. 2013. “Ethnic Inequality and the

Ethnification of Political Parties: Evidence from India.” Working paper, Dept.Polit. Sci., Columbia Univ.



http://www.journals.uchicago.edu/action/showLinks?crossref=10.1016%2Fj.jdeveco.2006.11.002




http://www.journals.uchicago.edu/action/showLinks?crossref=10.1257%2Faer.90.2.322

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1017%2FS0003055410000419




http://www.journals.uchicago.edu/action/showLinks?crossref=10.1073%2Fpnas.1017031108

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1093%2Fqje%2Fqju004

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1016%2Fj.worlddev.2013.07.002


All

Kanbur, Ravi, and Anthony J. Venables. 2008. “Spatial Inequality and Develop-ment—Overview of UNU-WIDER Project.” Manuscript, Cornell Univ.

Kim, Sukkoo. 2009. “Spatial Inequality and Economic Development: Theories,Facts, and Policies.” In Urbanization and Growth, edited by P. Annez, R. Buckley,andM. Spence, 1–39. Washington, DC: Internat. Bank Reconstruction and De-velopment and World Bank.

Kuznets, Simon. 1955. “Economic Growth and Income Inequality.” A.E.R. 45 ð1Þ:1–28.

Kyriacou, Andreas P. 2013. “Ethnic Group Inequalities andGovernance: Evidencefrom Developing Countries.” Kyklos 66 ð1Þ: 78–101.

La Porta, Rafael, Florencio Lopez-de-Silanes, Andrei Shleifer, and Robert Vishny.1998. “Law and Finance.” J.P.E. 106 ð6Þ: 1113–55.

———. 1999. “TheQuality of Government.” J. Law, Econ., andOrg. 15 ð1Þ: 222–79.Lessmann, Christian. 2014. “Spatial Inequality and Development—Is There an

Inverted-U Relationship?” J. Development Econ. 106 ð3Þ: 35–51.Loury, Glenn C. 2002. The Anatomy of Racial Inequality. Cambridge, MA: Harvard

Univ. Press.Michalopoulos, Stelios. 2012. “The Origins of Ethnolinguistic Diversity.” A.E.R.

102 ð4Þ: 1508–39.Michalopoulos, Stelios, and Elias Papaioannou. 2013. “Pre-colonial Ethnic Insti-

tutions and Contemporary African Development.” Econometrica 81 ð1Þ: 113–52.———. 2014. “National Institutions and Subnational Development in Africa.”

Q. J.E. 129 ð1Þ: 151–213.Mitra, Anirban, and Debraj Ray. 2014. “Implications of an Economic Theory of

Conflict: Hindu-Muslim Violence in India.” J.P.E. 122 ð4Þ: 719–65.Montalvo, Jose-Garcia, and Marta Reynal-Querol. 2005. “Ethnic Polarization, Po-

tential Conflict, and Civil Wars.” A.E.R. 95 ð3Þ: 796–816.Nunn, Nathan, and Diego Puga. 2012. “Ruggedness: The Blessing of Bad Geog-

raphy in Africa.” Rev. Econ. and Statis. 94 ð1Þ: 20–36.Pinkovskiy, Maksim. 2013. “Economic Discontinuities at Borders: Evidence from

Satellite Data on Lights at Night.” Manuscript, Massachusetts Inst. Tech.Pinkovskiy, Maksim, and Xavier Sala-i-Martin. 2014. “Lights, Camera, . . . Income!

Estimating Poverty Using National Accounts, Survey Means, and Lights.”Working Paper no. 19831, NBER, Cambridge, MA.

Putterman, Louis. 2008. “Agriculture, Diffusion and Development: Ripple Ef-fects of the Neolithic Revolution.” Economica 75 ð300Þ: 729–48.

Putterman, Louis, and David N. Weil. 2010. “Post-1500 Population Flows and theLong-Run Determinants of Economic Growth and Inequality.” Q.J.E. 125 ð4Þ:1627–82.

Robinson, James A. 2001. “Social Identity, Inequality, and Conflict.” Econ. Gover-nance 2 ð2Þ: 85–99.

Stewart, Frances. 2002. “Horizontal Inequalities: A Neglected Dimension of De-velopment.” Annual Lecture 5, World Inst. Development Econ. Res, UnitedNational Univ., Helsinki.

Tesei, Andrea. 2014. “Trust, Racial Fragmentation, and Income Inequality: NewEvidence from the US.” Manuscript, Queen Mary Coll., Univ. London.

Weidmann, Nils B., Jan Ketil Rod, and Lars-Erik Cederman. 2010. “RepresentingEthnic Groups in Space: A New Dataset.” J. Peace Res. 47 ð4Þ: 491–99.

Williamson, Jeffrey G. 1965. “Regional Inequality and the Process of National De-velopment: A Description of the Patterns.” Econ. Development and CulturalChange 13 ð4Þ: 1–84.

Wooldridge, Jeffrey M. 2002. Econometric Analysis of Cross Section and Panel Data.Cambridge, MA: MIT Press.


http://www.journals.uchicago.edu/action/showLinks?system=10.1086%2F450136




http://www.journals.uchicago.edu/action/showLinks?crossref=10.1111%2Fj.1468-0335.2007.00652.x

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1093%2Fjleo%2F15.1.222

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1162%2FREST_a_00161

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1162%2Fqjec.2010.125.4.1627

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1111%2Fkykl.12012



http://www.journals.uchicago.edu/action/showLinks?crossref=10.1093%2Fqje%2Fqjt029

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1007%2FPL00011021

http://www.journals.uchicago.edu/action/showLinks?crossref=10.1007%2FPL00011021

ethnic inequality - accueilpiketty.pse.ens.fr/files/alesinaetal2016.pdf · fort. consistent with...

Documents