NATIONAL INSTITUTIONS AND SUBNATIONAL DEVELOPMENT IN AFRICA

Abstract

We investigate the role of national institutions on subnational African development in a novel framework that accounts for both local geography and cultural-genetic traits. We exploit the fact that the political boundaries on the eve of African independence partitioned more than 200 ethnic groups across adjacent countries subjecting similar cultures, residing in homogeneous geographic areas, to different formal institutions. Using both a matching type and a spatial regression discontinuity approach we show that differences in countrywide institutional structures across the national border do not explain within-ethnicity differences in economic performance, as captured by satellite images of light density. The average noneffect of national institutions on ethnic development masks considerable heterogeneity partially driven by the diminishing role of national institutions in areas further from the capital cities.

I. Introduction

Few other issues have received more inquiry in the social sciences than “what are the fundamental determinants of comparative development?” The institutional view asserts that poorly performing institutional structures, such as lack of constraints on the executive, poor property rights protection, and inefficient legal and court systems are the ultimate causes of underdevelopment (see Acemoglu, Johnson, and Robinson 2005 for a review). Yet other works downplay the role of formal institutions, emphasizing instead the importance of geographical features, informal cultural norms, genetic, and epidemiological traits (see Spolaore and Wacziarg 2013 for a review).

This article investigates the role of national institutions in comparative development in Africa, developing a methodology that exploits in a “quasi-experimental” setting the artificial drawing of African borders that took place in the European capitals in the mid- to late nineteenth century, well before African independence and at a time when Europeans had hardly settled in the regions whose borders they were designing. The drawing of colonial boundaries partitioned on the eve of African independence more than 200 ethnicities across two (or more) countries. Taking advantage of this historical accident, we compare economic performance across adjacent regions belonging to the historical homeland of the same ethnic group but falling in different countries and hence are subject to different formal institutions. This identification strategy accounts both for differences in the natural environment (as the two neighboring regions have similar geography-ecology) and for ethnic-specific cultural and anthropological traits (since people from the same background reside on both sides of the border).

Our identification strategy entails two major challenges. First, we need to identify in a systematic way ethnic homelands partitioned by national borders. The second challenge derives from the sparsity of economic performance indicators at the local (country-ethnicity) level. On the first issue, we exploit anthropological data from the pioneering work of George Peter Murdock (1959), who has produced a map portraying the spatial distribution of African ethnicities in the mid- to late nineteenth century (Figure IA). Projecting Murdock’s map on contemporary national boundaries (Figure IB), allows us to identify in a systematic manner ethnic groups that have been split by the national borders. On the second issue, to overcome the paucity of economic data across African regions, we build on the recent contribution of Henderson, Storeygard, and Weil (2012) and measure development at the ethnicity-country level using satellite images of light density, which are available at a fine grid.

Figure I.

A Ethnic Boundaries

B Ethnic and Country Boundaries

Figures IIA–B illustrate our approach using as example the Ambo, an ethnic group partitioned between Angola (a country scoring quite low in most proxies of national institutions) in the north and Namibia (an institutionally advanced African country) in the south. In our empirical analysis we investigate whether differences in national institutions between adjacent countries explain differences in economic performance (as proxied by luminosity) within the historical homeland of the same ethnic group. We perform our analysis at two levels of aggregation: across partitioned ethnic areas in each country (Figure IIA) and across pixels falling in the homeland of split groups (Figure IIB). For the Ambo, luminosity is higher in Namibia, the more institutionally developed country.

Figure II.

A Example: Unit of Analysis I

B Example: Unit of Analysis II

C Pixel-Level Luminosity within the Anyi

D Pixel-Level Luminosity within the Kaonde

Yet this is not the case for the Kaonde and the Anyi, groups split by the Zambia–Democratic Republic of Congo border and the Ghana–Ivory Coast border, respectively. As Figures IIC and IID show, a larger faction of the respective ethnic homeland of both groups is lit in the Democratic Republic of Congo and in Ivory Coast, although national institutional quality is significantly higher on the other side of the border (in Zambia and Ghana, respectively).

Our analysis that spans all African partitioned ethnicities reveals new empirical regularities. First, we document that differences in national institutions across the border do not systematically translate into differences in economic performance within partitioned ethnicities. Although there is a significant positive correlation between national institutions and development across ethnic homelands, once we properly account for geographic-ecological and ethnic-specific differences via the inclusion of ethnicity fixed effects, the correlation weakens considerably and becomes statistically indistinguishable from zero. This pattern obtains both when the unit of analysis is a partition of an ethnic group (see Figure IIA) and when we take advantage of the finer structure of the luminosity data to obtain multiple observations within each partition (see Figures IIB, C, D); moreover, when we conduct a spatial regression discontinuity (RD) analysis that identifies the (local) effect of national institutions at the border we also find a close-to-zero effect. The lack of a systematic association between national institutions and regional development within partitioned African homelands cautions against extrapolating from the positive cross-country correlations.

Our results go against the conventional wisdom in economics on the causal effect of national institutions on development, in Africa at least; yet they are consistent with the African historiography that deemphasizes the importance of colonial and contemporary countrywide institutions in the hinterland (e.g., Davidson 1992; Herbst 2000). Moreover, our findings are in line with recent works that emphasize traits besides formal institutions, such as cultural norms, family ties, or important historical episodes (see Alesina and Giuliano 2013, Algan and Cahuc 2013, and Nunn 2013 for recent reviews).

Second, we show that the insignificant correlation between national institutions and ethnic development masks considerable heterogeneity. We find that the average effect of institutions is economically negligible and statistically insignificant for approximately 60% of partitioned ethnicities. Yet for some groups, consisting of approximately 20%–25% of the sample, a significant positive association emerges, whereas for the remaining ones the within-ethnicity association turns negative. Overall, the uncovered heterogeneity provides a useful reminder that generalizing from the findings of case studies focusing on a single border discontinuity can be quite misleading. Meanwhile, it suggests that case studies may be useful for shedding light on the specific circumstances that allow countrywide policies to shape regional development.

Third, building on insights from the African historiography that stress the inability of states to broadcast power in regions far from the capital (e.g., Bates 1983; Herbst 2000), we examine the spatial distribution of the uncovered heterogeneity. Contrary to the overall pattern, we find that national institutions do correlate with subnational development, but only when both partitions are close to the respective capital cities. We also present complementary evidence on the limited penetration of national institutions, expanding our analysis to the universe of African groups; in particular, we show that the explanatory power of national institutions on regional development decays for ethnic homelands (and pixels within groups) further from the capital centers.

Fourth, we complement our analysis on the differential role of national institutions far from the capital using individual-level data from the Afrobarometer Surveys on law enforcement perceptions and self-identification with the nation versus the ethnicity. Our within-country analysis shows that law enforcement weakens monotonically in areas further from the capitals. Moreover, national (as opposed to ethnic) identification is also inversely related to distance to the capital. These results thus further illustrate the limited penetration of country-wide institutions in the hinterland.

Our research nests and advances over several strands of the literature that examine the political economy of contemporary comparative development. First, an influential body of research asserts that through persistence, the institutions that European powers established during the colonial era are the deep roots of contemporary economic performance (e.g., La Porta et al. 1997, 1998; Acemoglu, Johnson, and Robinson 2001, 2002). Despite the implementation of ingenious instrumental variable approaches employed in the cross-country literature, omitted variables and estimate stability remain major concerns (see Glaeser et al. 2004; La Porta, Lopez-de-Silanes, and Shleifer 2008; Nunn 2012, among others). By exploiting within-country variation, our study circumvents some inherent limitations of the cross-country framework;¹ moreover, we move beyond average effects and uncover the interplay between national institutions, state presence, and proximity to the capital.

Second, our identification scheme that exploits border discontinuities in institutional arrangements relates to works that study the role of national policies across a particular border. In an early contribution Miguel (2004) compares public policies in health and education across the Kenya–Tanzania border to examine the effect of Tanzanian nation-building efforts. Bubb (2012) investigates how differences in de jure property rights between Ghana and Ivory Coast affect development in border areas, finding that in spite of large differences in formal institutions, there are no differences in the actual enforcement of property rights. Cogneau and Moradi (2011) examine differences in education and religiosity across the Ghana–Togo border and link them to different colonial investments; whereas Cogneau, Mespl-Somps, and Spielvogel (2012) examine access to electricity, consumption, and health at the borders of Ivory Coast with Ghana, Mali, and Burkina Faso, finding mixed evidence on the role of economic and institutional development. Finally, Miles (1994) studies the development of the Hausa after their partitioning (at the Niger–Nigeria border), documenting that differences between French and British policies (mainly regarding the role of local chiefs) endured after independence and had long-lasting effects.

Rather than focusing on the role of national features across a single border or within a single group, our study examines the role of national institutions on development across all partitioned ethnicities in Africa. Thus our results are less sensitive to the usual “external validity” arguments readily invoked in the context of case studies or works exploiting a single discontinuity. Moreover, our expansive focus sheds light on the underlying heterogeneity and uncovers the limited penetration of national institutions in areas far from the capital. The latter finding is related to parallel works of Campante and Do (2013) and Campante, Do, and Guimaraes (2013), who link capital city isolation with misgovernance and economic inefficiency.

Third, our results relate to the large literature in African political economy (see Bates 2008 and Dowden 2008 for eloquent narratives) that stresses features beyond national institutions; these include colonial investments in education, health and infrastructure (e.g., Huillery 2009; Jedwab and Moradi 2012), ethnic partitioning (e.g., Englebert, Tarango, and Carter 2002; Michalopoulos and Papaioannou 2013a), fractionalization (e.g., Easterly and Levine 1997; Alesina et al. 2003), inequality across ethnic lines (e.g., Alesina, Michalopoulos, and Papaioannou 2013), and geography-health (e.g., Alsan 2012; Nunn and Puga 2012).

Of most relevance from this body of research is our companion paper (Michalopoulos and Papaioannou 2013c), where we explore the long-lasting importance of ethnic (as opposed to national) political/institutional traits for contemporary African development. The main finding is that among numerous ethnic-specific features (related, for example, to the type of subsistence economy, occupational specialization, slavery), the legacy of pre-colonial political centralization appears to be a robust correlate of contemporary regional economic performance (Gennaioli and Rainer 2006, 2007 provide similar cross-country evidence). In contrast to these works that highlight the role of historical persistence at the ethnic level, in this article we examine the role of national institutions, which have been a dominant theme in the growth literature and in policy circles alike. In addition, our finding that national institutions wield significant explanatory power near the capitals, which rapidly diminishes for regions in the hinterland, highlights the importance of ethnic norms and reveals the coexistence of a dual institutional framework within African countries (Lewis 1954; Migdal 1988). Moreover, a methodological innovation of the present study is that we identify the role of national institutions within partitioned ethnic groups taking advantage of the arbitrary drawing of African borders and thus accounting for any (un)observable ethnic-specific differences across the diverse African tribal landscape. This has significant advantages over works exploiting cross-cultural variation and is motivated by the standard-textbook introduction on the role of national institutions showing the divergence in economic performance within culturally and geographically homogeneous entities, like North and South Korea or East and West Germany (see Weil 2008).

In this regard our article is related to subsequent work by Pinkovskiy (2013), who studies discontinuities in satellite-recorded light density at night across the globe and links them to country-level economic development and its main correlates (e.g., human capital, institutional capacity). In line with our findings, Pinkovskiy (2013) shows that borders of African nations tend to be relatively underdeveloped, with limited penetration of government activities and substantial economic activity concentrated around capital cities.² Moreover, our focus is on discontinuities within partitioned ethnic homelands; as such we effectively account for both geography-ecology and ethnic-specific features, related to history, social structure, and genetics. In fact, the uncovered weak and insignificant correlation between national institutions and subnational development within geographically and culturally homogeneous territories contributes to an emerging body of work that emphasizes the importance of cultural norms, historical persistence, and human and geographic traits for comparative development (see Diamond 1997; Landes 1998; Guiso, Sapienza, and Zingales 2006; Spolaore and Wacziarg 2009; Putterman and Weil 2010; Ashraf and Galor 2013; Michalopoulos, Naghavi, and Prarolo 2012, among others).

In the next section we describe how we identify partitioned groups and discuss the luminosity data that we use to proxy subnational development. In Section III we discuss the estimating framework and detail the identification design. In Section IV we report the baseline estimates on the average effect of national institutions on economic development within partitioned ethnic homelands. We also present spatial regression discontinuity (RD) estimates that quantify the effect of national institutions at the border. In Section V we estimate the role of national institutions separately for each split group mapping the overall heterogeneity. In an effort to understand the latter, in Section VI we explore how proximity to the capital shapes the explanatory power of national institutions on regional development. In Section VII we summarize discussing avenues for future research.

II. Data

II.A. Identifying Partitioned Ethnicities

Murdock’s map (Figure IA) portrays the spatial distribution of ethnicities across Africa in the mid- to late nineteenth century; it depicts 826 ethnic areas inhabited upon colonization (in total there are 834 polygons, but 8 regions are classified as “uninhabited”). Intersecting the pre-colonial map with the 2000 Digital Chart of the World (Figure IB) that portrays contemporary national boundaries, we obtain 1,247 country-ethnicity observations (e.g., the Ambo in Angola, the Ewe in Togo, and the Zulu in South Africa). We classify an ethnicity as partitioned when at least 10% of the historical homeland belongs to more than one contemporary state. We drop partitioned areas of less than 100 square kilometers as tiny partitions are most likely due to the lack of precision and projection error. Our procedure identifies 526 partitions that belong to the historical homeland of 227 ethnic groups. We report estimates for 220 partitioned ethnicities (507 country-ethnicity observations), because for six groups (the Ifora, Kunta, Berabish, Atta, Teda, and Asben) the local population estimates indicate zero population (in one of the two partitioned areas); moreover, we lose information from the Imragen—which are partitioned between Western Sahara and Mauritania—because we do not have information on national institutions for Western Sahara.

Online Appendix Table A lists the partitioned ethnicities. Examples include the Ewe that have been partitioned between Ghana and Togo (shares 44% and 56%, respectively), the Esa between Ethiopia and Somalia (shares 52% and 44%, respectively, while a small fraction falls in Djibouti), and the Yao (Wayao) between Mozambique (65%), Malawi (13%), and Tanzania (22%). Our procedure reveals that the median country in Africa has 43% of its population residing in the homeland of partitioned ethnicities. This estimate is similar to that of Asiwaju (1985) and Alesina, Easterly, and Matuszeski (2011), who use alternative sources and techniques to estimate that on average 40% of the African population belongs to partitioned ethnic groups. Thus our analysis, when focusing on partitioned homelands, captures a significant fraction of the African population (Figure IIIA).

Figure III.

A Partitioned Ethnic Homelands

B Luminosity across Partitioned Ethnic Homelands

Case study and anecdotal evidence suggest that in spite of population movements ethnic populations tend to reside in their respective historical homelands. Nunn and Wantchekon (2011) document that close to 55% of respondents in the Afrobarometer Surveys currently live in their ethnic group’s ancestral homeland. In the same vein, Glennerster, Miguel, and Rothenberg (2013) document in Sierra Leone that after the massive displacement of the 1991–2002 civil war there has been a systematic movement of individuals toward the areas of their ethnic group’s historical homeland.

II.B. Satellite Light Density at Night

The nature of our study requires data on economic development at the country-ethnic homeland level. Because there are limited geocoded measures of economic development in Africa, we build on the recent contribution of Henderson, Storeygard, and Weil (2012) and use satellite images on light density to proxy local economic activity.

Data come from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) that reports time-stable images of the Earth at night captured between 20:00 and 21:30. The measure ranges from 0 to 63 and is available for every 30-second area pixel (approximately 1 square kilometer). The satellite detects lights from human settlements, fires, gas flares, lightning, and the aurora. The annual composite measure is created by overlaying all daily images captured in a year, dropping images where lights are shrouded by cloud cover or overpowered by the aurora or solar glare (near the poles), and after removing ephemeral lights (like fires, lightning, and other noise).³ Using these data we construct light density per square kilometer for 2007 and 2008 averaging across pixels that fall within the historical homeland of each ethnic group in each country.

Besides its availability at a very fine geographic level, luminosity is well suited to spatial analyses of development in Africa for some additional reasons. First, most African countries have low-quality income statistics, both at the national and the subnational level (Jerven 2013). Second, we lack data on regional incomes for most African countries; although there are some local proxies of poverty and health, these do not map to our ethnic homeland-level unit of analysis.

Despite a series of works that establish a strong within-country correlation (both across time and across regions) between light density at night and GDP in the Online Appendix (Section 1) we provide further cross-validation checks of the luminosity data in our context.⁴ Specifically, using individual-level data from the Afrobarometer Surveys on access to electrification, presence of a sewage system, access to piped water, and education, we show that light density correlates strongly with these proxies of public goods and development both across Afrobarometer enumeration areas within countries and within ethnic homelands as well as across the border within partitioned ethnicities.

II.C. National Institutions

We measure national institutions using data from World Bank’s Governance Matters Database (Kaufmann, Kraay, and Mastruzzi 2008). The World Bank assembles numerous de facto institutional quality measures (originally compiled by various nongovernmental organizations and risk assessment agencies) and aggregates them into six categories via principal components aiming to minimize measurement error. We mainly use the rule of law index that reflects the effectiveness of the judiciary and the quality of property rights protection. Because many studies on African development focus on graft, we also report results using the control of corruption index. Both variables range theoretically between −2.5 and +2.5 with higher values indicating better functioning institutions and less corruption. To account for reverse causality and measurement error, we use predetermined values of the institutional quality measures, taking the mean over the period 1996–2006 (the results are robust to the choice of year). In our sample the countries with the lowest rule of law are Somalia (−2.195), Democratic Republic of Congo (−1.88), and Liberia (−1.73), whereas South Africa (0.156), Namibia (0.1675), and Botswana (0.615) are the most institutionally developed.

II.D. Data Patterns

There is significant variation both in national institutions and luminosity across African borders (Figure IIIB). Sharp border discontinuities in rule of law appear in several parts of Africa. For example, in the Botswana and Zimbabwe border (where the Hiechware, the Subia, and the Tlokwa are partitioned), across the Namibia and Angola border (where the Ambo are split), between Kenya and Somalia (where the Bararetta and other Somali tribes reside), and between Gabon and Congo (where the Duma live). Likewise, there are evident changes in luminosity across the border within the homeland of the same ethnic group. On one hand, despite differences in national institutions, in around 25% of the sample there are negligible differences in light density across the border. On the other hand, in about 40% there are more than one log point differences in luminosity. For example, sizable jumps in luminosity appear in the Egypt–Sudan border (where the Ababda and the Barabra are split), in the border between Ghana and the Ivory Coast (where the Assini reside), and between Benin and Togo (where the Popo are split).

III. Identification

III.A. Estimating Framework

1. Specification at the Country-Ethnic Homeland Level

Our analysis on the relationship between national institutions and regional development is based on variants of the following specification:

$$y_{i,c}=a_0+\mathit{\gamma IQ}{L}_c+{\mathrm{\lambda }}_{1}P{D}_{i,c}+{\mathrm{\lambda }}_2{\mathit{AREA}}_{i,c}+X_{i,c}^{\prime }\mathrm{\Phi }+a_i+{\epsilon }_{i,c}.$$ (1)

The dependent variable, y_i,c, reflects the level of economic activity in the historical homeland of ethnic group i in country c, as proxied by light density. Because a significant fraction (around 30%) of the (country-ethnicity) observations takes on the value of zero, we use as dependent variable the log of light density, adding a small number (y_{i, c} ≡ ln(0.01 + LightDensity_{i, c})).⁵ The logarithmic transformation is useful because we use all observations and because we account for some extreme values in luminosity (outliers).

2. Specification at the Pixel Level

We also exploit the fine structure of the luminosity data to obtain multiple observations (pixels) within a partitioned ethnic area in each country. At this level of aggregation we employ the following regression equation (variant of equation (1)):

$$\begin{array}{lll}{y}_{p,i,c}& =& a_0+\mathit{\gamma IQ}{L}_c+{\mathrm{\lambda }}_{1}P{D}_{p,i,c}+{\mathrm{\lambda }}_2{\mathit{AREA}}_{p,i,c}+X_{p,i,c}^{\prime }\mathrm{\Phi }\\ & & +a_i+{\zeta }_{p,i,c}.\end{array}$$ (2)

The dependent variable, y_p,i,c, is a dummy that takes on the value 1 if pixel p is lit and 0 otherwise. Each pixel, p, falls in the historical homeland of partitioned ethnicity i located in country c (see Figures IIB–D for examples). In our analysis we use pixels of 0.125 × 0.125 decimal degrees (approximately 12.5 km × 12.5 km). Because there are several unpopulated pixels (in the Sahara or in the rainforests) and to make sure that we examine the role of national institutions on development properly, we exclude pixels with zero population.⁶ This results in more than 120,000 observations across all African ethnic homelands, with 42,710 of those belonging to partitioned ethnicities.

Although conceptually the unit of our analysis is a partition of a group (Figure IIA), the pixel-level specifications (Figure IIB) are useful for several reasons. First, by looking at the extensive margin of light density, we account for the nonlinear nature of luminosity. Second, the linear probability models facilitate the interpretation of the estimates. Third, we are able to control for various relevant-for-development- characteristics at a fine level. Fourth, it is straightforward to augment the pixel-level specification with RD polynomials on distance to the national border, so as to estimate the effect of national institutions at the border. Fifth, we can estimate the pixel-level specification for each partitioned ethnicity separately and explore potential heterogeneity on the effect of institutions on development.

3. Independent Variables

IQL_c denotes institutional quality of country c, as reflected in the rule of law and the control of corruption measures. For split groups, each partition is assigned to the corresponding country c. For example, regional light density in the part of the Egba in Benin is matched to the institutional quality of Benin, whereas the adjacent area of the Egba in Nigeria is assigned the value of Nigeria.

In most specifications we control for log population density, PD_(p),i,c, because the association between luminosity and economic development strengthens and because by doing so the estimates capture the role of institutions beyond population density. We also control for the log of land area, AREA_(p)i,c; although most pixels are of the same size, those intersected by the coastline and national and ethnic boundaries are smaller (see Figures IIB, C, D).

Vector X_(p),i,c includes additional controls at the ethnic-homeland level (in equation (1)) or at the pixel level (in equation (2)).⁷ The Data Appendix provides variable definitions and data sources. Table I reports summary statistics in the sample of partitioned ethnicities (at the country-ethnic homeland level in Panel A and at the pixel level in Panel B).⁸

TABLE I.

Summary Statistics

Variable	Obs.	Mean	St. Dev.	p25	Median	p75	Min	Max
	Panel A: Partitioned ethnic homelands sample
Light density	507	0.228	0.768	0.000	0.015	0.095	0.000	8.561
Ln (0.01 + Light density)	507	−3.168	1.586	−4.605	−3.696	−2.258	−4.605	2.148
Ln (0.01 + Population density)	507	2.650	1.929	1.647	2.936	3.945	−5.069	6.386
Ln (1 + Water area)	507	0.242	0.407	0.000	0.075	0.264	0.000	2.464
Ln (Area)	507	1.876	1.490	0.798	1.906	2.963	−2.170	5.665
Mean elevation	507	0.605	0.434	0.281	0.469	0.938	0.000	2.181
Land suitability for agriculture	507	0.415	0.224	0.273	0.439	0.573	0.002	0.959
Malaria stability index	507	0.747	0.323	0.548	0.905	1.000	0.000	1.000
Oil deposit indicator	507	0.063	0.243	0.000	0.000	0.000	0.000	1.000
Diamond mine indicator	507	0.101	0.301	0.000	0.000	0.000	0.000	1.000
Distance to the capital city	507	0.524	0.371	0.280	0.429	0.685	0.013	1.882
Distance to the sea coast	507	0.591	0.425	0.225	0.552	0.931	0.000	1.739
Distance to the border	507	0.039	0.039	0.012	0.029	0.052	0.000	0.250
Rule of law	507	−0.914	0.573	−1.355	−1.021	−0.543	−2.198	0.615
Control of corruption	507	−0.797	0.501	−1.091	−0.864	−0.651	−1.664	0.814
	Panel B: Pixel-level, partitioned ethnic gomelands sample
Light density	42,710	0.312	2.335	0.000	0.000	0.000	0.000	62.857
Ln (0.01 + Light density)	42,710	−4.111	1.465	−4.605	−4.605	−4.605	−4.605	4.141
Light Density 0-1	42,710	0.124	0.329	0.000	0.000	0.000	0.000	1.000
Ln (0.01 + Population density)	42,710	2.074	1.771	0.734	2.143	3.328	−5.094	9.121
Water area indicator	42,710	0.129	0.335	0.000	0.000	0.000	0.000	1.000
Ln (area)	42,710	4.885	0.650	4.857	5.211	5.243	2.303	5.259
Mean elevation	42,566	641.612	470.408	292.111	526.048	969.556	−406.000	4623.750
Land suitability for agriculture	41,149	0.341	0.259	0.103	0.325	0.511	0.001	0.999
Malaria stability index	42,710	0.710	0.395	0.365	0.989	1.000	0.000	1.000
Oil deposit indicator	42,710	0.018	0.135	0.000	0.000	0.000	0.000	1.000
Diamond mine indicator	42,710	0.004	0.062	0.000	0.000	0.000	0.000	1.000
Distance to the capital city	42,710	0.532	0.369	0.271	0.439	0.707	0.001	1.933
Distance to the sea coast	42,710	581.363	414.433	233.430	507.473	896.488	0.286	1735.680
Distance to the border	42,710	65.938	68.091	15.754	44.255	93.799	0.000	448.539
Rule of law	42,710	−0.909	0.606	−1.355	−0.820	−0.506	−2.198	0.615
Control of corruption	42,710	−0.779	0.534	−1.141	−0.864	−0.651	−1.664	0.814

Notes. The table reports descriptive statistics for all variables employed in the empirical analysis for the sample of partitioned ethnic homelands in Africa. Panel A reports summary statistics at the country-ethnicity level. Panel B reports summary statistics at the pixel level. The Data Appendix gives detailed variable definitions and data sources.

4. Inference

In all specifications we employ the approach of Cameron, Gelbach, and Miller (2011) and cluster standard errors along the country and the ethnic-family dimension. Murdock assigns the 826 groups into 96 ethnolinguistic clusters. This double-clustering parameterization accounts for two main concerns related to nonadjusted standard errors. First, within each country we have several ethnic homelands (or pixels) where the country-level rule of law and the control-of-corruption measures take the same value and thus clustering at the country-level is required (Moulton 1990). Likewise, clustering at the ethnic family accounts for intra-ethnic-family correlation in (un)observed features. Second, the multiway clustering method allows for arbitrary residual correlation within both dimensions and thus accounts for spatial correlation; Cameron, Gelbach, and Miller (2011) explicitly cite spatial correlation as an application of the multi-clustering approach (see Nunn and Wantchekon 2011 and Michalopoulos and Papaioannou 2013c for applications of the multiway clustering method in a similar context). We also estimated standard errors accounting for spatial correlation of an unknown form using Conley’s (1999) generalized method of moments method. The two approaches yield similar standard errors.

III.B. Validity of Identification Design

For our identification strategy that compares economic development at the homelands of the same ethnicity in adjacent countries to be valid, one needs border drawing to not have been influenced by local circumstances and factors that shape or reflect economic well-being (see Angrist and Pischke 2008). Moreover, the ethnic areas across the border need to be similar across all relevant for development dimensions. Below we investigate these issues in detail.

First, all the anecdotal evidence suggests that colonizers drew African borders in an arbitrary manner (see Asiwaju 1985; Wesseling 1996; Herbst 2000; Englebert 2009 for eloquent reviews); the delineation of African borders took place in a period (late nineteenth century) when Europeans had not yet settled (most of) Africa and had limited knowledge of its political and economic geography; moreover, at that time colonizers were mostly assigning spheres of influence via the establishment of protectorates, colonies, and free-trade areas, let alone the borders of future African states. Consistent with the historical narratives on the artificiality of African borders, Alesina, Easterly, and Matuszeski (2011) document that close to 80% of African borders follow latitudinal or longitudinal lines, the highest percentage across all continents. Moreover, Michalopoulos and Papaioannou (2013a) show that besides landmass and water area, there are no systematic differences between nonsplit and partitioned ethnicities, across dozens of potentially relevant-for-development factors, related to geography, natural resources, the disease environment, and historical traits (capturing the intensity of pre-colonial conflict, political centralization, the type of subsistence economy, the intensity of the slave trades, etc.). Yet since some borders may have been drawn taking into account local conditions, in the Online Appendix we report estimates restricting attention to ethnicities partitioned by borders that follow relatively straight lines (as identified by the ‘box-count” method of Alesina, Easterly, and Matuszeski 2011). The results are virtually unchanged (see the associated discussion in the Online Appendix and Appendix Tables 9A and 9B).

Second, although research shows that ethnic groups tend to occupy geographically homogenous territories (Michalopoulos 2012), we investigated whether differences in institutional quality across the border correlate with differences in various characteristics. Table II, Panels A and B report ethnicity fixed effects estimates (“similarity” regressions) that associate geographic, ecological, and natural resource measures with the rule of law and the control of corruption index, respectively. In Panels C and D we focus on the two largest partitions of the 220 split groups (440 observations) and examine whether there are systematic differences between regions falling in the relatively high and the relatively low institutional quality country. The results are supportive of our identification strategy, as they point out that the two (or more) partitions of split groups are appropriate counterfactuals. Differences in geography-ecology, location, and natural resources across the border within partitioned ethnic homelands are small and not systematically linked to differences in national institutions.⁹

TABLE II.

Validity of Identification Design at the Country-Ethnicity Homeland Level

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
	Dependent variable is:
	Ln (land area)	Ln (area under water)	Mean elevation	Land suitability	Malaria stability	Oil indicator	Diamond indicator	Distance to the capital	Distance to the sea coast	Distance to the border
	Panel A: Rule of law
Rule of law	−0.0474	0.0480	0.0041	0.0091	−0.015	0.0365*	−0.0161	−0.1887	0.0001	−0.0037
Double-clustered std. err.	(0.2058)	(0.0453)	(0.0217)	(0.0184)	(0.0230)	(0.0199)	(0.0433)	(0.1233)	(0.0178)	(0.0049)
Adjusted R-squared	0.682	0.796	0.964	0.935	0.941	0.753	0.547	0.582	0.986	0.639
Observations	507	507	507	507	507	507	507	507	507	507
	Panel B: Control of corruption
Control of corruption	−0.0686	0.037	0.0119	−0.0052	−0.0324	0.0431	−0.0199	−0.1753	0.0037	−0.007
Double-clustered std. err.	(0.2229)	(0.0610)	(0.0222)	(0.0243)	(0.0233)	(0.0294)	(0.0612)	(0.1490)	(0.0176)	(0.0077)
Adjusted R-squared	0.683	0.795	0.964	0.935	0.941	0.753	0.547	0.565	0.986	0.551
Observations	507	507	507	507	507	507	507	507	507	507
	Panel C: High rule of law—two major partitioned ethnic areas
High rule of law	0.0696	0.0475	−0.0013	0.0129	−0.0103	0.0409	−0.0273	−0.1584	0.0036	0.0031
Double-clustered std. err.	(0.1111)	(0.0346)	(0.0171)	(0.0132)	(0.0195)	(0.0271)	(0.0414)	(0.1029)	(0.0111)	(0.0049)
Adjusted R-squared	0.818	0.866	0.972	0.952	0.953	0.804	0.629	0.612	0.991	0.639
Observations	440	440	440	440	440	440	440	440	440
	Panel D: High control of corruption—two major partitioned ethnic areas
High control of corruption	0.0262	0.0547	−0.0074	0.0105	−0.0025	0.0318	−0.0091	−0.0304	−0.0032	−0.0002
Double-clustered std. err.	(0.1259)	(0.0351)	(0.0185)	(0.0160)	(0.0217)	(0.0262)	(0.0455)	(0.1182)	(0.0103)	(0.0058)
Adjusted R-squared	0.817	0.867	0.972	0.952	0.953	0.802	0.627	0.569	0.991	0.638
Observations	440	440	440	440	440	440	440	440	440	440
Ethnicities	220	220	220	220	220	220	220	220	220	220
Ethnicity Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports within-ethnicity OLS estimates associating various geographical, ecological, and other characteristics with contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index (in Panel A) and control of corruption index (in Panel B) in areas of partitioned ethnicities. In Panels C and D we focus on the two major partitions of an ethnic group (in total 440 ethnic areas of 220 partitioned ethnicities). The high institutional quality indicator takes on the value 1 for the areas falling in the country with the higher value in the rule of law index (in Panel C) or in the control of corruption index (in Panel D). The dependent variable in column (1) is the log (land area); in column (2) is the log (1 + land area under water [lakes, rivers, and other streams]); in column (3) is mean elevation; in column (4) is an index of land suitability for agriculture; in column (5) is the average of a malaria stability index; in column (6) is an indicator for country-ethnic areas with an oil field; in column (7) is an indicator for country-ethnic areas with a diamond mine. In columns (8)–(10) the dependent variable is the distance of the centroid of each country-ethnic area from the capital city, the sea coast, and the national border, respectively. The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and the ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

IV. Baseline Results

IV.A. Preliminary Evidence: Cross-Sectional Analysis

Before we report the baseline ethnicity fixed effects estimates, we examine the cross-sectional patterns running simple least squares (LS) specifications that associate regional development with national institutions. Due to reverse causation and omitted variables, these specifications do not identify causal effects, yet they are useful for understanding the raw data patterns. Table III, Panel A reports estimates at the countryethnicity level (1,209 observations), and Panel B reports similar estimates at the pixel level (120,501 observations). Specification (1) shows that there is a positive and significant correlation between the rule of law index and regional development. In column (2) we control for log population density and log land area, whereas in column (3) we control for distance of each ethnic area (in Panel A) or each pixel (in Panel B) to the capital city, the national border, and the coast (“location controls”). Although all distance terms enter with significant coefficients, the estimate on the rule of law index retains its economic and statistical significance. In column (4) we augment the model with a rich set of geographic-ecological and natural-resource controls. Conditioning on geography reduces the coefficient on national institutions (approximately by 35%) though the estimate retains its statistical significance.¹⁰ The results are similar when we use the control of corruption index (in columns (5)–(8)).

TABLE III.

Cross-Sectional Analysis National Institutions and Regional Development

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	Rule of law				Control of corruption
	Panel A: Country-ethnic homeland level
Institutional quality	0.8229***	0.8017***	0.7646***	0.6513***	1.0328***	1.0351***	0.9945***	0.8218***
Double-clustered std. err.	(0.2839)	(0.2451)	(0.2507)	(0.1962)	(0.3421)	(0.2517)	(0.2334)	(0.2158)
Adjusted R-squared	0.080	0.314	0.379	0.468	0.0950	0.3340	0.3950	0.475
Observations	1,209	1,209	1,209	1,209	1,209	1,209	1,209	1,209
	Panel B: Pixel level
Institutional quality	0.1248***	0.1199***	0.0988***	0.0765**	0.1470***	0.1467***	0.1223***	0.0968**
Double-clustered std. err.	(0.0443)	(0.0415)	(0.0380)	(0.0334)	(0.0478)	(0.0431)	(0.0404)	(0.0409)
Adjusted R-squared	0.049	0.175	0.192	0.256	0.056	0.185	0.198	0.260
Observations	120,501	120,501	120,501	116,412	120,501	120,501	120,501	116,412
Pop. dens. and area	No	Yes	Yes	Yes	No	Yes	Yes	Yes
Location controls	No	No	Yes	Yes	No	No	Yes	Yes
Geographic controls	No	No	No	Yes	No	No	No	Yes

Notes. The table reports OLS estimates associating regional development with contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index (in columns (1)–(4)) and control of corruption index (in columns (5)–(8)) at the country-ethnicity level (in Panel A) and at the pixel level (in Panel B). In Panel A the dependent variable is log (0.01 + light density at night from satellite) at the ethnicity-country level. In Panel B the dependent variable is an indicator that takes on the value of 1 if the pixel (of 0.125 × 0.125 decimal degrees) is lit and 0 otherwise. In the specifications in columns (2)–(4) and (6)–(8) we control for log (population density in 2000) and log land area at the country-ethnicity level (in Panel A) or at the pixel-level (in Panel B). In columns (3), (4), (7), and (8) we control for location characteristics augmenting the specification with distance of the centroid of each ethnicity-country area or each pixel from the capital city of each country, distance from the closest sea coast, and distance from the national border. The set of geographic controls in columns (4) and (8) includes log (1 + area under water [lakes, rivers, and other streams]) for the models in Panel A and an indicator that takes on the value 1 if there is some water body in the pixel for the models in Panel B, land suitability for agriculture, elevation, a malaria stability index, a diamond mine indicator, and an oil field indicator. The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and at the ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

The coefficient in column (4) of Panel A implies that a 1-point-increase in the rule of law index (roughly 2 standard deviations), which corresponds to moving approximately from the institutional quality of Angola to that of Gabon, is associated with a 0.70 log point increase in regional luminosity (approximately half a standard deviation) and an 8% higher likelihood that a pixel is lit.

The correlations in Table III echo the findings of cross-country works; although the association between institutional quality and development weakens when one accounts for geography, it remains significant. Yet the fact that a small set of observable geographic features explains a moderate fraction of the cross-sectional correlation between institutional quality and ethnic development suggests that other unobserved or hard-to-measure aspects of geography-ecology and culture may further weaken the association.

IV.B. Ethnicity-Fixed-Effects Estimates

Having established that regions across the national border that are part of the historical homeland of the same ethnicity are reasonable counterfactuals (in Table II) and having explored the cross-sectional patterns (in Table III), we now examine the correlation between national institutions and luminosity within partitioned groups. Table IV, Panel A gives the results at the country-ethnic homeland level (equation (1)), and Panel B reports the results at the pixel level (equation (2)). For comparability, in odd-numbered columns we report cross-sectional estimates, and in even-numbered specifications we add ethnicity fixed effects. The cross-sectional estimates in columns (1) and (5) suggest that border areas that belong to countries with higher institutional quality display higher levels of development. For example, the estimate of the linear probability model in column (1) implies that a 1-point increase in the rule of law index increases the likelihood that a populated pixel is lit by approximately 11%.

TABLE IV.

National Institutions and Regional Development Across and within Partitioned Ethnic Groups

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	Rule of law				Control of corruption
	Panel A: Country-ethnic homeland level
Institutional quality	0.6510***	0.1943	0.5150**	0.2159	0.7904***	0.2566	0.6019***	0.2675
Double-clustered std. err.	(0.1951)	(0.1898)	(0.2024)	(0.2135)	(0.2268)	(0.2197)	(0.2329)	(0.2439)
Adjusted R-squared	0.292	0.792	0.392	0.798	0.298	0.792	0.393	0.798
Within R-squared	__	0.061	__	0.067	__	0.062	__	0.067
Observations	507	507	507	507	507	507	507	507
	Panel B: Pixel level
Institutional quality	0.1072***	0.0246	0.0834**	0.0278	0.1371***	0.0370	0.1097***	0.0403
Double-clustered std. err.	(0.0400)	(0.0165)	(0.0324)	(0.0181)	(0.0464)	(0.0273)	(0.0415)	(0.0290)
Adjusted R-squared	0.149	0.331	0.202	0.340	0.160	0.331	0.209	0.340
Within R-squared	__	0.059	__	0.066	__	0.059	__	0.066
Observations	42,710	42,710	41,025	41,025	42,710	42,710	41,025	41,025
Ethnicity fixed effects	No	Yes	No	Yes	No	Yes	No	Yes
Pop. dens. and area	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Location controls	No	No	Yes	Yes	No	No	Yes	Yes
Geographic controls	No	No	Yes	Yes	No	No	Yes	Yes

Notes. The table reports cross-sectional and within-ethnicity OLS estimates associating regional development with contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index (in columns (1)–(4)) and control of corruption index (in columns (5)–(8)) in areas of partitioned ethnicities at the country-ethnicity level (in Panel A) and at the pixel level (in Panel B). In Panel A the dependent variable is log (0.01 + light density at night from satellite) at the ethnicity-country level. In Panel B the dependent variable is an indicator that takes the value of 1 if the pixel (of 0.125 × 0.125 decimal degrees) is lit and 0 otherwise. Odd-numbered columns report cross-sectional specifications. Even-numbered columns report within-ethnicity estimates, where we include a vector of ethnicity fixed effects (constants not reported). In all specifications we control for log (population density in 2000) and log land area at the country-ethnicity level (in Panel A) or at the pixel level (in Panel B). In columns (3), (4), (7), and (8) we control for location and geography, augmenting the specification with distance of the centroid of each ethnicity-country (in Panel A) or each pixel (in Panel B) from the capital city of each country, distance from the closest sea coast, and distance from the national border. The set of geographic controls includes log (1 + area under water [lakes, rivers, and other streams]) for the models in Panel A and an indicator that takes the value 1 if there is some water body in the pixel for the models in Panel B, land suitability for agriculture, elevation, a malaria stability index, a diamond mine indicator, and an oil field indicator. Besides the adjusted R-squared, the table also reports the within R-squared (defined as the difference of the overall R-squared minus the overall R-squared of a model simply with ethnicity fixed effects). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

When we solely exploit within-ethnicity variation, the coefficients on rule of law and control of corruption drop sizably and become statistically insignificant. This is a uniform pattern across all permutations. At the country-ethnic homeland level the estimates fall from 0.65 to 0.19 and from 0.79 to 0.26. The coefficient drop in the linear probability models is starker; the estimate drops from 0.11 to 0.025 in the case of rule of law and from 0.14 to 0.037 in the case of the control of corruption index.¹¹ We also investigated formally whether the estimates on national institutions are statistically different between the cross-sectional and the within-ethnicity models performing a seemingly unrelated regression (SUR) Hausman-like test; the latter suggests that the difference in the coefficients is statistically significant at standard confidence levels. The insignificance of the ethnicity fixed effects estimates is not driven by a decrease in the precision of the coefficients, because the standard errors remain largely unchanged (in Panel A) or even fall (in Panel B); the drop in the coefficient magnitude is substantial.

Figures IVA and B illustrate the lack of a systematic within-ethnicity correlation between light density and institutional quality at the national level. The insignificance is not driven by a few (influential) observations; the within R² of the unconditional model is tiny, less than .02, implying a negligible economic effect.

Figure IV.

A Rule of Law and Subnational Development within Partitioned Ethnic Homelands

B Control of Corruption and Subnational Development within Partitioned Ethnic Homelands

In Table V we restrict estimation to areas close to the national border focusing in columns (1)–(2) and (5)–(6) on areas within 50 km from each side of the border (total 100 km), whereas in columns (3), (4), (7), and (8) we focus on regions within 25 km of each side of the border (total 50 km). The idea of this approach (which is similar in spirit to local regressions in regression discontinuity designs) is that by focusing on areas very close to the border, we further neutralize the role of unobservable factors. Although in the cross-section there is a strong positive association between national institutions and regional development, the within-ethnicity coefficients of rule of law and control of corruption are small and statistically indistinguishable from zero. This applies both at the country-ethnicity sample (in Panel A) and at the pixel level (in Panel B). The lack of significance is not driven by an increase in the standard errors; in all permutations 2-standard-error bands in the within-ethnicity estimates exclude the analogous cross-sectional ones. The cross-sectional estimates fall by more than two thirds once we add ethnicity fixed effects (from 0.68 to 0.083 in column (1) and from 0.86 to 0.09 in column (4)). Similarly, the linear probability estimates drop from 0.10–0.14 to 0.02, implying tiny economic effects. The statistical and economic insignificance of the coefficient is therefore casting doubt on the causal interpretation of the simple cross-sectional association between national institutional quality and subnational development.¹²

TABLE V.

National Institutions and Regional Development across and within Partitioned Ethnic Groups in Areas Close to the Border, Local Regressions

Bandwidth	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	Rule of law				Control of corruption
	100 km		50 km		100 km		50 km
	Panel A: Country-ethnic homeland level
Institutional quality	0.6788***	0.0827	0.5975***	0.1503	0.8585***	0.0866	0.7458***	0.1765
Double-clustered std. err.	(0.2116)	(0.2318)	(0.2098)	(0.2480)	(0.2465)	(0.2602)	(0.2591)	(0.2951)
Adjusted R-squared	0.311	0.842	0.285	0.848	0.324	0.842	0.293	0.848
Within R-squared	__	0.027	__	0.033	__	0.027	__	0.033
Observations	432	432	432	432	432	432	432	432
	Panel B: Pixel level
Institutional quality	0.0994***	0.0168	0.1031**	0.0159	0.1292***	0.0269	0.1346***	0.0197
Double-clustered std. err.	(0.0384)	(0.0181)	(0.0407)	(0.0163)	(0.0471)	(0.0259)	(0.0492)	(0.0210)
Adjusted R-squared	0.134	0.319	0.137	0.345	0.144	0.319	0.149	0.345
Within R-squared	__	0.051	__	0.046	__	0.051	__	0.046
Observations	21,289	21,289	13,408	13,408	21,289	21,289	13,408	13,408
Ethnicity fixed effects	No	Yes	No	Yes	No	Yes	No	Yes
Pop. dens. and area	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports cross-sectional and within-ethnicity OLS estimates associating regional development with contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index (in columns (1)–(4)) and control of corruption index (in columns (5)–(8)) in areas of partitioned ethnicities at the country-ethnicity level (in Panel A) and at the pixel level (in Panel B). In Panel A the dependent variable is log (0.01 + light density at night from satellite) at the ethnicity-country level. In Panel B the dependent variable is an indicator that takes the value of 1 if the pixel (of 0.125 × 0.125 decimal degrees) is lit and 0 otherwise. In columns (1), (2), (5), and (6) we focus on ethnic areas within 50 km of each side of the national border (total 100 km). In columns (3), (4), (7), and (8) we focus on ethnic areas within 25 km of each side of the national border (total 50 km). Odd-numbered columns report cross-sectional specifications. Even-numbered columns report within-ethnicity estimates, where we include a vector of ethnicity fixed effects (constants not reported). In all specifications we control for log (population density in 2000) and log land area at the country-ethnic homeland level (in Panel A) and in the pixel-level (in Panel B). Besides the adjusted R-squared, the table also reports the within R-squared (defined as the difference of the overall R-squared minus the overall R-squared of a model simply with ethnicity fixed effects). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

1. A Note on Ethnicity Fixed Effects

Across all model permutations in Tables IV and V, the ethnicity fixed effects are jointly significant at the 99% confidence level. Moreover, most of the ethnicity constants are individually highly significant. This suggests that African subnational development has a strong local geographic-ecological or/and ethnic-specific component. Because our focus is on isolating the effect of national institutions from local geographic and cultural traits, we do not examine in detail which aspects of the local environment the ethnicity fixed effects capture. Nevertheless, given the importance of this issue, in the Online Appendix (Section 3) we conduct a preliminary analysis using anthropological data from George Peter Murdock’s (1967) Ethnographic Atlas that records various quantitative measures of cultural, economic, and political traits for a large sample of African ethnicities. The cross-ethnicity specifications reveal that simply conditioning on the degree of pre-colonial political centralization (a key correlate of both economic and institutional development in Africa; see Gennaioli and Rainer 2007, 2006 and Michalopoulos and Papaioannou 2013c) and/or the type of the subsistence economy (dependence on agriculture and pastoralism, in particular) considerably weakens the cross-sectional coefficient on the national institutions proxy that becomes statistically indistinguishable from zero. These preliminary findings are in accord with recent work on the deep origins of African development that emphasize the key role of ethnic traits, as opposed to national features (see Michalopoulos and Papaioannou 2013b for a review of recent studies).

IV.C. Regression Discontinuity (RD) Estimates

We now report the results from the RD approach that identifies the (average) effect of national institutions at the border.

Our RD specification within the two major partitions of each ethnic group takes the following form:

$$\begin{array}{lll}{y}_{p,i,c}& =& a_0+\mathit{\gamma IQ}{L}_c^{\mathit{HIGH}}+f(B{D}_{p,i,c})+{\mathrm{\lambda }}_{1}P{D}_{p,i,c}\\ & & +{\mathrm{\lambda }}_2{\mathit{AREA}}_{p,i,c}+X_{p,i,c}^{\prime }\mathrm{\Phi }+a_i+{\zeta }_{p,i,c}.\end{array}$$ (3)

The difference from specification (2) is that since we aim at identifying the effect of national institutions on regional development at the border, we add RD-polynomials of the distance from the centroid of each pixel to the national border (f(BD_p,i,c)), allowing the coefficients on the polynomial terms to be different for each side of the boundary. $\mathit{IQ}{L}_c^{\mathit{HIGH}}$ is a dummy variable that takes the value 1 for pixels falling in the country with the relatively better institutions, as reflected either in the rule of law or in the control of corruption index.¹³ This RD-type design (see Imbens and Lemieux 2008; Lee and Lemieux 2010) exploits the discontinuity in the quality of national institutions at the border to identify institutions’ local average (treatment) effect.

Previous research has employed variants of regression equation (3) using different control functions of the running variable (distance to the border in our application) and limiting estimation close to the discontinuity using different bandwidths. Some researchers control for high-order polynomials in the forcing/running variable using all observations (both far and close to the discontinuity), because this approach maximizes the sample and is more efficient (e.g., Lee, Moretti, and Butler 2004). Others estimate local regressions limiting their analysis to an area close to the discontinuity either including or excluding the polynomial control function (e.g., Angrist and Lavy 2002; Dell 2010). For completeness, we employ all approaches and experiment with different bandwidths.

1. Baseline Results

Table VI, Panel A reports the RD results that identify the effect of national institutions on development exactly at the border. In columns (1) and (2) we use all pixels across the two major partitions of split-by-the-border ethnic homelands, controlling flexibly for the distance to the discontinuity using a cubic RD polynomial and a fourth-order RD polynomial, respectively. On average the likelihood that a pixel is lit is just 1 percentage point higher in the country with the relatively higher institutional quality. The coefficient on the dummy that switches to 1 for ethnic regions falling in the relatively less corrupt country is even smaller (in specifications (7)–(12)). Specifications (3)–(6) and (9)–(12) report local-linear regression results narrowing estimation to pixels within 50 or 25 km from each side of the border (total bandwidth 100 km and 50 km, respectively), while including the RD polynomial terms. The coefficient on the high institutional quality country is close to 0 (in some specifications is negative) and statistically insignificant. In Panel B, rather than including a common to all partitioned ethnicities RD polynomial, we include ethnic-specific polynomial terms on distance to the border (i.e., allowing for the effect of the running variable to be different for each group). The RD estimates on the national institutions measures are small and statistically indistinguishable from zero across all perturbations.

TABLE VI.

National Institutions and Regional Development at the Border Regression Discontinuity (RD) Estimates

Bandwidth RD polynomial	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)
	Rule of law						Control of corruption
	All pixels		100 km		50 km		All observations		100 km		50 km
	3rd-order	4th-order	3rd-order	4th-order	3rd-order	4th-order	3rd-order	4th-order	3rd-order	4th-order	3rd-order	4th-order
	Panel A: Global RD polynomial
Institutional quality	0.0166	0.0058	0.0123	0.0050	−0.0010	0.0122	0.0038	−0.0086	0.0116	0.0123	−0.0030	0.0079
Double-clustered std. err.	(0.0138)	(0.0135)	(0.0165)	(0.0185)	(0.0240)	(0.0272)	(0.0154)	(0.0137)	(0.0166)	(0.0176)	(0.0223)	(0.0276)
Adjusted R-squared	0.342	0.342	0.320	0.320	0.347	0.347	0.342	0.343	0.320	0.320	0.347	0.347
Within R-squared	0.059	0.059	0.037	0.037	0.064	0.064	0.059	0.060	0.037	0.037	0.064	0.064
	Panel B: Ethnic-specific RD polynomial
Institutional quality	0.0175	0.0203	0.0037	0.0135	0.0068	0.0185	0.0020	0.0114	0.0011	0.0094	0.0124	0.0128
Double-clustered std. err.	(0.0125)	(0.0154)	(0.0124)	(0.0183)	(0.0176)	(0.0318)	(0.0119)	(0.0138)	(0.0136)	(0.0181)	(0.0189)	(0.0283)
Adjusted R-squared	0.416	0.427	0.412	0.423	0.441	0.459	0.416	0.427	0.412	0.423	0.441	0.459
Within R-squared	0.133	0.144	0.129	0.140	0.158	0.176	0.133	0.144	0.129	0.140	0.158	0.176
Observations	40,209	40,209	21,289	21,289	13,408	13,408	40,209	40,209	21,289	21,289	13,408	13,408
Ethnicity fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Pixel area & pop. dens.	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports RD estimates. Estimation is performed across the two major partitions of split ethnic group. In Panel A we include a global (common to all partitioned ethnicities) RD polynomial in distance of the centroid of each pixel to the national border, allowing the polynomial terms to differ on the two sides of the border. In Panel B we include ethnic-specific RD polynomials in distance of the centroid of each pixel to the national border, allowing the polynomial terms to differ on the two sides of the border. In odd-numbered columns we include a third-order RD polynomial, and in even-numbered columns we include a fourth-order RD polynomial. In columns (3)–(6) and (9)–(12) we restrict estimation to pixels close to the national border using two different bandwidths. In columns (3), (4), (9), and (10) we focus on pixels within 50 km on each side of the national border (total 100 km). In columns (5), (6), (11), and (12) we focus on pixels within 25 km of each side of the national border (total 50 km). The high institutional quality indicator takes the value of 1 for pixels falling in the country with the higher value in the rule of law index (in columns (1)–(6)) or in the control of corruption index (in columns (7)–(12)). All specifications include ethnicity fixed effects (constants not reported), the log of pixel-level population density and the log of pixel surface area. Besides the adjusted R-squared, the table also reports the within R-squared (defined as the difference of the overall R-squared minus the overall R-squared of a model simply with ethnicity fixed effects). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and at the ethnolinguistic level dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

2. Graphical Illustration

Figures VA and VB provide a visual illustration of the RD results. The figures plot the average likelihood that a pixel is lit for bins of 5 km width, conditioning on log pixel area and log population density in areas close to the national border (the vertical line marks the national border). We report results using a 100 km and a 50 km bandwidth, respectively. The figures also plot predicted pixel-level luminosity from a regression that includes a third-order RD polynomial on distance to the border fitted separately for pixels in the partitions that fall in the relatively high institutional quality country (where distance takes on positive values) and pixels falling in the relatively low institutional quality country (where distance takes on negative values). In line with the RD estimates in Table VI, there are no discernible differences in luminosity crossing the national boundary towards the more institutionally developed country.

Figure V.

A National Institutions and Subnational Development at the Border: Regression Discontinuity Estimates A

B National Institutions and Subnational Development at the Border: Regression Discontinuity Estimates B

3. Large Discontinuities

In Table VII we report RD specifications (with a third- and fourth-order polynomial in distance to the national border) as well as simple ethnicity fixed effects estimates for the two major partitions of groups split between countries with large, defined as higher than the median or larger than the 75th percentile, differences in rule of law or in graft. By focusing on adjacent countries exhibiting substantial differences in national institutions (e.g., Botswana–Zimbabwe, Kenya–Somalia, Ghana–Ivory Coast), we minimize concerns that the in-significance of the estimates is due to measurement error and/or is driven by small differences in national institutions across the border. The estimated coefficients on the two measures of national institutions are quantitatively small (very close to zero) and statistically insignificant across all permutations.

TABLE VII.

National Institutions and Regional Development at the Border Regression Discontinuity (RD) Estimates, Large Discontinuities

Cut-off Global RD polynomial	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)
	Rule of Law						Control of Corruption
	median			75th percentile			median			75th percentile
	No	3rd-order	4th-order	No	3rd-order	4th-order	No	3rd-order	4th-order	No	3rd-order	4th-order
Institutional quality	0.0159	0.0116	−0.0027	0.0134	0.0049	−0.0207	0.009	0.0133	−0.003	0.041	0.0367	0.0024
	(0.0186)	(0.0199)	(0.0212)	(0.0311)	(0.0376)	(0.0330)	(0.0204)	(0.0215)	(0.0201)	(0.0346)	(0.0336)	(0.0251)
Adjusted R-squared	0.398	0.400	0.400	0.458	0.461	0.462	0.399	0.399	0.400	0.457	0.460	0.461
Within R-squared	0.072	0.074	0.075	0.075	0.078	0.080	0.064	0.065	0.065	0.078	0.080	0.082
Observations	19,349	19,349	19,349	10,031	10,031	10,031	19,814	19,814	19,814	10,353	10,353	10,353
Ethnicities	110	110	110	51	51	51	108	108	108	46	46	46
Ethnicity fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Pixel area & pop. dens.	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports simple fixed effects as well as RD estimates focusing on large discontinuities in national institutions across the border. Estimation is performed across the two major partitions of each ethnic group. In columns (2), (5), (8), and (11) we include a third-order RD polynomial on distance to the national border. In columns (3), (6), (9), and (12) we include a fourth-order RD polynomial on distance to the national border. In columns (1)–(3) and columns (7)–(9) we use the median value of differences in the rule of law index across adjacent countries to identify large discontinuities. In columns (4)–(6) and columns (10)–(12) we use the 75th percentile of differences in the rule of law index across adjacent countries to identify large discontinuities. The high institutional quality indicator takes the value of 1 for pixels falling in the country with the higher value in the rule of law index (in columns (1)–(6)) or in the control of corruption index (in columns (7)–(12)). All specifications include ethnicity fixed effects (constants not reported), the log of pixel-level population density, and the log of pixel surface area. Besides the adjusted R-squared, the table also reports the within R-squared (defined as the difference of the overall R-squared minus the overall R-squared of a model simply with ethnicity fixed effects). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and at the ethnolinguistic level dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

IV.D. Sensitivity Checks

We have performed numerous perturbations to investigate the robustness of the lack of systematic association between national institutions and regional development within partitioned African ethnicities. We report and discuss in detail these robustness checks in the Online Appendix (in Section 4). Specifically, we show that the results are similar when:

1. we express luminosity in levels and estimate the specifications with either ordinary least squares (OLS) or the Poisson pseudo-maximum likelihood estimator that accounts for the nonlinear nature of the luminosity (see Silva and Tenreyro 2006);

2. we use the alternative institutional quality measures from the World Bank’s Governance Database;

3. we use the 2007 values of the rule of law or the control of corruption measures;

4. we restrict estimation to ethnic homelands partitioned by straight line–like borders;

5. we proxy ethnic development with log population density;

6. we account for population clustering near the border, temporary migration, and bleeding of the lights data by excluding areas close to the border;

7. we drop observations iteratively from northern, southern, central, western, and eastern Africa;

8. we replace ethnicity fixed effects with border fixed effects;

9. we experiment with a two-dimensional RD polynomial on latitude and longitude (see Dell 2010);

10. we perform our analysis in a smaller sample of 32 partitioned ethnicities using data from the 2005 Afrobarometer household surveys on public goods provision and education as the dependent variable.

V. Heterogeneity

Our estimates so far summarize the average effect of national institutions on development across all split ethnicities; conducting the analysis for the universe of split groups is useful as it accounts for idiosyncratic features across specific borders and ethnicities. Moreover, studying the link between national institutions and ethnic development across all partitioned homelands makes the estimates less prone to ‘external validity’ concerns that characterize case study evidence. Nevertheless, the richness of the quasinatural experiment of ethnic partitioning in Africa allows us to estimate the role of national institutions separately for each split group. Doing so will help reveal whether the weak correlation between national institutions and subnational development within groups reflects a generalized phenomenon or conceals considerable heterogeneity. Finding the latter will set the stage for exploring the factors that give rise to the observed pattern.

As a first step of moving beyond average effects, we repeated our estimation for groups split within each of the main African regions (using Nunn’s 2008 classification). The results (reproduced in Online Appendix Table 16) suggest that the relationship between national institutions and ethnic development varies across African regions. On average for 28 groups split within countries in southern Africa, those partitions belonging to the more institutionally developed nations are more economically developed. However, the opposite is true when we focus within North Africa (though the number of split groups is quite small). In eastern and western Africa, differences in national institutions have no bearing on the within-ethnicity economic performance, whereas in central Africa there is some evidence of a significantly positive association.

Perhaps more important, we estimate the correlation between national institutions and pixel-level development separately for each split group. Figure VIA reports the distribution of the unconditional estimates on the high rule of law binary index; this is similar to a simple test of means of the likelihood that a pixel is lit on the two sides of the border. Figure VIB plots the distribution of the estimate when we condition on log population density and log pixel area.¹⁴ (The results are similar when we include additional controls and/or RD polynomials.)

Figure VI.

A Heterogeneity A: National Institutions and Subnational Development within Each Partitioned Ethnic Homeland

B Heterogeneity B: National Institutions and Subnational Development within Each Partitioned Ethnic Homeland

A couple of interesting patterns emerge. First, both the average and the median value of the ethnic-specific estimates are 0; this applies to the unconditional and the conditional coefficients. Moreover, the mass of the distribution is centered on 0. Not only is the coefficient small, but for about 120 groups (62.5%) the absolute value of the t-statistic in the conditional estimates is less than 1.60, implying statistically insignificant patterns.

Second, better national institutions translate into a significantly higher regional development for approximately 38 cases (around 20% of the sample). The Ambo, depicted in Figure IIB, is one such example. Other examples where differences in rule of law translate into increased luminosity include the Thonga (split by the border of South Africa and Mozambique), the Seke (split between Gabon and Equatorial Guinea), and the Tukulor (split by the Senegal–Mauritania border).

Third, in approximately 30–34 cases the coefficient for the high institutional quality country is significant, but with the “wrong” sign; so in roughly 17% of the sample of partitioned groups, we obtain the counterintuitive pattern of within-ethnicity development being higher in the country with the relatively worse national institutions. For example, in the homeland of the Anyi in Ivory Coast (rule of law index −1.25) 60% of populated pixels (132) are lit, whereas in the more institutionally advanced Ghana (rule of law index −0.23), only 30% of all (89) populated pixels are lit (see Figure IIC).¹⁵

Overall, the uncovered heterogeneity provides a useful reminder that extrapolating from the findings of an analysis that focuses on a single border discontinuity can be quite misleading. At the same time it raises the question about the determinants of the observed variation.

VI. Penetration of National Institutions

In this section we link the heterogeneous relationship between national institutions and subnational development with the limited reach of African governments in areas far from the capital. First, we discuss the argument of the African historiography on the limited presence of the colonial and national governments in the hinterland. Second, we estimate the baseline empirical model (equation (2)) linking national institutions to development separately for pixels close to and far from the respective capitals. Third, we explore the interaction between capital distance and national institutions in shaping regional development for all African ethnic homelands. Fourth, we test the conjecture regarding the limited reach of the state in remote areas using direct proxies of law enforcement at the regional level.

VI.A. The Argument

Europeans’ presence in Africa was (with some exceptions) limited to the coastline and the capital cities (see Herbst 2000 for a summary). Hence, colonial institutional arrangements, reflected through persistence on today’s national institutions, would have limited reach far from the capital cities. Similarly, several African scholars have argued that due to the lack of infrastructure (roads, transportation systems) and the unfavorable geography (desert areas, rugged terrains, rainforests), colonial and postcolonial institutions have minimal effect on the hinterland (Mamdani 1996). Herbst (2000), for example, reproduces a quote of a senior official in Central African Republic saying that “the State stops at PK 12, twelve kilometers from the capital, Bangui.” In his influential work Bates (1983) argues that African national leaders did not even try to extend the state’s power in the countryside, focusing instead on consolidating power in the capital and the few large urban centers.

Moreover, the scramble for Africa resulted in many countries having peculiar shapes, further inhibiting the presence of national governments in remote areas (Herbst 2000; Englebert 2009). For example, the Casamance region in southern Senegal (where the partitioned Diola-Jola reside) is isolated from Dakar, as Gambia effectively cuts Senegal into two parts. Likewise, the rainforest of central Africa limits the presence of the government of Democratic Republic of Congo in the eastern provinces of North and South Kivu, located hundreds of miles away from Kinshasa. The latter are quite often ruled by local militias and rebel groups.

Furthermore, Europeans mostly ruled from the capitals that were located close to the coast, as colonizers had little interest in settling the hinterland that was controlled with the assistance of local chiefs. (Herbst 2000, p. 16) notes that “rather systematically, Europeans created capitals that moved power toward the ocean and away from the interior centers of power that Africans had slowly created.” He lists many examples where colonizers decided to ignore local needs and established capital cities outside preexisting polities. As extreme examples he lists Mauritania and Bechuanaland (Botswana) that were ruled during colonization by capitals outside their nominal territories (Saint-Louis and Mafeking, respectively).¹⁶

VI.B. National Institutions and Proximity to the Capital: Evidence from Partitioned Ethnicities

1. Baseline Estimates

The key hypothesis that emerges from this narrative is that colonial and contemporary national institutions are less likely to be influential in areas far from the capital centers. Yet in the vicinity of the latter, national institutions are likely to be important. In this section we formally explore this conjecture. In Table VIII in odd-numbered columns we estimate the baseline ethnicity fixed effects specification for pixels of split groups that are close to the capitals; in even-numbered columns we examine the within-ethnicity association between national institutions and development across pixels that are far from the capital city.

TABLE VIII.

Heterogeneity: National Institutions and Regional Development within Partitioned Ethnicities Close to and Far from the Capital Cities

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	Rule of law		Control of corruption		Rule of law		Control of corruption
	Absolute distance to the capital				Relative distance to the capital
	Close	Far	Close	Far	Close	Far	Close	Far
	Panel A: All observations/pixels
Institutional quality	0.0675	0.0206	0.1220**	0.0330	0.1064*	0.0127	0.1455**	0.0218
Double-clustered std. err.	(0.0438)	(0.0233)	(0.0581)	(0.0258)	(0.0585)	(0.0225)	(0.0675)	(0.0237)
Adjusted R-squared	0.421	0.210	0.423	0.211	0.451	0.245	0.452	0.246
Within R-squared	0.089	0.062	0.091	0.063	0.084	0.046	0.085	0.047
Observations	12,546	15,225	12,546	15,225	10,612	19,385	10,612	19,385
Number of ethnicities	85	95	85	95	66	139	66	139
	Panel B: Local regressions (100 km bandwidth)
Institutional quality	0.0450	0.0040	0.1111*	0.0153	0.1140*	−0.0005	0.1564**	0.0112
Double-clustered std. err.	(0.0618)	(0.0239)	(0.0651)	(0.0269)	(0.0641)	(0.0235)	(0.0620)	(0.0258)
Adjusted R-squared	0.407	0.195	0.410	0.196	0.445	0.225	0.447	0.225
Within R-squared	0.085	0.044	0.087	0.044	0.090	0.036	0.092	0.036
Observations	6,129	8,120	6,129	8,120	4,655	10,987	4,655	10,987
Number of ethnicities	81	94	81	94	60	137	60	137
Ethnicity fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Pop. dens. and area	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Location and geography	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports within-ethnicity OLS estimates associating regional development with contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index (in columns (1), (2), (5), and (6)) and control of corruption index (in columns (3), (4), (7), and (8)) in areas of partitioned ethnicities. The unit of analysis is a pixel of 0.125 × 0.125 decimal degrees (around 12 × 12 km) within partitioned ethnicities. The dependent variable is a dummy variable that takes the value 1 if the pixel is lit and 0 otherwise. In odd-numbered columns we restrict estimation to pixels of partitioned ethnic areas that are close to the capital city. In even-numbered columns we restrict estimation to pixels of partitioned ethnic areas that are far from the capital city. In columns (1)–(4) we use as a cut-off the median value of absolute distance to the capital across all African pixels, while in columns (5)–(8) we use as a cut-off the median value of the within-country distance to the respective capital city. In Panel A we use all pixels of partitioned ethnicities. In Panel B we report local linear regressions restricting estimation to pixels of partitioned ethnicities within 50 km from each side of the national border (total 100 km). All specifications include a vector of ethnicity fixed effects (constants not reported). In all specifications we control for log (population density), log (land area), distance of the centroid of each pixel from the capital city of each country (in columns (5)–(8) we use relative distance to the capital city of each country), distance from the closest sea coast, distance from the national border, an indicator for pixels with a water body (lakes, rivers, and other streams), land suitability for agriculture, elevation, a malaria stability index, a diamond mine indicator, and an oil field indicator. Besides the adjusted R-squared, the table also reports the within R-squared (defined as the difference of the overall R-squared minus the overall R-squared of a model simply with ethnicity fixed effects). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

Panel A reports results using all pixels, and Panel B gives local linear regression estimates restricting estimation to pixels within 50 km from each side of the border. In columns (1)–(4) we use as a cutoff the median distance to the capital across all pixels (367.4 km). The estimates on the rule of law and the control of corruption measures in columns (1) and (3) are positive (0.07 and 0.12) and considerably larger than the baseline estimates in the full sample of split ethnicities (where the estimate was around 0.03 for both measures; see Table IV). The coefficient on the control of corruption index is significant at standard confidence levels, whereas the estimate on the rule of law index is weakly insignificant (p-value .15). In contrast, when we focus on pixels that are far from the respective capital city on both sides of the national border, the coefficients on the rule of law and the control of corruption index in columns (2) and (4) become much smaller (0.02–0.03); both estimates are statistically indistinguishable from 0.

In columns (5)–(8) we use the median relative distance to the capital within each country to distinguish between pixels that are relatively close and relatively far from the capital.¹⁷ The estimates in columns (5) and (7) on the two proxies of national institutions are positive (0.105 and 0.146) and statistically significant, suggesting that better-functioning national institutions translate into a higher development within partitioned ethnic homelands, but only when we look at pixels that are relatively close to the respective capital cities. In contrast, the small (0.01–0.02) and statistically insignificant estimates on national institutions in columns (6) and (8) suggest that within partitioned groups for pixels that are relatively far from the capital on both sides of the border, there seems to be no association between institutional quality and subnational development. The coefficient in column (7) implies that for pixels of partitioned ethnic homelands that are relatively close to the respective capitals, a 1 standard deviation increase in the control of corruption index (approximately 0.5) increases the probability that a pixel is lit by approximately 7% (more than half of the average likelihood that a pixel is lit; see Table I).

A couple of examples help illustrate these results. Both partitions of the Fon, a group split between Benin and Togo, are close to the corresponding capitals (on average pixels are 100 km away from Porto-Novo and Lome, respectively). Luminosity in the Fonland in Benin, where the rule of law index equals −0.43, is on average 53.8%, whereas luminosity is 24% in the less institutionally developed Togo, where the rule of law index equals −0.94. Another example is the Chaga, split by the Tanzanian–Kenyan border; all pixels on both sides of the border are close to Nairobi and Dodoma, respectively. Of populated pixels of the Chaga in Tanzania, 37.5% are lit (rule of law index equals −0.45) whereas luminosity at the Kenyan side of the border is significantly lower with only 22.8% of pixels lit (rule of law index in Kenya is −1.02). Another interesting case is the Ronga, a Shona group split between Mozambique (rule of law equals −0.745) and South Africa (rule of law index equals 0.156). Ronga’s homeland in Mozambique has 41.5% of its pixels lit. Incidentally the capital Maputo falls within its boundaries hence all pixels are classified as close. Yet in the more institutionally advanced South Africa, 64% of the pixels in the Ronga partition that are close to Pretoria are lit.

It is instructive to put the magnitudes of the coefficients in perspective and obtain a welfare assessment in terms of the affected population. Across African countries (as of 2000) there were roughly 200 million people that reside in the homelands of partitioned ethnicities. Of those 200 million individuals, 45% live relatively far from the respective capitals with the remaining 55% reside relatively close to the capitals, where national institutions seem to have a bearing on regional economic performance.

2. RD Visualization

Figures VIIA and VIIB provide a graphical illustration of the differential association between national institutions and regional development within partitioned groups for pixels that are relatively close (on both sides of the border) and pixels that are relatively far from the capital city (on both sides of the border), respectively.

Figure VII.

A Regression Discontinuity Graphs within Partitioned Ethnic Homelands Close to the Capital Cities

B Regression Discontinuity Graphs within Partitioned Ethnic Homelands Far from the Capital Cities

The figure plots the average likelihood that pixels within 5 km bins are lit, conditional on log population density and log land area; the figure also plots predicted luminosity using cubic RD polynomials on distance to the border. When we focus on pixels of partitioned ethnicities that are relatively close to the respective capitals on both sides of the border (Figure VIIA), pixel-level luminosity at the side of the border with the better-functioning national institutions is significantly higher. In contrast, when we zoom in on pixels that are relatively far from the respective capital cities (Figure VIIB), there are no differences in luminosity, in spite of differences in national institutions.

3. Robustness

In the Online Appendix (Section 5) we report several sensitivity checks showing that regional development is significantly higher in the part of split homelands falling in the country with better functioning institutions, but only for pixels that are (relatively) close the capital. Specifically, this pattern obtains when: (i) we use information from all split groups by pooling all pixels (40,209 observations) and interact the national institutions measures (as well as all controls) with a dummy variable indicating pixels that are close (or relatively close) to the respective capital; (ii) we drop iteratively each of the five main African regions; (iii) we proxy national institutions with the alternative composite measures of World Bank’s Governance Database; (iv) following Acemoglu and Johnson (2005) we proxy “property rights” institutions with Polity Project’s executive constraints measure or we proxy “contractual” institutions with the legal formalism indicators of Djankov et al. (2003); and (v) we use proxies of “early” political institutions (on African independence).

VI.C. Evidence from All Groups

In principle, the conjecture of the African historiography that the role of national institutions dissipates as one moves further from the capital regards the universe of ethnic homelands. In Table IX we augment the baseline model (equation (2)) with the (absolute or relative) distance of each pixel to the capital city and an interaction term between distance to the capital and national institutions.

TABLE IX.

National Institutions, Distance to the Capital, and Regional Development Pixel-Level Estimates in the Full Sample

	(1)	(2)	(3)	(4)	(5)	(6)
	Absolute distance to the capital			Relative distance to the capital
Rule of law	0.1434***	__	__	0.1790***	__	__
Double-clustered std. err.	(0.0426)			(0.0472)
Distance to the capital	−0.1352***	−0.2103***	−0.1097**	−0.3098***	−0.2530***	−0.1457***
Double-clustered std. err.	(0.0511)	(0.0404)	(0.0534)	(0.0846)	(0.0502)	(0.0499)
Rule of law ×	−0.1243***	−0.1513***	−0.0850**	−0.2297***	−0.1850***	−0.1042**
Distance to the capital	(0.0363)	(0.0315)	(0.0352)	(0.0624)	(0.0461)	(0.0439)
Adjusted R-squared	0.261	0.362	0.429	0.267	0.362	0.429
Observations	116412	116412	116412	116412	116412	116412
Pixel area and pop. dens.	Yes	Yes	Yes	Yes	Yes	Yes
Location controls	Yes	Yes	Yes	Yes	Yes	Yes
Geographic controls	Yes	Yes	Yes	Yes	Yes	Yes
Country fixed effects	No	Yes	No	No	Yes	No
Country-ethnicity FE	No	No	Yes	No	No	Yes

Notes. The table reports OLS estimates associating regional development with distance to the capital city and contemporary national institutions, as reflected in World Bank’s Governance Matters rule of law index across the full sample of country-ethnic homelands in Africa. The unit of analysis is pixels of 0.125 × 0.125 decimal degrees (around 12 × 12 km). The dependent variable is a dummy variable that takes the value 1 if the pixel is lit and 0 otherwise. Columns (1) and (5) give unconditional estimates. In columns (2)–(4) and (6)–(8) we control for log (population density), log (land area), distance of the centroid of each pixel from the capital city of each country, distance from the closest sea coast, distance from the national border, an indicator for pixels with a water body (lakes, rivers, and other streams), land suitability for agriculture, elevation, a malaria stability index, a diamond mine indicator, and an oil field indicator. The specifications in columns (3) and (7) include a vector of country fixed effects (constants not reported). The specifications in columns (4) and (8) include a vector of country-ethnicity fixed effects (constants not reported). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the country and ethno-linguistic family dimensions. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

In columns (1) and (4) we report cross-sectional specifications. Due to omitted variables, reverse causation, and other forms of endogeneity, these specifications do not identify casual relationships; yet these models are useful to quantify the correlation between institutional and economic development at the capital; the estimate in column (4) suggests that a 1 point increase in the rule of law index increases the likelihood that a pixel at the capital city is lit by approximately 18%; for comparability the analogous cross-sectional estimate in Table III was approximately half, 0.095. Moreover, the negative and significant coefficients on both absolute and relative distance to the capital indicate that subnational development declines as one moves further from the capital. Yet the most interesting finding is the significantly negative coefficient on the interaction term between distance to the capital and rule of law; this implies that the correlation between national institutions and regional development declines for regions further from the capital centers.

In columns (2) and (5) we include a vector of country-specific constants because this allows us to account for time-invariant country-level characteristics. The coefficient on distance to the capital and, most importantly, the coefficient on the interaction term between distance to the capital and the rule of law index remains significant at conventional levels.

Columns (3) and (6) report our preferred country-ethnicity fixed effect specifications; estimating these quite restrictive models—which include a vector of 1,209 ethnic constants—is feasible as proximity to the capital changes not only across but also within ethnic homelands. As in our baseline estimates in the sample of partitioned ethnic groups (in Tables IV-VIII), by exploiting solely within ethnic homeland variation, we account for ethnic-specific features, related to either geography or culture; moreover, the inclusion of ethnicity fixed effects also account for any relation between pre-colonial ethnic features and the placement of the capital by colonizers.¹⁸ The coefficient on the interaction term between national institutions and regional development continues to enter with a negative sign; moreover, the coefficient is statistically significant at standard confidence levels. Thus, the within-ethnicity results in the full sample of African ethnic homelands are quite similar to the pattern obtained within split ethnicities (shown in Table VIII). The statistically negative estimate on the interaction in columns (3) and (6) implies that the decaying role of countrywide institutional structures far from the capitals is a general phenomenon and not an artifact of country-specific and more important ethnic-specific differences within countries. Figures VIIIA and VIIIB provide a graphical illustration of how (relative) distance to the capital mediates the within-ethnicity correlation between rule of law and subnational development in the African continent.¹⁹

Figure VIII.

A Absolute Distance to the Capital and the Decaying Role of National Institutions

B Relative Distance to the Capital and the Decaying Role of National Institutions

VI.B. The Reach of the Law beyond the Capital Cities

The pattern uncovered in Tables VIII and IX points out that the influence of national institutions on subnational development diminishes in areas further from the capital cities. These findings are consistent with the view of the African historiography that both the colonial and the national state is becoming increasingly absent in remote areas with a concomitant decline in law enforcement. In an attempt to provide direct evidence of the differential state presence in areas far and close to the capital, we used individual-level data from the third round of the Afrobarometer Surveys, conducted in 2005 across a representative sample of either 1,200 or 2,400 respondents in 17 Sub-Saharan countries.²⁰

The Afrobarometer Surveys include two questions closely related to law enforcement. The first question regards law enforcement of a serious crime and reads: “How likely do you think it would be that the authorities could enforce the law if a person like you committed a serious crime?” The second question regards tax evasion and reads: “How likely do you think it would be that the authorities could enforce the law if a person like you did not pay a tax on some of the income you earned?” Both measures range from 1 to 4, with higher scores indicating that law enforcement is more likely.

To examine whether perceived law enforcement decays in remote areas, we regressed these two proxies on (absolute and relative) distance to the capital. In columns (1)–(8) in Table X we report the results. In all specifications we exploit within-country variation, including a vector of country fixed effects. Odd-numbered columns report unconditional estimates, and even-numbered columns include a rich set of individual controls.²¹ The coefficient on absolute distance to the capital (in columns (1), (2), (5), and (6)) is negative and significant, suggesting that the likelihood of law enforcement declines for respondents residing further from the capital city. The same pattern applies when we use relative distance to the capital (in columns (3), (4), (7), and (8)). Individuals located in enumeration areas (villages/towns/cities) relatively further from the respective capitals answer that it is less likely for the law to be enforced. These results clearly indicate a declining presence of authorities in the hinterland.

TABLE X.

Law Enforcement and Ethnic Identification as Function of Distance to the Capital Afrobarometer Sample, Country-Fixed-Effects Estimates

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)	(12)
	Law enforcement 1: Serious crime				Law enforcement 2: Tax evasion				National versus ethnic identification
Absolute distance	−0.1238***	−0.1027**			−0.1118***	−0.1006***			−0.0850*	−0.0740*
Double-clustered std. err.	(0.0446)	(0.0408)			(0.0410)	(0.0386)			(0.0414)	(0.0382)
Relative distance			−0.0860**	−0.0665*			−0.0655*	−0.0563*			−0.0949**	−0.0819**
Double-clustered std. err.			(0.0375)	(0.0354)			(0.0341)	(0.0326)			(0.0411)	(0.0379)
Adjusted R-squared	0.069	0.075	0.068	0.075	0.053	0.062	0.052	0.062	0.096	0.108	0.097	0.109
Observations	20,466	19,862	20,466	19,862	20,842	20,230	20,842	20,230	20,125	19,503	20,125	19,503
Individual controls	No	Yes	No	Yes	No	Yes	No	Yes	No	Yes	No	Yes
Country FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes. The table reports country fixed effects OLS estimates, associating legal enforcement (in columns (1)–(8)) and ethnic identification (in columns (9)–(12)) as reflected by individual responses in the Afrobarometer Surveys (3rd round) at the individual level with distance to the capital. The dependent variable is columns (1)–(4) is an ordered index (range 1–4) based on individual responses on how likely it is that authorities will enforce the law if a person like the respondent committed a serious crime. The dependent variable in columns (5)–(8) is an ordered index (range 1–4) based on individual responses on how likely it is that authorities will enforce the law if a person like the respondent did not pay taxes on income earned. For both measures a score of 1 indicates “not at all likely”, a score of 2 indicates “not very likely”, a score of 3 indicates “likely”, and a score of 4 indicates “very likely.” The dependent variable in columns (9)–(12) is a dummy variable that takes the value of 1 if the respondent replies that she identifies more strongly with her nation than the ethnicity to which she belongs. Even-numbered columns include a rich set of individual controls. The individual-level conditioning set includes age, age squared, a gender indicator variable, an urban indicator, 22 religion fixed effects, 25 occupation fixed effects, 5 living conditions fixed effects, and 9 education fixed effects. All specifications include a set of country fixed effects (constants not reported). The Data Appendix gives detailed variable definitions and data sources. Below the estimates we report in parentheses double-clustered standard errors at the enumeration area and at the ethnicity level. ***, **, and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively.

The specifications reported in columns (9)–(12) provide auxiliary evidence on the declining presence of the state in areas far from the capital using survey data on individual’s self-identification. In these linear probability models, the dependent variable is a dummy equal to 1 if the respondent identifies either exclusively or predominantly with the nation rather than his or her ethnic group. Across all specifications both absolute and relative distance enter negatively, implying that in areas far from the capital people are more likely to identify with their ethnicity, as compared to the state.²²

It should be acknowledged that systematic measurement error in the institutional quality proxies may also explain the weakening correlation between rule of law and regional development in areas far from the capital. Because most institutional variables are measured in the capital cities, capturing the rules governing activities of the formal economy, they might not reflect accurately the institutional features in rural areas (Pande and Udry 2006). Yet the evidence uncovered in Table X clearly indicates that law enforcement and state presence more generally decline with distance to the capital centers.

VII. Conclusion

We study the role of national institutions in shaping regional development captured by satellite images on light density at night by employing a novel approach that accounts both for (un)observed geographical-ecological features and ethnic-specific (cultural, historical) factors. We take advantage of the fact that the arbitrarily drawn colonial borders that endured after African independence partitioned several ethnic groups in different countries, subjecting identical cultures residing in geographically homogeneous territories to different country-level institutions.

We present four sets of results. First, our analysis documents that—in contrast to previous cross-sectional works—differences in economic performance within ethnicities partitioned between neighboring countries cannot be explained by differences in national institutions. The lack of within-ethnicity association between formal institutions and subnational development casts doubt on the causal interpretation of the cross-sectional strong correlation, for Africa at least. This finding nevertheless is in line with studies of the African historiography that downplays the role of national institutional structures, stressing instead the key role of ethnic-specific traits, related to the role of chiefs, culture, and pre-colonial organization (see Michalopoulos and Papaioannou 2013b for a literature review). Moreover, our results are in accord with a growing body of research that emphasizes the persistent effect of geographical features, cultural and ethnic norms, epidemiological traits, as well as history on long-run comparative development (see Spolaore and Wacziarg 2013; Nunn 2013 for reviews).

Second, we move beyond average effects by examining the association between national institutions and regional development for each partitioned group in Africa. The analysis reveals considerable heterogeneity. Third, in an effort to shed light on the factors shaping heterogeneity and motivated by studies in the African historiography that emphasize the inability of many African governments to broadcast power outside the capitals (Herbst 2000), we find that unlike split groups in the hinterland, differences in national institutions map into within-ethnicity differences in luminosity for those located close to the respective capitals. This phenomenon is not specific to split groups. In fact, it reflects a generalized pattern. Looking at the universe of African groups reveals that the explanatory power of national institutions monotonically declines with distance to the capital city. Fourth, we use individual-level data from the Afrobarometer Surveys and show that law enforcement decays with distance from the capital. Moreover, ethnic (as opposed to national) identification becomes stronger for individuals in the hinterland.

These correlations vividly illustrate the limited penetration of the state in remote areas, suggesting that at least within Africa, treating countries as homogeneous entities where nationwide characteristics (in our case the quality of national institutions) exert a uniform influence across regions may be quite misleading. In fact, the uncovered patterns are supportive to an old idea among development scholars (e.g., Lewis 1954) on the coexistence in Africa (and other parts of the developing world) of a “dual” economic-institutional framework with customary rules being dominant in the countryside and colonial-national institutions becoming relevant for regions closer to the capitals (see also Herbst 2000).

Our article calls for future research. One could employ our methodology that compares well-being in border areas using high resolution proxies of development (such as satellite light density) to shed light on the debate regarding the determinants (correlates) of comparative economic development; examining, for example, the effect of financial development, public policies, foreign aid, and fragmentation. Moreover, subsequent works could build on our analysis regarding the declining role of national institutions in areas far from the capital and explore the effect of other geographic, cultural, national, or ethnic-specific features that may attenuate or strengthen the importance of national policies in promoting subnational development. Finally, our evidence suggests that research on the role of institutions on development should move beyond average effects and examine the delicate interplay between formal institutions, state capacity, and ethnic traits.

Data Appendix

Light Density at Night

Light Density is calculated averaging light density observations across pixels that fall within the unit of analysis. We use the average of the values in 2007 and 2008. In the regressions at the ethnic homeland we use Log (0.01 + Average Luminosity). In the Online Appendix we also use the level of luminosity as the dependent variable. In the regressions at the pixel level we use as the dependent variable a dummy variable that takes on the value of one if the area is lit and zero otherwise. Source: Available at http://www.ngdc.noaa.gov/dmsp/global_composites_v2.html.

Population Density

Log (population density per sq. km. in 2000). Source: Nelson, African Population Database Documentation, UNEP GRID Sioux Falls (2004), available at http://na.unep.net/siouxfalls/datasets/datalist.php.

Area

Log (land area) at the country-ethnic homeland level or at the pixel level. Although most pixels are of the same surface area (2.5 (2.5 degrees), pixels intersected by the coastline and borders, and lakes are smaller.

Water Area

In the regressions at the country-ethnicity level we use the Log (1 + total area covered by rivers or lakes in sq. km.). In the regressions at the pixel level we use a dummy variable that takes on the value of 1 when some water body (river, lake, or other stream) falls in each pixel and 0 otherwise. Source: Constructed using the “Inland water area features” data set from Global Mapping International, Colorado Springs, CO, Global Ministry Mapping System.

Elevation

Average elevation in km of each ethnicity-country or each pixel. Source: National Oceanic and Atmospheric Administration (NOAA) and U.S. National Geophysical Data Center, Ter-rainBase, release 1.0 (CD-ROM), Boulder, CO, http://www.sage.wisc.edu/atlas/data.php?incdataset=Topography.

Land Suitability for Agriculture

Average land quality for cultivation within each country-ethnicity homeland or within each pixel. The index is the product of two components capturing the climatic and soil suitability for farming. Source: Michalopoulos (2012); original source: Atlas of the Biosphere, available at http://www.sage.wisc.edu/iamdata/grid_data_sel.php.

Malaria Stability Index

Average prevalence of malaria within each ethnicity-country or within each pixel. The index takes into account the prevalence and type of mosquitoes indigenous to a region, their human biting rate, their daily survival rate, and their incubation period. Source: Kiszewski et al. (2004).

Distance to the Capital City

The geodesic distance from the centroid of each ethnicity-country/pixel to the capital city of the country it belongs to, measured in thousands of kilometers. Source: Calculated using the Haversine formula.

Relative Distance to the Capital City

Distance to the capital city divided by the maximum distance to the capital in the same country.

Distance to the Sea Coast

The geodesic distance to the nearest coastline from the centroid of each ethnicity-country/pixel, measured in thousands of kilometers. Source: Global Mapping International, Colorado Springs, CO, Global Ministry Mapping System, version 3.0.

Distance to the National Border

The geodesic distance to the nearest national border from the centroid of each ethnicity-country/pixel, measured in thousands of kilometers. Source: Calculated using ArcGis.

Petroleum

Indicator variable that equals 1 if an oil field is found in the region (or pixel) of ethnic group i in country c. Source: The Petroleum Dataset v.1.1 contains information on all known on-shore oil and gas deposits throughout the world, available at http://www.prio.no/CSCW/Datasets/Geographical-and-Resource/Petroleum-Dataset/Petroleum-Dataset-v11/.

Diamond

Indicator variable that takes on the value of 1 if a diamond mine is found in the region (or pixel) of ethnic group i in country c. Source: Map of Diamond Resources, available at http://www.prio.no/CSCW/Datasets/Geographical-and-Resource/Diamond-Resources/.

Rule of Law

The index is “capturing perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, property rights, the police, and the courts, as well as the likelihood of crime and violence.” The standardized index ranges from −2.5 to +2.5 with higher values indicating better-functioning institutions. We use the average value over the period 1996–2006. In the RD analysis we also define a dummy (indicator) variable that takes on the value 1 for pixels falling in the country with the higher level of the rule of law index. Source: World Bank Governance Matters Indicators Database (Kaufmann, Kraay, and Mastruzzi 2008), available at http://info.worldbank.org/governance/wgi/index.asp.

Control of Corruption

Index on the control of corruption “capturing perceptions of the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as capture of the state by elites and private interests.” The standardized index ranges from −2.5 to +2.5 with lower values indicating a higher degree of corruption. We use the average value over the period 1996–2006. In the RD analysis we also define a dummy (indicator) variable that takes on the value 1 for pixels falling in the country with the higher level of the control of corruption index. Source: World Bank Governance Matters Indicators Database (Kaufmann, Kraay, and Mastruzzi 2008), available at http://info.worldbank.org/governance/wgi/index.asp.

Law Enforcement 1

Ordered (range 1–4) variable capturing law enforcement regarding committing a serious crime based on individuals’ responses to the following question: “How likely do you think it would be that the authorities could enforce the law if a person like you committed a serious crime?” A score of 1 indicates “not at all likely”; a score of 2 suggests “not very likely”; a score of 3 indicates “likely”; and a score of 4 indicates “very likely.” Source: Afrobarometer Surveys, 3rd Round, Question Q70B, available at http://www.afrobarometer.org/data/round-3-merged.

Law Enforcement 2

Ordered (range 1–4) variable capturing law enforcement regarding tax evasion based on individuals’ responses to the following question: “How likely do you think it would be that the authorities could enforce the law if a person like you did not pay a tax on some of the income you earned?” A score of 1 indicates “not at all likely”; a score of 2 suggests “not very likely”; a score of 3 indicates “likely”; and a score of 4 indicates “very likely.” Source: Afrobarometer Surveys, 3rd Round, Question Q70D, available at http://www.afrobarometer.org/data/round-3-merged.

National Identification

Dummy variable that takes on the value of 1 if the respondent replies that he or she identifies more strongly with the nation rather than his or her ethnicity. The question reads: “Let us suppose that you had to choose between being a [Ghanaian/Kenyan/etc.] and being a ______ [respondent’s identity group]. Which of these two groups do you feel most strongly attached to?” The variable equals 1 if the respondent replies that he identifies solely or predominantly with his nation and 0 otherwise. Source: Afrobarometer Surveys, 3rd Round, Question Q82, available at http://www.afrobarometer.org/data/round-3-merged.