Ethnolinguistic diversity and urban agglomeration

Our urban concentration measures capture the extent to which provincial populations concentrate into cities ($Urban$), and the extent to which that urbanized population is found in just one city ($Primate$). To construct them, we classify each grid cell in the categories city cores ($core$), dense towns ($dense$), semidense towns ($semi$), suburban ($sub$), and rural area (see SI Appendix for a detailed description of definitions and algorithms). Given this classification, our dependent variables are defined aswhere $Po{p}_i$ is the total population of province $i$ in 2015, $Po{p}_i^{1st}$ is the population in the largest city core in province $i$, and $Po{p}_i^{core}$, $Po{p}_i^{dense}$, $Po{p}_i^{semi}$, and $Po{p}_i^{sub}$ correspond to the total population of all grid cells in province $i$ of the respective category. For the urban share equation, we note that urban in the numerator is broadly defined. The GHS project has a low density threshold as part of its urban definitions of semidense towns and suburbs (300 people per square km), meaning that, in general, it reports higher urban shares worldwide than the United Nations World Urbanization Prospects data. However, we are only interested in relative comparisons across provinces within countries. For the primate share equation, we note that, for any specific city, the GHS project only identifies the dense $Po{p}_i^{core}$ population; suburban populations are not assigned to specific core cites. Thus, to have a denominator consistent with the numerator in Eq. 2, for all cities in a province, we include only dense urban populations, $Po{p}_i^{core}$ and $Po{p}_i^{dense}$. Later, as robustness checks, we will employ a stricter definition of urban share limited to core cities and dense towns in the numerator of Eq. 1, and we will use a measure of primate city size that attempts to incorporate commuting zones around cities in Eq. 2.

As noted above, we match the grid cells with fine-grained language information. Our language data from the WLMS are arguably the most precise source currently available, and have recently been used by refs. 22–24. The need to disentangle subtle differences in urbanization patterns has required us to construct our data at a more fine-grained level (1-km grid cells) than previous publications. Moreover, we apply the algorithm pioneered by ref. 24 for allocating languages to population in multilinguistic areas, which further increases precision. These features and the use of consistent definitions and data sources for urbanization and linguistic measures account for our dataset being the most precise of its kind currently available.

To compute measures of ethnolinguistic diversity, we use the $Fractionalization$ measure capturing the degree to which the population is segmented into many different groups at a provincial level. We also show, in SI Appendix, results for the $Polarization$ index capturing the extent to which the population is divided into two equal-sized and potentially opposing groups.

The reason we focus on ethnic $Fractionalization$ as the main measure is that it has been linked to both small-scale frictions in public good provision (25, 26) and large-scale social conflict and civil wars (2, 27, 28), whereas the use of ethnic $Polarization$ has been more confined to the study of large-scale wars (e.g., in refs. 1, 2, and 28), making the concept arguably narrower and, in our view, slightly less relevant than $Fractionalization$ for studying urbanization outcomes. Thus, we use $Polarization$ as alternative measure and relegate it to SI Appendix. Formally, the two measures are defined in the literature (1) aswhere $M_i$ designates the total number of groups $m=1,\dots ,M_i$ in province $i$, and ${\pi }_i^m$ corresponds to the population share of a group $m$ in the province’s total population.

We populate the language map with 1975 GHS population numbers (29), so as to represent language diversity historically. Ethnologue has up to 15 levels of distinction yielding 6,208 country–language pairs (e.g., “French–Canada” and “French–Switzerland” are two country–language pairs) when applying the finest level of language distinction. The information of Ethnologue and WMLS allows us to distinguish ethnolinguistic groups at different levels of language affinity, and these indices can be computed at any of the 15 levels. High levels of aggregation distinguish only major language families, while low levels of aggregation, for example, level 15, result in distinguishing very fine-grained differences between similar languages. Some countries such as India have enormous diversity, with 391 languages distinguished at the most disaggregated level, and 18 already at level 2.

As an example, in Fig. 3, we graphed the language structure for Himachal Pradesh, the above-mentioned province of about 7.5 million in northwest India. The figure illustrates the branches of its language tree, showing, for each branch, the highest level of disaggregation. The province starts on level 1 with two languages and then proceeds down to its finest division at level 8 with 18 final languages and ethnic groups.

In the main analysis, as in ref. 24, we shall focus on level 15, the highest disaggregation level worldwide. For most states in India, like Himachal Pradesh, the branches of the tree end at levels 6 through 8 (denoted by the underlining end language). When looking at level 15, branches ending sooner (say, level 6 or 8) are accounted as level 15 language affinity. In SI Appendix, Fig. S2, we show a similar graph for Switzerland. In the regression analysis, we demonstrate robustness at more-aggregated levels, where related languages in the tree are lumped together.

This section systematically studies the association between ethnolinguistic factors and urbanization patterns by regressing contemporary measures of urban concentration on historical measures of ethnolinguistic diversity, as well as initial urban concentration levels from four decades ago, using data from provinces across the world.

Table 1 displays our results. It is divided into three parts: the top two rows are a cross-sectional analysis, the next four rows are longitudinal, by additionally controlling for the past (1975) value of the dependent variable, while the last five rows show the number of observations and the controls used in both types of analyses. Columns are in pairs for different samples and outcomes, and, within each pair, columns are distinguished by the set of controls.

Column 1 in Table 1 for the cross-sectional rows regresses the $Urban\hspace{0.17em}share$ in a given province in 2015 on presample ethnolinguistic fractionalization in 1975, yielding a coefficient of –0.126 that is statistically significant at the 1% level. To give perspective, this means that moving from a perfectly ethnolinguistically homogeneous province (i.e., with ethnolinguistic fractionalization of 0) to a perfectly diverse one (i.e., with ethnolinguistic fractionalization of 1) would be associated with a 13% lower share of the urban population in the province. This change in urbanization corresponds to about half of a standard deviation of the $Urban\hspace{0.17em}share$ measure, or the difference between the very urbanized Netherlands and the less urbanized United States, which contains more rural area population. Note that this specification controls for country fixed effects, which means that the estimation is based solely on within-country comparisons of provinces, filtering out unobserved between-country heterogeneity. There is a concern, however, that estimates in the cross-sectional regressions could be biased because of omitted variables and reverse causality. For example, urbanization over long periods of time could influence fractionalization.

To deal with this, we move, in Table 1, column 1 of the longitudinal regressions, to a more demanding specification where we also control for 1975 values of urban share, in which we investigate the impact of fractionalization on the evolution of urbanization over the following four decades. A control for the 1975 urban share also controls for the influence of omitted variables, at least on historical urbanization, a topic we return to below. Of course, it also sweeps up any impact of ethnolinguistic fractionalization on historical urban share, leading us to potentially understate the total effect of fractionalization on urban share in 2015. However, conditioning on base period urbanization tells us, more unambiguously, how subsequent urbanization is influenced by baseline fractionalization. When controlling at the province level for urban share in 1975, in the longitudinal regressions, we still find a statistically significant negative coefficient, albeit its magnitude is reduced by half compared to the cross-sectional regressions. Of note is that the coefficient of past urban share is sizeable and highly significant, pointing toward a large persistence of urbanization patterns over time. Overall, it is reassuring that, in the longitudinal rows, we continue to find evidence of ethnic fractionalization slowing down the pace of urbanization, after controlling for presample urbanization.

In column 2 of Table 1, we estimate the analogous specifications as in column 1, but controlling, in addition, for terrain ruggedness and population density in 1975 (see SI Appendix, Data, for a detailed description of these control variables, and see SI Appendix, Table S2 for all estimated coefficients). The results remain very similar, and the coefficients of interest remain statistically significant at the 1% level.

With regard to the measure of urban concentration, we estimate the same specifications for the share of the primate city in total urban population ($Primate$). Note that, unlike the 1975 urban share, the past primate share from 1975 is only observable for a restricted sample, since some provinces in 1975 did not have a core city ($Po{p}_i^{core}$). Hence, we run the regressions of primate share in the cross-sectional rows of Table 1 on fractionalization first on the full sample (columns 3 and 4) and then on the restricted sample (columns 5 and 6) to improve comparability. We find that the importance of the biggest city among urbanized areas is considerably reduced in the face of ethnolinguistic fractionalization. Put differently, ethnolinguistic diversity is associated with having several smaller cities instead of a single megacity. Quantitatively, moving from a fully homogeneous to a fully heterogeneous society (i.e., moving ethnolinguistic fractionalization from 0 to 1) would be associated with an at least 8% lower $Primate\hspace{0.17em}share$ in columns 5 and 6 in the longitudinal rows, equal to about a quarter of a standard deviation of this variable.

Note that we also carry out a regression analysis linking ethnic diversity to conflict. In the interest of space, this investigation is relegated to SI Appendix. In SI Appendix, Table S8, we show that there is a strong and statistically significant association between ethnolinguistic fractionalization in 1975 at the province level and several measures of armed conflict between 1975 and 2015 at the province level.

How robust are our results to various considerations? The first concern is omitted variables. In SI Appendix, Table S2, our results are robust to including further control variables that could potentially influence the spread of cities. In particular, we control for square and cubic terms of population density, for distance to coast, elevation, latitude, and provincial GDP, and for whether the national capital is located in the given province. We also control for the degree of historical conflict from 1946–1974 to address concerns that initial antagonism may have shaped diversity and urbanization in 1975. SI Appendix, Data, contains a detailed description of these control variables. Note that these robustness checks can reduce sample size, as the additional information is not observed in all countries. Coefficients on ethnolinguistic fractionalization move very little in response to varying the sets of controls. Finally, we assess the maximum potential remaining bias from omitted (unobserved) variables by performing a test following Altonji et al. (30) and Oster (31). In our specification with most controls for observables, that is, the longitudinal rows in Table 1, we calculate an estimate of the extent of possible bias for the effect of fractionalization of $+0.020$ for urban share and $+0.022$ for primate share.^† Hence our point estimates remain substantially below zero, even allowing for such potential bias.

Next, for robustness, we show that the overall stability of estimated coefficients remains when varying the threshold levels in the language tree for distinguishing different languages. As explained above, our data allow us to compute ethnolinguistic diversity measures for different definitions of what constitutes distinct languages. When using an aggregation level of 1, we only distinguish the most fundamental differences in the language tree, such as the difference between Indo-European and Sino-Tibetan language families, but lump together distinctions, such as Italian and German, into the Indo-European group. In contrast, as we move down the tree, the distinctions become more fine-grained, where local dialects are distinct, such as Kangri, Hinduri, and Dogri as dialects of Western Paharai which, in turn, is related to Punjabi in Fig. 3 above, or, say, Arpitan, Romansch, Lombard, and French in non-German Switzerland (SI Appendix, Fig. S2).

We graph the pattern of coefficients and their significance in Fig. 4, linking ethnic diversity to urban share, primate share, and conflict. Overall, the results of Fig. 4 highlight the stability of estimated coefficients over a range of possible aggregation levels of the language data. In particular, we observe a statistically significant negative association between ethnic fractionalization, on the one hand, and urban and primate shares, on the other hand, across a wide range of possible language aggregation levels. Moreover, the positive correlation between ethnic fractionalization and conflict is found across the board of different aggregation levels. We note that explanatory power of the regressions across all these graphed levels varies minimally.^‡

Next, we turn to our alternative measure of ethnolinguistic diversity. While the fractionalization measure takes high values for areas with a large number of groups, the main alternative diversity measure defined above, ethnolinguist polarization, reaches high values for situations closer to bimodal distributions of a small number of sizeable groups. As discussed above, we prefer fractionalization—the arguably somewhat broader concept, fitting better the context of urbanization—and have relegated polarization to SI Appendix.

The relationship in the data between our fractionalization and polarization measures is displayed in SI Appendix, Fig. S4. After filtering out country averages (SI Appendix, Fig. S4B), the two diversity measures are highly correlated, although the correlation is far from perfect. It is therefore useful to replicate our baseline Table 1 using polarization measures instead of fractionalization. Studying the role of ethnolinguistic polarization also provides a different perspective on diversity—the effect of being more bimodal versus simply more diverse. The results of the baseline specification using polarization instead of fractionalization are displayed in SI Appendix, Table S3 with very similar results for primacy and somewhat weaker results for urban share.

Further, we consider alternative measures for the outcome variables urban share and primate share reported in SI Appendix, Table S4. First we consider, in columns 1 and 2, a narrower measure of the degree of urbanization, by only considering city cores and dense towns in Eq. 1, leading to similar results for both fractionalization and polarization. Then we consider an alternative definition of primate share. We draw on data from a joint Organisation for Economic Co-operation and Development (OECD) and European Commission (EC) project described in ref. 32 which offers a globally harmonized definition of commuting zones called functional urban areas (FUA). We measure primate city share as the FUA population divided by the broad definition of urban population in the numerator in Eq. 1. We use the broad definition, since FUAs contain population in less dense areas. Using this definition for primate city share in columns 3 and 4 again yields very similar results for both fractionalization and polarization.

Last, we explore the “modifiable areal unit problem” (MAUP) (33, 34) and ecological correlations (35), which could arise if results at the levels of (large) provinces do not carry over to smaller spatial units. Put differently, our results could be sensitive to the definition and scale of units for which data are collected. One way to investigate this is to split our provincial sample in two, according to the scales of provinces (area or population), and then check whether the findings hold similarly for the samples of countries with smaller versus larger provinces. This is what we do in SI Appendix, Tables S5 and S6. In the former table, we split the sample according to average population area (unweighted and population-weighted), while, in the latter, we split according to average province population and the number of provinces in a country. For both small and large province samples, in all cases, we continue to find large negative effects of ethnic fractionalization on urban share and primate share, with no clear pattern of whether results are stronger for the small or large province samples. We conclude that the MAUP is not driving our results.

The above results tell a stark story of ethnolinguistic diversity slowing down urbanization and urban concentration, and hence potentially affecting economic development. Still, there may be room for policies to dampen the extent of this relationship. One natural candidate for a policy dimension that may be able to modulate ethnic tensions is democracy. In particular, there exists evidence that, while full, consolidated democracy reduces the risk of ethnic tensions and conflict, nascent or fragile/intermediate democracies may bear higher risks of political violence than autocracies (3, 36).^§ Hence, in what follows, we shall investigate whether the impact of ethnic fractionalization is magnified in countries with intermediate democracy levels.

In particular, we interact our fractionalization measure with three regime types: full democracy, intermediate regime, and full autocracy. We control for the full set of fixed effects and other baseline controls (ruggedness, population density), including the 1975 levels of the urban variables in the longitudinal regressions. Results are reported in Table 2. In columns 1 and 2, the democracy measure is taken from the Polity IV project (38), while, in columns 3 and 4, we rely on democracy scores from Freedom House (39). The overall picture emerging from Table 2 is that, indeed, the impact of ethnic diversity on urban share and primate share tends to be distinctly magnified in intermediate regimes. However, the differences in coefficient magnitudes in many cases are statistically weak and stronger for primacy than for urban share (see tests in the bottom three rows of the cross-sectional and the longitudinal regressions for details). Hence, these results need to be interpreted with caution. We find similar patterns in SI Appendix, Table S7 for ethnic polarization as for fractionalization.

All data used in this study are from public and commercial data sources as described in SI Appendix. Generated data and code to generate variables and results are publicly available at the Harvard Dataverse (https://doi.org/10.7910/DVN/PLDXPD) (40).

We are grateful for helpful comments from participants at the 13th Meeting of the Urban Economics Association and the 9th European Meeting of the Urban Economics Association. We thank Yannis Ioannides for comments on a formative version of the project and the suggestion that urban share should be a measure of urban concentration. We thank the editor and two referees for very helpful comments. U.J.E. gratefully acknowledges financial support from the Swiss National Science Foundation Doc.Mobility Fellowship P1LAP1_181253 and especially thanks the Centre for Economic Performance at the London School of Economics and Political Science for hosting him during 2018–2020. D.R. gratefully acknowledges financial support from the European Research Council Starting Grant POLICIES FOR PEACE-677595 and warmly thanks University of British Columbia for the hospitality during 2018–2019 when the first draft of this paper was written. K.S. is grateful to University of California, Berkeley for the hospitality during the academic year 2017–2018 when this project took shape.