Empirical strategies

The aim of this paper is to study the causal impact of political institutions and international trade on the concentration of people in the biggest cities. However, before getting into the detail of these two determinants, let's first analyze the impact of institutions in general, considering all the formal and informal constraints and rules (the different codes of conduct, customs, laws, governance) that define political institutions and also the economic interactions necessary to long-distance trade) that causes the concentration of a nation's urban population in its first city. The most simple equation allowing to describe key identification issues of this question may be the following:Footnote ¹²

(1)$$u_{it} = \alpha I_{it} + f_i + f_t + \varepsilon _{it}\; $$

where u _it is the population in the biggest metropolis,Footnote ¹³ f _t are time-fixed effects that control for shocks affecting all individuals at the same time, f _i are country-fixed effects aimed at capturing the characteristics of countries and I _it the variable of institutions of the country i at the period t.

The main parameter of interest is α, the elasticity of agglomeration with respect to institutions. An unbiased OLS estimate for this parameter is obtained under the following assumption:

(2)$${Cov}\lpar {{I}_{{i},{t}},{\,f}_{i} + {\,f}_{t} + {\varepsilon}_{{it}}} \rpar = {Cov}\lpar {{I}_{{i},{t}},{\,f}_{i}} \rpar + {Cov}\lpar {{I}_{{i},{t}},{\,f}_{t}} \rpar + {Cov}\lpar {{I}_{{i},{t}},{\varepsilon}_{{it}}} \rpar = 0$$

As is well known, when the covariance on the left hand side of this equation differs from zero, any of the three covariances on the right hand side is a source of bias. We discuss these three terms in turn in order to build our empirical strategy.

Concerning the first term, Cov(I _i,t, f _i), there are some reasons to consider that time-varying institutions depend on invariant country characteristics: for instance, natural resources or the level of development may have an impact on institutions. A related concern is the fact that the degree of urban primacy may influence institutions. Agglomeration economies may be higher in countries with urban giants and incomes generated from spillover, and/or pecuniary externalities can influence institutional change. This relationship is even more obvious concerning international trade: greater productivity in a particular sector (e.g. in finance) due to a concentration of activities in one place may influence institutions, at least the formal institutions, of a country. To analyze this bias, consider the following reversed causality between institutions and concentration:

(3)$${I}_{{it}} = {\beta} {u}_{{it}} + {\eta} _{i}$$

From classic sociological theory to political modernization theory (Lerner, Reference Lerner1958; Lipset, Reference Lipset1959; Weber, Reference Weber and Runciman1909), there is a vast literature that defends this kind of relationship, considering that the urban structure affects political mobilization and political change (see Anthony (Reference Anthony2014) for a survey).

Using equation (1) without time effects to simplify, yields:

(4)$${I}_{{it}} = \displaystyle{{{\beta} {\,f}_{i} + {\eta} _{i}} \over {1-{\beta \alpha}}}, $$

which allows us to rewrite the previous covariance as follows:

(5)$$\lpar {1-{\beta \alpha}} \rpar {Cov}\lpar {{I}_{{i},{t}},{\,f}_{i}} \rpar = {\beta Var}\lpar {{\,f}_{i}} \rpar + {Cov}\lpar {{\eta}_{i},{\,f}_{i}} \rpar $$

Hence the first origin of the bias comes from the variance of individual specific characteristics Var(f _i), magnified by the elasticity of institutions with respect to the concentration of people β. The second channel of an upward bias is the correlation between individual characteristics that explain institutions and urban primacy (one can think respectively of culture and productivity), which lead one to conclude that such a bias is likely.Footnote ¹⁴

The other origins of bias in equation (2) are simple to understand: the covariance between institutions and temporal chocks (e.g. a global financial crisis), Cov(I _i,t, f _i), may differ from zero. Finally a covariance between institutions and the error term is the usual definition of endogeneity.

To reduce each of these problems, we first look for a lagged measure of institutions, or more precisely a measure of past institutional factors, that explains the sustainability of urban agglomeration before the industrial revolution. We choose the density in 1500 to approximate political institutions around that period. Density in 1500 is a good proxy of the current institutions according to AJR, who show that countries that were relatively rich in 1500 (density is considered as a proxy of economic prosperity) have fostered predatory behaviors and extractive institutions on the part of colonizers. We denote this variable I _i,1500. We also use a proxy for the current state of political institutions by using the binary view of Acemoglu, Naidu, Restrepo and Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017), who classify a country as democratic when different databases of political institutions (Freedom House, Polity IV) conclude as such, and dictatorial in the reverse case. This variable is denoted I _it.

To take into account institutions that play a major role in explaining international trade, we regress bilateral trade flows on distance, GDPs and fixed effects and we use the estimates of fixed effects (which typically capture unobservable institutions that explain trade) and distance to build an indicator of market access (also called market potential), hereafter denoted as Ω_it. To reduce the bias of an omitted variable, several variables of control are also introduced, including a proxy of the income gap between the biggest city and the rest of the country. While it seems trivial to say that people move from one place to another according to income difference, this variable has never been used to analyze urban primacy until now, mainly due to a lack of data at the regional level in developing countries. To approximate this spatial income gap in each country we use the differential between the intensity of night lights in the biggest city and the intensity of lights in the rest of the country from satellite pictures.Footnote ¹⁵ We also add classical controls (mainly climate data) for characteristics that are specific to country and time. All these controls are represented in the following equations by a vector C.

We start our analysis with a simple cross-sectional analysis (for the year 2010), running ordinary least squares on the following equation:

(6)$${u}_{i} = {\alpha} {I}_{{i},{t}-{x}} + {\delta} {\Omega} _{\bf i} + {\gamma} {C}_{i} + {\varepsilon} _{i}$$

where I _i,t−x represent past institutions when the indicator of density in 1500 is used (t − x = 1500, or the current level when we use a binary variable of democracy (t − x = 0) Many papers on the impact of institutions on urban primacy are based on this type of cross-sectional analysis. This naive regression presents all the aforementioned bias (and many more), and the problem of endogeneity due to omitting variables is particularly serious. However, the advantage of this estimation is that it allows us to include the invariant historical measure of institutions, I _i,1500. This variable, which is also a measure of the urban primacy in 1500, helps us to understand the evolution of the spatial economy of nations over a long period of time. The coefficientα, if negative, indicates a dispersion of the population from the biggest city, i.e. a reversal in the attractiveness of the biggest city.

We then rely on the panel structure of the data to tackle endogeneity issues (due to omitting variables) by estimating:

(7)$${u}_{{it}} = {\alpha} {I}_{{i},{t}-{x}} + {\delta} {\Omega} _{{\bi it}} + {\gamma} {C}_{{it}} + {\,f}_{i} + {\,f}_{t} + {\varepsilon} _{{it}}$$

This equation is analyzed through the lens of the Fixed Effect (FE) estimator over the period 1992–2010. This FE estimator is useful to reduce the endogeneity bias of omitted variables. However, the FE procedure is useless to obtain consistent estimates of the time-invariant proxy of institution I _i,1500, since there is a perfect colinearity between fixed effects and institutions in 1500. As a result, the FE model constrains us to use exclusively time-varying institutions I _it, which potentially introduce endogeneity from reversed causality.

To carry out inference on historical density, we thus use the Fixed Effects Filtered (FEF) estimator of Pesaran and Zhou (Reference Pesaran and Zhou2016). This estimation is in two steps, first a FE estimation is used to obtain coefficients of the time-varying variables and these estimates are used to filter out the time-varying effects. Then the residuals from the first step are averaged over time and computed as a dependent variable in a cross-section OLS estimation that includes a historical density and trade network.

To obtain an unbiased estimate of the elasticity of agglomeration with respect to the current institutions due to reverse causality, we propose two different Instrumental Variables (IV) estimations.

Data

Trade

The indicator of market access is built directly from economic geography theory. We modify market access of Redding and Venables (Reference Redding and Venables2004) to take into account the specificity of our research. The first step is to estimate the following trade gravity equation on bilateral exports, coming from COMTRADE, with individual fixed effects ($f_i^x, f_i^m$) and distance using a pseudo-maximum likelihood (PPML) estimator:Footnote ¹⁸

(8)$${X}_{{ij}} = -{\varepsilon ln}{\tau} _{{ij}} + {a}_{i}{\,f}_{i}^{x} + {a}_{\,j}{\,f}_{\,j}^{m} + {v}_{{ij}}$$

where bilateral trade costs τ _ij take the following common specification:

$${\tau} _{{ij}} = {dis}{t}_{{ij}}{exp}\lpar {{\delta}_1{b}_{{ij}} + {\delta}_2{c}_{{ij}} + {\delta}_3{l}_{{ij}}} \rpar$$

With b _ij, c _ij, l _ij are dummies characterizing bilateral trade barriers, respectively a border, past colony links and a common language; dist _ij is a weighted measure of distance between cities. The literature usually uses two kind of distance, the simple distance between capitals and the distance between capitals weighted by the population of each country. Since simple distances between capitals represent poorly the geography of trade costs (in particular for large countries) and introduce bias in the estimation (e.g. over-estimation of the border effect), many articles routinely use the weighted measure, proposed by Head and Mayer (Reference Head and Mayer2002) and freely available from the GeoDist database of the CEPII. This distance between country i and j is calculated as follows:

$${dis}{t}_{{ij}} = \sum \nolimits_{k} \lpar {{\,po}{\,p}_{k}/{\,po}{\,p}_{i}} \rpar \sum \nolimits_{l} \lpar {{\,po}{\,p}_{l}/{\,po}{\,p}_{\,j}} \rpar {d}_{{kl}},$$

where pop _k and pop _l are the population of the largest cities k and l. In brief, distances are weighted by the share of the biggest city in the total population, which is also our dependent variable. Thus it seems obvious that we cannot use such a measure to explain the concentration of the population in the nation's largest city without introducing endogeneity bias. We thus deviate from the literature by building our own measure using a weighting scheme based on internal distance between the capital and its cities:

$${dist}_{{ij}} = \displaystyle{{{d}_{{ij}}} \over {{d}_{i}{d}_{\,j}}}{\;with\;} {d}_{i} = \sum \nolimits_{k} {dis}{t}_{{ik}},{\;\;} {d}_{\,j} = \sum \nolimits_{l} {dist}_{{\,jl}}{\;,}$$

where d _ij is the distance in kilometres between the biggest cities i and j of each partner, and d _i is the internal distance, i.e. the sum of the distance between the biggest city i and all other cities k belonging to the same country.

Last, an important part of the gravity equation making the link between theory and empiric are fixed effects $f_i^x$ and $f_j^x$ (and their symmetric expressions) that allow us to obtain the predicted value of supply and market capacity.Footnote ¹⁹ Using these predicted values, with our weighted measure of bilateral distance, and estimates of the trade elasticity ε give market access:

(9)$${\Omega} _{i} = \lpar {{exp}\lpar {{\,f}_{i}^{m}} \rpar } \rpar ^{{{\hat{\alpha}}} _{i}} + \sum \nolimits_{{\,j}\ne {i}} \Bigg [ {{\lpar {\exp \lpar {{\,f}_{\,j}^{m}} \rpar } \rpar }^{{{\hat{\alpha}}}_{\,j}}{dis}{t}_{{ij}}^{-{\hat{\varepsilon}}} {\;b}_{{ij}}^{\widehat{{{\delta}_1}}} {\;c}_{{ij}}^{\widehat{{{\delta}_2}}} {\;l}_{{ij}}^{\widehat{{{\delta}_3}}}} \Bigg ] $$

The gravity equation is estimated in cross-sections to obtain a market access that varies over time Ω_it.

To get rid of multicollinearity problems in estimating (7) and also to dispose of the fact that market potential can be endogenous to the size of mega-cities, we also use an historical measure of market access that is linked not to trade flow but to trade relationship: we calculate an indicator of networks, the so-called out-degree of trade, measuring the number of arcs pointing to partners with respect to the total number of nodes in the bilateral trade network of countries in 1900.Footnote ²⁰

Institutions

Concerning institutions, one of the most recent dichotomous measure of democracy proposed by Acemoglu, Naidu, Restrepo and Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017) (hereafter ANRR) is chosen. This index, running from 1960 to 2010, combines information from several sources including Freedom House and Polity IV databases. ANRR classify a country as democratic when several databases conclude as such. We also use the instrument of ANRR, who exploit regional waves of democratization and transitions to non-democracy as a source of exogenous variation in democracy. More precisely ANRR divide the world into seven regions, R _i (where i is a country in Africa, East Asia and the Pacific, Eastern Europe and Central Asia, Western Europe and other developed countries, Latin America and the Caribbean, the Middle East, North Africa, and South Asia) and consider a dummy $D_{it_0}$ taking 1 when the country i is a democracy at the start of the period. Finally democracy in country i is influenced by democracy in the set of countries $J_i = \lcub {{i}^{\prime}:{i}^{\prime}\ne i,\; R_{{i}^{\prime}} = R_i,\; D_{{i}^{\prime}t_0} = D_{it_0}} \rcub$, which gives the following instrument of political governance:

(10)$${G}_{{it}} = \displaystyle{1 \over {\vert {{J}_{i}} \vert }} \mathop \sum \limits_{{{i}^{\prime}}\in {J}_{i}} {G}_{{{i}^{\prime}t}}$$

In the words of ANRR, G _it is the jack-knifed average of democracy in a region × initial regime cell, i.e. a distance-weighted average of democracy among “neighbors.” In order to avoid confusion, it is noteworthy that the “geographical components” of this instrument are radically different from the ones included in market access, geographical distance is not taken into account, and the neighborhood has a wide definition (e.g. every country in Africa).

For robustness checks regarding the variable of democracy, we also use the Polity 2 variable, from the Polity IV project and the Freedom House indices of political rights (PR) and civil liberty (CL).

The mortality rates of soldiers, laborers, and bishops used during colonial times used as an instrument of institutions, come from Albouy (Reference Albouy2012).

First results

Table 1, columns (1) and (2), examines the cross-sectional patterns running ordinary least squares specification as described in equation (6) to illustrate the raw data patterns by regressing the log share of urban primacy in 2010 on two different indicators of democracy. In column (1), based on AJR, this indicator is the log of density in 1500. The bias of reversed causality is reduced by using this historical variable, but this advantage comes at a cost: the effects of the current political institutions cannot be analyzed. In column (2), we use the indicator of democracy of ANRR. These two columns make our results comparable with the literature on urban primacy that often uses cross-sectional data, but as discussed at length in Section 1, this specification may provide biased estimates of the variables of interest. Then, in columns (3), (4), (5) and (6), we exploit the panel structure of the data to tackle endogeneity issues by estimating equation (7) using the Fixed Effects (FE) and the Fixed Effects Filtered (FEF) models.

Table 1.

Urban reversal

Note: OLS estimator in columns (1) and (2), Fixed effects (FE) in column (3), Fixed Effects Filtered (FEF) in columns (4)–(6). All variables are in Log, ^a: significant at 1%, ^b: significant at 5%. Data on urban primacy are from World Development Indicators. Data on density in 1500 are from the Clio-infra database on urban settlement sizes, data on urbanization in 1500 from Acemoglu Johnson and Robinson (Reference Acemoglu, Johnson and Robinson2002). The binary variable of democracy comes from Acemoglu, Naidu, Restrepo and Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017). The income gap between the biggest city and other cities is approximated by the differential between the intensity of night lights in the biggest city and the intensity of lights in other cities obtained from satellite pictures. Market access is computed from the estimation of a gravity equation using bilateral exportation from COMTRADE, distance and geographical variables from the CEPII. FE estimations include a full set of year and country fixed effects.

In column (1), the elasticity of urban primacy to past density is significant, high, and negative, revealing a reversion in the size of the urban concentration relative to its total population. On average a country that was twice more densely populated in 1500 is now a country where the share of its biggest city in the total population is 23% smaller than in other countries. We argue that the level of density in 1500, being a proxy of bad governance in the following centuries (countries densely populated in 1500 were richer, which was an incentive to set extractive institutions),Footnote ²² represents a repulsive force for the native population. This mechanism is consistent with the finding of Nunn and Puga (Reference Nunn and Puga2012) that rugged terrains were chosen in Africa because these places afforded protection to those being raided during the slave trade. The other side of this result is that countries with a weak concentration of the population in the nation's largest city in 1500 have benefited from inclusive institutions, such as the development of institutions of private property, which made the growth of urban giants profitable by developing agglomeration economies. As in the cross-section analysis, the impact of past density is still negative and significant (columns (4) and (5)) using the Fixed Effects Filtered (FEF) estimator of Pesaran and Zhou (Reference Pesaran and Zhou2016). Finally in columns (2), (3) and (6), the indicator of democracy is always significant and influences positively the population growth in the biggest city. All these results go against the finding of Ades and Glaeser (Reference Ades and Glaeser1995) that bad governance fosters population concentration at the top of the urban hierarchy.

In column (1), the past network of trade is also significant and negative. This negative impact of the past network of international trade may be related to the particular specialization of some countries, the first globalization has favored land tenets in Latin America for instance (Findlay and O'Rourke, Reference Findlay and O'Rourke2009) and thus urbanization was possibly not profitable for this elite. As a result this group may have put in place policies against cities and in particular against large cities that were a fertile land for industrialization. The introduction of fixed effects (column (5)) confirms this result, since the coefficient of the past network, albeit smaller, is still significant. However the introduction of an indicator of the current democracy raises some doubt about the validity of this variable of past network which is no longer significant (column (6)) and thus no longer used in the rest of the paper. Indeed this indicator of past network, reflect more past political institutions and networks between kings than between traders.

Columns (3) and (4) present the positive impact of market access. This result supports the thesis that a decrease in trade costs, at least in the first step of development, exacerbates regional disparities (Krugman, Reference Krugman1991). Activities and people tend to concentrate where there is a large market and the market is large where activities are concentrated; trade openness fosters this circular causality of agglomeration in the biggest city.

Regarding controls, the income gap approximated by the difference in night lights between the biggest cities and other cities has the expected positive impact.

The negative impact of temperature in the cross-section analysis (columns (1) and (2)) may be driven in some countries by the indirect effect of new technologies such as air conditioning or large investment infrastructures (e.g. dams) that have improved conditions of life in many cities that were initially inhospitable. Once individual fixed effects are introduced capturing the heterogeneity of countries on these aspects, temperatures are no longer significant (columns (3)–(6)).

Two-step SLS identification strategy with a zero-stage

The zero-stage is based on the finding that the mortality rate faced by settlers is a good instrument of the current institution. Since this proposition has stirred up questions (Albouy, Reference Albouy2012),Footnote ²³ we also take a more contemporaneous variable by using the regional waves of democratization and transitions to non-democracy as an instrument (this is the instrument used by ANRR, already described in the previous section).

Denoting H _i the historical mortality rates of soldiers, laborers, and bishops during colonial times (source: Albouy, Reference Albouy2012), I _i, the indicator of democracy (source: ANRR), and G _i, the measure of political governance based on waves of democratization (defined by equation (10)), we estimate the following equation:

(11)$${I}_{i} = {\beta} {H}_{i} + {\gamma} {G}_{i} + {v}_{i}$$

This equation is simply an extension of AJR using a different dependent variable (democracy instead of expropriation risks), and an additional variable, G _i. Finally while AJR estimates this equation for one year only, we regress it on each year of our sample which gives an historical-geographical instrument, denoted $\widehat{{H_{it}^g}}$ that varies over time.Footnote ²⁴ Democracy effect identification will come from this time-varying predicted instrument $\widehat{{H_{it}^g}},$ which captures discretionary variations in institutions. It reflects the exogenous component of the democracy variable as a consequence of distant past shocks in the contemporaneous institutions trend.

Concerning the endogeneity of trade, we follow the same strategy by using the ancestral genetic distance (Spolaore and Wacziarg, Reference Spolaore and Wacziarg2018). The relevance condition asserting that this variable is a good instrument to predict the endogenous dependent variable, i.e. trade, is well established. For instance, Guiso, Sapienza and Zingales (Reference Guiso, Sapienza and Zingales2009) find that genetic distance is a good proxy for cultural proximity, which impacts on the trust between two partners and then on trade exchanges. Spolaore and Wacziard (Reference Spolaore and Wacziarg2018) present clear evidence that genetic distance acts as a temporary barrier to the diffusion of innovations and development. Giuliano, Spilimbergo and Tonon (Reference Giuliano, Spilimbergo and Tonon2014) find that genetic distance is a good proxy for transportation costs between countries because of international migrations. The variation in genetic distance is partially explained by international migrations, which partially cause international trade via business and social networks. See for instance Rauch and Trindade (Reference Rauch and Trindade2002) on the impact of ethnic networks on trade. Market access is then regressed on ancestral genetic distance, denoted $D_i^g$, for each year of our sample, which enables us to get an instrument of market access based on genetic distance that varies over time, hereafter denoted $\widehat{{D_{it}^g}}$.

The first step explains the variation of institutions by using the exogenous variable of institutions defined in the previous step:

(12)$${I}_{{it}} = {\varsigma} \widehat{{{H}_{{it}}^{g}}} + {\,f}_{i} + {\,f}_{t} + {C}_{{it}} + {\chi} _{{\bi it}}$$

and in a similar way the variation over time in market access:Footnote ²⁵

(13)$${\Omega} _{{it}} = {\varrho} \widehat{{{D}_{{it}}^{g}}} + {\,f}_{i} + {\,f}_{t} + {C}_{{it}} + {v}_{{\bi it}}$$

In this first step,$\; \varsigma$ is estimated from within changes in $\widehat{{H_{it}^g}}$ which, by depending on the distance-weighted average of democracy among neighbors, is an exogenous variable of the current urban primacy. In a similar way ϱ is not estimated by cross-country difference but by internal changes in genetic distance between partners, which is certainly not a function of trade. Individual fixed effects in the two equations are particularly important in controlling for differences in the level of development. The term C _it represents three variables of control that are used in the two stages: precipitation, temperature, and a gap in the intensity of night lights.

Finally the second step uses the predicted values of the “historical-geographical” institutions and of the predicted value of “genetic” market access as follows:

(14)$${u}_{{it}} = {\;\hat{\varsigma}} \widehat{{{H}_{{it}}^{g}}} + {\;} {\hat{\varrho}} \widehat{{{D}_{{it}}^{g}}} + {\,f}_{i} + {\,f}_{t} + {C}_{{it}} + {\epsilon} _{{it}}$$

Results of the zero-stage

Table 2 confirms the positive impact of waves of democratization on institutions. Results are less clear concerning the effect of the mortality rate of colonizers. Coefficients associated with the colonizers’ mortality rate are significant only between 1962 and 1972, which may suggest an erosion of the bad influence of past colonial institutions. However, additional estimations provided in the online appendix show that the initial result of ARJ is verified over the whole period when we use a different indicator of democracy, i.e. the mortality rate of colonizers explains institutions as measured by the Polity 2 index of democracy, the Freedom House indices of political rights (PR) and civil liberty (CL). It is also noteworthy that this indicator of mortality has a small impact on institutions in comparison with waves of democratization.Footnote ²⁶

Table 2.

Democracy (ANRR index), zero-stage

Note: OLS estimator. Mortality rate variable is in Log, ^a: significant at 1%, ^b: significant at 5%. The mortality rate of colonizers (in log) comes from Albouy (Reference Albouy2012). Data on democracy are from Acemoglu, Naidu, Restrepo, Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017), denoted ANRR in the title of this table, and instrumented with their instrument based on waves of democratization.

In Table 3 we present some of the zero-stage of market access. Throughout the period we find a negative impact of genetic distance on market access.

Table 3.

Market access, zero-stage

Note: OLS estimator. All variables are in Log, ^a: significant at 1%, ^b: significant at 5%. Data on genetic distance in 1500 are from Spolaore and Wacziarg (Reference Spolaore and Wacziarg2018). Market access is computed from the estimation of a gravity equation using bilateral exportation from COMTRADE, distance, and geographical variables from the CEPII.

Results of the first stage

Table 4 (column (1)) presents the first stage described by equation (12). As expected, our historical and geographical instrument, based on mortality rates and on waves of globalization, significantly explains the level of democracy. This first stage confirms the different results obtained by AJR and ANRR. The coefficients on the instrument variables are highly significant. We present the F-test statistics that reject the joint null effect hypothesis of these IVs and they are above 10, the threshold proposed by Staiger and Stock (Reference Staiger and Stock1997). As a robustness check we also run an estimation by excluding the mortality rate in the zero-stage, and thus by using an instrument in this first step only based on the waves of democratization. We do not report this estimation since the results are almost identical (we find a coefficient of 0.871 with a Robust Standard Error (RSE) of 0.052).

Table 4.

Democracy and trade, first stage

Note: 2SLS procedure, ^a: significant at 1%, ^b: significant 5%. Market access is computed from the estimation of a gravity equation using bilateral exportation from COMTRADE, distance and geographical variables from the CEPII. This market access is instrumented using genetic distance in 1500 from Spolaore and Wacziarg (Reference Spolaore and Wacziarg2018). Data on democracy are from Acemoglu, Naidu, Restrepo, Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017) and instrumented with their instrument, denoted ANRR, based on waves of democratization and with the mortality rate of colonizers provided by Albouy (Reference Albouy2012). All estimations include time effects and individual fixed effects.

Column (3) presents the first stage regressing market access on its instrument with various controls (see equation (13)). As required, the instrument of trade openness based on genetic distance significantly explains market access. This result confirms, in a different way, the main finding of Giuliano, Spilimbergo and Tonon (Reference Giuliano, Spilimbergo and Tonon2014) that genetic distance captures transportation costs between countries.

Columns (2) and (4) propose a simple test to verify that there are no other problems of endogeneity in this strategy by analyzing whether market access explains the current level of democracy. Indeed the causal link of globalization on democracy has been much discussed by economists and political scientists (e.g. Milner and Mukherjee, Reference Milner and Mukherjee2009) and not directly addressed in this paper until now. Our result confirms the finding of this literature that trade openness does not explain the process of democratization (when adequate IV are used).

Column (3) shows that the instrument of market access has no impact on institutions, while column (4) shows that the instrument of institutions does not impact on market access.

Columns (5), (6), (7) and (8) present exactly the same specification with a different measure of democracy. We use the variable Polity 2 instead of the democracy variable of ANRR. Results are not affected by this change.

Results, second stage

Table 5 (column (1)) shows the results obtained by estimating the second stage described by equation (14). Column (2) presents a “half-naive” estimation, where democracy is instrumented but market access is not. Column (3) presents the symmetrical estimation, where market access is instrumented while democracy is not.

Table 5.

Trade, institutions and urban primacy

Note: 2SLS procedure, ^a: significant at 1%, ^b: significant at 5%, ^c: significant at 10%. Market access is computed from the estimation of a gravity equation using bilateral exportation from COMTRADE, distance and geographical variables from the CEPII. This market access is instrumented using genetic distance in 1500 from Spolaore and Wacziarg (Reference Spolaore and Wacziarg2018). Data on democracy are from Acemoglu, Naidu, Restrepo, Robinson (Reference Acemoglu, Naidu, Restrepo and Robinson2017) and instrumented with their instrument, denoted ANRR, based on waves of democratization and with the mortality rate of colonizers provided by Albouy (Reference Albouy2012). All estimations include time effects and individual fixed effects.

In all estimations reported in Table 5, whatever the specification used, we find that the indicator of democracy, instrumented by past mortality rates and by waves of democratization, favors agglomeration in the biggest city.Footnote ²⁷ These results indicate that the negative impact of democracy found by Ades and Glaeser (Reference Ades and Glaeser1995) and Henderson and Wang (Reference Henderson and Wang2007) may be due to a lack of control and/or to endogeneity bias. Another possibility, which is more interesting, is that there is something specific in our sample of countries. Ades and Glaeser (Reference Ades and Glaeser1995) for instance consider a sample of 70 countries without considering the history of these states. Our analysis is slightly different, since we focus our research exclusively on countries colonized by Europeans 500 years ago, following the arguments of AJR and of many historians that the long-run effects of colonization are still unclear. As a result, it is possible that countries not colonized by Europeans have developed institutions that shape their economic geographies differently.

Concerning the second hypothesis, market access is never significant, which reveals how the previous results arrived at without controls on time-varying institutions and without an exogenous measure of trade openness were misleading. This is in accordance with the skeptical view of Ades and Glaeser (Reference Ades and Glaeser1995) and Henderson and Wang (Reference Henderson and Wang2007) about the impact of international trade on the urban pattern. The same estimations are done in columns (4), (5) and (6) but with the variable Polity 2 and based on the first step of columns (5) and (6) of Table 4. The estimation of equation (14) in column (4), which is the most careful estimation correcting for the endogeneity of market access and of institutions, provides the same result as column (1).Footnote ²⁸