Journal Issue
Share
Article

Income and Democracy

Author(s):
Anke Hoeffler, Robert H. Bates, and Ghada Fayad
Published Date:
December 2012
Share
  • ShareShare
Show Summary Details

Introduction

Writing in 1959, Seymour Martin Lipset reported a strong and positive correlation between income per capita and democracy in a global cross section of nations (Lipset, 1959). Doing so, he not only laid the foundations of modernization theory in comparative politics but also defined a major portion of the contemporary agenda in political economy, with its focus on the relationship between political institutions and economic development.

Lipset’s finding invites a dynamic and causal interpretation. It was therefore startling when estimating Markov transition models, Przeworski et al. (2000) failed to find a significant relationship between the level of income per capita and the likelihood of transition to democracy. While Boix and Stokes (2003) and Epstein et al. (2006) have challenged Przeworski et al.’s finding, it has subsequently been replicated by (Acemoglu, Johnson, Robinson and Yared, 2008) (henceforth AJRY). In our contribution, we focus on this last article and challenge their result.

In doing so, we report the presence, rather than the absence, of a significant statistical relationship between income and democracy for a group of countries that have been neither consistent democracies or non-democracies over the past 50 years. And the relationship we find is negative. To be noted is that AJRY also reported, but failed to comment upon, negative coefficients in their estimates.2

We take as our point of departure the important critique of (Gundlach and Paldam, 2008) (GP hereafter), who argue that by including year and country fixed effects AJRY purged from their panels useful information, thereby predisposing them to fail in their search for a relationship between income and democracy. In mounting this critique, GP highlight an important methodological dilemma: Including country specific fixed effects purges informative variation from the data; but excluding them introduces omitted variable bias. We confront – and surmount – this dilemma. Employing an augmented version of the Pooled Mean Group (PMG) estimator of Pesaran et al. (1999) we account for both country and year effects while relaxing the assumption of cross-sectional parameter homogeneity. Even while controlling for (time-invariant) omitted variables, we thereby extract information from sources of variation that were previously ignored.

In addition, inspired by the literature on the “rentier state” (Mahdavy, 1970; Beblawi and Giacomi, 1987; Chaudry, 1994) and the “political resource curse” (Barro, 1996; Barro, 1999, Ross, 1999, Ross, 2001), we decompose overall per capita income into its resource and non-resource components and find that the source as well as the level of income matters: the larger the portion originating from natural resource rents, the lower the level of democracy.

I. Additional Background

Given its central place in the literature, Lipset’s law claims the central position in this paper. Our research resonates with two other literatures as well. One, which has already been mentioned, addresses the resource curse; the second focuses on the relationship between economic growth and political instability.

As noted by Ross (2001), in ″rentier states″, governments with large oil revenues reduce the level of political discontent by maintaining low taxes and high levels of public benefits. In addition, they spend more on internal security, forestalling the formation of political organizations inclined to demand political rights. Lastly, as, Ross states: ″if resource-led growth does not lead to higher education levels and greater occupational specialization”, as has commonly been the case,” it should also fail to bring about democracy″ (Ross, 2001, p.336/337).

High income might co-vary with low levels of democratization for a second reason: growth reversals might spark political protests which result in the overthrow of incumbent regimes (Burke and Leigh, 2010). In the case of democracies, one party might supplant another in power; but in the case of authoritarian governments, the authoritarian regime might give way to a democracy, leading to an emerging relationship between low national income and democratic governance.

Haggard and Kaufman (1997) provide qualitative evidence that growth declines weaken authoritarian regimes, reducing their ability to trade political benefits for political deference (see also Geddes, 1999). Burke and Leigh (2010) develop a game theoretic model that illustrates how economic contractions can reduce the opportunity costs of political protest, including demands for democratic reforms. And in their survey of the literature on growth, Alesina and Perrotti (1994) use data on the average annual per capita rate of growth of GDP with and without government changes, documenting that income growth is lower in years with government change, still lower in years with major change, and lowest in years with coups. They conclude that ″transitions from dictatorship to democracy, being associated with sociopolitical instability, should typically be periods of low growth″ (Alesina and Perotti 1994, p. 359). Londregan and Poole (1990) estimate a simultaneous equation model and similarly find that income per capita and economic growth are important determinants of coups: poverty and low growth increase the likelihood of coups. Over time, this channel too would yield a relationship between low income and democracy.

While our work targets Lipset’s law and thus the long-run relationship between income per capita and democracy, it relates to these literatures as well.

II. Data and Methods

Before testing Lipset’s law, we first introduce our data and our methods.

A. Data

We use the Penn World Tables’ (PWT 6.3) chain weighted real GDP per capita series and the Polity IV democracy index which distributes over a range spanning the interval between perfect autocracies (score of-10) and perfect democracies (score of 10). Based on a sample of 105 countries, Figure 1 shows that on average incomes and polity scores have risen over time. While incomes have grown relatively smoothly, in the late 1980s, the polity index jumped discontinuously from -0.4 in 1989 to 1.9 in 1992. As seen in Figure 2, which is based on the same sample, there are important regional differences in the movement toward democracy. While Asia and Latin America’s Polity scores started on an upward trend in the late 1970s, Latin America/Asia democratized prior/post to the fall of Communism. Africa and the Middle East both democratized after 1990; their polity scores then diverged, with those in Sub-Saharan Africa improving more rapidly.

Figure 1.Democracy and Income Global Averages, 1960-2008

Sources: Penn World Tables 6.3, Polity IV Project and authors’ calculations.

Figure 2:Regional Average Polity Scores, 1960–2008

Source: Polity IV Project and authors’ calculations.

Our sample includes 105 countries. Its size and composition is limited by the method we employ. Because the heterogeneous PMG estimator (discussed below) computes coefficients for each country separately, we can include only countries with long time series 3 and must exclude countries with no time variation in the dependent variable.4

When we disaggregate income per capita by sources of income (resource and non-resource), we use the United Nations Statistics Division National Accounts Database. The data are available for 104 out of the 105 countries in our sample, but run only from 1970–2007. The data classify GDP into several categories, one of which includes Mining and Quarrying. We use this category as a measure of resource wealth. Data description in terms of definition and sources are provided in the Appendix.

B. Methods

Estimation technique

The PMG estimator allows intercepts, slope coefficients and error variances to vary across panel members. More specifically, it allows the short-run coefficients to vary across countries, while restricting long-run relationships to be homogeneous. In the context of this research, the estimator “assumes” that in the short run—or while adjusting to a common long-run equilibrium—each country’s political institutions respond differently to income shocks.

Because it allows for heterogeneous intercepts, the PMG estimator can incorporate country-specific fixed effects. But because it estimates the model for each country separately, it cannot allow the inclusion of year fixed effects. To correct for potential cross-section dependence in the estimated errors, we—as do Binder and Offermanns (2007)—therefore augment the model with the cross-sectional averages of the dependent variable and regressors.

We are particularly concerned with the possibility of error arising from two sources: unobserved sources of time specific heterogeneity and endogeneity bias.

Time Specific Heterogeneity: Given rapid increases in the global integration of financial and product markets, an economic shock could affect all units of a given cross section of a global sample. Heterogeneity would then introduce cross-section dependence in the errors of panel members, i.e. the errors would become contemporaneously correlated. In addition, if the same latent factor affects regressors and errors, the shock could be correlated with regressors, rendering conventional regressors inconsistent (Coakley et al., 2006).

Given the difficulty of modeling error arising from cross-section dependence in panel data, few have attempted to do so (Phillips and Moon, 1999). As the PMG estimator estimates the model separately for each cross-section, we cannot employ time dummies to account for common temporal effects. Given our focus on the importance of heterogeneity, we are reluctant to assume common country responses to global factors through simple cross-sectional demeaning which attempts to extract the factors prior to estimation. Instead, we wish to include proxies of these common factors as right-side variables and employ PMG estimation to capture country-specific differences in short-run adjustments to them. We therefore follow Binders and Offermanns (2007) by allowing short-run heterogeneity in responses to both observed and unobserved factors, and thus augment the PMG model by proxies for global factors: the cross-sectional averages of all model’s variables.

Endogeneity. In Table 1 we present Granger causality tests. The data suggest the presence of reverse causality for our sample, which, if unattended, would introduce endogeneity bias in our estimates of the coefficient on income.

Table 1:Granger Causality Tests
Null hypothesisObservationsLagsF-statProbability
Democracy does not Granger cause income453235.470.001
Income does not Granger cause democracy453236.870.000
Source: Authors’ calculations.Note: In testing whether democracy Granger causes income, income is regressed on lags of income and democracy, and the reported F-stat is a Wald-type test of the joint significance of all estimated coefficients on such lags. We also report the probability of rejecting the null hypothesis.
Source: Authors’ calculations.Note: In testing whether democracy Granger causes income, income is regressed on lags of income and democracy, and the reported F-stat is a Wald-type test of the joint significance of all estimated coefficients on such lags. We also report the probability of rejecting the null hypothesis.

We address endogeneity concerns in two ways. First, our preferred PMG estimator model is augmented with lags of the regressors and dependent variable to minimize the resultant bias and to ensure that the regression residuals are serially uncorrelated. Pesaran (1997) and Pesaran and Shin (1999) show that, for inference on the long-run parameters, sufficient augmentation of the order of the ARDL model can simultaneously correct for the problem of residual serial correlation and endogenous regressors. In choosing the optimal lag structure, we apply the Akaike Information Criterion (AIC) or the Schwartz Bayesian Criterion (SBC)). In doing so, we are constrained to a maximum of three lags by our time series dimension and number of our regressors.5 Note that any remaining bias works against our conclusion: since our estimates yield negative signs, the coefficients should be even more negative once this bias is taken into account.

Second, we use alternative heterogeneous panel cointegration techniques that correct for endogeneity, namely the group mean Fully Modified OLS (FMOLS) of Pedroni (2000) and Dynamic OLS (DOLS) of Pedroni (2001).6 Both methods allow for regressors’ complete endogeneity, and treat all parameters, i.e. dynamics and cointegrating vectors, as heterogeneous across panel members. With group mean DOLS/FMOLS, the sample coefficients are simple averages of the member-specific DOLS/FMOLS estimators. More specifically, FMOLS runs a static OLS with fixed effects for each panel member individually and uses the estimated residuals to (non-parametrically) build member-specific adjustment terms, which are then used to correct for each member’s endogeneity. Differences in regressors are used as internal instruments. Instead, DOLS individually corrects for endogeneity parametrically by running OLS with fixed effects for each panel member including leads and lags of differenced regressors.7

The Model

To summarize formally, let dit represent democracy and yit represent income per capita for country i at time t, and y¯t=N1Σi=1Nyit,d¯t=N1Σi=1Ndit respectively represent their cross-sectional averages. The ECM with p lags on both the dependent and explanatory variables then is:

Crucially, the error term εit is identically and independently distributed across i and t even in the presence of common time effects. Country intercepts -- unobserved country heterogeneity – are captured by the term μi.

The second part of equation (1) includes the lagged changes of income and democracy; the coefficients represent the short-run adjustment terms and are assumed to vary across countries. We do not report the short-run coefficients below. The first part of equation (1) captures the common long-run relationship between income and democracy. The slope coefficients—β, η, and α—measure the long-run response of democracy to income, world income and world democracy. φ is the error correction coefficient and indicates the speed of adjustment If the system is dynamically stable and converges to a long-run equilibrium, then this coefficient will be negative and less than one in absolute value. We report these long-run coefficients below.

Starting with an initial estimate of the long-run parameters, the PMG estimator calculated estimates of error-correction and other short-run coefficients (including country-specific intercepts and error variances) as the averages of the estimated parameters for each cross-section.

It then employs these average estimates to update its estimates of the long-run parameters, repeating the process until convergence is achieved.

Note that we also report (in Table 3) the mean group (MG) estimator (Pesaran and Smith, 1995) which allows for complete (short-run and long-run) parameter heterogeneity across panel cross-sections. If the slope coefficients are heterogeneous, the MG estimator is consistent. Since our cross-sectional dimension is large, the MG estimator is less likely to be biased by outliers. The mean group estimator does not take into account that some economic conditions tend to be common across countries in the long run, however. The PMG estimator does, and so captures efficiency gains from assuming common long-run relationships while at the same time allowing for heterogeneous short-run dynamics. Using the difference between the two sets of estimates, we employ a Hausman-type test to assess the assumption of long-run homogeneity.

Table 2:Reproducing AJRY with Pooled OLS and Fixed Effects; Annual Data 1960–2000
Dependent variable: DemtAJRY SamplePMG Sample
Pooled

1
FE

2
FE

3
Pooled

4
FE

5
FE

6
Demt-10.961***

(0.004)
0.897***

(0.011)
0.863***

(0.013)
0.949***

(0.005)
0.898***

(0.011)
0.854***

(0.013)
Log income per capitat-10.114***

(0.027)
0.407***

(0.112)
-0.110

(0.116)
0.161***

(0.038)
0.522***

(0.140)
0.002

(0.130)
Obs493349334933383638363836
Countries153153153105105105
R-squared0.940.940.940.910.800.81
Country FENOYESYESNOYESYES
Time FENONOYESNONOYES
Source: Authors’ calculations.Notes: Robust standard errors (clustered at the country level) are reported in parentheses. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels All regressions include a constant. AJRY baselines specification includes year dummies and their distinction between pooled and fixed effects estimation is based on whether country fixed effects are included or not. We do not report regression results for when only annual dummies are accounted for but find that just like the AJRY results, the coefficient on lagged income per capita in this case is positive and significant and does not significantly differ in magnitude from its counterpart with no year and country dummies.
Source: Authors’ calculations.Notes: Robust standard errors (clustered at the country level) are reported in parentheses. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels All regressions include a constant. AJRY baselines specification includes year dummies and their distinction between pooled and fixed effects estimation is based on whether country fixed effects are included or not. We do not report regression results for when only annual dummies are accounted for but find that just like the AJRY results, the coefficient on lagged income per capita in this case is positive and significant and does not significantly differ in magnitude from its counterpart with no year and country dummies.
Table 3:Augmented PMG Estimation; Overall Sample (N=105); 1955–2007
Dependent variable: Polity IV measure of democracy
Long-run coefficientsPMG

1
MG

2
Hausman Test

3
Log income per Capita-1.239***

(0.153)
0.39

(1.368)
1.44

[0.23]
World democracy0.800***

(0.029)
0.926***

(0.143)
0.80

[0.37]
World output3.059***

(0.547)
0.293

(2.958)
0.90

[0.34]
Joint Hausman test2.39

[0.50]
Error correction coefficient-0.264***

(0.029)
-0.469***

(0.034)
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion subject to a maximum lag of three. World democracy and world output are respectively the cross-sectional averages of democracy and output, which we take as proxies of the common unobserved global shocks.
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion subject to a maximum lag of three. World democracy and world output are respectively the cross-sectional averages of democracy and output, which we take as proxies of the common unobserved global shocks.

III. Estimation and Results

We begin by employing an extended version of our dataset to reproduce the results of AJRY and GP. This dataset includes, in addition to our overall sample, all the countries that were dropped due to the restrictions imposed by PMG (as discussed above). This results in a sample of 153 countries for the annual data panel, and 129 countries for the five- and ten-year data panels, over the 1960–2000 period. As did AJRY, we find (Table 2, columns 1-3) that the coefficient on the income variable is positive and significant, when estimated from pooled data using ordinary least squares, but does not significantly differ from zero when including time and country fixed effects.8 We also find that when we estimate their model employing our smaller PMG sample (Table 2, columns 4-6), their findings remain unchanged. Insofar as our results differ from those of earlier researchers, then, it is not because we are making use of different data. We next provide econometric evidence in support of that.

Table 3 presents the major (PMG) results derived from our model. Our results are reported in the first column of Table 3, while MG estimates appear in the second. The Hausman test in column 3 result testifies to the validity of the long-run homogeneity restrictions imposed by the PMG estimator.9 The coefficients generated by the pooled mean estimator suggest that income is negatively and significantly related to democracy. Given that the model is linear log, they suggest that a 10 percent increase in per capita income leads in the long run to a roughly 0.12 unit decrease on the polity scale. While AJRY report negative coefficients for the relationship between income and democracy, they refrain from commenting upon them. We instead conclude that not only is there no positive relationship between income and democracy; the relationship is negative.

In contrast to the coefficients on per capita income, those on global changes in output and democratization over the sample period positively affected the level of democracy: both are significant and large. When the global democracy score increases by one unit, the democracy score improves by an average of 0.8 units; and on average, a 10 percent increase in world income improves the democracy score by 0.3 units.

The error correction coefficient is significant; it suggests about 26 percent of error correction in the single-period response of democracy to a departure from its long-run equilibrium value as predicted by the level of per capita income. These results are robust to the optimal lag selection criterion (AIC vs. SBC), to the number of lags, to the cross-sectional demeaning of the data, and to whether Sub-Saharan Africa is included or not in the overall sample.

Recall that GP found that when country fixed effects alone were included in the model, the coefficient on lagged income per capita was significant and positive. AJRY also reported negative and often significant coefficients for income in models which include both country and annual dummies.10,11

As discussed earlier, we also run the FMOLS and DOLS estimators on our sample. Results, shown in Table 4, point in the same direction: coefficients on income per capita are yet again found to be negative and statistically significant.12

Table 4:FMOLS and DOLS; Overall Sample (1955–2007; N=105)
Dependent variable: DemocracyDOLSFMOLS
Log of income per capita-2.293***

(0.189)
-0.503**

(0.197)
Countries105105
Country and year FEYESYES
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Variables were cross-sectionally demeaned prior to estimation to account to ensure error cross-sectional independence. Numbers reported in parentheses are standard errors.***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels.
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Variables were cross-sectionally demeaned prior to estimation to account to ensure error cross-sectional independence. Numbers reported in parentheses are standard errors.***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels.

Discussion

When analyzing the difference between our findings and those of AJRY, it is useful to turn to Table 5, which reports the results we secure when we employ the pooled error correction OLS model to regress democracy on its lags and on the level of income per capita (also with lags, and country and year fixed effects) while using the PMG sample. As can be seen, we then get negative and significant long-run coefficients on income per capita in our PMG sample, and the magnitudes are similar to our long-run PMG coefficients (discussed below). However, estimating the pooled error correction model while using the AJRY (bigger) sample 13 yields long-run coefficients on income per capita that are insignificant, regardless of the number of lags.14

Table 5:Pooled Error Correction Model by OLS With Country and Year Fixed Effects on the PMG Sample (1960–2007; N=105)
Dependent variable: Democracy Longrun coefficientLags=2Lags=3Lags=4Lags=5
1234
Log of income per capita-1.325*-1.236-1.595*-1.683*
(p-value=0.07)(p-value=0.11)(p-value=0.07)(p-value=0.06)
Error correction coefficient-0.141***-0.145***-0.142***-0.144***
R-squared0.850.850.850.85
Countries105105105105
Observations4453434842434138
Country and year FEYESYESYESYES
Source: Authors’ calculations.Note: we do not report the short-coefficients on income per capita (which are statistically insignificant). The p-values for the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance at which we can reject that the long run-coefficient is zero.
Source: Authors’ calculations.Note: we do not report the short-coefficients on income per capita (which are statistically insignificant). The p-values for the long-run coefficients are calculated with the non-linear test procedure “testnl” in Stata, and indicate the level of significance at which we can reject that the long run-coefficient is zero.

The difference between our results and those of AJRY thus arise from 1) our estimation methods, which exploit both the dynamic and heterogeneous properties of the data and 2) our samples, which exclude both consistent autocracies and consistent democracies, as noted above. The differences in the samples prove consequential: by restricting our sample to country-years that experienced changes in both their incomes and polity scores, i.e. countries that witnessed movements either away or towards more democracy, we are able to detect the relationship between income and democracy, one that turned out to be significantly negative.15 Both the sample choice and the methodology thus led us to our results.

It is also worth noting that in relation to Figure 2, where global income and democracy appear to be negatively correlated up to 1985, we check whether our results on the negative relationship between income and democracy are driven by the pre-1985 period. Estimating both the PMG and the pooled OLS ECM for the sub-period 1985-2007, results (not reported here) maintain the same negative relationship, with the only difference that global output in the PMG model is now significantly negatively related to democracy.

IV. Digging Deeper

In this section, we explore the possible impact of additional sources of variation: variation in the composition of the national income and regional relationships between income and democracy.

The sectoral composition of output: Returning to the literature on the “rentier state” (Mahdavy, 1970; Beblawi and Giacomi, 1987; Chaudry, 1994) and the “resource curse” (Barro, 1996, Barro, 1999; Ross 1999, Ross, 2001), we augment our baseline regression with the World Bank measure of natural resource rents as a percent of GDP. Doing so reduces our overall sample to 98 countries over the period 1970-2007. As seen in Table 6, we too find a negative and significant coefficient for the relationship between resource rents and democracy.

Table 6:Augmented PMG Estimation with Resource Rents; N=98; 1970–2007
Dependent variable: Polity IV measure of democracy Long-run coefficientsPMGMGHausman Test
123
Log income per capita-0.651***

(0.157)
-2.500

(4.726)
0.19

[0.67]
Natural resource rents to GDP-0.021***

(0.005)
-0.141

(1.440)
0.01

[0.93]
World democracy1.211***

(0.035)
1.225***

(0.322)
0.00

[0.97]
World output-8.554***

(1.268)
-4.031

(5.799)
0.64

[0.42]
World rents0.201***

(0.038)
-0.241

(0.382)
1.35

[0.25]
Joint Hausman test4.2

[0.52]
Error correction coefficient-0.355***

(0.039)
-0.664***

(0.057)
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion subject to a maximum lag of three. World democracy, world output and world rents are respectively the cross-sectional averages of democracy, output, and natural resource rents to GDP which we take as proxies of the common unobserved global shocks.
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values. ***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion subject to a maximum lag of three. World democracy, world output and world rents are respectively the cross-sectional averages of democracy, output, and natural resource rents to GDP which we take as proxies of the common unobserved global shocks.

Table 7 decomposes national income into two components: that deriving from natural resources and that deriving from other sources. Our sample now consists of the 102 countries over the period 1970-2007. To highlight the results of interest, we refrain from reporting the coefficients on the cross-sectional averages. Column 1 of Table 7 reproduces the specification employed in Table 3, but estimated from the current sample. The coefficients of interest remain roughly the same as that in Table 3. Columns 2, 3, and 4 report the PMG coefficient on resource and non-resource GDP per capita, first separately and then combined. The results confirm that it is only the resource proportion of income per capita that is negatively and significantly related to democracy.

Table 7:Augmented PMG Estimation with Sectoral Output; N=102; 1970–2007
Dependent variable: Polity IV measure of democracy
Long-run PMG coefficients1234
Log Income per capita-1.228***

(0.259)
Log of resource GDP per capita-0.306***

(0.081)
-0.295***

(0.081)
Log of non-resource GDP per capita0.710**

(0.271)
0.518*

(0.272)
Error correction coefficient-0.305***

(0.034)
-0.248***

(0.029)
0.245***

(0.030)
-0.307***

(0.034)
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values.***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion. All regressions include cross-sectional averages of the dependent variable and all regressors. Hausman test results for the coefficients of interest, not reported here, fail to reject the null of long-run cross-section parameter homogeneity.
Source: Authors’ calculations.Notes: All equations include a constant country-specific term. Numbers reported in parentheses are standard errors. Numbers reported in brackets are p-values.***, **, and * indicate significance respectively at the 1, 5, and 10 percent levels. We use the Schwartz Bayesian optimal lag selection Criterion. All regressions include cross-sectional averages of the dependent variable and all regressors. Hausman test results for the coefficients of interest, not reported here, fail to reject the null of long-run cross-section parameter homogeneity.

V. Conclusion

In this paper we have scrutinized Lipset’s law, which states that there is a positive and significant relationship between income and democracy and that higher incomes lead to democratization (Lipset, 1959). Lipset’s law is commonly regarded as the foundation of modernization theory in comparative politics and the debate over it shapes the contemporary agenda in development economics, with its focus on the relationship between political institutions and economic development.

We revisit this debate and use dynamic panel data methods to examine the direction of causation as well as the short and long run relationship between democracy and income. Our tests fail to establish the direction of causality: income ‘Granger’ causes democracy and vice versa. With this in mind, we turned to modelling democracy as a function of income. We use the augmented PMG estimator (Pesaran et al., 1999) which enables us to examine the short/long run relationship while allowing for country and year effects as well as for parameter heterogeneity across panel members. Accounting for the dynamics, we then find a significant and negative relationship between income and democracy for the group of countries that have been neither consistent democracies nor consistent non-democracies over the sample period. Doing so, we draw information from two sources of variation overlooked by previous scholars. One is cross country variation in the short term responses to income shocks; another is variation in the structure and composition of income. In countries that receive little or no income from resources the relationship between democracy and income is positive and significant. In resource rich countries, the reverse is true: higher incomes bear a negative relationship with democracy.

Appendix
Table A:List of Countries and Time Periods with Available Polity and GDP per Capita Data
Afghanistan1970–2000Kenya1963–2007
Albania1970–2007Korea, Rep.1955–2007
Algeria1962–2007Kuwait1970–2007
Angola1975–2007Laos1970–2007
Argentina1955–2007Lebanon1970–2007
Bahrain1971–2006Lesotho1966–2007
Bangladesh1972–2007Liberia1955–2007
Benin1960–2007Madagascar1960–2007
Bolivia1955–2007Malawi1964–2007
Botswana1966–2007Malaysia1957–2007
Brazil1955–2007Mali1960–2007
Bulgaria1955–2007Mauritania1960–2007
Burkina Faso1960–2007Mauritius1968–2007
Burundi1962–2007Mexico1955–2007
Cambodia1970–2007Mongolia1955–2007
Cameroon1960–2007Morocco1956–2007
Central African Rep1960–2007Mozambique1975–2007
Chad1960–2007Nepal1955–2007
Chile1955–2007Nicaragua1955–2007
China1955–2007Niger1960–2007
Colombia1955–2007Nigeria1960–2007
Comoros1975–2007Oman1955–2007
Congo, Dem. Rep.1960–2007Pakistan1955–2007
Congo, Rep.1960–2007Panama1955–2007
Cote d’Ivoire1960–2007Paraguay1955–2007
Cyprus1960–2007Peru1955–2007
Djibouti1977–2007Philippines1955–2007
Dominican Republic1955–2007Poland1970–2007
Ecuador1955–2007Portugal1955–2007
Egypt, Arab Rep.1955–2007Romania1955–2007
El Salvador1955–2007Rwanda1961–2007
Equatorial Guinea1968–2007Senegal1960–2007
Ethiopia1955–2007Sierra Leone1961–2007
Fiji1970–2007Solomon Islands1978–2007
France1955–2007Somalia1970–2007
Gabon1960–2007South Africa1955–2007
Gambia1965–2007Spain1955–2007
Ghana1960–2007Sri Lanka1955–2007
Greece1955–2007Sudan1956–2007
Guatemala1955–2007Swaziland1970–2007
Guinea1958–2007Syrian Arab Rep1961–2007
Guinea-Bissau1974–2007Tanzania1961–2007
Guyana1966–2007Thailand1955–2007
Haiti1955–2007Togo1960–2007
Honduras1955–2007Trinidad & Tobago1962–2007
Hungary1957–2007Tunisia1961–2007
India1955–2007Turkey1955–2007
Indonesia1955–2007Uganda1962–2007
Iraq1970–2002Uruguay1955–2007
Iran, Islamic Rep.1955–2007Venezuela, RB1955–2007
Israel1955–2007Zambia1964–2007
Jamaica1959–2007Zimbabwe1970–2007
Jordan1955–2007
Source: Penn World Tables 6.3 and Polity IV Project.
Source: Penn World Tables 6.3 and Polity IV Project.
Table B:Descriptive Statistics
VariablesObsCountriesMeanStd. Dev.MinMax
Tables 3, 4
Democracy5085105-1.06.9-10.010.0
Log of Real Income per Capita50761058.01.05.011.5
Table 5
Democracy369398-0.47.0-10.010.0
Log of Real Income per Capita3723988.11.05.011.5
Rents to GDP3724987.713.50.0116.5
Table 6
Democracy3834102-0.47.0-10.010.0
Log of Real Income per Capita38751028.11.05.011.5
Log of Resource Income per Capita38691025.21.8-2.411.1
Log of non-Resource Income per Capita38751028.01.05.010.5
Share of Resource Income387610210.013.90.077.5
Source: Authors’ calculations.
Source: Authors’ calculations.
Table C:Share of Resource GDP in Overall GDP
Albania22.1Korea, Rep.3.0
Algeria31.4Kuwait48.0
Angola39.5Laos4.6
Argentina4.7Lebanon3.5
Bahrain22.4Lesotho2.9
Bangladesh1.4Liberia12.1
Benin1.5Madagascar1.4
Bolivia12.7Malawi3.3
Botswana35.6Malaysia12.4
Brazil3.8Mali4.3
Bulgaria9.8Mauritania15.2
Burkina Faso2.2Mauritius2.2
Burundi0.7Mexico9.8
Cambodia2.3Mongolia15.5
Cameroon6.9Morocco5.8
Central African Rep6.1Mozambique2.4
Chad5.6Nepal1.3
Chile13.4Nicaragua2.7
Colombia6.3Niger7.3
Comoros1.2Nigeria27.2
Congo, Dem. Rep.12.7Oman47.6
Congo, Rep.34.8Pakistan4.7
Cote d'Ivoire2.8Panama3.1
Cyprus3.0Paraguay1.6
Djibouti5.3Peru12.5
Dominican Republic3.3Philippines3.7
Ecuador13.3Poland8.2
Egypt, Arab Rep.8.9Portugal2.8
El Salvador1.9Romania6.4
Equatorial Guinea29.0Rwanda1.3
Fiji3.7Senegal3.4
France2.1Sierra Leone13.2
Gabon42.2Solomon Islands1.0
Gambia1.3Somalia0.7
Ghana4.9South Africa13.3
Greece3.3Spain2.9
Guatemala2.5Sri Lanka2.6
Guinea18.2Sudan3.1
Guinea-Bissau0.7Swaziland4.8
Guyana13.1Syria13.0
Haiti1.5Tanzania8.5
Honduras2.1Thailand4.0
Hungary7.7Togo8.4
India4.3Trinidad & Tobago22.2
Indonesia12.7Tunisia7.9
Iran, Islamic Rep.20.2Turkey2.9
Iraq77.5Uganda2.6
Israel2.5Uruguay2.7
Jamaica9.0Venezuela, RB19.5
Jordan5.3Zambia16.5
Kenya3.8Zimbabwe6.7
Source: United Nations Statistics Division National Accounts Database and authors’ calculations.UN Statistics Division National Accounts Database which provides data from 1970–2007 on sectoral GDP shares for the following overall categories: 1. Agriculture, hunting, forestry, fishing ; 2. Mining, Manufacturing, Utilities; 3. Manufacturing; 4. Construction; 5. Wholesale, retail trade, restaurants and hotels; 6. Transport, storage and communication; 7. Other Activities. Category 2 (Mining, manufacturing and utilities) is an aggregation of economic activities of a. Mining and quarrying, b. Manufacturing and c. Utilities. The data available allows us to compute Mining, Quarrying and Utilities by subtracting Category 3 (Manufacturing) from Category 2 (Mining, Manufacturing, Utilities). We take this as our proxy of resource GDP. Unfortunately UN data on Mining and Quarrying alone involve short time series dimensions for the countries in the sample which does not allow us to estimate using our PMG method.
Source: United Nations Statistics Division National Accounts Database and authors’ calculations.UN Statistics Division National Accounts Database which provides data from 1970–2007 on sectoral GDP shares for the following overall categories: 1. Agriculture, hunting, forestry, fishing ; 2. Mining, Manufacturing, Utilities; 3. Manufacturing; 4. Construction; 5. Wholesale, retail trade, restaurants and hotels; 6. Transport, storage and communication; 7. Other Activities. Category 2 (Mining, manufacturing and utilities) is an aggregation of economic activities of a. Mining and quarrying, b. Manufacturing and c. Utilities. The data available allows us to compute Mining, Quarrying and Utilities by subtracting Category 3 (Manufacturing) from Category 2 (Mining, Manufacturing, Utilities). We take this as our proxy of resource GDP. Unfortunately UN data on Mining and Quarrying alone involve short time series dimensions for the countries in the sample which does not allow us to estimate using our PMG method.
Table D:Data Description and Sources
VariableDescriptionSource
Income per capitaData measured as log real GDP per capita (chain weighted method, income measured in constant USD) from Penn World Tables 6.3.http://pwt.econ.upenn.edu/
DemocracyPolity IV index ranging from-10 for perfect autocracies to +10 for perfect democracies.http://www.systemicpeace.org/polity/polity4.htm
Natural resource rentsExpressed in percent of GDP. Rents are measured as the market value of extracted material minus the average extraction cost. Natural resources include bauxite, coal, copper, forest, gold, iron, lead, lignite, natural gas, nickel, oil, phosphates, silver, tin and zinc.World Bank data:http://go.worldbank.org/OV4R25M150
Resource and non-resource income sharesResource income share is defined as the share of Mining and Quarrying in GDP. Non-resource income constitutes the rest.United Nations Statistics Division National Accounts Database.

Figure A:Country-Specific FMOLS Income Per Capita Coefficients, N=105

Source: Authors’ calculations.

References

    AcemogluDJohnsonSRobinsonJAYaredP. 2008. Income and Democracy. American Economic Review 98:808842.

    AlesinaAPerottiR. 1994. The Political Economy of Growth: A Critical Survey of the Recent Literature. World Bank Economic Review 8: 351371.

    BarroR. 1996. Democracy and Growth. Journal of Economic Growth 1: 127.

    BarroR. 1999. Determinants of Democracy. Journal of Political Economy 107: 158182.

    BeblawiHGiacomiL. 1987. The Rentier State.London, Croom Helm.

    BinderMOffermannsCJ. 2007. International Investment Positions and Exchange Rate Dynamics: a dynamic panel analysis. CFS Working Paper No. 2007/23.

    • Search Google Scholar
    • Export Citation

    BoixCStokesS. 2003. Endogenous Democratization. World Politics 55: 517549. Burke PGLeigh AK. 2010. Do Output Contractions Trigger Democratic Change? American Economic Journal: Macroeconomics 2124157.

    • Search Google Scholar
    • Export Citation

    ChaudhryKA. 1994. Economic Liberalization and the Lineages of the Rentier State. Comparative Politics 27125.

    CoakleyJFuertesAMSmithR. 2006. Unobserved Heterogeneity in Panel Time Series Models. Computational Statistics & Data Analysis 50: 23612380.

    • Crossref
    • Search Google Scholar
    • Export Citation

    EpsteinDBatesRHGoldstoneJKristensenIO’Halloran. S2006. Democratic Transitions. American Journal of Political Science 50: 551569.

    • Crossref
    • Search Google Scholar
    • Export Citation

    GeddesB. 1999. What Do We Know About Democratization After Twenty Years? Annual Review of Political Science 2: 115144.

    GundlachEPaldamM. 2008. Income and Democracy: A Comment on Acemoglu, Johnson, Robinson, and Yared, Kiel Working Paper No. 1458.

    HaggardSKaufmanR. 1997. The Political Economy of Democratic Transitions. Comparative Politics 29: 263283.

    HausmanJ. 1978. Specification Tests in Econometrics. Econometrica 46: 12511271. ImKSPesaranMHShinY. 2003. Testing for Unit Roots in Heterogeneous Panels. Journal of Econometrics 115: 5374.

    • Crossref
    • Search Google Scholar
    • Export Citation

    LipsetSM. 1959. Some Social Requisites of Democracy: Economic Development and Political Legitimacy. The American Political Science Review 53: 69105.

    LondreganJPooleK. 1990. Poverty, the Coup Trap, and the Seizure of Executive Power. World Politics 42: 151183.

    MahdavyH. 1970. The Patterns and Problems of Economic Development in Rentier States: The Case of Iran.In M. A. Cook (Ed.)Studies in Economic History of the Middle East.London: Oxford University Press1970.

    • Search Google Scholar
    • Export Citation

    NickellS. 1981. Biases in Dynamic Models with Fixed Effects. Econometrica 49: 14171426.

    PedroniP. 2000. Fully Modified OLS for Heterogeneous Cointegrated Panels. In B.H. Baltagi (Ed.)Nonstationary Panels Cointegration in Panels and Dynamic Panels. Amsterdam: Elsevier.

    • Search Google Scholar
    • Export Citation

    Pedroni P. 2001. Purchasing Power Parity Tests in Cointegrated Panels. Review of Economics and Statistics 83: 727731.

    PedroniP. 2004. Panel cointegration: aymptotic and finite sample properties of pooled time series tests with an application to the PPP hypothesis. Econometric Theory 20: 579625.

    • Search Google Scholar
    • Export Citation

    PesaranMH. 1997. The Role of Economic Theory in Modelling the Long Run. The Economic Journal 107: 178191.

    PesaranMHSmithR. 1995. Estimating Long-run Relationships from Dynamic Heterogeneous Panels. Journal of Econometrics 68: 79113.

    PesaranMHShinY. 1999. An Autoregressive Distributed Lag Modelling Approach to Cointegration Analysis. In Strom S. (Ed.) Econometrics and Economic Theory in the 20th Century: the Ragnar-Frisch Centennial Symposium.Cambridge University Press.

    • Search Google Scholar
    • Export Citation

    PhillipsP.C.B.HansenB.E. 1990. Statistical Inference in Instrumental Variables Regression with I (1) Processes. Review of Economic Studies 57: 99125.

    • Crossref
    • Search Google Scholar
    • Export Citation

    PesaranMHShinYSmithR. 1999. Pooled Mean Group Estimation of Dynamic Heterogeneous Panels. Journal of the American Statistical Association 94: 621634.

    • Crossref
    • Search Google Scholar
    • Export Citation

    PhillipsPCBMoonHR. 1999. Linear Regression Limit Theory for Nonstationary Panel Data. Econometrica 67: 10571112.

    PrzeworskiA. AlvarezM.A. CheibubJ.A. LimongiF. 2000. Democracy and Development: Political Institutions and Well-Being in the World 1950–1990 Cambridge University Press: Cambridge UK.

    • Crossref
    • Search Google Scholar
    • Export Citation

    RossML. 1999. The Political Economy of the Resource Curse. World Politics 51:297322.

    RossML. 2001. Does Oil Hinder Democracy?World Politics 53: 325361.

    StockJ.H.WatsonM. 1993. A Simple Estimator of Cointegrating Vectors in Higher Order Integrated Systems. Econometrica, 614: 783820.

    • Search Google Scholar
    • Export Citation
1We thank Tim Callen, Andrew Berg, Camelia Minoiu, Carlo Sdralevich, and participants at the MCD Discussion Forum for helpful comments and discussions. We are also grateful for comments on a previous version of this paper by Tim Belsey, Rick van der Ploeg, and Steven Poelhekke. All remaining errors are our own.
2Their Table 2, cols. 3, 4 8, and 9; Table 3, cols 2, 3, 4, 8; and Table 4, cols 1, 2, 4,5,8, and 9.
3The countries we lose in this respect are: Armenia, Azerbaijan, Belarus, Croatia, Czech Republic, Eritrea, Estonia, Georgia, Kyrgyzstan, Kazakhstan, Latvia, Lithuania, Macedonia, Moldova, Namibia, Russia, Slovak Republic, Slovenia, Tajikistan, Turkmenistan, Ukraine, Uzbekistan, and Yemen.
4The countries dropped because they are either consistent democracies or consistent non-democracies are Australia, Austria, Belgium, Bhutan, Canada, Costa Rica, Cuba, Denmark, Finland, Ireland, Italy, Japan, Libya, Netherlands, New Zealand, Norway, Qatar, Saudi Arabia, Singapore, Sweden, Switzerland, United Arab Emirates, United Kingdom, United States, and Vietnam.
5To illustrate: Using SBC, we determine the lag order for each country, subject to a maximum lag of three; we then impose a homogeneous lag structure, using the most common of the country-specific lag orders. Note that another advantage of using the PMG ARDL approach is that there is no need for pre-testing our variables for the presence of unit roots. Pesaran et al. (1999) show the consistency of the PMG estimator in the case of I (0) and I (1) regressors.
6Developed as panel analogues of their time series versions respectively by Phillips and Hansen (1990) and Stock and Watson (1993).
7It is worth noting that under endogeneity and no cointegration, using OLS produces first-order bias, and external instruments are needed. However, under endogeneity and cointegration, OLS produces superconsistent estimates and second-order endogeneity bias (i.e. inconsistent estimates of standard errors). Internal instruments are then used (such as in FMOLS-DOLS).
8We only report results from the annual sample. We also reproduced but chose not to report their results with 5-year and 10-year data, because when allowing for fixed effects in short panels, the lagged dependent variable bias is large (Nickell, 1981).
9More specifically, the difference between both MG and PMG estimators is used to compute a Hausman-type statistic. Under the null hypothesis of long-run parameter homogeneity, both estimators are consistent, but the PMG is more efficient. When the true long-run parameters are instead heterogeneous, the MG estimator remains consistent while the PMG loses consistency.
10Their Table 2, cols. 3, 4 and 8; Table 3, cols 2, 3, 4, 8; and Table 4, cols 1, 2, 8, 4.
11When we run the PMG without accounting for time effects, the coefficient on income per capita is instead positive and significant.
12Please see Figure A in the Appendix for the country-specific coefficients on income per capita in the FMOLS regression.
13Which, unlike our own, includes countries with no time-variation in democracy variable;
14There is also evidence in this sample that democracy and income per capita are I (1) and cointegrated using Im, Pesaran and Shin (2003) unit root tests and Pedroni (1999, 2004) cointegration tests.
15This relationship was also picked up by a simple pooled (non-heterogeneous) OLS error correction model.

Other Resources Citing This Publication