Acemoglu, et al (2008) cast doubt on the robustness of the cross-country empirical relationship between income and democracy. They demonstrate that the strong positive correlation between income and democracy levels disappears when one controls for country fixed effects. In this comment, we reexamine the robustness of the income-democracy relationship. We make use of newer income data and employ estimation methods that explicitly allow for censored or binary measures of democracy. Our results show that when one uses both the new income data available and a properly nonlinear estimator, a statistically significant positive income-democracy relationship is robust to the inclusion of country fixed effects. However, we obtain notably different coefficient estimates using binary versus continuous or more finely graded democracy measures. Income has an economically important effect only for the binary measure of democracy. The continuous democracy measures represent weighted averages of a broad range of institutional and political processes associated with well-functioning democracies, while the binary measure essentially requires periodic elections of the legislature and chief executive, more than one party, and some alternation of power. This less exacting binary measure responds much more readily to income than the continuous measures. 0
We confront two primary issues: first, more and better-measured data on income has become available since the publication of AJRY (2008). This development is crucial because the inclusion of country fixed effects limits inference to that based on within-country variation in the data. The Penn World Tables 6.1 [Heston, et al, (2002)] (PWT 6.1) data used in AJRY (2008) results in smaller samples than those available in either the Penn World Tables 7.0 data set [Heston, et al, (2011)] (PWT 7.0), or the alternative Maddison (2003) data set. 1
Second, the measures of democracy used in AJRY (2008) are censored, violating the maintained assumptions under OLS. We respond to this issue by using the two-sided Tobit specification with the Chamberlain (1980) random effects correction used by AJRY (2009), as well as the Wooldridge (2005) method, which generalizes the Chamberlain estimator and parameterizes the fixed effects as well as the initial conditions in a dynamic panel. Finally, we report results using two measures of democracy. First, we report results for the continuous Polity measure used in AJRY (2008). Second, we also examine a binary variable proposed by Alvarez, et al (1996), and used in AJRY (2009) henceforth called Democracy/Dictatorship. 2
Our results with both democracy measures indicate a statistically significant contribution from income to democracy, even in the presence of country fixed effects. However, our estimated coefficient values suggest that the importance of this contribution differs considerably for the continuous and the binary measures; our continuous measure suggests a modest role for income in influencing democracy levels, while our binary measure indicates that income difference can have a strong impact on the probability that a country will be democratic or not. As we discuss below, we attribute this discrepancy to the fact that these measures used in the literature represent substantially different views of democracy. 3
II. Data and empirics We consider two measures of income: First we use the PWT (7.0) [Heston, et al (2011)] data set, which has greater coverage than the PWT 6.1 data set (189 countries vs 168 in PWT 6.1), and whose times series extend farther for most countries. We also use the Maddison data of per Capita GDP for 202 countries [Maddison (2003)]. Our Maddison data set also tends to have greater coverage per country than the Penn World Tables, facilitating the estimation of within-country variations. 2.2 Specification We consider the following specification d = α + β d + β logy + δ + θ + ε (1) it 1 it 1 2 it 1 t i it where d it and logy it 1 are democracy and log of GDP per capita for country i at period t respectively, δ t and θ i represent time and country fixed effects respectively, and ε it is a disturbance term, clustered by country. 4
Another source of concern is that our sample is a dynamic panel which includes lagged values of democracy. For the linear case, this problem can be treated with panel-gmm, such as the Arellano and Bond (1991) method used by AJRY (2008). However, our nonlinear specification presents a more challenging problem, as we cannot use differencing to eliminate the impact of country fixed effects. Taking the initial condition of the dependent variable as independent of unobserved heterogeneity will bias estimates in dynamic panels that are short in the time dimension. We therefore also follow Wooldridge (2005) in generalizing the Chamberlain approach by assuming that country fixed effects can be specified as a linear function of the mean sample value of the observable independent variables, the initial condition for the lagged dependent variables and country specific random effects. 5
For our dichotomous democracy measure, first we use a probit estimation using the Chamberlain (1980) random effects correction in order to compare our results with the ones in AJRY (2009). We also deal with dynamic panel issues for the dichotomous measure by again using the Wooldridge (2005) method described above. 6
3.1 Continuous measure of democracy (Polity) TABLE 1: Results with Polity measure of democracy OLS FE OLS Chamberlain Woolridge penn7.0 maddison penn7.0 maddison penn7.0 maddison penn7.0 maddison Dependent variable is the Polity measure of democracy dem (t-1) 0.785*** 0.795*** 0.435*** 0.459*** 0.906*** 0.917*** 0.805*** 0.573*** (0.03) (0.03) (0.05) (0.05) (0.02) (0.02) (0.03) (0.04) inc (t-1) 0.034*** 0.038*** 0.009 0.017 0.076*** 0.051* 0.078*** 0.065** (0.01) (0.01) (0.03) (0.03) (0.03) (0.03) (0.03) (0.03) inc(mean) -0.019 0.013-0.023 0.028 (0.03) (0.03) (0.03) (0.03) dem(t=0) 0.149*** 0.358*** (0.03) (0.05) Observations 998 1,067 998 1,067 998 1,067 998 1,067 R-squared 0.76 0.763 0.82 0.822 # of cty 158 156 158 156 Note: Estimation results using PWT 7.0 and Maddison income data as indicated, with Polity democracy measure. Estimation by ordinary least squares, with country fixed effects excluded and included, Tobit regression with Chamberlain (1980) adjustment, and Wooldridge dynamic panel estimator as indicated. See text for details. Standard errors in parentheses: errors are robust and clustered by countries. *** indicates statistical significance at 1% confidence level, ** at 5% confidence level, and *** at 10% confidence level. Country and time fixed effect estimates suppressed, but available upon request. We conclude that the relationship between income and democracy regains its statistical significance in the presence of country fixed effects when we use both the newer richer data sets and a nonlinear estimator appropriate for the censored measures of democracy. 7
In terms of the magnitude of the effects, consider that the mean value of the log of lagged income in our sample using the Wooldridge estimator is 8.14 for the PWT 7.0 data set and 7.87 for the Maddison data set, with standard deviations of 1.25 and 1.04 respectively. It follows that our point estimates using that estimator indicate that a two standard deviation increase in lagged income would result in an increase of democracy of 0.20 for the PWT 7.0 sample and 0.14 for the Maddison samples. These results indicate that substantial increases in income alone are unlikely to result in transitions from autocracy to values close to pure democracy using the Polity measure. For example, if a country with average income and a Polity score of 0 on our 0-to-1 scale doubled its income, its Polity score would increase only by 0.05 according to our point estimates. Such a score would still correspond to a strong autocracy. As such, while our results using the Polity measure indicate that very wealthy countries are likely to be substantively more democratic, it is unlikely that even rapid growth could lead a country from autocracy to democracy in a short period of time. 8
3.2 Dichotomous measures of democracy The results for the dichotomous measure of democracy are shown in Table 2. TABLE 2: Results with binary democracy/dictatorship measure of democracy OLS FE OLS Chamberlain Woolridge penn7.0 maddison penn7.0 maddison penn7.0 maddison penn7.0 maddison Dependent variable is Dem/Dict measure of democracy dem (t-1) 0.750*** 0.734*** 0.342*** 0.381*** 0.802*** 0.787*** 0.745*** 0.746*** (0.03) (0.04) (0.06) (0.06) (0.03) (0.03) (0.04) (0.04) inc (t-1) 0.045*** 0.061*** 0.058 0.07 0.205*** 0.145** 0.222*** 0.155** (0.01) (0.01) (0.05) (0.05) (0.07) (0.07) (0.08) (0.07) inc(mean) -0.098-0.005-0.13-0.022 (0.08) (0.07) (0.08) (0.08) dem(t=0) 0.243*** 0.163*** (0.05) (0.05) Observations 1,124 1,093 1,124 1,093 1,124 1,093 1,124 1,093 R-squared 0.683 0.662 0.769 0.747 # of cty 184 160 184 160 Note: Estimation results using PWT 7.0, and Maddison income data as indicated with binary Democracy-dictatorship measure. Estimation by ordinary least squares, with country fixed effects excluded, as well as Chamberlain-adjusted probit estimation and Wooldridge dynamic panel estimator, as indicated. Coefficient estimates for Chamberlain and Wooldridge estimations correspond to marginal effects. See text for details. Robust standard errors clustered by country in parentheses. *** indicates statistical significance at 1% confidence level, ** at 5% confidence level, and *** at 10% confidence level. Country and time fixed effect estimates suppressed, but available upon request. 9
However, using either our nonlinear Chamberlain-adjusted probit estimator or the Wooldridge estimator, lagged income again enters positively at least at a 5% confidence level for both the PWT 7.0 and Maddison income measures. Moreover, our point estimates are substantially larger than those that we obtained for our continuous Polity measure above. Our estimated coefficients indicate that the presence or absence of democracy is extremely sensitive to income: a two standard deviation increase in income would lead to 52% and 30% increases in the probability of democracy using the Chamberlain-adjusted probit estimator, and by 55% and 32% respectively using the Wooldridge estimator. For a country with average income, our Wooldridge estimation point estimates indicate that doubling income would result in a 15% increase in the probability of being a democracy using the PWT 7.0 data, and an 11% increase in the probability of being a democracy using the Maddison data. 10
These results suggest a far stronger role for income in the determination of democracy than those obtained for the nonlinear Chamberlain random effects estimator used by AJRY (2009). One reason for this discrepancy is the restriction in AJRY (2009) to the use of balanced panels, while our reported results above are for unbalanced panels. However, the Wooldridge estimator has been shown to perform well numerically in unbalanced panel estimation [Akay (2009)], leaving us confident in using our baseline results above. 11
In any event, the poor performances of income in explaining democracy in AJRY (2009) appears not to be associated with the restriction to balanced panels per se, but rather with the number of countries for which data is available for balanced panels of the mandated duration. Table 3 repeats the results for the binary measure using balanced panels from 1965 through 2000 for the PWT 7.0 and Maddison data sets, together with our nonlinear Chamberlain- and Wooldridge estimators. It can be seen, as in AJRY (2009), that lagged income enters insignificantly in both. As such, it would appear that our results above for income are not robust to the use of balanced panels. However, the use of balanced panels of this duration substantially reduces the number of countries in our sample. For the PWT 7.0 income measure, the number of countries is reduced from 158 to 79, while for the Maddison income measure, the number is reduced from 156 to 95. As such, the weaker balanced panel results may be attributed to the change in the sample, rather than the balanced panels themselves. To verify this conjecture, we examined a balanced panel of shorter duration, from 1980 through 2000. This shorter panel allows for the inclusion of a TABLE 3: Long and short balanced panel results with binary democracy/dictatorship measure larger set of countries, 135 using the PWT 7.0 sample and 127 using the Maddison data. 12
TABLE 3 Panel length: 1965-2000 Panel length: 1980-2000 Chamberlain Wooldridge Chamberlain Wooldridge penn7.0 maddison penn7.0 maddison penn7.0 maddison penn7.0 maddison Dependent variable is Dem/Dict measure of democracy dem(t-1) 0.718*** 0.757*** 0.665*** 0.717*** 0.762*** 0.786*** 0.673*** 0.702*** -0.053-0.044-0.051-0.047-0.037-0.035-0.044-0.043 inc(t-1) 0.076 0.099 0.077 0.108 0.206** 0.156 0.227** 0.186* -0.109-0.104-0.114-0.11-0.105-0.099-0.115-0.11 inc(mean) 0.109 0.099 0.093 0.076-0.093-0.056-0.138-0.112-0.112-0.108-0.119-0.116-0.107-0.103-0.121-0.119 dem(t=0) 0.187*** 0.148** 0.331*** 0.351*** -0.069-0.068-0.075-0.083 Observations 632 760 632 760 675 635 675 635 # of cty 79 95 79 95 135 127 135 127 Note: Estimation results using PWT 7.0, and Maddison income data as indicated with binary Democracy-dictatorship measure. Estimation by ordinary least squares, with country fixed effects excluded, as well as Chamberlain-adjusted probit estimation and Wooldridge dynamic panel estimator, as indicated. Coefficient estimates for Chamberlain and Wooldridge estimations correspond to marginal effects. See text for details. Robust standard errors clustered by country in parentheses. *** indicates statistical significance at 1% confidence level, ** at 5% confidence level, and *** at 10% confidence level. Country and time fixed effect estimates suppressed, but available upon request. 13
VI. Conclusion The introduction of fixed effects sets a high bar for establishing that income is a significant explanatory variable for democracy in crosscountry panel data, as inference is restricted to information available from within-country variations in the data. We find that the use of larger income data sets, such as PWT 7.0 or Maddison, combined with appropriate nonlinear estimators to account for the censoring of democracy measures can reestablish a statistically significant positive relationship between income and democracy in the presence of country fixed effects using either our continuous or discrete measures of democracy. Moreover, the statistical significance of income was maintained in balanced panels with sufficiently short duration to retain large cross-sections. 14
Still, while our results resurrect a statistically significant relationship between income and democracy, we also found that the importance of this relationship can be sensitive to the measure of democracy used. While we found that changes in income can have a substantive impact on the probability of being a democracy using Prezworski s dichotomous democracy measure, our coefficient estimates suggested that changes in income were unlikely to have large impacts on democracy as measured by the continuous Polity index. 15
The discrepancy between these results is not so surprising when one acknowledges that they measure different concepts. The probability of being a democracy in terms of having leaders chosen in meaningful elections, as in the binary measure of democracy, may more readily respond to income and wealth levels than the more inertial institutional and political processes associated with the quality and effective functioning of democracy, as measured by the Polity index. Our results suggest that economic development may have substantive impact on the probability of holding competitive elections, while changes in income may have less power to explain changes in the political institutions and processes that affect the degree of democracy, either among democracies or autocracies. 16