Democracy and Economic Growth
A statistical analysis of GDP per capita growth’s effects on a nation’s democratic strength
Introduction:
Over the past 200 years, democracy has become the most prominent form of governance in the world. There can be many potential causes for this phenomenon. Understanding these causes can help shed light on how the geopolitical state of the world has been, and continues to be, shaped by various conditions. In this paper, panel data from Acemoglu, et al. (2008), the CEPII, and the Augmented Freedom House Political Rights Index is examined via simple and multiple regressions to determine whether there is a relationship between democracy (measured using the Political Rights Index) and the growth of GDP per capita.
Examination of Data:
The panel data, which comes from Acemoglu, et al. (2008), the CEPII, and the Augmented Freedom House Political Rights Index (PRI), tracks several variables of interest for a random sample 147 countries between the years 1950 and 2000. For each country, the data was recorded every 5 years. This results in 11 total observations for each country. Tables 1, 1.1, 1.2, 1.3, and 1.4 summarize the relevant variables regarding each observation. Each variable has a different number of total observations since different countries were sampled over a period of 50 years. Nevertheless, the available data is sufficiently expository for an analysis. There are a few variables of interest: the PRI, the log of the real GDP per capita, the proportion of the population in different age groups, average schooling years, log of the total population (in thousands), whether the country was part of the Soviet Bloc, and whether the country was a former colony or not.
There are 1194 PRI observations, with mean PRI of 0.568. The standard deviation for the PRI is 0.358, which means that about 68% of all the PRIs lie between 0.21 and 0.926. Furthermore, the mean logarithm of real GDP per capita, which has 1199 observations, is 8.165 with a standard deviation of 1.037, which means that 68% of all the observed logs of real GDP per capita are between 7.128 and 9.202. Also, the average number of schooling years across 722 available observations was 4.458 and the standard deviation was 2.874, and the range of observations where 68% of observations lie is the same as before, with the bounds being the mean in addition to or without the standard deviation. In addition, the logarithm of a country’s total population in thousands is 8.682 with a standard deviation of 1.907 out of 1308 available observations.
Some other relevant data concerns a country’s status as a former colony and as a former Soviet Bloc member. These factors are pertinent since they relate to the two greatest geopolitical phenomena of the second half of the 20th century: the fragmentation of Europe’s colonial empires and the Cold War. Of 1617 available observations, 15% of them show affiliation with the Soviet Union and 70.1% of the observations were former colonies. Tables 1.1 and 1.2 show the mean and standard deviation differences between former colonies and non-former colonies. On average, former colonies display lower PRIs, lower logs of real GDP per capita, lower average schooling years, and lower log of total population than countries that were never colonies. Furthermore, countries that have been associated with the Soviet Union on average have lower PRIs, higher logs of real GDP per capita, higher average number of schooling years, and higher log of total population than countries that haven’t been associated with the Soviet Union. This data can be found in Tables 1.3 and 1.4.
Regression Analysis:
In total, ten regressions have been run on the data. All of them fundamentally examine the relationship between the PRIs and the lagging logarithm in real GDP per capita growth. The first six, which can be found in Table 2, gradually introduce new constraints in addition to the logarithm of real GDP per capita. The constraints that become gradually introduced are clustered standard errors, country fixed effects, year fixed effects, and demographic controls such as proportion of population in a particular age group, average schooling years, and the log of total population. Table 3 builds on four of these regressions by adding a colony interaction and a Soviet Union interaction to the lag of log real GDP per capita to see whether these groups of countries are impacted differently by changes in the lag of log real GDP per capita.
Table 2’s first regression is a simple pooled regression where no clustering of standard errors occurs. For this regression, and most of the following regressions, there are 147 examined countries with 1023 total observations. The coefficient here is 0.23, which means that a 1% increase in the lag of GDP per capita is associated with, on average, a PRI increase of 0.0023. Furthermore, this coefficient is significant at the 1% level, which means that this result is less than 1% likely if the lag of log real GDP per capita had no impact on the PRI. Also, the confidence interval shows that there is 95% confidence that the effect for a 1% increase in the lag of real GDP per capita is between 0.00211 and 0.0024.
Table 2’s second regression is a simple pooled regression but where the standard errors are clustered, which is the way all the other standard errors are treated in all the remaining regressions. The coefficient is still 0.23, with the same interpretation as before. Also, this result is still significant at the 1% level. However, the confidence interval has widened in that there is 95% confidence that the effect for a 1% increase is between 0.00201 and 0.0025. The confidence interval has widened because the clustered standard errors have resulted in more variance. Despite the estimate being less concentrated than before, it is still good to have clustered standard errors. This is because regressions are meant to make statements about the general population, which are countries in this case. Clustering the standard errors by country means that the results are more reflective of what happens to each country. Despite this improvement, there is still room to increase the accuracy of this regression. This regression does suffer from omitted variable bias. For instance, geography is a time-invariant factor that can affect both income (due to resource and trade availability) and democracy (due to proximity with other democratic countries, which can have a spillover political effect). Thus, a regression where the country is fixed might omit some of the bias.
Table 2’s third regression is a fixed effects regression where the fixed effects are considered by country. The coefficient there is 0.06, which means that a 1% increase in the lag of GDP per capita is associated with, on average, a PRI increase of 0.0006. Furthermore, this coefficient is only significant at the 10% level. Also, the confidence interval shows that there is 95% confidence that the effect for a 1% increase in the lag of real GDP per capita is between -0.00005 and 0.00115. Even though we have controlled for the country, there may be some factors that do not vary much by country but rather vary across time. An example may be the state of the Cold War as a whole, and how this impacts trade negotiations across the world (which affects income) as well as democracy (since the Cold War was, fundamentally, a political conflict between the democratic U.S.A. and the authoritarian USSR). Thus, time controls are also a good way to reduce omitted variable bias and produce a more causal regression.
The fourth regression in Table 2 is the same as the third regression but with year fixed effects in addition to country fixed effects. The coefficient here is still 0.06. However, this coefficient is not significant even at the 10% level. Also, the confidence interval shows that there is 95% confidence that the effect for a 1% increase in the lag of real GDP per capita is between -0.00024 and 0.00142. Furthermore, the exclusion of the year fixed effects yields an F-statistic of 6.76, with a p-value of 0 when rounded to 2 decimal places. Thus, the exclusion of the year fixed effects is significant even at the 1% level.
Table 2’s fifth regression is identical to the fourth except that fewer observations and countries are covered. More specifically, there are now 94 countries and 685 observations. This subsample was created by removing all the observations with no information on age groups, education, and the log of total population. With these controls, the coefficient for the log of lag real GDP per capita is -0.02. However, it is not significant at the 10% level, which means that this coefficient is more than 10% likely if the lag of income had no impact on the PRI. However, the year fixed effect’s exclusion is still significant since the F-statistic is 5.8 and it has a rounded p-value of 0.
The sixth and final regression in Table 2 takes the subsample defined in the fifth regression and regresses on the lag of log real GDP per capita and all the controls that were used to define the subsample. All the previous fixed effects still remain, and F-tests are conducted on the exclusion of year fixed effects, age controls, and the age groups, education, and log of total population (or the demographic controls). The lag of log real GDP per capita coefficient is now 0.02, which means that a 1% increase in the lag of GDP per capita is, on average, associated with a 0.0002 increase in the PRI. This estimate is lower than the fourth regression’s estimate of 0.06. However, the sixth regression’s estimate is also not significant at the 10% level. The smaller sample, coupled with various additional controls, are the most likely reason for why the estimate has decreased compared to the fourth regression. The F-statistic of the year fixed effects is 6.46, a small drop compared to the fourth regression’s coefficient of 6.76. However, the rounded p-value is still 0, which means that the exclusion of the year fixed effects is still significant at the 1% level. The F-statistic of the age controls is 2.45, which yields a rounded p-value of 0.05. Thus, the exclusion of age controls in general is only significant at the 5% level. Furthermore, the F-statistic for the exclusion of the demographic controls is 1.72, with a p-value of 0.12. Thus, the exclusion of the demographic controls is not significant even at the 10% level. The estimated coefficient of the average number of schooling years is 0, which means that an additional year of average education has no impact on the PRI. Furthermore, this is not significant even at the 10% level. The coefficient belonging to the log of population is -0.1, which can be interpreted as being that a 1% increase in population is, on average, associated with a -0.001 decrease in the PRI. However, this coefficient is also not significant at the 10% level, so its veracity is questionable. As for the various age group coefficients, the interpretation of one of them is that an increase in one of the groups of 1% is associated with, on average, a change in the PRI of that coefficient divided by 100. Of course, the percentages of all the age groups have to sum to 100%, which is why the age group of 0–15 has been excluded from the regression. This means that all the coefficients for the remaining age groups describe the expected PRI change when the group’s proportion increases, holding the other age groups constant, and subtracting that percentage from the 0–15 age group. Of the four coefficients, only one of them is statistically significant at the 10% level or below. That is the coefficient for the age group of 30–45, which has a coefficient of -2.62, and is significant at the 1% level. This means that an increase of 1% in the 30–45 age group is associated with an expected decrease of -0.0262 in the PRI when deducting that 1% from the 0–15 age group and holding all the other coefficients constant.
Table 3 takes four of the regressions in Table 2 and augments them by introducing a country’s status as a colony and a country’s status as a Soviet Bloc member as variables that interact with the lag of log real GDP per capita. In other words, these regressions show whether the aforementioned geopolitical markers impact how the lag of log real GDP per capita changes the PRI. The first regression is Table 2’s second regression, which is still a pooled regression but with clustered standard errors. The coefficient for the lag of log real GDP per capita is 0.25 instead of 0.23. Furthermore, it is significant at the 1% level as well. Countries that were former colonies have an expected coefficient on the lag of log real GDP per capita that is lower by -0.04 compared to countries that were not colonies. This is significant only at the 10% level. Also, countries that were part of the Soviet Bloc have an expected coefficient that is 0.15 higher than countries that were not, and this is significant at the 1% level.
The second regression is essentially the third regression in Table 2 but with the colony and Soviet Bloc controls. The coefficient here is 0.14 instead of 0.06, and it is significant at the 1% level instead of only at the 10% level. Furthermore, the interaction coefficient for former colonies is -0.17, and is significant at the 1% level. Also, the interaction coefficient for Soviet Bloc members here is 0.17 and is not significant even at the 10% level. The F-statistic for colony exclusion here is 2.24, which is insignificant even at the 10% level. However, the F-statistic for the Soviet Bloc exclusion is 10.37, which is significant at the 1% level.
Table 3’s third regression modified Table 2’s fourth regression. The lag of log real GDP per capita coefficient here is 0.15 instead of 0.06 and is significant at the 1% level instead of being insignificant even at the 10% level. Also, the interaction coefficient for former colonies is -0.18, which is significant at the 1% level. Furthermore, the Soviet Bloc interaction coefficient is 0.13, and it is insignificant even at the 10% level. The F-statistic for the colony coefficient exclusion is 9.6, with and is also significant at the 1% level. The Soviet Bloc’s coefficient exclusion F-statistic is 2.40 and is insignificant at the 10% level. Furthermore, controlling for year fixed effects yields an F-statistic of 6.89, which is significant at the 1% level.
The final regression in Table 3 takes the final regression in Table 2 and includes the interactions between lag of log real GDP per capita and the geopolitical markers. Compared to Table 2’s regression, the lag of log real GDP per capita coefficient is 0.13 instead of 0.02. However, it is insignificant even at the 10% level both here and in Table 2. The colony interaction coefficient here is -0.15, and it is significant only at the 10% level. In addition, the Soviet Bloc interaction is 0.99 in this regression and is significant at the 1% level. However, the size of the coefficient is questionable because the PRI only takes on values between 0 and 1. Furthermore, the only significant demographic control coefficient is the one that pertains to the 30–45 age group, which is a similar finding to Table 2’s final regression. The coefficient here is -2.22 instead if -2.62, although it is still significant at the 1% level. On the whole, the various F-statistics of the year fixed effect, age controls, and demographic controls exclusions are similar in significance to the regression found in Table 2. Furthermore, the colony control exclusion’s F-statistic is 3.58 here with a rounded p-value of 0.06, which means that its exclusion is significant only at the 10% level. Also, the F-statistic for the Soviet Bloc control is 25.92, which has a rounded p-value of 0, meaning that it is significant even at the 1% level.
On the whole, the lag of log real GDP per capita initially appears to have a substantial and significant impact on the PRIs. However, the size of this estimate, and its significance, decrease substantially when new controls are added. Therefore, income growth in a country is not a strong determinant of democratic growth. Rather, it is correlated with other features that determine the growth of democracy. From Table 2, the most prominent determinants of the growth of democracy in a country from the years 1950 to 2000 were the year fixed effects and the age controls. More specifically, an increase in the proportion of the population that is between 30 and 45 is associated with a significant decrease in PRI. The inclusion of colony controls and Soviet Bloc controls in Table 3 does not change the veracity of Table 2’s findings with regards to the importance of year fixed effects and age controls. Furthermore, the inclusion of Soviet Bloc controls to the regression has a very strong impact on how income increases fuel PRI growth. It seems to be the case that countries with USSR affiliations are more receptive to democratic growth when income increases compared to other countries.
Limitations and Conclusions:
These findings, especially once controls have been applied, make a relatively convincing case for the conclusion that income growth is only correlated with the growth of democratic institutions with a country and is not a strong determinant of democratic growth in itself. However, this analysis is by no means perfect. For one, some of the individual time observations for certain countries are blank. This means that some observations had to be omitted, which means that the regressions occurred on a relatively small dataset. Another potential issue is that the data only covers the period between 1950 and 2000 in 5-year intervals. Many events, both domestic and international, can occur in a 5-year period. This increases the potential for omitted variable bias. More observations via a shorter span of years are likely to make the analysis more causal than descriptive.
Overall, the analysis suggests a significant relation between income growth and democratic growth that loses its size and significance when other factors are considered. The most pertinent factors in this case are the year, the age groups, and the country’s affiliation with the USSR. However, a causal interpretation of these findings is precarious since the dataset suffers from some missing observations and a long span of time between each recorded observation.
References:
Acemoglu, Daron, Simon Johnson, James A. Robinson, and Pierre Yared. 2008. “Income and Democracy.” American Economic Review, 98 (3): 808–42.DOI: 10.1257/aer.98.3.808
Tables and Figures:
