Chapter Fourteen:
The Statistical Findings
This book has doggedly pursued a hypothesis. The impact of regime history on development, we conjectured, is the product of a country’s regime history, not simply its present regime status. Two species of evidence have been introduced to test this argument, the first relies on crossnational statistical tests and the second on country-focused case studies (Brazil, India, and Mauritius). In this chapter, we review the statistical findings, discuss the potential policy significance of these results, address threats to validity, and interrogate the data for clues that might help to clarify the theory. (Chapter fifteen surveys the evidence provided by the case studies and reflects on the question of causal mechanisms.)
Table 14.1 summarizes the results of the empirical tests performed in chapters 3-10. For each outcome-measure of development – growth, economic policy, infrastructure, policy continuity, environmental policy, education, public health, and gender equality – we assess the causal effects of democracy, as measured by the traditional “level” indicator (a country’s score on the Polity2 variable) and our “stock” indicator (a country’s accumulated democratic experience over the past century). If the relationship is shown to be statistically significant (at standard thresholds of significance), we also calculate an estimate of its causal impact on the outcome in question. Impact is understood as the cumulative effect of two decades of high-quality democracy (Polity2=10), compared with two decades of thorough autocracy (Polity2=-10).
[Discuss results, once Table 14.1 is completed.]
The final column in Table 14.1 refers to the overall plausibility of the finding. A number of threats to inference must be carefully considered: a) identification problems, b) error in the measurement of democratic stock, and c) sample bias. These will be discussed at length in the following sections. Our interest in these matters extends not simply to the causal relationship but also, equally important, to the source of this relationship. Here, covariational statistical modeling is employed to light on causal mechanisms.
Space restrictions prevent us from presenting results for all of the developmental outcomes listed in Table 14.1. We focus here on per capita GDP growth because of its prominence in the literature, its secondary causal effects on all other outcomes, and the excellent quality and coverage of the data. [Tests for other outcomes may be listed in an appendix or on a web site.]
Table 14.1:
Summary of Results
Concept / Outcome measure / Democracy Level / Democracy StockSig. / Impact / Specification / Plausibility / Sig. / Impact / Specification / Plausibility
Economic growth / GDPpc growth rate / ? / Model 1, Table 3.1 / ++ / Model 3, Table 3.1 / high
Economic policy / Trade openness
Foreign direct investment
Investment risk
Infrastructure / Electricity
Policy continuity
Environmental policy / CO2
Sulfur emissions
Education / School attainment
Public health / Infant mortality
Gender equality / Female population (% of total)
Life expectancy ratio (F/M)
Female labor force participation
Fertility
Schooling gap (F/M)
“Statistical significance” refers to the t statistic for the variable (democracy level or democracy stock) in the specification noted. The following shorthand applies: ? = null hypothesis cannot be rejected; + = democracy is positively correlated with this developmental outcome at .10 level (two-tailed test); ++ = democracy is positively correlated with this developmental outcome at .05 level (two-tailed test). “Impact” refers to the estimated causal effect of 20 years high-quality democracy (Polity2=10), relative to 20 years of autocracy (Polity2=-10). Units or scale of dependent variable are indicated in parentheses. Effects for logged dependent variables reflect the effect as a percentage change in the dependent variable. Impact is measured only if the variable is statistically significant. “Plausibility” refers to the robustness of the finding in a variety of specification tests as well as our sense of the identification problems posed by the analysis.
Identification Problems
The most troublesome sort of identification problem concerns selection effects – in this case, the process by which countries become democratic and remain democratic (i.e., accumulate democratic stock). It is readily apparent that this process is not independent of the outcomes under study, as is common in observational research.
Conventionally, this sort of problem is handled with instrumental variables -- variables that model the selection process and allow for a purged set of regressors in a two-stage procedure. Unfortunately, there are few predictors of democracy (i.e., instruments highly correlated with regime-type) that are not also causally related to development, violating the assumptions of IV analysis. Note that although such factors as colonial status (English or other) or religion (Muslim or other) have been commonly employed as instruments for democracy, these factors are also commonly thought to have a direct effect on the quality of governance, and from thence on long-term economic and social development. We are loath to perform two-stage analyses with such poor instruments, which in our view would constitute a misleading statistical adventure (Dunning 2005; Heckman et al. 1998). Note that the task of finding good instruments in this study is further complicated by our theoretical concern with a country’s stock of democracy (a continuous and continuously changing indicator) and our preference for time-series (rather than cross-sectional) analysis. We have discovered no satisfactory solution to this dilemma.
Even so, we are confident that results reported here represent a reasonably accurate portrayal of democracy’s effect on various developmental outcomes. Note that the fixed-effect framework allows us to neutralize any underlying factor that affects either a country’s regime trajectory and/or its developmental trajectory so long as that underlying factor holds relatively constant across the period of analysis (1950-2003). This means that geographic factors, disease vectors, cultural/sociological factors, colonial histories, as well as other deeply-rooted historical factors should not bias the results of our analyses.
It is still conceivable that a factor that varies with changes in democracy (stock or level) might be driving the analysis, in which case our measure of regime-type might be serving as a proxy for this un-measured factor. To solve this potential problem we include a wide range of controls that should be sensitive to the overall health of the polity including GDP per capita, growth (in some models), measures of instability, and year fixed-effects, as shown in Table 14.2 (see also Tables 3.1 and 3.2). These variables should serve as proxies for whatever additional -- unmeasured and time-varying -- factors might be driving developmental patterns across countries.
A particular concern is the interaction of GDP per capita and democracy. The current consensus among scholars is that richer countries are more likely to preserve democracy; dispute persists over whether richer countries are also more likely to democratize in the first place.[1] In any case, there is every reason to expect that development (as measured by GDP per capita) has a causal effect on regime-type. Our argument, of course, is the reverse: that democracy (over time) causes development. This presents a classic identification problem.
However, it is perhaps not as worrisome as it may appear. Note, first, that richer countries tend to grow at a slower rate (a phenomenon known as “convergence”). This means that insofar as economic development affects regime-type – helping to create more rich democracies -- this should bias results against our hypothesis (when the per capita GDP control is removed). Note, second, that the outcome of interest is not per capita GDP but rather growth, i.e., annual percentage changes in per capita GDP. There is no established view as to whether short-term growth performance has a causal effect on regime-type; therefore, the problem of causal independence between left- and right-side variables is somewhat vitiated. Of course, accumulated patterns of growth constitute GDP, so the two factors are not entirely independent of one another. However, it would require a very substantial change in GDP to significantly affect a country’s regime-type (even if one is satisfied that there is such a relationship). The short-term changes that we are trying to explain are scarcely affected by this very attenuated feedback loop. To sum up, there is a problem of endogeneity, but it is of slight importance to an analysis focused on short-term changes in GDP per capita.
This is confirmed by a test with ten-year lags imposed on the democracy variable, which is normally lagged only one year behind the dependent variable (see models 1 and 2 in Table 14.2). Remarkably, the value of the coefficient for democracy stock diminishes only slightly, while the standard errors are stable (relative to the benchmark equation), confirming that the relationship between democracy stock and growth is probably not an instance of reverse causality.
A second concern is that our stock measure of democracy may be proxying for a country’s overall level of political stability. Our argument, it should be emphasized, is not about regime stability, but rather about democratic stability, specifically the maintenance of high-quality democracy.[2] The point to note is that the construction of a country’s democratic stock registers continuity in regime-type, not the continuity of the regime. Authoritarian regimes may rise and fall, and democratically elected governments change hands, without altering a country’s score on the Polity2 scale. Thus, in principle, there should be no problem with proxy effects of this sort. In order to assure that this is indeed the case we include a set of variables measuring regime durability, regime fluctuation, political instability, and conflict in models 3 and 4 in Table 14.2. (Variables are defined in the chapter appendix.) [Comment on results]
Table 14.2:
Robustness Tests
1 / 2 / 3 / 4Democracy stock / X
Democracy stock / 0.004**
(t-10) / (0.001)
Democracy level / X
Democracy level / X
(t-10)
Regime durability / X / X
(Polity2)
Regime fluctuation / X / X
(Polity2)
Instability / X / X
(Banks)
Conflict1 / X / X
Conflict2 / X / X
Conflict3 / X / X
Conflict4 / X / X
GDP pc (ln) / -2.851*** / X / X / X
(0.561)
Constant / 23.241***
(4.271)
Observations / 5574
Countries / 178
Sample Period / 1950-00 / 1950-00 / 1950-00 / 1950-00
R-squared (within) / 0.03
Prob > F / 0.0000
Fixed effect regressions with AR(1) disturbance. Units of analysis: country-year. Dependent variable: annual per capita growth rate. All predictors lagged one year. Newey-West standard errors in parentheses. *** p<.01 ** p<.05 *p<.10 (two-tailed test)
Democratic Stock Reconsidered
Measuring a country’s level of democracy is complicated, as noted in chapter two. Measuring its accumulated stock is even more troublesome, for another layer of complexity is added to an already vexing concept. In order to assure that our results are not the product of measurement error in the key theoretical concept, and in order to probe the meaning of this concept, we include a series of tests focused on changing definitions and measurement of the stock concept.
In chapter two we described how the stock variable was constructed. Calculating a “stock” of democracy for a given country depends upon some underlying measure of how democratic that country was throughout its (modern) history. Our principal data source is the Democracy/Autocracy variable -- “Polity2” – drawn from the Polity IV dataset. This is supplemented with data from other democracy indicators (e.g., Freedom House) so as to include micro-states (e.g., Sao Tome) and semisovereign nations (e.g., Hong Kong). For lands that did not enjoy sovereignty in any degree (and therefore are not included in any extant democracy index) but were part of a contiguous empire such as the Russian Empire or the USSR, we applied the latter’s democracy score. Thus, the contemporary state of Belarus receives the same score as Russia prior to 1919, and the same score as the USSR until 1991.
Model 1 in Table 14.3 depicts the benchmark model (drawn from Table 3.1). In order to assure that these results are not subject to the authors’ re-codings, as described above, we run the same model on a subset of countries for which no additional data was imputed. Results are displayed in model 2. This measurement of democratic stock is drawn directly from the Polity IV data set (Polity2). Not surprisingly—since we have lost only 697 observations—the coefficients and standard errors are fairly stable.
It is possible, however, that democracy stock is measuring something other than democracy per se. One possibility is that this variable is acting as a proxy for the length of time a country has been autonomous, that is, the duration of a nation-state. It could be, in other words, that a high score on democracy stock is indicative of a long historical experience of sovereignty, rather than (or in addition to) a long experience with democratic elections. As a test of this hypothesis, in model 3 we recalculate the benchmark regression including only those countries for which complete data is available for the entire twentieth century (including scores based on the Polity2 values of empires of which a country was, for a time, encapsulated within, as explained above). Although the number of countries in the data set, as well as the corresponding number of observations, drops to less than one-third, the coefficient and standard errors are fairly stable, suggesting that our results are indeed driven by differences in democracy and not by sovereignty or a lack thereof or by arbitrary coding decisions with respect to missing data.
[Describe other reformulations of democratic stock and the results.]
Table 14.3:
Alternate Measures of Democracy Stock
1 / 2 / 3 / 4 / 5 / 6 / 71% depreciation / 0.006***
(Polity2, the benchmark var.) / (0.001)
1% depreciation / 0.007***
(Polity2, no imputations) / (0.002)
1% depreciation / 0.006**
(Polity2, no imputations
or incomplete cases) / (0.003)
20-year moving ave. / 0.005**
(Polity2) / (0.002)
50-year moving ave. / 0.005**
(Polity2) / (0.002)
10% depreciation / X
(Polity2)
1% depreciation / X
(Moon)
GDP pc (ln) (WDI) / -2.961*** / -2.854*** / -3.195*** / -2.885*** / -3.096*** / X / X
(0.488) / (0.524) / (0.887) / (0.693) / (0.900)
Constant / 23.885*** / 22.985*** / 27.513*** / 23.784*** / 26.555***
(3.670) / (3.926) / (7.163) / (5.332) / (7.253)
Observations / 6264 / 5567 / 2884 / 4657 / 2851
Countries / 180 / 156 / 78 / 176 / 95
Sample Period / 1950-00 / 1950-00 / 1950-00 / 1950-00 / 1950-00
R squared (within) / 0.03 / 0.03 / 0.04 / 0.03 / 0.03
Prob > F / 0.0000 / 0.0000 / 0.0004 / 0.0001 / 0.0027
Fixed effect regressions with AR(1) disturbance. Units of analysis: country-year. Dependent variable: annual per capita growth rate. All predictors lagged one year. Newey-West standard errors in parentheses. Variables and procedures defined in the text. *** p<.01 ** p<.05 *p<.10 (two-tailed tests)
Heretofore, all measures of democracy stock have been based on a continuous measure of a country’s democratic quality from year to year (the Polity2 variable, supplemented by various additional measures). Yet, the stock concept could also be derived from a dichotomous measure of country-level democracy. That is, rather than attempting to determine how democratic/autocratic a country is in at time T, one could attempt to determine whether certain minimal attributes of democracy existed at that time (usually, these center on the existence of multi-party elections).[3] In our coding, a country-year is democratic only if it falls above 4 on the Polity2 scale (from –10 to +10).