Abstract
Democratization might be instrumental in addressing climate change on a global scale but the effect of democracy on carbon dioxide emissions is unclear. Treadmill of production (TOP) theory implies that democracy may increase emissions as publics pressure governments for increased economic growth. On the other hand, theories aligned with a rational choice perspective and ecological modernization perspectives suggest that democracy will reduce emissions as autocratic leaders have little incentive to address environmental policy. In this article, we clarify the role of democracy in emissions by using data that disaggregate different characteristics of democracies. Using novel causal inference techniques and generalized linear models, we allow the effect of democratic characteristics to vary across countries. Results indicate that democracy tends to reduce emissions, but the effect is modest, and its intensity varies significantly across countries.
Notes
Many scholars have been highly critical of the notion of a “tragedy of commons” and claims made by Hardin (Citation1968). For particularly notable critiques see Ostrom et al. (Citation2002) and Ostrom (Citation2009).
Indeed, with the notable exception of Chen et al. (Citation2012) we are unaware of any other studies in environmental sociology that use causal inference methods.
An alternative strategy is to recode Polity2 into an ordinal variable with several categories and then use multi-valued propensity score matching techniques (e.g., Imbens Citation2000). Matching in the context of a multi-value (i.e., multi-category) treatment is far more experimental and contested than more standard techniques for binary treatments and require much more data to produce unbiased estimates. Given these constraints we opt for our dichotomization strategy—which allows for the use of entropy balancing. Another challenge to causal inference into the effect of democracy on emissions arises because democracy is relatively time invariant within panels—countries rarely switch from autocracy to democracy and vice versa. As such, dynamic panel and related models to assess causality in a panel data framework are not applicable.
There is a range of available indicators of economic development including GNI per capita and GDP per capita adjusted via various means. We accessed the World Bank database for a range of GNI per capita and GDP per capita variables and found that they correlated quite strongly (r > .85), suggesting that the choice of indicator will not likely alter the results.
As of this writing entropy balancing is a relatively new innovation in causal inference methods. However, it has been quickly adopted by applied researchers as an alternative to more established techniques (i.e., propensity scores). For instance, Marcus (Citation2013) used entropy balancing to balance between unemployed and employed individuals to understand the effect of plant closure on health while Neuenkirch and Neumeier (Citation2016) balanced several covariates between countries facing U.S. economic sanctions and those not under sanction.
We use a GLM approach instead of other panel data estimators (such as the well-known fixed effects estimator executed with Stata’s xtreg command) because most panel data estimators require that weights remain constant within each panel. Entropy balancing weights individual observations, not entire panels, and common panel data estimators cannot accommodate observation-specific weights. Another advantage of the GLM approach is its ability to estimate country-specific coefficients.
The empirical Bayes method uses parameter estimates (i.e., regression coefficients) to estimate the random slopes.
Based on the literature we also included other control variables. In unreported models we controlled for income inequality, measured using the Gini index. In each model the coefficient for this variable was substantively small and data were very sparse, leading to a significant loss of sample size. Thus, it was not included in the reported models. A handful of studies has connected human capital, such as the level of education in the population, to environmental outcomes at the national level—though results are conflicting (Authors 2013; Jorgenson 2005; Kinda Citation2011; McHenry Citation2007). In other unreported models we included a measure of literacy rates; the coefficient for this variable was not large and the available data were sparse, causing a sharp reduction in the number of valid cases. For this reason, we do not control for literacy rates.
Additional information
Notes on contributors
Adam Mayer
Adam Mayer is a PhD student in environmental sociology at Colorado State University where he is affiliated with the Cooperative Institute for Research in the Atmosphere, the Center for Disaster and Risk Analysis, and the Center for the Study of Crime and Justice.