Explaining Causal Findings Without Bias: Detecting and Assessing Direct Effects

Abstract

Researchers seeking to establish causal relationships frequently control for variables on the purported causal pathway, checking whether the original treatment effect then disappears. Unfortunately, this common approach may lead to biased estimates. In this article, we show that the bias can be avoided by focusing on a quantity of interest called the controlled direct effect. Under certain conditions, the controlled direct effect enables researchers to rule out competing explanations—an important objective for political scientists. To estimate the controlled direct effect without bias, we describe an easy-to-implement estimation strategy from the biostatistics literature. We extend this approach by deriving a consistent variance estimator and demonstrating how to conduct a sensitivity analysis. Two examples—one on ethnic fractionalization’s effect on civil war and one on the impact of historical plough use on contemporary female political participation—illustrate the framework and methodology.

The enduring influence of institutions on universal health coverage: An empirical investigation of 62 former colonies

“…Institutions affect economic development, which may also affect health. Conditioning on a post-treatment variable (in this case, GDP) can produce bias (Acharya et al 2016a) and so to address this we estimated the following models:…”

“…Crucially, by parsing out its association with institutions we thereby stop GDP acting as a collider variable (Morgan and Winship 2007). As an additional check on this approach, we also re-estimate our models -following Acharya et al -using a g-estimation procedure (Acharya et al 2016a). Not only does this model avoid the problems of post-treatment bias but it also tests whether any association between institutions and our health outcomes persists that does not operate through GDP.…”

“…Institutions affect economic development, which may also affect health. Conditioning on a post-treatment variable (in this case, GDP) can produce bias (Acharya et al 2016a) and so to address this we estimated the following models:In short, we estimate a two-stage least squares model with GDP as the outcome, institutions as the endogenous variable, and our two instruments along with our control variables. This model approximates Acemoglu, Johnson, and Robinson's model but with the addition of our extra instrument (Distance in KM).…”

“…We do not include these variables in our main models in part due to sample size and in part because there is good evidence that, like GDP, these variables sit on the causal pathway between institutions and health, and would potentially bias our results. However, even when we control for these variables, our two measures of institutions remain correlated with our measures of health outcomes in 16 almost all instances (see Web Appendix 3).We also re-estimate our main instrumental regression models using a g-estimation procedure (see Table 6) (Acharya et al 2016a). The g-estimation procedures 'transforms (or demediates) the dependent variable by removing from it the effect of the mediator and then estimating the effect of the treatment on this demediated outcome ' (Acharya et al 2016a).…”

“…We also re-estimate our main instrumental regression models using a g-estimation procedure (see Table 6) (Acharya et al 2016a). The g-estimation procedures 'transforms (or demediates) the dependent variable by removing from it the effect of the mediator and then estimating the effect of the treatment on this demediated outcome ' (Acharya et al 2016a).…”