Tracing Causal Paths from Experimental and Observational Data

• In Event: Experiments

Abstract

Social scientists have long been interested in tracing multiple causal pathways through which a treatment affects an outcome. Most applications of causal mediation analysis, however, have focused on only one causal pathway while treating all other possible pathways as a nuisance (i.e., lumping them together under the umbrella of “direct effect”). In fact, in the presence of multiple, potentially overlapping, causal pathways, the total effect of the treatment on the outcome can be decomposed into a set of path-specific effects that are identified under the standard assumptions of causal mediation analysis. This article introduces a simple imputation approach for estimating path-specific causal effects from experimental and observational data. In contrast to existing methods for causal mediation analysis, this approach does not require any model for the mediator(s) of interest. All we need is to model how the outcome depends on the treatment and varying sets of mediators, which can be implemented via highly nonparametric methods such as the Bayesian Additive Regression Trees (BART). We illustrate this approach using three empirical examples.

Author

• Xiang Zhou, Harvard University