New Perspectives on Causal Mediation Analysis

Abstract

Social and behavioral scientists have long used path analysis and related linear structural equation models (SEMs) to decompose parameters of ordered systems of equations into “direct effects” and “indirect effects” through mediating variables. These decompositions have been used to address substantive questions of fundamental interest, for example, how a person’s social background affects his/her earnings through education. However, in general, the “direct effects” and “indirect effects” defined in and estimated from these models should not be given causal interpretations, even in randomized experiments. To illustrate this, we first define various direct and indirect effects using potential outcomes notation and discuss situations where an investigator might want to consider these. Second, we consider identification of these effects: the required identifying assumptions are more often than not implausible for the kinds of data collected and questions considered in social and behavioral research. Third, we present other identifying assumptions that might be used to identify direct and indirect effects, and then briefly discuss different methods to estimate these effects, including regression, instrumental variables, marginal structural models, and weighting methods.Finally, we introduce an alternative approach to mediation (principal stratification), define several possible effects of interest, and briefly discuss identifying assumptions and estimation.KeywordsStructural Equation ModelCausal ModelControl EffectMarginal Structural ModelCausal ParameterThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.