Causal inference in psychosocial epidemiology

Abstract

Methods to help researchers draw causal inferences based on analyses of observational data are increasingly embraced throughout epidemiology. Clearly stating causal questions, adopting rigorous methods to evaluate these questions, and triangulating across analytic methods, data sources and research designs will substantially strengthen research in psychosocial epidemiology. Directed acyclic graphs offer a convenient tool to help researchers communicate assumptions, specify hypotheses and guide analyses. A handful of methods, including marginal structural models, fixed effects models and instrumental variables analyses, are quickly growing in popularity because of their promise to deliver causal effect estimates under more plausible assumptions than conventional epidemiologic research designs. Each of these designs has advantages only in select settings and under strong assumptions, and therefore are likely to be helpful, but not panaceas, for research on the most challenging causal questions.