Abstract
IMPORTANCE: Associations between putative risk factors and psychiatric and substance use disorders are widespread in the literature. Basing prevention efforts on such findings is hazardous. Applying causal inference methods, while challenging, is central to developing realistic and potentially actionable etiologic models for psychopathology. OBSERVATIONS: Causal methods can be divided into randomized clinical trials (RCTs), natural experiments, and statistical models. The first 2 approaches can potentially control for both known and unknown confounders, while statistical methods control only for known and measured confounders. The criterion standard, RCTs, can have important limitations, especially regarding generalizability. Furthermore, for ethical reasons, many critical questions in psychiatric epidemiology cannot be addressed by RCTs. We review, with examples, methods that try to meet as-if randomization assumptions, use instrumental variables, or use pre-post designs, regression discontinuity designs, or co-relative designs. Each method has strengths and limitations, especially the plausibility of as-if randomization and generalizability. Of the large family of statistical methods for causal inference, we examine propensity scoring and marginal models, which are best applied to samples with strong predictors of risk factor exposure. CONCLUSIONS AND RELEVANCE: Causal inference is important because it informs etiologic models and prevention efforts. The view that causation can be definitively resolved only with RCTs and that no other method can provide potentially useful inferences is simplistic. Rather, each method has varying strengths and limitations. We need to avoid the extremes of overzealous causal claims and the cynical view that potential causal information is unattainable when RCTs are infeasible. Triangulation, which applies different methods for elucidating causal inferences to address to the same question, may increase confidence in the resulting causal claims.