Dos and Don'ts for Conducting Mediation Analysis: A Commentary With Practical Tips to Avoid Common Problems

Purpose: Mediation analyses allow for exploration of causal mechanisms that explain how a predictor is related to an outcome. Tests of mediation are fundamental to addressing some of the most consequential questions in rehabilitation science. In recent decades, the development of easy-to-use analytical tools has made conducting statistical tests of mediation more accessible to researchers. Unfortunately, there are persistent problems in the conceptual underpinning of many tests of mediation. Even in cases where the statistical analyses are correctly run, problems with the underlying rationale for the mediational analysis will render the results inconsequential, in the best case, or misleading, in the worst case. Method: In this commentary, I summarize the uses of mediation analysis and through a series of six main types of errors provide practical, plain language guidance ("Dos and Don'ts") for conducting a conceptually robust mediation analysis. Results: The "Dos and Don'ts" laid out in this commentary highlight that there are persistent issues with lack of understanding of mediation, confusion about the differences between moderation, mediation, and covariates, lack of strong theoretical justification for mediation, and lack of attention to methodological issues (e.g., measurement) in many mediation analyses. Conclusions: Promoting the use of mediation analysis in rehabilitation research will advance theory and effective practice in our field. Researchers undertaking mediation analysis are encouraged to prioritize developing a strong theoretical framework that justifies use of mediation analysis, ensuring study methodology supports and enables tests of mediation, as well emphasizing a strong statistical approach to conducting the test of mediation. Impact and ImplicationsDevelopment of new and easy-to-use analytical tools has made conducting statistical tests of mediation more accessible to researchers. The ease with which statistical tests of mediation can now be completed makes it easier for those without a clear understanding of mediation analyses to make errors; most common errors in tests of mediation are not related to the mathematical aspects of running the statistical models, but rather of developing coherent and compelling theoretical models that call for tests of mediation and attending to nonstatistical aspects of the study method, such as measurement of model constructs and sample size/power to detect the mediated effect. Increasing the rigor