Statistical Methods for Mediation Analysis

• Activity Number: 94
• Type: Roundtables
• Date/Time: Monday, August 5, 2013 : 7:00 AM to 8:15 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Douglas Gunzler*+
• Companies: Case Western Reserve University
• Keywords: mediation analysis ; structural equation modeling ; causal inference ; indirect effect ; direct effect ; temporal order

Abstract

Does a tobacco prevention program change social norms for tobacco use leading to a reduction of cigarette smoking? The question suggests a chain of relations in which an antecedent measure affects a mediating measure that then affects an outcome. We can address the question, or more generally any question involving a mediation process, which health policy researchers routinely encounter, by performing mediation analysis. Different approaches, starting with Baron and Kenny (1986) and including structural equation modeling (SEM)-based and causal inference methods, have been proposed for mediation analysis. Further, there are methodological issues to consider for a given data set and study such as making reasonable statistical assumptions and handling where appropriate noncontinuous outcomes or mediators, multiple mediators, latent constructs, confounders, missing data, or repeated measures. This roundtable discusses the advantages and disadvantages of the approaches for mediation analysis for investigating mediation in real-world problems.