2011 HSR&D National Meeting Abstract
2015 — Mediation Analysis in Health Services Research
Zhou A (VA Puget Sound Health Care System, Seattle; University of Washington), Atkins D
(University of Washington), Taylor L
(VA Puget Sound Health Care System, Seattle)
In testing whether or not a randomized intervention has an effect on outcome, oftentimes we seek to understand how such effects come to be. What kinds of causal sequences does the intervention initiate? What are the causal pathways through which the intervention exerts its effect? Over the years, methods used to test such models have grown in sophistication, yet frequently, researchers are using methods that are out of step with advances made in the statistical methods literature. The goal of this workshop is to: (1) clarify the assumptions used in popular methods like the Barron and Kenny approach, exposing their pitfalls; (2) update researchers on some of the more recent causal mediation methods; and (3) provide examples and code for implementing these newer methods.
This 90-minute workshop will be presented in lecture format. Part I will define mediation analysis in the context of randomized studies and provide a brief overview of the current methods, making the general distinction between the Barron and Kenny approach, and more recent causal inference methods. We will then outline the Barron and Kenny approach, using real-world examples, and the pitfalls of this approach when assumptions are not met. In Part II we provide an overview of more recent causal inference methods, and implementation of one of these methods using package in R, a free software environment for statistical computing. Ample time will be allowed for audience questions.
Clinical researchers, methodologists, statisticians
Assumed Audience Familiarity with Topic:
The audience should have a basic understanding of randomized controlled trials and statistical regression modeling.