Invited Symposium

The Statistical Analysis of Mediation Effects

Time and Location
Saturday May 24, 2008, 9:00 AM - 10:20 AM
Sheraton 1

Abstract

Co-Chair: Dave MacKinnon
Arizona State University

Mediation occurs when the effect of one variable on another is explained by an intervening variable called a mediator. Many methods have been suggested for statistically evaluating mediation effects in a variety of settings. This symposium provides a forum for presenting, comparing, and contrasting some of these methods.

Kristopher J. Preacher

Kristopher J. Preacher (Chair)
University of Kansas

Kristopher J. Preacher

Kristopher J. Preacher
University of Kansas
Bootstrapping Null-Centered Confidence Intervals for Mediation Effects

Amanda Fairchild
Arizona State University
An Illustrative Framework for Examining Mediated-Moderation
Relations between variables are often more complex than bivariate associations between a predictor and criterion. Rather, these relations may involve third variable effects. Recent research has examined the influence of more than one third variable effect in a research design. The investigation of mediation and moderation together is newer relative to either analysis alone, but work in the area is increasing. Sources have described how mediation and moderation work together, and have proposed methods to analyze them jointly. This presentation considers the case of mediated-moderation, describing the effect in detail. A point estimator for the effect is presented, and simple slopes of the model are explored with example data.

Co-Author:
David P. MacKinnon, Arizona State University

Patrick Shrout
New York University
Adjusting for Baseline Effects in Mediation Analyses
Many experimental and quasi-experimental studies of an intervention X are enhanced by mediation analyses that show that X's effect on Y is carried through a mediating variable M. The best-designed studies measure X at time 1, M at a later time 2, and Y at a still later time 3. Often it is possible to obtain baseline measures of M and Y prior to the action of X at time 1, and in this case methodologists recommend that adjustments be made for these baseline variables. In practice such adjustments are rarely done. In this talk, we review the amount of bias that can be introduced by failing to make these adjustments, and compare two different adjustment strategies: regression adjustment vs. difference scores. Results can vary depending on whether M and Y represent quantitative measures or binary indicators.

Co-Author:
Margarita Krochik, New York University

JeeWon Cheong
University of Pittsburgh
Accuracy of Estimates and Statistical Power for Testing Mediation in Latent Growth Modeling
Longitudinal mediation can be evaluated in latent growth models (LGM) by modeling the repeated measures of the mediator and the outcome as distinctive parallel processes influenced by the independent variable (Cheong, MacKinnon, & Khoo, 2003). Researchers have applied the LGM approach to mediation; however, little is known about the accuracy of the estimates and statistical power of different methods when mediation is assessed in the LGM framework. This talk presents the findings from a simulation study to address these issues under various conditions including sample size, effect size of the mediated effect, number of measurements, and the proportion of variance of the repeated measures explained by the growth factors.

Co-Author:
David P. MacKinnon, Arizona State University

Go back

GLN Consulting