181 – Statistical Modeling, Missing Data, and Observational Studies
Causal Mediation Analysis in Multi-site Trials: An Application of Ratio-of-Mediator-Probability Weighting to the Head Start Impact Study
Xu Qin
University of Chicago
Guanglei Hong
University of Chicago
This study focuses on developing methods for causal mediation analysis in multisite trials and uses the national Head Start Impact Study as a motivating example. The causal effects of interest, defined in terms of potential outcomes, include the indirect effect of assignment to Head Start programs on child vocabulary learning mediated by a program-induced increase in parent reading to child and the direct effect of Head Start programs. The goal is to reveal not only the prevalent causal mechanism but also how the mechanism may vary across sites. Extending the ratio-of-mediator-probability weighting (RMPW) approach to causal mediation analysis in multi-site trials, we estimate the average direct effect, the average indirect effect, and the between-site variance and covariance of these causal effects. This strategy allows for treatment-by-mediator interaction. It greatly simplifies the outcome model specification and therefore avoids possible model misspecifications. The performance of the approach is assessed across a range of multi-site designs that differ in the number of sites and the sample size per site. We investigate the relative strengths and limitations of the RMPW strategy through simulations.