Primary Submission Category: Mediation Analysis, Mechanisms
Multiply Robust Estimation with Machine Learning for Causal Mediation Analysis with Clustered Data and Unmeasured Cluster-Level Confounders
Authors: Cameron McCann, Xiao Liu,
Presenting Author: Cameron McCann*
Mediation analyses in behavioral research often involve clustered data; however, existing methods for causal mediation in this context are relatively limited. In this study, we extend a multiply robust estimation method to estimate causal mediation effects in clustered data while accounting for unmeasured pre-treatment confounders at the cluster level. To control for unmeasured cluster-level confounders in the nuisance model estimation, we perform cluster-mean centering and include cluster means and cluster dummies; for comparison, we also examined the performance when excluding the cluster dummies. For statistical inference, we consider within-cluster correlation in calculating standard errors and confidence intervals. Through simulations, we assess the performance for inference of the cluster-average and individual-average causal mediation effects, comparing nuisance model estimation with (1) parametric models (fixed or random-effects regressions) versus (2) machine learning prediction models (a super learner ensemble of parametric and nonparametric regressions). Finally, we illustrate the method using data from the National Longitudinal Study of Adolescent to Adult Health (Add Health; Harris & Udry, 2008).