B0814
Title: Causal mediation analysis based on partial linear models
Authors: Xizhen Cai - Williams College (United States) [presenting]
Yeying Zhu - University of Waterloo (Canada)
Yuan Huang - Yale University (United States)
Debashis Ghosh - University of Colorado Anschutz Medical Campus (United States)
Abstract: The focus is on estimating the direct and indirect effects of mediation analysis based on a set of partial linear regression models. We allow a nonlinear relationship among the baseline covariates and the response variables in each model. Since we are only interested in estimating the coefficients for the treatment and the mediator in the structural models, we assume partial linear models where the baseline covariates are regarded as a nuisance. The estimates can be interpreted as causal effects without the linearity assumption. We also propose variable selection procedures when the set of mediators is high-dimensional. Simulation results show the superior performance of our proposed method and a data application is conducted when the set of candidate mediators are high-dimensional methylations.
