A1805
Title: Semiparametric models for dynamic treatment effects and mediation analyses with observational data
Authors: Shenshen Yang - Tianjin University (China) [presenting]
Sukjin Han - University of Bristol (United Kingdom)
Sungwon Lee - Sogang University (Korea, South)
Abstract: A semiparametric model is proposed that captures how a sequence of interventions interacts with a sequence of outcomes. In such settings, the outcome at a given period is affected by the history of treatments and outcomes, directly or indirectly through mediators. The main challenge in understanding various channels of dynamic effects is that, in observational settings, individuals make dynamically endogenous decisions regarding treatment selection. A model with an AR(1) process on the latent variable across stages is constructed. Using instrumental variables methods, it is shown how average and quantile dynamic treatment effects and mediation effects can be point identified and efficiently estimated in a class of semiparametric models under treatment endogeneity and flexible heterogeneity. The proposed procedure requires only binary instruments. As a byproduct of the semiparametric specification, parameters reflecting the degree of endogenous selection and time invariant heterogeneity are also identified and estimated.
