Longitudinal Mediation Analysis with Missing Not at Random Data
Longitudinal mediation analysis suffers from at least two challenges.First,the mediators and outcome at a given time point affect those at later time points and they act as post-treatment confounders(or time-varying confounders).Second,Missing Not At Random(MNAR)data are commonly seen in longitudinal studies,which can lead to substantial systematic bias in the estimation of model parameters if not properly justified.To our knowledge,there is no literature simultaneously justifying time-varying confounders and MNAR data in longitudinal mediation analysis and we aimed to fill this gap.Based on potential outcome framework,we identified some regularity conditions for the identifiability of causal effects in longitudinal mediation analysis with MNAR data,and proposed a new statistical procedure based on estimating equations,multiple imputation and natural effect model to infer the causal effects.Large sample properties were established for the proposed method,and its finite sample performance was evaluated through simulations and real data analysis.
longitudinal mediation analysismissing not at randomtime-varying confoundersmultiple imputationnatural effect model
万方数据
维普
