协变量、中介变量和结局均存在不可忽略缺失时的因果中介分析

摘要

Causal mediation analysis typically assumes that all variables are fully observed, yet missingness is common and often nonignorable. We develop a framework for mediation analysis when pretreatment covariates, the mediator, and the outcome are all subject to missing not at random (MNAR). We propose three interpretable MNAR mechanisms, namely response continuation, self selection, and missingness driven by covariates and the mediator. Under each, we establish nonparametric identification of mediation effects using a shadow variable and completeness conditions. Empirical results show that subjective well-being partially mediates the association between job satisfaction and depressive symptoms.