Causal Inference in the Presence of Missing Outcome and Treatment Variables: Triply Robust Estimator and Sensitivity...

Estimating causal effects from observational studies with missing data typically requires the no unmeasured confounder (NUC) assumption and the missing at random (MAR) assumption. However, correctly specifying all relevant models is often challenging in practice, and model misspecification can lead to biased estimates. In this study, we develop a triply robust estimator for causal effects when the outcome or treatment variable is partially observed. The proposed estimator incorporates outcome, treatment, and missing-data models, and remains consistent as long as at least two of the three models are correctly specified. Although this approach provides robustness against model misspecification, its validity still depends on the NUC and MAR assumptions, which are untestable from the observed data alone and may be frequently violated in the presence of unmeasured confounding. To address this limitation, we introduce a novel sensitivity analysis framework to evaluate the potential impact of unmeasured confounding on causal effect estimates and demonstrate its practical usefulness through an application to real data.