Doubly Robust Inference for Targeted Minimum Loss Based Estimation in Randomized Trials with Missing Outcome Data

摘要

Missing outcome data is one of the principal threats to the validity oftreatment effect estimates from randomized trials. The outcome distributions ofparticipants with missing and observed data are often different, whichincreases the risk of bias. Causal inference methods may aid in reducing thebias and improving efficiency by incorporating baseline variables into theanalysis. In particular, doubly robust estimators incorporate estimates of twonuisance parameters: the outcome regression and the missingness mechanism, toadjust for differences in the observed and unobserved groups that can beexplained by observed covariates. Such nuisance parameters are traditionallyestimated using parametric models, which generally preclude consistentestimation, particularly in moderate to high dimensions. Recent research onmissing data has focused on data-adaptive estimation of the nuisance parametersin order to achieve consistency, but the large sample properties of suchestimators are poorly understood. In this article we discuss a doubly robustestimator that is consistent and asymptotically normal (CAN) underdata-adaptive consistent estimation of the outcome regression or themissingness mechanism. We provide a formula for an asymptotically validconfidence interval under minimal assumptions. We show that our proposedestimator has smaller finite-sample bias compared to standard doubly robustestimators. We present a simulation study demonstrating the enhancedperformance of our estimators in terms of bias, efficiency, and coverage of theconfidence intervals. We present the results of an illustrative example: arandomized, double-blind phase II/III trial of antiretroviral therapy inHIV-infected persons, and provide R code implementing our proposed estimators.