A1215
Title: Matching Estimators to censoring rate improves inference with outcome-dependent right-censored covariates
Authors: Jesus Vazquez - Johns Hopkins University (United States) [presenting]
Yanyuan Ma - Pennsylvania State University (United States)
Karen Marder - Columbia University (United States)
Tanya Garcia - University of North Carolina at Chapel Hill (United States)
Abstract: Disease progression studies often model health outcomes as a function of time to a clinical eventfor example, motor and cognitive health scores as a function of time to disease onset. For participants who have not yet experienced the event, this time-to-event becomes a right-censored covariate, and censoring often depends on the health outcome being modeled, violating standard regression assumptions. Three consistent estimators for the outcome-dependent right-censored covariate problem are developed: two augmented inverse probability weighted (AIPW) estimators and one maximum likelihood estimator (MLE). A central contribution demonstrates that estimator performance depends on censoring rate: under high censoring, the AIPW estimator leverages frequently-observed censoring times that improve estimation and statistical inference while the MLE struggles with rarely-observed event times; under low censoring, the pattern reverses. Applied to Huntington disease data characterizing decline as individuals approach mild cognitive impairment, the AIPW estimator revealed biologically plausible patterns with 25\% greater precision than the inverse probability weighting estimator, while the MLE produced counterintuitive covariate effects at 78\% censoring. An R package implementing all estimators with guidance for selecting among them based on censoring rate enables researchers to use event-based timescales for more clinically meaningful estimation and statistical inference.
