IASS - Correcting Bias from Covariate-Dependent Censoring in Survival Disparity Decomposition
Free
- Length
- 1 hr 30 min
- Platform
- Zoom
- Topic
- Other
About this webinar
Speaker: Yan Li
Abstract: Research on racial and socioeconomic disparities in time-to-event outcomes is central to public health, but common decomposition tools can be biased when censoring depends on baseline covariates. Traditional Peters-Belson (PB) survival decompositions typically combine Kaplan-Meier (KM) estimates for observed group survival with Cox-model predictions for counterfactual survival. This workflow creates two problems. First, KM targets marginal survival only under censoring independent of all covariates; this is implausible when attrition, dropout, linkage failure, or competing risks are related to demographic, socioeconomic, and behavioural characteristics. Second, averaged Cox predictions are conditional-model quantities and need not reproduce a marginal survival curve. The observed and counterfactual pieces of the decomposition may therefore be calibrated to different targets.
This extended abstract presents a semiparametric inverse-probability-of-censoring-weighted propensity-score (IPCW-PS) framework for survival disparity decomposition., The method estimates observed and counterfactual survival with the same weighted estimating logic, corrects covariate-dependent censoring, accommodates sampling weights, and preserves the interpretation of overall, unexplained, and explained disparity over follow-up time.