Abstract:
|
Standardization estimates the counterfactual risks in a target population, which may be exposed, unexposed, or total (exposed + unexposed) population. For each target, it is typically demonstrated nonparametrically, whereas two model-based methods exist for commonly-seen epidemiologic data with high-dimensional confounders: regression standardization featuring outcome regression (OR) model, and inverse probability of exposure weighted-reweighted estimators featuring propensity score (PS) model. Our aim is to address the challenges shared by these estimators that arise in practice: model-dependency for OR and PS, and unavailability for censored event data. For total, exposed, and unexposed targets, we provide doubly robust estimators of standardized risks from censored events, which only require correct specification of at least one of the OR or PS models. With dependently censored events, the method further requires the model for censoring, but the following double robustness still holds: consistent under correct specification of either (a) the model for event hazard, or (b) the models for censoring hazard and PS. The estimators are evaluated in simulated and empirical datasets.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.