Ordinal outcome regression with censored covariates and cured fraction

Thomas M. Menino Convention & Exhibition Center 

This work develops a statistical framework for regression analysis with ordinal outcomes, where the important covariates are subject to right censoring and a portion of the population is cured, meaning that the cured individuals will never experience an event. We propose a two-stage regression analysis method for cross-sectionally sampled data with an ordinal outcome, measured at the time of sampling. The proposed approach, first, models the probability of being cured and the distribution of the time-to-event covariate for the uncured individuals. In the second stage, we estimate the ordinal outcome regression, using either a proportional odds regression model or an adjacent categories regression. Our approach accounts for censored covariates and for the presence of cured subjects in the data, which improves estimation efficiency compared to the complete case analysis and removes bias compared to naive analysis. We provide identifiability conditions for our approach, develop asymptotic distribution of our estimator and investigate its performance in simulations. Finally, we apply it to the Adult Perthes' Disease data. An R package implementing our approach is also developed.

selection bias

cross-sectional sampling

complete case analysis

mixture model

cure survival model

identifiability 

Biometrics Section