Statistical Methods for Interval Censored Data with Informative Sampling Weights and Unknown Subpopulation Not at Risk for the Event

Aiko Hattori, University of North Carolina at Chapel Hill
Chirayath Suchindran, University of North Carolina at Chapel Hill

Research on interval censored time to event using complex survey data faces three challenges: clustered data with informative sampling weights, interval-censored event times, and an unknown subpopulation not at risk. We use non-parametric maximum likelihood methods using Turnbull algorithm to estimate the Kaplan-Meier analog of survival function. We further extend the Pseudo Maximum Likelihood method to a random effect, mixture distribution model within the framework of Accelerated Failure Time model with particular attention to the scaling of sample weights. Using the National Longitudinal Study of Adolescent Health data, we confirm that regression coefficient estimates can be biased when informative sampling design is ignored. Also variances are underestimated when clustering is ignored. When two subpopulations, one at risk and the other not at risk for the event, are not addressed in a model, regression coefficient estimates may be biased and mask the true association between covariates and time to event.