Nonparametric Bayesian Methods for Prediction of Event Times for Analysis with Interval-Censored Data

• Activity Number: 498
• Type: Topic Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Government Statistics
• Author(s): Stephanie Lustgarten*+ and Gheorghe Doros
• Companies: Boston University and Boston University
• Keywords: Bayesian ; survival ; Gibbs sampler ; non-parametric ; interval-censored ; prediction

Abstract

In trials with failure-time primary outcomes, statistical information is determined by accumulated events. Interim and final analyses are performed after observing a specified number of events. It is of interest to predict when such events will be observed based on accumulating data. It is often impractical to assess patients continuously for the primary endpoint, and data are interval-censored. To predict the time of events, we propose a flexible fully Bayesian non-parametric approach in modeling survival probabilities that generalizes to interval-censored data. We use a Gibbs sampler to sample from the survival distribution posterior to obtain point and interval estimates for the time of events. Parametric and semi-parametric methods have been proposed for such prediction. In cases when intervals are wide relative to true failure time, such methods may not provide accurate, efficient prediction. Accuracy and precision of our proposed non-parametric approach is compared to parametric and semi-parametric methods that treat interval-censored data as right-censored data. Our proposed method offers greater flexibility and based on our studies can match or outperform existing methods.