On Regression Models When the Predictor Is Subject to Censoring

• Activity Number: 7
• Type: Invited
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): David Oakes*
• Companies: University of Rochester
• Keywords: mean residual life, survival analysis

Abstract

There is now a voluminous literature on regression models when the response variable is subject to right censoring, but relatively little on problems when the response variable is available for all individuals but the predictor is subject to right censoring. An exception is the "limit of detection problem" which corresponds to left censoring at a fixed point, often known as Type I censoring in the survival analysis literature. The importance of the general problem is due to the use of variables subject to possible right censoring as biomarkers in clinical studies. (Cai et al. Biostatistics, 7, 187-197, 2006). Omission of observations with censored predictors does not typically lead to bias but may cause substantial efficiency loss. Not only is the sample size reduced but the omitted observations will tend to be those of high leverage in the regression. We consider ways of recovering information available from the observations with censored predictors. We review some recent work using parametric models and propose a nonparametric approach based on the mean residual life function.