Methods for Cox Regression with Nonclassical Measurement Error in the Covariates with an Application to Nutritional Epidemiology

• Activity Number: 420
• Type: Contributed
• Date/Time: Wednesday, August 1, 2007 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Pamela A. Shaw*+ and Ross L. Prentice
• Companies: National Institutes of Health and Fred Hutchinson Cancer Research Center
• Keywords: Measurement Error ; Survival Analysis ; Nutritional Epidemiology ; Cox Regression

Abstract

Self-reported measures of diet are error prone, yet most historical analyses relating diet to disease have not adjusted for systematic aspects of measurement error. We develop methods to analyze censored failure time data with Cox regression when the covariate of interest has both subject-specific and systematic measurement error. Three techniques for the classical error model, riskset regression calibration (Xie, Wang, and Prentice, JRSSB, 2001), conditional score (Tsiatis and Davidian, Biometrika, 2001), and non-parametric corrected score (Huang and Wang, JASA, 2000), are extended to the case where the covariate of interest follows a generalized measurement error model and a covariate measured with unbiased error is available on a subset. An example from the Women's Health Initiative motivates this more flexible model and provides the basis of a simulation study to compare the methods.