A Study of Grouped Failure Time Data: Misapplication of Repeated Binary Responses

Abstract

When epidemiological studies assess terminal binary responses at pre-determined follow-up times, researchers must decide which observations will be included at each follow-up time. Some recent studies of bone erosion in rheumatoid arthritis (RA) patients have recorded subjects' responses at follow-up times after the terminal response is detected. The objective of this thesis is to examine how models for binary data perform when a terminal response continues to be observed at follow-up times after it has been recorded. A Monte-Carlo simulation study and analysis of data from a study of AIDS onset are performed to compare logistic regression, generalized linear mixed models (GLMMs), a generalized estimating equations (GEE) approach, and models for grouped failure time data. The comparison of regression coefficient estimates, their associated variance, empirical type I error, and power are discussed. This study concludes that certain GLMMs and GEE models are invalid when applied to grouped failure time data.