Failure time models with matched data

Abstract

SUMMARY Holt & Prentice (1974) adapted the proportional hazards model to give failure time models for matched pairs in which a member of the i th pair with covariate z has hazard function λ oi ( t ) exp(β z ). With full data, more efficient estimators of β can be obtained if a Weibull specialization, λ i nt nt –1 , of the baseline, hazard can be assumed but these estimators are unavailable for censored data. Estimates of β formed by assuming the λ i 's are independent gamma variates are investigated. The resulting estimators are found to be more efficient than those obtained by Holt & Prentice when the λ i ,'s are in fact gamma, to be available for censored data, and to perform well under a number of very different unimodal alternatives to a gamma distribution for the λ i 's.