No label defined (Q4026593)

[For the entire collection see Zbl 0745.00009.] In the usual analysis of survival data by the logrank test and the proportional hazards model it is assumed that the treatment effects are the same for all individuals and that all relevant covariates are measured and affect the failure rate by multiplying the hazard function by a constant that does not change with follow up time. Here we focus on the situation in which a treatment effect need not be the same for all individuals in a study. This is modeled by an interaction between the treatment indicator and an unobserved random effect, or frailty. We first study this question in a parametric framework in which the baseline hazard is assumed known. Here we compute a score test, examine its properties, and also study estimation by maximum likelihood. In the subsequent section we consider the problem in a semiparametric set up where the baseline hazard is assumed unknown. We develop an analog of the score test and study its properties. An asymptotically equivalent approximation is also considered which is more amenable to mathematical analysis and which behaves better in small samples. Estimation is much more difficult in this context and is only briefly considered.