ML and REML estimation in survival analysis with time dependent correlated frailty

In the study of multiple failure times for the same subjects, for example, recurrent infections for patients with a given disease, there are often subject effects, that is, subjects have different risks that cannot be explained by known covariates. Standard methods, which ignore subject effects, lead to overestimation of precision. The frailty model for subject effects is better, but can be insufficient, because it assumes that subject effects are constant over time. Experience has shown that the dependence between different time periods often decreases with distance in time. Such a model is presented here, assuming that the frailty is no longer constant, but time varying, with one value for each spell. The main example is a first-order autoregressive process. This is applied to a data set of 128 patients with chronic granulomatous disease (CGD), participating in a placebo controlled randomized trial of gamma interferon (γ-IFN), suffering between 0 and 7 infections. It is shown that the time varying frailty model gives a significantly better fit than the constant frailty model.