A diagnostic plot for guiding the choice of the frailty distribution in clustered survival data

Clustered survival data are often analysed using frailty models. The frailty distribution provides a way to model the type of dependence between event times within a cluster. In this paper, we propose a new diagnostic plot to guide the choice of the frailty distribution. Our approach is based on the idea that although frailties are unobservable the dependence structure that they impose on the data can be observed. We estimate the observed association and we compare it with model-based structures obtained from different frailty distributions. To measure association, we use quantile dependence coefficients. The method easily accommodates any cluster sizes, non-ordered observations within clusters, and various censoring schemes.