Comparison of semiparametric regression models for correlated survival data
Résumé
Statistical models such as the semi-parametric Cox's model allow to analyse failure time data in prospective studies. One key assumption of these models is the independence between observations. In fact, this assumption is not fulfilled in the case of clustered samples, where measurements on subjects (individual level) belonging to a same group (cluster level) may be correlated. Simulations have been managed to study the estimation accuracy and test power of four models for correlated survival data: naive Cox's model, marginal Cox's model, Cox's model including fixed cluster effect and frailty Cox's model. The most important results are the following: (i) the frailty model presented the best fit and was able to correctly hold severe censoring, e.g., until 60%, (ii) gamma and normal distribution assumption for the frailty led to similar estimation and test results for the regression coefficient but not for the variance estimate of the frailty and (iii) the benefits of using models taking into account the correlation were more important in the case of a small number of big clusters than in presence of many little clusters.