Proceedings of the 2013 annual meeting of the Netherlands Epidemiology Society
Volume 1 Issue S1 Abstract 13
H.E. Smedinga, Department of Public Health, Erasmus MC, Rotterdam
E.W. Steyerberg, Department of Public Health, Erasmus MC, Rotterdam
E.C. Zwarthoff, Department of Pathology, Josephine Nefkens Institute, Erasmus MC, Rotterdam
Y. Vergouwe, Department of Public Health, Erasmus MC, Rotterdam
Several survival models have been proposed for recurrent event data. The use of these models for clinical predictions has not been extensively examined, yet. Here, we compare variance correction models with shared frailty models in terms of their discriminative ability.
We analyzed a series of 615 patients with bladder cancer. Information was present for primary tumors and recurrent tumors (n = 456) with a maximum of three recurrences per patient. We created four prognostic models that differed on how they dealt with within- subjects correlation and event dependence. Within-subjects correlation was taken into account either by correcting the variance of the estimated regression coefficients of the predictors or by adding a frailty at the patient level. Event dependence was incorporated by either adding event number as covariate (non-stratified) or by stratifying the regression model on event number. To quantify discrimination, we estimated the c(oncordance)-index.
The discriminative ability of the variance correction models was relatively low with a c-index of 0.63 for the non-stratified model and 0.61 for the stratified model. The frailty models showed much higher discriminative ability with c-indices of 0.83 for both the non-stratified and stratified models. If we also included the individual frailty terms in the patient specific predictions, the discriminative ability was even higher.
Our results suggest that predictor effects estimated conditional on the frailties of the individuals have higher discriminative ability in a recurrent event context than the unconditional predictor effects. The way event dependence was taken into account was less important.
Published: 06 Jun, 2013