Please activate JavaScript!
Please install Adobe Flash Player, click here for download

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Poster Presentations 127Wednesday, 27th August 2014 • 11:00-11:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust P3.5.175 A censoring-robust concordance measure for proportional hazards regression models in external validation data: the calibrated Gönen and Heller estimator D van Klaveren1 , M Gönen2 , EW Steyerberg1 , Y Vergouwe1 1 Department of Public Health, Erasmus MC, Rotterdam, The Netherlands, 2 Epidemiology & Biostatistics, Memorial Sloan Kettering, New York, United States   For validation of proportional hazards regression models within the model development data, the Gönen and Heller (GH) concordance probability estimator is a censoring-robust alternative to Harrell’s concordance-index (c-index). In an external validation population it merely assesses the influ- ence of case-mix heterogeneity since it uses the regression coefficients of the development population. To estimate the concordance probability in external validation data we propose to apply the GH estimator to predic- tions with the regression slope calibrated to the external validation data, i.e. the calibrated GH estimator. We aimed to study the behaviour of the calibrated GH estimator in external validation settings with a focus on its sensitivity to censoring. In a simulation study we compared the calibrated GH estimator with Harrell’s c-index. We first generated hypothetical samples of patients (n=400 or n=1,600) for fitting Cox regression models. To mimic different external validation settings, we simulated new patient data with different true regression coefficients and different case-mix heterogeneity (10.000 replications per setting). In each setting we varied the amount of censor- ing from 0% to 70%. The calibrated GH estimates remained stable at the true value with increasing proportions of censoring, while the c-index in- creased unfavourably. The calibrated GH estimator was further illustrated in a clinical example. We conclude that the calibrated GH-estimator is an attractive, censoring- robust alternative for the c-index when assessing the discriminative ability of a proportional hazards regression model in external validation data.

Pages Overview