An application of Harrell's C-index to PH frailty models

Authors

  • R. Van Oirbeek,

    Corresponding author
    1. Interuniversity Institute for Biostatistics and Statistical Bioinformatics, Katholieke Universiteit Leuven, Kapucijnenvoer 35, Blok D, bus 7001, B3000 Leuven, Belgium, and Universiteit Hasselt, Belgium
    • Interuniversity Institute for Biostatistics and Statistical Bioinformatics, Katholieke Universiteit Leuven, Kapucijnenvoer 35, Blok D, bus 7001, B3000 Leuven, Belgium, and Universiteit Hasselt, Belgium
    Search for more papers by this author
  • E. Lesaffre

    1. Interuniversity Institute for Biostatistics and Statistical Bioinformatics, Katholieke Universiteit Leuven, Kapucijnenvoer 35, Blok D, bus 7001, B3000 Leuven, Belgium, and Universiteit Hasselt, Belgium
    2. Department of Biostatistics, Erasmus Medical Center, PO Box 2040, 3000 CA Rotterdam, The Netherlands
    Search for more papers by this author

Abstract

Frailty models are encountered in many medical applications, yet little research has been devoted to develop measures that quantify the predictive ability of these models. In this paper, we elaborate on the concept of the concordance probability to clustered data, resulting in an ‘Overall Conditional C-index’ or bfCO, C and an ‘Overall Marginal C-index’ or CO, M. Both Overall C-indices can be split up into a ‘Between Conditional’ or CB, C and a ‘Between Marginal C-index’ or CB, M and into a ‘Within Conditional’ or CW, C and a ‘Within Marginal C-index’ or CW, M. For PH frailty models of the power variance family, CW, C and CW, M are equivalent resulting in one ‘Within C-index’ CW. We propose an application of Harrell's C-index to estimate the proposed indices within a likelihood and a Bayesian context and the performances of their point estimates and confidence/credible intervals are compared in an extensive simulation study. This simulation study shows that the point estimates of CW and CB, M perform good within both a likelihood and Bayesian context but that the point estimates of CB, C show less bias for the Bayesian approach than for the likelihood approach. The 95 per cent confidence/credible intervals also possess good coverage properties, given that the point estimates perform good. The performance of the C-indices is evaluated on a real data set. Copyright © 2010 John Wiley & Sons, Ltd.

Ancillary