Premium
A general, prediction error‐based criterion for selecting model complexity for high‐dimensional survival models
Author(s) -
Porzelius Christine,
Schumacher Martin,
Binder Harald
Publication year - 2010
Publication title -
statistics in medicine
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.996
H-Index - 183
eISSN - 1097-0258
pISSN - 0277-6715
DOI - 10.1002/sim.3765
Subject(s) - overfitting , model selection , computer science , resampling , cross validation , selection (genetic algorithm) , predictive modelling , data mining , data set , statistics , set (abstract data type) , prediction interval , machine learning , algorithm , artificial intelligence , mathematics , artificial neural network , programming language
When fitting predictive survival models to high‐dimensional data, an adequate criterion for selecting model complexity is needed to avoid overfitting. The complexity parameter is typically selected by the predictive partial log‐likelihood (PLL) estimated via cross‐validation. As an alternative criterion, we propose a relative version of the integrated prediction error curve (IPEC), which can be stably estimated via bootstrap resampling. The IPEC has the advantage of being applicable for models and fitting techniques where the PLL is not available. To investigate the performance of this new criterion, a simulation study is carried out, mimicking microarray survival data. Additionally, model selection by predictive PLL, estimated via bootstrap resampling instead of cross‐validation, is examined. It is seen that this mostly results in similar prediction performance of the selected models, compared to estimates based on cross‐validation. Model selection by bootstrap estimates of the IPEC performs about as well as selection by cross‐validation estimates of the PLL. Therefore, it is expected to be a reasonable alternative in cases where there is no PLL. Similar results are seen in the analysis of a microarray survival data set from patients with diffuse large‐B‐cell lymphoma. Copyright © 2010 John Wiley & Sons, Ltd.