Regression modelling strategies for improved prognostic prediction

• DOI: 10.1002/sim.4780030207
• Publication Date: 2007-04-23T13:50:00Z
• Authors: Frank E. Harrell; Kerry L. Lee; Robert M. Califf; David B. Pryor; Robert A. Rosati
• Keywords: Overfitting; Regression diagnostic; Principal component regression

Abstract Regression models such as the Cox proportional hazards model have had increasing use in modelling and estimating prognosis of patients with a variety diseases. Many applications involve large number variables to be modelled using relatively small patient sample. Problems overfitting identifying important covariates are exacerbated analysing because accuracy is more function events than sample size. We used general index predictive discrimination measure ability developed on training samples varying sizes predict survival an independent test suspected having coronary artery disease. compared three methods fitting: (1) standard ‘step‐up’ variable selection, (2) incomplete principal components regression, (3) regression after developing clinical indices from clusters. found offer superior predictions sample, whereas offers easily interpretable nearly good models. Standard selection has deficiencies.