A method for checking regression models in survival analysis based on the risk score (Q1366899)

Regression models, especially Cox' proportional hazards model, have become increasingly popular in survival analysis. Recently linear alternatives have been proposed; both non-parametric and semiparametric. The available methods for checking model fit are, however, still insufficient, even for the Cox model. For linear normal models, residual methods are the primary tools in the assessment of the model fit. For regression models in survival analysis diagnostic methods based on (martingale) residuals have been proposed, but the interpretation of the individual residuals is difficult. We discuss methods for checking the overall fit of the regression models based on martingale residual processes grouped after the risk score. We consider both graphical methods and formal goodness of fit tests. By stochastic simulation, we study how well the methods are able to detect errors in the model specification. The simulations are supplemented by a real data example of patients with primary biliary cirrhosis (PBC) of the liver.