Prediction in Survival Analysis: Model or Medic?

Abstract

Subjective survival time predictions were obtained from experienced physicians for two groups of lung cancer patients. Predictions are compared with outcome by means of a non-standard loss function, which is also used to assess the accuracy of objective predictions based on proportional hazards models. Neither subjective nor objective predictions are particularly impressive. It is shown that the proportion of variation which can be explained by a proportional hazards model will invariably be relatively low as a result of the underlying assumptions, and hence point predictions should be expected to be relatively inaccurate for this family of models.