ABSTRACT
Threshold for decision making shapes the busy clinical routine. Many clinicians and statisticians adamantly demand that confidence intervals are placed around predicted probabilities to indicate the uncertainty in the prediction. The main reason for the limited usefulness of confidence intervals for predicted probabilities lies with interpretation. It is not uncommon to see a Hosmer-Lemeshow test reported to assess the calibration of a risk prediction model. We strongly discourage analysts from using backward elimination, forward selection, and any other form of stepwise selection for the reasons. As an alternative, one should use a more sophisticated algorithm, such as penalized regression controlled by cross-validation. An extremely popular measure for the performance of a prediction model applied to survival data is a rank correlation coefficient called the c-index. With time-to-event outcome, the integrated Brier score is obtained by cumulating the Brier score across a sequence of prediction time horizons.
