On the Problem of Using ‘Optimal’ Cutpoints in the Assessment of Quantitative Prognostic Factors

The identification and assessment of prognostic factors is an important task in clinical cancer research. Quantitative prognostic factors are often categorized by using one or several cutpoints to obtain an easier interpretation with respect to prognosis of the resulting patients’ subgroups. Considering the selection of a data-driven ‘optimal’ cutpoint as a ‘prototype’ of statistical model building, we demonstrate that prognostic relevance of a single factor with no effect can solely be produced by the statistical model building process. Furthermore, we show how to overcome these problems by using corrected P-values and shrinkage methods. The problem and its solutions are illustrated by using the data of two breast cancer studies.