Dangers of Using ‘Optimal ’ Cutpoints in the Evaluation of Cyclical Prognostic Factors

It is common strategy in medical research to categorize a continuous covariable before evaluating its prognostic impact on clinical outcome. In most cases the covariable is divided into just two groups. The chosen cutpoint is either a value already published in other studies, or a certain sample quantile like the median, or a so-called ‘optimal cutpoint’, that is the value which corresponds to the most significant relation with outcome. Because the multiple testing problem is often ignored, the term ‘optimal ’ is misleading in this context. Altman et al. (1994) suggest that the method be called the ‘minimum P-value approach ’ instead, and present simulation and asymptotic results of the inflation of the type I error rate. Recently the influence of menstrual status at the time of surgery on the prognosis of women suffering from breast cancer was discussed in the medical literature. Although the paper which triggered the discussion, reported a high relative risk for death in patients who underwent breast cancer surgery during the perimenstrual period, almost all of the subsequently published work could not confirm this result in retrospective studies. The menstrual status at the time of breast cancer surgery is a cyclical covariable. Its splitting into two segments is a similar strategy of analysis like the categorization of a continuous covariable. In the case that this splitting is based on a minimum P-value search, the problem of multiple testing has to be taken into account, too. Following Altman et al. (1994), a simulation study was performed to gain some insight into the relation of the actual versus nominal type I error rate with regard to the breast cancer example