A statistical model that takes into account patient heterogeneity in decision making

Statistical evaluation of clinical treatments or preventive medicine has profoundly contributed to decision making in medical fields such as with the acceptance of new treatment methods and health promotion policies. It is crucial in such decision making to find a correct statistical model to treat a surprisingly large variety of patients, or a heterogeneous group of patients, even with the same diagnosis. In diseases such as cancer, cardiovascular disease or diabetes, patients are often followed up to certain endpoints and these data are frequently analyzed by logrank tests or Cox-models to evaluate treatment effects. Although these methods have been widely accepted and extensively studied, we are sometimes faced with problems in applying these methods when the heterogeneity of patients is large and a lot of prognostic factors affecting the endpoints have to be considered. Based on the results of the analyses of survival data from more than 6,000 gastric cancer patients, it is revealed that the stratified logrank test may suffer serious power loss, even though primary prognostic factors are used as stratified factors. A so-called 'piecewise linear Cox regression method' for properly treating the heterogeneity of patients is introduced and extensively studied. This method is shown to be appropriate for patient groups with a high degree of heterogeneity such as the gastric cancer patients. The same method is, in principle, applicable to patients of other diseases, too, using statistical software such as SAS, BMDP and etc.