Prognostic factors in clinical cancer trials.

Differences in nontreatment-related covariates ("prognostic factors") often account for differences between treatments in clinical cancer trials. This is true even in randomized trials in which the number of patients randomized is less than 200. Hence analysis of prognostic factors is crucial in historically controlled and small randomized trials. However, it is known that factors found prognostic in one series are often not prognostic in another. This is a result of the increased type I error inherent in most methods used to identify the optimal cutpoint of a potential prognostic factor. This report describes new graphical methods, based on Martingale residuals, that can be used to better identify the relationship between outcome and a covariate. For example, use of these methods indicated that the effect of antecedent hematological disorder on survival in acute myelogenous leukemia/myelodysplastic syndrome is continuous (the longer the length of antecedent hematological disorder, the shorter the survival) rather than dichotomous (antecedent hematological disorder present = unfavorable). This report also discusses the use of graphical methods to test the assumption of proportional hazards crucial to the Cox model and to account for any time-varying effects of a prognostic factor. The graphical methods discussed here provide a better fit of a statistical model to the data and provide more reliable estimates of the effect of a particular variable.