Statistical aspects of prognostic factor studies in oncology

Abstract

Studies of new prognostic factors form an extensive part of the literature of oncology. Recent examples published in this journal include studies of patients with primary cerebral lymphoma (Blay et al., 1993), head and neck sarcomas (Eeles et al., 1993) and testicular cancer (Steyerberg et al., 1993). In some cases such studies may offer insight into the molecular pathogenesis of the disease. In others they may help in medical decision making; for example, in identifying which patients are at sufficiently high risk of recurrence to warrant a toxic or expensive treatment. Identification of major prognostic determinants can facilitate the design of further clinical trials, aid in inter-trial comparisons and guide the counselling of individual patients. Unfortunately, however, the results of different prognostic factor studies are often inconsistent or contradictory. Whereas widely accepted methodological principles have evolved to guide the design, conduct, analysis and reporting of clinical trials, no similar guidelines exist for prognostic factor studies. McGuire (1991), Levine et al. (1991) and Gasparini et al. (1993), discussed some general methodological problems with prognostic factor studies. Here we will consider some of the statistical problems common to prognostic factor studies as a step towards the development of broadly acceptable principles that might guide their conduct and reporting.