Many studies have attempted to define useful prognostic and predictive fact
ors in cancer but few have achieved acceptance in clinical practice because
of methodological weaknesses. These include failure to test clearly formul
ated hypotheses, inadequate sample size, inappropriate multiple significanc
e testing, arbitrary definition of patient groups, inadequately reproducibl
e assays, and failure to verify prognostic factors with data independent of
the data which suggested the original hypothesis. This unsatisfactory situ
ation will persist until critical attention is routinely paid to study desi
gn and prospective validation of supposed prognostic and predictive factors
, without which classical approaches will be suboptimally exploited and the
flood of data from new molecular technologies will not be used effectively
. We propose that prognostic factors should be evaluated in three phases: I
, assay definition; II, retrospective testing; III, prospective testing, id
eally as a designed part of clinical trials.