Simons R, Altman DG. Statistical aspects of prognostic factor studies in oncology. Br J Cancer 1994;69:979-985.

  • A principle of good research requires that hypotheses be stated in advance.
  • Study design. Prognostic factors studies can be prospective, in which newly diagnosed patients are entered and followed up for an adequate length of time to allow a comparison of outcome for groups with different baseline values fo the factors of interest.
  • Inclusion and exclusion criteria should be carefully defined.
  • For retrospective studies there is a risk of bias because stored smaples are likely to include a disproportionate number of larger tumours.
  • For regression modelling the number of events (e.g. death) should be at least ten times the number of potential prognostic variables that could be included in the model.
  • Regression models such as Cox's proportional hazards regression model are often used to study the joint influence of several prognostic factors.
  • forward selection, backword elimination, stepwise regression
  • The fact that a marker is statistically significantly associated with outcome does not necessarily mean that it is important. Importance depends on the degree to which the marker influences patient outcome.
  • When a model contains several prognostic variables it is sometimes useful to construct a prognostic index, which is a new variable combining the information from all the prognostic factors.
  • reporting prognostic studies
    • inclusion and exclusion criteria
    • the relevant clinical and demographic characteristics of the sample shoudl be described
    • authors should describe all analyses performed, not just those analysses and variables selected for inclusion in the reports.