Prognostic assessment from studies with non-randomized treatment assignment.

Abstract

In clinical practice, prognostic factors are primarily used to identify and select patients with a relatively poor prognosis who may need more aggressive treatment in order to improve their prognosis. Preferably, assessment of the ability of prognostic factors to distinguish these high risk patients from low risk patients should take place in the absence of such prognosis-modifying adjuvant therapy. Adjuvant therapy may dilute the effect of prognosticators (covariate-treatment interactions) and such a situation may go unnoticed in tests for interaction because of their low statistical power. These undetected covariate-treatment interactions will bias the estimated effects of prognostic factors in both studies with randomized and non-randomized treatment assignment but the bias will usually be greater in studies when treatment was not randomized. For the evaluation of prognostic factors, it is therefore argued that the study population should be restricted to patients who do not received adjuvant therapy. This restriction might result in a less powerful analysis, but it carries the advantage that undetectable biases are avoided. Non-random treatment assignment will not invalidate this restriction analysis provided that all the prognostic factors on which treatment choice was based, are known and incorporated into the analysis.