[Roaming through the methodology. XX. Randomization as a means of avoiding confounding through indication]

Two methodological issues are central in discussions about the value of randomised studies for evaluation of the effectiveness of medical interventions: confounding by indication and generalizability. Confounding occurs when a factor with bearing on the outcome of disease, has a different distribution among intervention groups. Confounding by indication is a special case and results from the fact that in medical practice the choice between intervention alternatives is dependent on the perceived prognosis of the patient for whom these interventions are considered. As the precise arguments for this choice in general are not explicit and hence not measurable, correction for confounding by indication is impossible, and randomisation is applied. Surprisingly, systematic comparison of the results of randomised and non-randomised studies for the same interventions shows remarkable resemblance. This suggests that non-randomised studies may have value in specific situations, but unfortunately so far we do not know for which situations this may be the case.