[Roaming through methodology. XXIV. Problems in generalizing from randomized intervention studies with selected patient populations].

Abstract

The generalizability of intervention research is dependent of the degree of success in including patients who represent the disease category for which the intervention might be considered. To be relevant for medical practice, estimates of intervention effects need to be quantitatively generalizable. This requires specification of the characteristics of the intervention (e.g. dosage and timing) and of the disease outcome, and of socalled effect-modifying factors: disease severity, co-interventions and the characteristics of the patients which influence the magnitude of the intervention effect. Literature remains less clear about the necessity to specify these effect-modifying factors. If effect modification is seriously considered, the investigator must decide either to design the trial in such a way that the subgroup-specific effects may be estimated, or to restrict the study to one homogeneous stratum of patients. Despite many theoretical reasons and some empirical examples of substantial effect modification by characteristics of patients, systematic comparison of randomised (selected populations) and non-randomised studies (less selection) fail to show many differences. This suggests that differences between subgroups are not very frequent and that trial results, provided they are adequately specified, generally do have a high degree of generalizability.