Random assignment of patients to comparison groups stochastically tends, with increasing sample size or number of experiment replications, to minimize the confounding of treatment outcome differences by the effects of differences among these groups in unknown/unmeasured patient characteristics. To what degree such confounding is actually avoided we cannot know unless we have validly measured these patient variables, but completely avoiding it is quite unlikely. Even if this confounding were completely avoided, confounding by unmeasured Patient Variable x Treatment Variable interactions remains a possibility. And the causal power of the confounding variables is no less important for internal validity than the degree of confounding.
Copyright 2003 Wiley Periodicals, Inc. J Clin Psychol.