Virtual controls as an alternative to randomized controlled trials for assessing efficacy of interventions

Randomized controlled trials are ubiquitously spoken of as the “gold standard” for testing interventions and establishing causal relations. This article presents evidence for two premises. First: there are often major problems with randomized designs; it is by no means true that the only good design is a randomized design. Second: the method of virtual controls in some circumstances can and should replace randomized designs.

Randomized trials can present problems with external validity or generalizability; they can be unethical; they typically involve much time, effort, and expense; their assignments to treatment conditions often can be maintained only for limited time periods; examination of their track record reveals problems with reproducibility on the one hand, and lack of overwhelming superiority to observational methods on the other hand.

The method of virtual controls involves ongoing efforts to refine statistical models for prediction of outcomes from measurable variables, under conditions of no treatment or current standard of care. Research participants then join a single-arm study of a new intervention. Each participant’s data, together with the formulas previously generated, predict that participant’s outcome without the new intervention. These outcomes are the “virtual controls.” The actual outcomes with intervention are compared with the virtual control outcomes to estimate effect sizes. Part of the research product is the prediction equations themselves, so that in clinical practice, individual treatment decisions may be aided by quantitative answers to the questions, “What is estimated to happen to this particular patient with and without this treatment?”

The method of virtual controls is especially indicated when rapid results are of high priority, when withholding intervention is likely harmful, when adequate data exist for prediction of untreated or standard of care outcomes, when we want to let people choose the treatment they prefer, when tailoring treatment decisions to individuals is desirable, and when real-world clinical information can be harnessed for analysis.