Increasing the Expected Value to Society of Clinical Research Studies?

The value of a methodological paper generally depends on 1) the importance of the problem addressed, 2) the completeness of the presentation of the issues and prior progress, and 3) a new solution to the problem that is feasible and better than what has already been proposed. The Uyei and Braithwaite article (hereafter UB) clearly satisfies the first two criteria and is well worth contemplating. The article is thought-provoking, not only because it raises issues related to appropriate choice of control groups in randomized clinical trials (RCTs) on which many (I for one) have become complacent, but because it focuses attention on the societal impact of RCTs so often ignored. My focus here is on 3 issues: value to society, nonplacebo control groups, and benefit-to-harm balance, all central to UB arguments. Nowhere in UB is there a definition of the phrase ‘‘value to society.’’ Let me propose a definition: the value of any clinical research study to society depends on how much that study can contribute to improving medical decision making for patients like those in the research study. Clearly, major influences on the value of a study are how serious and/or how widespread is the disorder under study, the consequences of inadequate intervention, and the feasibility, costs, and risks of adequate intervention—all clinical issues. Here, and in UB, the focus is only on design decisions. An invalid design that misleads clinical decision making (type I error), a valid design with inadequate power that leaves the research question unanswered (type II error), or a valid and powerful design that answers a clinically irrelevant research question (sometimes called type III error) are all of little value to society. The RCT is one type of clinical research study, specifically meant to inform choice between treatments (treatment T v. control/comparison C) for patients with a certain disorder. For an RCT to have value to society, the research question must be important to clinicians, and the design decisions (sampling, measurement, design, analysis) must lead to an accurate and precise indication of the clinical effect size of T v. C, some indication of how much better patient outcomes would be if T were used rather than C in the population sampled. To be specific about such an effect size, if one patient were sampled from that population and given T and another given C, what is the probability that the one given T would have an outcome clinically preferable to the one given C (symbolically T . C) compared to the opposite (C . T): the success rate difference (SRD)? If the outcome measure were univariate and binary (success/fail), SRD equals p1 – p2, the effect size used in UB. If the outcome measure were univariate and normally distributed, SRD = 2F(d/O 2) – 1, where F() is the standard cumulative normal distribution and d a form of Cohen’s d. SRD can be estimated in an RCT for any outcome measures provided only that one could use them to compare the outcomes of 2 patients and decide which (if either) is clinically preferable, a necessary condition for an RCT outcome measure. Now, with estimation of SRD as the goal, let us now reconsider the UB arguments. The discussion in UB focused on a limited subset of such RCTs, only those when no proven intervention existed. The issue of placebo (C = P) or nocebo/ null treatment (C = N) control groups is more heatedly debated when effective interventions exist, for then, using either P or N might be viewed as ethically questionable, as essentially withholding effective treatment from patients for the duration of the RCT. In Received 7 August 2015 from the Stanford University, Palo Alto, CA, USA (HCK). Revision accepted for publication 8 August 2015.