Commentary on Hey and Kimmelman.

Randomization is firmly established as a cornerstone of clinical trial methodology. Yet, the ethics of randomization con-tinues to generate controversy. The default, and most efficient, allocation scheme randomizes patients equally (1:1)across all arms of study. However, many randomized trials are using outcome-adaptive allocation schemes, which dyna-mically adjust the allocation ratio in favor of the better performing treatment arm. Advocates of outcome-adaptive allo-cation contend that it better accommodates clinical equipoise and promotes informed consent, since such trials limitpatient-subject exposure to sub-optimal care. In this essay, we argue that this purported ethical advantage of outcome-adaptive allocation does not stand up to careful scrutiny in the setting of two-armed studies and/or early-phase research. KeywordsAdaptive randomization, ethics, equipoise, therapeutic misconception Randomization is firmly established as a cornerstone ofclinical trial methodology. Yet, the ethics of randomiza-tion continues to generate controversy. On its face, ran-dom allocation appears to conflict with both a patient’sbest interest and the ethics of clinical practice. Whywould a patient ever agree to receive a randomly chosentreatment?The concept of clinical equipoise provides an answerand helps to resolve this fundamental tension. Clinicalequipoise demands that there exists a state of commu-nity uncertainty about the relative therapeutic meritsacross all arms of a trial. Insofar as this conditionholds, all patient-subjects enrolled in a trial can beassured of receiving nothing less than competent medi-cal care. Generally speaking, the randomized controlledtrials (RCTs) for serious illnesses are ethically accepta-ble only where conditions of clinical equipoise hold(for a survey of debates over clinical equipoise, seeLondon).However, clinical equipoise is silent on the nature ofthe random allocation scheme. The default, and mostefficient, allocation scheme randomizes patients equally(1:1) across all arms of study. But many RCTs useunequal allocation schemes (e.g. 2:1 or 3:1), whichassign more patients to the experimental intervention,or outcome-adaptive allocation schemes, which dyna-mically adjust the allocation ratio in favor of the betterperforming treatment arm. These alternative alloca-tion schemes are often justified by appeals to increasingpatient-subject benefit. That is, by weighting the allo-cation ratio in favor of the experimental interventionor ‘‘better performing’’ arm, the trial increases the num-ber of patients who receive the presumed superior treat-ment. Therefore, advocates contend that unequalallocation schemes can better accommodate clinicalequipoise, since such trials limit patient-subject expo-sure to sub-optimal care.While this argument is appealing, it deserves carefulscrutiny. Elsewhere, we argued that unequal allocationratios in confirmatory trials are ethically problematic,since they fail to minimize patient burden, leveragepatients’ therapeutic misestimations, and potentiallyintroduce new internal validity threats. What abouttrials that use outcome-adaptive allocation?In what follows, we consider whether criticisms ofunequal allocation apply to two-armed studies usingadaptive allocation (see, for example, Maki et al. andGarcia-Manero etal.). These studies—which are themost contentious adaptive allocation study types—generally begin by randomizing patients on a 1:1 basis Studies of Translation, Ethics and Medicine (STREAM), Biomedical Ethics Unit, McGill University, Montréal, QC, Canada Corresponding author: Jonathan Kimmelman, Studies of Translation, Ethics and Medicine (STREAM), Biomedical Ethics Unit, McGill University, 3647 Peel Street, Montréal, QC H3A 1X1, Canada. Email: [email protected] to treatment or control. As preliminary patient-response data emerge, the allocation ratio is adjusted infavor of the better performing arm. Once the probabil-ity of accepting or rejecting the null hypothesis hasdropped below a pre-specified threshold, the trial isclosed. After reviewing the ethics of two-armed adap-tive allocation trials, we briefly consider the ethics of aless contentious study design: multi-arm adaptive allo-cation (see, for example, Giles et al. and Kim et al.). Minimizing patient burden A foundational ethical principle is that trials shouldminimize research burdens. Proponents of outcome-adaptive allocation designs argue that they advance thischarge, since they maximize the number of patientsreceiving superior treatments (and consequently mini-mize the number receiving inferior ones).This argument holds little water when consideredagainst the background of drug development. First, theprimary objective of early-phase clinical testing is toestablish the evidentiary grounds for conducting confir-matory trials. Whereas confirmatory trials employ clin-ical endpoints (e.g. survival) and large sample sizes,early-phase trials typically use surrogate endpoints (e.g.tumor response) that can be collected in a shorter timeframe, and smaller sample sizes. The fact that surrogatebenefit and underpowered studies are fickle guides forclinical benefit is indicated by the frequent discordancebetween effect sizes in phase 3 studies and those inphase 2. Indeed, the average probability that an inter-vention reaching phase 3 testing will advance to regula-tory licensure—a good proxy for clinical utility—isapproximately 50%. This probability is considerablylower in realms like neurology and oncology, whereadaptive allocation is often employed; not infrequently,agents that look very promising in early-phase trialsactually make patients worse off against comparatorsin late phase trials. This means that in a best case sce-nario where one drug outperforms another in anoutcome-adaptive study, patients receiving the betterperforming drug still only have a 50% probability ofreceiving a drug that is competitive with or better thanstandard of care.Second, new treatments tend to deliver only smallimprovements over standard ones. Several systematicreviews of RCTs in cancer, for example, show that theodds ratios cluster close to unity. However, adaptiverandomization tends to offer advantages only wheneffect differences between two interventions are large.As a consequence, the purported therapeutic advan-tages of adaptive randomization will rarely be realized,yet their disadvantages in terms of requiring larger sam-ple sizes will often be encountered.These arguments take on greater force in the contextof phase 3 trials. In favor of adaptive allocation in thiscontext is the fact that allocation adjustment will bebased not on surrogate, but rather clinical endpoints.This means that better performing arms are much morelikely to be clinically superior than they would be inphase 2 studies. This advantage, however, will often—though not always—be nullified by the lengthy periodbetween enrollment and observation of clinical out-comes. These trials must therefore be designed to eitherrecruit very slowly, in order to gain the benefits ofadaptive allocation, or else they may have already com-pleted enrollment by the time sufficient outcome evi-dence is available to adjust the allocation ratio.Advocates of adaptive allocation might offer tworejoinders. First, they might argue that inefficienciesdescribed above work to the advantage of patients,since larger sample sizes afford more opportunities forpatients to enter trials and access superior treatment(e.g. Berry seems to suggest this point). Leaving asidethe fact that this advantage only holds if standard ofcare proves inferior, this reasoning neglects the econo-mies that arise in dividing labor. Research systems areset up to resolve scientific uncertainties; care systemsare established to deliver therapies based on that resolu-tion. Running larger trials because they offer patientsbetter care options asks research systems to shouldertasks that care systems are designed to carry. It is farbetter to resolve scientific questions as efficiently aspossible, so that on the one hand research resources canbe channeled towards resolving other uncertainties,while on the other, the therapeutic ‘‘baton’’ can be expe-ditiously passed to systems that are designed to delivercare.Concerning the efficiency of the research enterpriseas a whole, all commentators agree that outcome-adaptive trials are more complicated and expensive toplan and coordinate. While some research centers maybe able to implement outcome-adaptive allocation as arule, the requisite funding and stakeholder support can-not be assumed to hold across the research enterprise.To be sure, the current unavailability of support doesnot mitigate the force of proponents who claim thattrials ought to be done this way, but it does underminethe claim that we can just ‘‘slot in’’ adaptive allocationinto the usual course of research. Absent this support,and in light of the Food and Drug Administration’s(FDA) ambivalence toward the evidence produced inadaptive trials, outcome-adaptive trials seem morelikely to make research less efficient on the whole.A second argument in defense of two-armed adap-tive randomization might be made for trials employingplacebo comparators in spite of the existence of stan-dard care options. In such cases, adaptive randomiza-tion will indeed decrease the number of patientsreceiving placebo, should the novel agent show activity.However, we regard such departures from clinical equi-poise as unethical—at least in the context of serious ill-nesses. The appropriate way to resolve the ethics hereHey and Kimmelman103