Commentary on Hey and Kimmelman

In classical randomized controlled trials (RCTs), randomization probabilities remain fixed throughout the duration of the study. However, in any trial that ultimately shows one arm to be superior, data will at some point provide mounting though not yet definitive evidence in favor of the better performing arm. In trials with serious, irreversible outcomes, a data monitoring committee is typically charged with regular review of emerging data and with making recommendations about early termination if extreme effects are seen. Still, commentators struggle over the necessity of randomizing participants to the arm that appears inferior based on emerging data. Some investigators and ethicists have advocated response-adaptive over fixed randomization to maximize the proportion of participants assigned to the better performing arm. In this issue of Clinical Trials, Hey and Kimmelman advocate a ‘‘rebuttable presumption’’ against adaptive randomization, at least in the setting of two-arm trials comparing a novel intervention to a control. They offer a range of arguments to buttress their conclusion. First, they contend that advocates of adaptive randomization overestimate the likelihood and magnitude of benefits that it offers participants. Studies may adjust randomization probabilities based on surrogate endpoints that don’t correlate with clinical benefit; clinical endpoints may be too delayed to allow timely adjustment; the magnitude of incremental benefit associated with novel therapies is often small; and the complexities of implementing adaptive designs may preclude their use in most settings. Second, they note that because 1:1 randomization represents the most efficient allocation ratio, adaptive trials are likely to take longer than conventional trials, require more participants, or both. Third, they claim that adaptive trials encourage rather than dispel participants’ unrealistic expectations about the benefits of experimental therapies, thereby failing to advance, or even undermining, informed consent. Finally, they note that changes over time in the proportions of individuals assigned to each arm might interact with temporal trends in background conditions or participant characteristics to threaten the validity of the trial. In drawing their measured conclusion against the use of adaptive randomization, Hey and Kimmelman do not conclude that such trials are unethical. In this, they are surely correct. But adaptive trials do involve important tradeoffs among core scientific and ethical values that investigators must weigh as they consider alternative research designs. A widely recognized framework holds that ethical research fulfills seven criteria: social value, scientific validity, fair subject selection, favorable benefit/risk ratio, independent review, informed consent and respect for participants. Three of these criteria—scientific validity, favorable benefit/risk ratio and informed consent—are particularly relevant to assessing adaptive designs. Scientific validity is a basic condition for ethical research. Research aims to produce generalizable knowledge; studies that are unlikely to permit valid conclusions cannot justify exposing participants to risk. In addition, the prospect of valid conclusions is what legitimizes the use of resources in the research, including money, facilities, personnel and—most important— participants. The main challenge to validity noted by Hey and Kimmelman is the possibility that subjects who enter later in the trial may differ systematically— for example, by prognosis—from those who enter earlier. Adjustment for all known and unknown factors associated with prognosis would eliminate this problem, but such adjustment is generally impossible. Thus, imbalanced allocation could lead to systematic between-group differences in prognosis that would bias the analysis of the trial. In addition to this and other challenges to validity that Hey and Kimmelman identify, open-label adaptive trials face two further threats. First, most phase III randomized trials strictly maintain the confidentiality of