Learning health systems, clinical equipoise and the ethics of response adaptive randomisation

Abstract

To give substance to the rhetoric of 'learning health systems', a variety of novel trial designs are being explored to more seamlessly integrate research with medical practice, reduce study duration and reduce the number of participants allocated to ineffective interventions. Many of these designs rely on response adaptive randomisation (RAR). However, critics charge that RAR is unethical on the grounds that it violates the principle of equipoise. In this paper, I reconstruct critiques of RAR as holding that it is inconsistent with five important ethical principles. I then argue that these criticisms rest on a faulty view of equipoise encouraged by the idea that a RAR study models the beliefs of a single rational agent about the relative merits of the interventions being studied. I outline a view in which RAR models an idealised health system in which diverse communities of fully informed experts shrink or grow as their constituent members update their expert opinions in light of reliable medical evidence. I show how a proper understanding of clinical equipoise can reconcile this conception of RAR with these five ethical principles. This analysis removes an in-principle objection to RAR and sheds important light on the relationship between clinical equipoise and transient diversity in the scientific community.

The ProBio Trial: Molecular Biomarkers for Advancing Personalized Treatment Decision in Patients with Metastatic Castration-Resistant Prostate Cancer

“…Multiple simulation studies have shown, however, that multi-arm designs employing adaptive outcome-randomisation strategies that protect control allocation over time, such as the one used in ProBio, proved to be superior to designs using fixed randomisation probabilities, including multi-arm multi-stage designs (54)(55)(56). Lastly, while a discussion of the ethical objections to outcome-adaptive randomisation is outside the scope of this paper, we note that strong counterarguments to those objections have been put forward (57,58).…”

Comparison of response adaptive randomization features in multiarm clinical trials with control

“…RAR is currently controversial, with good performance observed in some contexts, such as comparisons to MAMS and to arm dropping and has theoretical optimality properties in certain settings. Criticisms of RAR, particularly in the two‐arm setting, have also emerged on the basis of potential power loss, statistical biases, the presence of “tail risks” where a subset of RAR trials performs poorly, as well as ethical controversy …”

Outcome-adaptive randomization in clinical trials: issues of participant welfare and autonomy

“…Thus, Scott Saxman notes: Outcome-adaptive randomized trials start out in equipoise, but equipoise is disturbed as soon as data are available from the first group of patients enrolled into the study and the randomization is adapted to favor the ‘better’ treatment arm. [ 9 , p. 63] The second reading of equipoise, which I label E2, is offered by Alex London [ 11 ] and Laura Bothwell and Aaron Kesselheim [ 12 ], who advance a different relationship between emerging data and the adjustment of randomization weights from that proposed within E1. London argues that OAR is compatible with clinical equipoise because the latter does not require that randomization probabilities should be equal: If it is consistent with concern for welfare for a patient to be directly treated with A or B or C (to receive that intervention with certainty), then it cannot violate concern for welfare if that patient is assigned to those interventions with any distribution of probabilities that sums to 1.…”

“…London argues that OAR is compatible with clinical equipoise because the latter does not require that randomization probabilities should be equal: If it is consistent with concern for welfare for a patient to be directly treated with A or B or C (to receive that intervention with certainty), then it cannot violate concern for welfare if that patient is assigned to those interventions with any distribution of probabilities that sums to 1. [ 11 , p. 412] This is persuasive. If k is the number of treatments under test, it does not matter ethically that some participants are randomized to treatment A with a probability less than 1/ k , since treatment A is regarded as optimum by a portion of the expert community (even if the other treatments under test are preferred by a larger portion of the community).…”

“…If it is consistent with concern for welfare for a patient to be directly treated with A or B or C (to receive that intervention with certainty), then it cannot violate concern for welfare if that patient is assigned to those interventions with any distribution of probabilities that sums to 1. [ 11 , p. 412]…”