Equipoise

Equipoise in Clinical Trials and Health Technology Assessment

Equipoise in its usual clinical form asks whether the expert community, rather than each individual clinician, is divided about the treatments a trial compares, and it is the standard ethical justification for randomisation. For health economists the idea matters well beyond research ethics, because the randomised controlled trial is the main source of relative effectiveness evidence in health technology assessment, and whether a trial is still ethical, and still able to recruit, shapes coverage decisions made under uncertainty. Value of information analysis asks a related question in decision-theoretic terms, namely whether resolving the uncertainty is worth what the research costs.

Why randomisation needs an ethical justification

London sets out the problem in a JAMA methods guide. Randomisation severs the link between what a participant receives and the recommendation of a treating clinician who has a duty to provide the best possible care for that person. Ashcroft describes the same tension from the critics' side, where entering a patient into a trial seems to involve knowingly failing to offer the treatment the doctor believes best.

The answer given in the trials literature is that randomised assignment can be justified when there is genuine uncertainty about which of the compared treatments is superior. Ashcroft notes that the many definitions of equipoise are attempts to say how much uncertainty is needed and whose uncertainty counts. In London's statement of the principle, when there is uncertainty or conflicting expert opinion about the relative merits of a set of interventions, a participant may be allocated to any of them so long as there is no consensus that an alternative would better advance that participant's interests.

Individual equipoise and clinical equipoise

The two main versions of the principle differ in whose uncertainty is required. Freedman's 1987 paper in the New England Journal of Medicine is the usual reference point. Freedman wrote that the then current understanding, which required the investigator to have no treatment preference throughout the trial, presented nearly insuperable obstacles to starting or completing a controlled trial. His alternative is satisfied if there is genuine uncertainty within the expert medical community, not necessarily on the part of the individual investigator.

Lilford gave the individual version a decision-analytic form. Decision analysis combines Bayesian probabilities of treatment effects with the values a patient places on the outcomes to calculate which treatment maximises expected utility, and a person for whom both treatments have equal expected utility can accept randomisation without loss. Because different people value outcomes differently, the same evidence can leave one patient equipoised and another not. London argues that this strict version would often make equipoise disappear even while experts still disagree, and that requiring each clinician to be uncertain gives too much ethical weight to personal judgement and hinders useful research.

The distinction is visible in trial recruitment. In a qualitative synthesis of trials comparing surgical and non-surgical care, Davies and colleagues found that clinicians could justify taking part through community equipoise, the collective uncertainty about which treatment is best, while their views about the best management of a particular patient, which they call individual equipoise, sometimes made it hard to recruit some individuals.

How equipoise is applied to trial design and choice of comparator

Research ethics guidance applies the principle mainly through the choice of control arm. The CIOMS international ethical guidelines (2016, Guideline 5) state that when there is credible uncertainty about whether an established effective intervention is superior to an investigational agent, which they call clinical equipoise, it is permissible to compare the two directly.

London links equipoise to the information a study produces. To disturb equipoise, a study must be designed to generate information that resolves the uncertainty or narrows the disagreement among qualified experts, and any risks to participants must be reasonable in light of the value of the information the study is likely to produce. London places that judgement with the institutional review board, before anyone is enrolled.

Why equipoise matters in health economics and HTA

Coverage decisions are taken on uncertain evidence, and the way a payer acts on that uncertainty can change whether the missing evidence is ever produced. Walker and colleagues set out a framework for choosing between coverage, no coverage and options that tie coverage to research, which they call only in research and only with research. They note that early coverage may affect whether the needed research happens at all, because manufacturers have less incentive to invest once a technology is covered, physicians might consider further trials unnecessary and unethical, and patients might be unwilling to take part if they already have access. Clinicians who regard a trial as unethical have in practice lost equipoise, and a randomised trial may then fail to recruit. They also describe Medicare's coverage with evidence development in the United States, a form of managed entry agreement that can restrict coverage to patients treated within a trial or registry.

The NICE manual (PMG36, section 6.4.14) builds the same concern into recommendations for more research. Before making one, the committee considers the impact of its recommendation on the feasibility of doing the research and the ethical or practical aspects of doing it, and balances the net benefits to current patients of an unrestricted recommendation against the net benefits to current and future patients of a more secure evidence base.

Claxton and colleagues list the key principles for these coverage decisions as whether the technology is expected to be cost-effective, whether it has significant irrecoverable costs, whether additional research is needed, whether research is possible with approval, whether approval would commit opportunity costs that cannot later be recovered, and whether effective price reductions are available. They conclude that cost-effectiveness is a necessary but not sufficient condition for approval.

How value of information differs from equipoise

Claxton argued in 1999 that adoption decisions should rest on expected net benefit, whether or not differences are statistically significant, and that the distribution of net benefit matters only for deciding whether more information is needed. The value of resolving all uncertainty is the expected value of perfect information.

$$\text{EVPI} = E_{\theta}\left[\max_{j} NB(j,\theta)\right] - \max_{j} E_{\theta}\left[NB(j,\theta)\right]$$

where $j$ indexes the treatment options, $\theta$ is the set of uncertain parameters, $NB(j,\theta)$ is the net benefit of option $j$ when the parameters take the value $\theta$, and $E_{\theta}$ is the expectation over current uncertainty. Walker and colleagues describe EVPI as an upper bound on the value of additional evidence, and treat research as potentially worthwhile only if the value of the uncertainty it could reduce exceeds its costs. The value of a specific trial design is measured by the expected value of sample information.

The two tests can disagree. Equipoise is a judgement about whether informed experts are divided, so that no participant is knowingly disadvantaged, while value of information weighs the expected gain to the whole patient population against the cost of research. A disagreement among experts about a rare condition may support an ethical trial that is not worth funding, and a decision with a large population may carry high value of information even when the expected net benefit clearly favours one option.

Worked example of equipoise, value of information and an only-in-research decision

All numbers in this example are illustrative. A new treatment B for a chronic condition costs the same as standard care A over the time horizon, so the difference in net health benefit equals the difference in QALYs. Current evidence allows two possibilities, a gain of 0.30 QALYs per patient with probability 0.4 or a loss of 0.15 QALYs through toxicity with probability 0.6, and expert clinicians are divided between them.

1. Expected gain on current evidence.

$$E[\Delta Q] = 0.4 \times 0.30 + 0.6 \times (-0.15) = 0.12 - 0.09 = 0.03$$

where $E[\Delta Q]$ is the expected QALY gain per patient from B compared with A. B is preferred on expected value although it is more likely to be worse than better.

2. Equipoise. Experts disagree about which scenario is true, so the condition for clinical equipoise is met. A decision maker who holds these probabilities is not strictly equipoised in Lilford's sense, because the expected gain is 0.03 QALYs and not zero.

3. Value of resolving the uncertainty. With perfect information B would be chosen only if it is beneficial, so the expected gain per patient is 0.4 × 0.30 + 0.6 × 0 = 0.12 QALYs. EVPI per patient is 0.12 − 0.03 = 0.09 QALYs, and for an illustrative 10,000 patients over the life of the decision, undiscounted, population EVPI is 900 QALYs, or 720 QALYs for the 8,000 patients treated after a trial could report. A trial whose cost displaces an illustrative 100 QALYs elsewhere in the health system passes the necessary condition that value exceeds cost.

4. Compare the coverage options. Suppose the trial takes long enough that 2,000 of the 10,000 patients are treated before it reports, and assume, as an upper bound, that it resolves the uncertainty fully. Trial participants are counted within the 2,000 and their own expected gain is ignored for simplicity. Totals are QALYs gained against treating everyone with A.

5. Interpretation. Approval with research would be best, at 920 QALYs, but only if the trial can still recruit after approval. If approval would make clinicians regard the trial as unnecessary and patients unwilling to be randomised, that option disappears and the choice is between approval without research, 300 QALYs, and only in research, 860 QALYs. Withholding B from current patients costs 920 − 860 = 60 QALYs, the price of keeping the evidence obtainable, against a gain of 960 − 240 − 100 = 620 QALYs, net of research cost, from better decisions for later patients, so only in research beats approval without research by 560 QALYs. Equipoise settles whether randomising patients in the research arm is ethically permissible, and value of information settles whether running the trial is worth it.

Limits of equipoise and common misunderstandings

Equipoise does not require the arms to be judged equally likely to be best. Freedman's version requires only genuine disagreement in the expert community, and in the example above clinical equipoise holds even though the probabilities are 0.4 and 0.6. A related error is to treat equipoise as sufficient. Ashcroft accepts that equipoise is neither necessary nor sufficient for a trial to be justified, while keeping it as the central practical test for randomisation, noting that phase I and II trials are rarely based on it and that a trial with equipoise can still be unethical.

Equipoise concerns the merits of treatments for participants. London states it in terms of therapeutic, prophylactic or diagnostic merits, so on that wording uncertainty about costs or cost-effectiveness alone does not create equipoise and belongs to value of information analysis and coverage policy. Ashcroft also notes that when sponsors measure outcomes that matter less to patients and clinicians, industry trials can fail to reflect the reasonable equipoise of working clinicians.

Finally, equipoise changes over time. Accumulating evidence, early adoption in practice and positive coverage decisions can all erode it. Ashcroft notes that designs which alter allocation in the light of new information, as some adaptive trial designs do, are hard to analyse and depend on contested priors.

Sources

• Ashcroft RE. Ethics of randomised controlled trials: not yet time to give up on equipoise. Arthritis Research and Therapy. 2004;6(6):237. https://doi.org/10.1186/ar1442
• Claxton K. The irrelevance of inference: a decision-making approach to the stochastic evaluation of health care technologies. Journal of Health Economics. 1999;18(3):341-364. https://doi.org/10.1016/S0167-6296(98)00039-3
• Claxton K, Palmer S, Longworth L, Bojke L, Griffin S, Soares M, Spackman E, Rothery C. A comprehensive algorithm for approval of health technologies with, without, or only in research: the key principles for informing coverage decisions. Value in Health. 2016;19(6):885-891. https://doi.org/10.1016/j.jval.2016.03.2003
• Council for International Organizations of Medical Sciences (CIOMS). International Ethical Guidelines for Health-related Research Involving Humans. Geneva: CIOMS; 2016. Guideline 5. https://cioms.ch/wp-content/uploads/2017/01/WEB-CIOMS-EthicalGuidelines.pdf
• Davies L, Beard D, Cook JA, Price A, Osbeck I, Toye F. The challenge of equipoise in trials with a surgical and non-surgical comparison: a qualitative synthesis using meta-ethnography. Trials. 2021;22:678. https://doi.org/10.1186/s13063-021-05403-5
• Freedman B. Equipoise and the ethics of clinical research. New England Journal of Medicine. 1987;317(3):141-145. https://doi.org/10.1056/NEJM198707163170304
• Lilford RJ. Ethics of clinical trials from a bayesian and decision analytic perspective: whose equipoise is it anyway? BMJ. 2003;326(7396):980-981. https://doi.org/10.1136/bmj.326.7396.980
• London AJ. Equipoise in research: integrating ethics and science in human research. JAMA. 2017;317(5):525-526. https://doi.org/10.1001/jama.2017.0016
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). London: NICE; 2022, last updated 13 October 2026. Section 6.4.14.
• Walker S, Sculpher M, Claxton K, Palmer S. Coverage with evidence development, only in research, risk sharing, or patient access scheme? A framework for coverage decisions. Value in Health. 2012;15(3):570-579. https://doi.org/10.1016/j.jval.2011.12.013