Adaptive Trial Design

Adaptive Trial Design for Planned Changes, Error Control and the Cost of Evidence

An adaptive trial lets a protocol change course at planned interim analyses, for example by stopping early, resizing, dropping arms or narrowing the population, under rules written before any comparative data are seen. For health economists the design matters twice, because the adaptations change the estimates that feed cost-effectiveness models, and they change how much the evidence costs to produce. Regulators, including the drafters of ICH E20, define which changes count as planned. Bayesian rules and borrowing are covered under Bayesian adaptive design.

What makes a design adaptive

The defining feature is prospective planning. The FDA guidance (2019) treats an adaptation as prospective only if it is specified before any comparative analysis of accumulating data. The ICH E20 draft defines an adaptive design in almost the same words but is stricter on timing, requiring that the possible adaptations must be prespecified in the protocol before the trial starts. It excludes unplanned modifications, such as an amendment proposed by a data monitoring committee after unexpected interim results.

The FDA guidance separates adaptations on non-comparative data, which ignore treatment assignment, from those on comparative data. The first kind have no or limited effect on the type I error probability; the second often increase it and bias estimates unless the analysis allows for the design. Its principles for confirmatory use are adequate control of erroneous conclusions, reliable estimation, complete prespecification and maintained trial integrity.

At the time of writing, in October 2026, ICH E20 was not yet final. It reached Step 2 in June 2025, the FDA issued it as draft guidance in September 2025, the EMA consultation closed on 30 November 2025, and the ICH work plan of 12 February 2026 schedules Step 4 sign-off for October 2026.

The main families of adaptation

Adaptive designs form a family, and one trial can combine several adaptations. Each leaves a different mark on the evidence an economic model later uses.

The FDA treats platform trials as a special case of arm selection, often with a common control arm.

Controlling type I error across interim looks

Repeated testing is the most familiar source of inflated error. The FDA guidance explains that testing at the one-sided 0.025 level at both an interim and a final analysis pushes the overall false-positive chance above 2.5 per cent. For a design with one interim look, the boundaries must satisfy the following condition.

$$\Pr_{H_0}(Z_1 \geq c_1 \ \text{or} \ Z_2 \geq c_2) = \alpha$$

where $Z_1$ and $Z_2$ are the standardised test statistics at the interim and final analyses, $c_1$ and $c_2$ are their critical values, $H_0$ is the null hypothesis of no treatment effect, and $\alpha$ is the one-sided type I error rate. Because the final analysis contains the interim data, the two statistics are correlated; with the interim at half the information their correlation is $\sqrt{0.5} = 0.707$.

Solving this condition numerically gives the values used in the worked example. Testing at 1.96 at both looks gives an overall error of about 0.042. A Pocock design uses a constant boundary of about 2.178, while an O'Brien-Fleming design uses about 2.797 at the interim and 1.977 at the end. O'Brien-Fleming boundaries need very persuasive early results, Pocock boundaries stop early more often, and alpha spending allows flexible timing of looks. Futility rules are commonly nonbinding; a binding rule can relax the efficacy threshold but controls the error rate only if it is always followed.

Other adaptations need other tools. Designs that resize on the interim effect combine stage-wise statistics, for example by the inverse normal method or conditional error functions. For complex designs, operating characteristics such as error rates, power, expected sample size and estimation bias are often established by Monte Carlo simulation.

What adaptation does to the economic evidence

The FDA guidance states that conventional end-of-trial estimates that ignore the adaptations tend to overestimate the true treatment effect, and that this also applies to secondary endpoints correlated with the primary endpoint. Trial costs and QALYs are usually secondary endpoints and, where correlated with the primary endpoint, a cost-effectiveness ratio built from naive estimates can inherit that bias. The ACE reporting guideline notes that secondary analyses such as health economic evaluation may be affected if appropriate adjustments are not made.

Three further consequences follow. A trial stopped early for efficacy has a smaller sample for safety and secondary endpoints, which for a model means shorter follow-up and fewer events for survival extrapolation. Enrichment can confine the evidence to a biomarker-defined subgroup, and Pallmann and colleagues warn that results may then apply only to patients with those characteristics. When allocation drifts towards the better arm, the comparator's costs and outcomes rest on a smaller group.

Practice has lagged behind the methods. Reviewing 37 adaptive trials with a health economic analysis, registered or published up to August 2016, Flight and colleagues could not tell in 17 whether the economic analysis allowed for the design where this was thought necessary.

Designing for value with EVSI and the cost of sampling

Value of information methods let the choice of design rest on expected payoff to the health system rather than power alone. Flight and colleagues (2022) extended the expected value of sample information to group sequential designs and compared a fixed design with Pocock and O'Brien-Fleming designs with two or five analyses; in their hypothetical case study the O'Brien-Fleming design with two analyses gave the highest expected net benefit of sampling. Their cost of sampling, simplified here to the terms that do not depend on arm or delay, is as follows.

$$TC = C_f + N_a C_a + n C_v$$

where $TC$ is the total cost of sampling, $C_f$ is the fixed cost, $N_a$ is the expected number of analyses, $C_a$ is the cost of each analysis, $n$ is the expected sample size and $C_v$ is the variable cost per patient recruited. The full formula adds arm-specific costs and an opportunity cost of delay.

Adaptive trials also need more support. In the Costing Adaptive Trials exercise, seven UK clinical trials units costed paired adaptive and fixed versions of five scenarios based on funded trials. Resources were always moderately higher for the adaptive version, with median increases of 2 to 4 per cent except for sample size re-estimation at 26.5 per cent, because that design could lengthen the study. Designs in which net benefit itself drives stopping are covered under Bayesian adaptive design.

Worked example costing three trial designs

The figures are illustrative. A publicly funded two-arm trial is planned with one-sided $\alpha = 0.025$ and 90 per cent power, needing 600 patients with a fixed design. The fixed cost is £400,000, each analysis costs £20,000 and each patient recruited costs £3,000. The two group sequential options have one interim look at half the maximum sample size and stop only for efficacy, using the boundaries computed above. Outcomes are assumed to be observed soon after recruitment, so no further patients are recruited while the interim analysis is done.

1. Fixed design. With one analysis and 600 patients the cost is as follows.

$$TC = 400000 + 20000 + 600 \times 3000 = 2220000$$

where the terms are the fixed cost, one analysis and 600 patients at £3,000, in pounds.

2. Maximum sample sizes. Keeping 90 per cent power needs inflation factors of about 1.007 for O'Brien-Fleming and 1.100 for Pocock, giving maxima of about 604 and 660 patients.

3. Expected sample size if the true effect equals the effect the trial is powered to detect. The trial stops at the interim with probability 0.310 under O'Brien-Fleming and 0.589 under Pocock, saving half the maximum sample each time. For O'Brien-Fleming the expected sample size is as follows.

$$n = 604.3 \times (1 - 0.5 \times 0.3099) = 510.7$$

where $n$ is the expected sample size, 604.3 is the unrounded maximum and 0.3099 is the probability of stopping at the interim. This is about 15 per cent below 600, the reduction the FDA guidance quotes for this design. For Pocock it is 465.6.

4. Expected cost of sampling. The expected number of analyses is two minus the probability of stopping at the interim. At that effect, the O'Brien-Fleming cost is as follows.

$$TC = 400000 + 1.690 \times 20000 + 510.7 \times 3000 = 1965900$$

where 1.690 is the expected number of analyses and 510.7 is the expected sample size; unrounded, £1,965,773. Repeating the steps gives the comparison.

5. Interpretation. With no true effect the trial rarely stops early, because these designs have no futility rule, so both adaptive designs cost more than the fixed design, Pocock most because of its larger maximum; adding a futility rule would cut the expected cost when the treatment does not work. At an illustrative 50:50 chance of the powered effect both are cheaper in expectation, O'Brien-Fleming slightly so. A full comparison would also weigh each design's expected value of sample information, since early stopping shortens follow-up for the model.

Reporting, appraisal and HTA acceptance

The Adaptive designs CONSORT Extension (ACE) adds seven new items to the CONSORT 2010 checklist, among them pre-planned adaptive features, estimation and inference methods and adaptation decisions, and modifies nine others. For an appraisal these items show which adaptations occurred and whether estimates were adjusted.

Adaptive designs get no specific treatment in the NICE manual (PMG36, last updated 13 October 2026), so a health technology assessment applies its general evidence and uncertainty requirements. The DSU checklist TSD 7 asks for early-stopped trials to enter a synthesis unadjusted when the stopping rules were prespecified. TSD 7's advice concerns pooling trials in a synthesis. Where an economic evaluation rests on a single adaptive trial, the FDA guidance and ACE point towards reporting estimates that allow for the design alongside the conventional ones.

Limits and common misreadings

Several errors recur. An adaptive trial need not be Bayesian, and an amendment made after seeing comparative results is an unplanned change, outside the E20 scope. Efficiency is not guaranteed, and the FDA guidance notes that an adaptive design can have a larger maximum sample size, needs longer planning and gains little when outcomes take longer to observe than recruitment takes. Comparative interim results need restricted access, since knowing that the sample size was increased can reveal the interim effect. A saving in patients is also not a gain in value if early stopping leaves the economic model short of long-term evidence.

Sources

• Ades AE, Caldwell DM, Reken S, Welton NJ, Sutton AJ, Dias S. NICE DSU Technical Support Document 7: Evidence synthesis of treatment efficacy in decision making: a reviewer's checklist. Sheffield: Decision Support Unit, ScHARR; 2012. Section A3.2.
• Dimairo M, Pallmann P, Wason J, Todd S, Jaki T, Julious SA, et al. The Adaptive designs CONSORT Extension (ACE) statement: a checklist with explanation and elaboration guideline for reporting randomised trials that use an adaptive design. BMJ. 2020;369:m115. https://doi.org/10.1136/bmj.m115
• European Medicines Agency. ICH E20 guideline on adaptive designs for clinical trials: Step 2b (EMA/CHMP/ICH/206586/2025). Amsterdam: EMA; 2025.
• Flight L, Arshad F, Barnsley R, Patel K, Julious S, Brennan A, Todd S. A review of clinical trials with an adaptive design and health economic analysis. Value in Health. 2019;22(4):391-398. https://doi.org/10.1016/j.jval.2018.11.008
• Flight L, Julious S, Brennan A, Todd S. Expected value of sample information to guide the design of group sequential clinical trials. Medical Decision Making. 2022;42(4):461-473. https://doi.org/10.1177/0272989X211045036
• International Council for Harmonisation. ICH E20 EWG Work Plan. Geneva: ICH; 12 February 2026.
• 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.
• Pallmann P, Bedding AW, Choodari-Oskooei B, Dimairo M, Flight L, Hampson LV, et al. Adaptive designs in clinical trials: why use them, and how to run and report them. BMC Medicine. 2018;16:29. https://doi.org/10.1186/s12916-018-1017-7
• US Food and Drug Administration. Adaptive Designs for Clinical Trials of Drugs and Biologics: Guidance for Industry. Silver Spring, MD: FDA; November 2019.
• US Food and Drug Administration. E20 Adaptive Designs for Clinical Trials: Draft Guidance for Industry. Silver Spring, MD: FDA; September 2025.
• Wilson N, Biggs K, Bowden S, Brown J, Dimairo M, Flight L, et al. Costs and staffing resource requirements for adaptive clinical trials: quantitative and qualitative results from the Costing Adaptive Trials project. BMC Medicine. 2021;19:251. https://doi.org/10.1186/s12916-021-02124-z