Concurrent Control

Concurrent Control in Trials and Health Technology Assessment

A concurrent control group is recruited from the same population as the patients receiving a new treatment and is followed alongside them in the same trial, usually after random allocation. Anything that changes during the trial, such as supportive care, the mix of patients referred or the way outcomes are assessed, then affects both groups alike, so the difference between them can be credited to the treatment. The idea matters in health economics because the relative effect estimated against the control group drives the cost-effectiveness model, and a control group out of step in time with the treated group shifts the QALY gain and the ICER without any sign in the probabilistic results. The question has become practical with platform trials, where one shared control arm can run for years while new treatments join and leave.

What makes a control group concurrent

The ICH E10 guideline on the choice of control group (2000) gives control groups one major purpose, to separate outcomes caused by the test treatment from those caused by other factors, such as the natural progression of the disease, the expectations of patients or observers, or other treatment. E10 accepts that a similar group of earlier patients might serve in the unusual case where the course of a disease is predictable, but states that in most situations a concurrent control group is needed because outcome cannot be predicted with adequate accuracy or certainty.

E10 defines a concurrent control group as one chosen from the same population as the test group, treated in a defined way as part of the same trial, and over the same period of time. The two groups should be similar in all baseline and on-treatment variables that could influence outcome apart from the study treatment, and failure to achieve this can introduce bias. E10 sorts control groups into five types, the first four concurrent and distinguished by what the control group receives.

The active control trial is the type closest to the needs of health technology assessment, because the active comparator can be the treatment the new technology would displace. E10 places external controls, historical ones included, in a separate fifth type because of doubts that such trials can ensure comparable groups and limit bias.

Concurrency and randomisation do different jobs

Concurrency deals with time, while random allocation deals with who ends up in each group. E10 states that random division of a single sample avoids systematic differences in known or unknown baseline variables, and calls the inability to eliminate such differences a major problem of studies without a concurrent randomised control. Blinding then keeps the groups treated and assessed alike during the trial.

A control group can therefore be concurrent without being randomised. The NICE manual for technology appraisal (PMG36) lists observational database studies with concurrent control among non-randomised studies, which tend to be at high risk of bias because the factors that influence treatment assignment may also predict outcome, which is confounding. Concurrency removes calendar time as an explanation for a difference but does nothing about patients being chosen for treatment because of their prognosis. The reverse case exists too, since E10 counts a group treated during the same period in another setting as an external control.

Shared control arms and non-concurrent controls in platform trials

A platform trial compares several experimental treatments with a shared control arm and lets new treatments enter as it runs. For a treatment that joins later, the control arm splits in two. Concurrent controls were randomised to control while that treatment was open to randomisation, and non-concurrent controls were randomised before it entered and could never have received it. Lee and Wason (2020) note that the split is specific to each arm, so the same patients can be concurrent for one treatment and non-concurrent for another added later.

Pooling all control patients enlarges the control group but compares the new treatment partly with patients from an earlier period. Lee and Wason give a learning curve among study staff, a shift in baseline characteristics of the population, and changes in the effect of the control treatment, for example through improved standard of care, as sources of time trends.

$$\hat{\Delta}_{pool} = \bar{Y}_D - \left[ w \bar{Y}_{C1} + (1-w) \bar{Y}_{C2} \right] = \Delta + w \left( \bar{Y}_{C2} - \bar{Y}_{C1} \right)$$

where $\bar{Y}_D$ is the mean outcome on the new drug, $\bar{Y}_{C1}$ is the mean outcome among non-concurrent controls, $\bar{Y}_{C2}$ is the mean outcome among concurrent controls, $w$ is the share of non-concurrent patients in the pooled control group, and $\Delta = \bar{Y}_D - \bar{Y}_{C2}$ is the concurrent comparison. The extra term is the change in control outcomes over time scaled by the share of borrowed patients, and it vanishes only when there is no trend.

Lee and Wason found by simulation that with a positive trend and all control data used, power and the type I error rate could exceed their nominal values, although the root mean squared error of the estimate could improve. They concluded that non-concurrent data should not be used where strict error control is required. Adjusting for a stage effect was equivalent to using concurrent controls only, and a linear time term did not guarantee valid inference under a non-linear trend. Bofill Roig and colleagues found that a step-function model fitted to all arms gave more power while controlling the type I error rate, but only when time trends were equal across arms and additive on the model scale.

What regulators expect from the control comparison

The US Food and Drug Administration draft guidance on master protocols (revision 1, June 2026) recommends that the primary comparison for each drug generally include only concurrently eligible controls, meaning participants who could have been randomised to that drug but were randomised to control. Non-concurrent controls, and concurrent controls who could not have received the drug, should not be included. The FDA cites temporal shifts in participant characteristics, trial conduct or standard of care, and notes that a shared control arm still brings considerable efficiency gains when only concurrently eligible controls are used.

Other control data may be justified in rare circumstances, such as early-phase exploratory trials and trials in rare diseases with feasibility constraints. The FDA asks for that decision to be agreed before the trial starts, to rule out borrowing after a poorly performing control arm has been seen, and suggests sensitivity analyses that include an estimate based on concurrent controls only.

In Europe, the Complex clinical trials questions and answers document (2022) treats concurrency as one attribute to weigh when trial data are proposed as control data, warns that non-concurrency of control and test arms may affect trial interpretability, and describes randomised trials with an internal concurrent control arm as the standard where appropriate and feasible. The EMA concept paper on platform trials (2022) describes non-concurrent controls as a feature not needed by design that can break key principles of confirmatory trials, such as strict type I error control, and proposes a reflection paper to discuss whether such use can be reasonable.

Why concurrency matters for cost-effectiveness evidence

For relative treatment effects PMG36 states a strong preference for high-quality randomised controlled trials, and the randomised controlled trial supplies that effect through its concurrent control. PMG36 also lists the year of the study among the factors that can affect the estimate of relative effectiveness. A comparison against controls from a different period mixes a calendar effect into the treatment effect, and the model carries that mixture over a lifetime horizon.

PMG36 describes probabilistic analysis as characterising the uncertainty in the precision of mean parameter values, so a calendar effect mixed into the relative effect passes into the cost-effectiveness analysis without widening its uncertainty intervals. When platform trial evidence reaches an appraisal, the analysis needs to show which control set produced the relative effect, and the concurrent-only estimate is the natural base case.

A network meta-analysis links treatments through comparisons made within randomised trials, and PMG36 requires trial randomisation to be preserved, treating comparisons of single arms from different trials as observational. Where non-concurrent control data are used, the FDA draft lists a network meta-analysis combining comparisons between concurrently randomised arms among the analyses that can limit confounding from changes over time.

Worked example with a shared control arm

The following figures are illustrative only. A platform trial measures response at 12 weeks. Before drug D joins, 200 patients are randomised to the shared control arm and 30% respond. After drug D joins, 200 patients receive drug D and 50% respond, while 200 concurrent control patients reach 40%, because supportive care improved between the periods.

The concurrent comparison gives 50% minus 40%, or 10 percentage points. Pooling all 400 control patients gives a control response of (60 + 80) / 400 = 35% and an estimated effect of 15 points. The equation gives the same answer, since $w = 200 / 400 = 0.5$ and the control trend is 10 points, so the bias is 0.5 times 10, or 5 points.

In the economic model each extra responder gains 0.40 QALYs, and drug D adds GBP 1,500 per patient after downstream costs, held fixed for simplicity. The concurrent estimate gives 0.10 times 0.40, or 0.04 QALYs, and an ICER of GBP 1,500 / 0.04 = GBP 37,500 per QALY gained. The pooled estimate gives 0.06 QALYs and an ICER of GBP 1,500 / 0.06 = GBP 25,000 per QALY gained.

Against an illustrative threshold of GBP 30,000 per QALY, the pooled ICER falls below it and the concurrent ICER does not. The ICER reaches the threshold at a QALY gain of 1,500 / 30,000, or 0.05, which needs a response difference of 0.05 / 0.40 = 0.125, or 12.5 points. The decision turns on whether 5 points of bias from the earlier control patients are credited to the drug, which only a concurrent-only scenario shows.

Concurrent control and neighbouring designs

A control arm is any group a trial uses for comparison, and a concurrent control is the kind that shares the trial's population and period. A historical control is drawn from an earlier period and is the main form of external control arm, the fifth E10 type. Bofill Roig and colleagues (2023) separate concurrent external controls, collected at the same time in another setting, from non-concurrent internal controls, the earlier patients in the same platform trial, and judge that the latter avoid much of the selection and measurement risk of external data but remain exposed to time trends.

Common misunderstandings about concurrent controls

Concurrent does not mean randomised. An observational comparison over the same years removes calendar time as an explanation but leaves selection by prognosis.

Membership of the same platform trial does not make every control patient concurrent. The FDA draft counts only controls randomised while the drug was open to them, and the European Q&A weighs how far control and test data were collected over the same period.

Adjusting for time does not fully restore concurrency. Validity then rests on assumptions about the shape and equality of time trends, which is why the FDA draft asks for those assumptions to be described and the analysis tested under temporal shifts.

A concurrent control does not guarantee a relevant comparator for an appraisal. The control reflects care at the time of the trial, while PMG36 asks the scope to identify comparators that are established practice in the NHS.

Sources

• Bofill Roig M, Krotka P, Burman CF, et al. On model-based time trend adjustments in platform trials with non-concurrent controls. BMC Medical Research Methodology. 2022;22:228. https://doi.org/10.1186/s12874-022-01683-w
• Bofill Roig M, Burgwinkel C, Garczarek U, Koenig F, Posch M, Nguyen Q, Hees K. On the use of non-concurrent controls in platform trials: a scoping review. Trials. 2023;24:408. https://doi.org/10.1186/s13063-023-07398-7
• European Medicines Agency, Committee for Medicinal Products for Human Use. Concept paper on platform trials (EMA/CHMP/840036/2022). Amsterdam: EMA; 2022. https://www.ema.europa.eu/en/platform-trials-scientific-guideline
• European Medicines Agency, ACT EU. Complex clinical trials: questions and answers (EMA/298712/2022), version 2022-05-23. https://health.ec.europa.eu/latest-updates/questions-and-answers-complex-clinical-trials-2022-06-02_en
• International Conference on Harmonisation. ICH E10: Choice of control group and related issues in clinical trials. Step 4 version, 20 July 2000. https://database.ich.org/sites/default/files/E10_Guideline.pdf
• Lee KM, Wason J. Including non-concurrent control patients in the analysis of platform trials: is it worth it? BMC Medical Research Methodology. 2020;20:165. https://doi.org/10.1186/s12874-020-01043-6
• 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. https://www.nice.org.uk/process/pmg36
• US Food and Drug Administration. Master protocols for drug and biological product development. Draft guidance for industry, revision 1. Silver Spring, MD: FDA; June 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/master-protocols-drug-and-biological-product-development