Platform trials lack stage-level reporting for network synthesis
Statement
A platform trial is a sequence of fixed designs, and only arms open for randomization in the same stage are concurrently randomized; a contrast assembled across stages is confounded with calendar time and with every design feature that changed between stages. Trials of this kind normally enter a systematic review or network meta-analysis as one trial-level effect per comparison, so a reviewer cannot tell which stages contributed or what share of the reported effect came from non-concurrent controls. Part of the necessary reporting already exists: item 14a of the Adaptive designs CONSORT Extension asks for the exact recruitment and follow-up dates of each treatment group and for a statement of whether all groups were recruited concurrently, with STAMPEDE as its worked exemplar. What no standard requires is the numbers, namely stage-specific treatment effects and their standard errors for each randomized comparison within each fixed design, which is what a reviewer needs to rebuild the within-trial network, reproduce the direct-evidence proportion, and test the trial’s own direct evidence against its own indirect evidence. No review-side standard requires concurrency to be recorded either: PRISMA-NMA adds items on network geometry but none on whether a contributing trial’s contrast was concurrently randomized. The stakes are measurable: in a network analysis of one STAMPEDE comparison the concurrent and non-concurrent evidence disagreed, with the test of inconsistency at p = 0.017, so a review that pooled them would have combined incompatible information without being able to notice.
Why it is open
The stage decomposition exists only inside the trial team’s own data, and the decision that destroys it is made at write-up, before any reviewer is involved and before any synthesis model is specified, so nothing fitted later can price what the report did not carry. Reporting guidance for adaptive trials is organized around making the adaptation transparent rather than around making the trial decomposable into concurrently randomized comparisons, which is why the recruitment dates are required and the stage-specific estimates are not. The analysis side moved first: network and time-adjustment methods for non-concurrent controls are published and implemented, so the obstacle is missing inputs rather than a missing model. Reconstructing individual data from published Kaplan-Meier plots recovers part of the information for time-to-event endpoints only, and it injects reconstruction error that then travels through the network unlabeled. Nobody has counted how many platform and multi-arm multi-stage trials now sit inside published networks, or how much of their contributed evidence is non-concurrent, so the size of the exposure is unmeasured and the case for a new item rests on a single demonstration.
What has been tried
| Work | What it contributes |
|---|---|
| Marschner and Schou 2022, Clinical Trials | Embeds stratified and time-adjusted analyses of a platform trial in a common network framework, shows that unadjusted analysis is biased under time trends in risk, and applies it to a STAMPEDE comparison where direct randomized and indirect non-concurrent evidence conflict at p = 0.017; recommends that platform trials report stage-specific treatment effects for use in later reviews and re-analyses. |
| Dimairo et al. 2020, BMJ (Adaptive designs CONSORT Extension) | Reporting checklist for adaptive trials whose item 14a requires exact recruitment and follow-up dates for each treatment group and an explicit statement of whether all groups were recruited concurrently, illustrated with STAMPEDE; item 17c requires interim results used for adaptation decisions. This supplies the concurrency timeline but not stage-specific effect estimates per fixed design. |
| Lee and Wason 2020, BMC Medical Research Methodology | Quantifies the precision gained and the bias risked by including non-concurrent control patients in platform-trial analysis; establishes that concurrency is a quantitative property of a comparison rather than a formality. |
| Bofill Roig et al. 2022, BMC Medical Research Methodology | Model-based time-trend adjustments for platform trials with non-concurrent controls; shows the analysis machinery for using non-concurrent evidence exists, which is why the missing input rather than the missing method is the binding constraint. |
| Sydes et al. 2012, Trials | Describes the STAMPEDE multi-arm multi-stage design with arms stopped for lack of benefit and research arms added mid-trial; the concrete stage structure that a downstream network needs and that a single aggregate effect per comparison hides. |
| Hutton et al. 2015, Annals of Internal Medicine (PRISMA-NMA) | The review-side reporting standard for network meta-analysis, adding items on network geometry and on the summary of the network; carries no item asking whether a contributing trial’s contrast was concurrently randomized. |
| Thorlund, Druyts, Toor, Jansen and Mills 2014, Clinical Epidemiology | Converts a re-randomization trial into parallel-design form so it can populate a network, imputing the arm-level quantities the publication did not report; a worked instance of the general pattern in which non-standard designs enter networks only after reviewer-side reconstruction. |
Probable solution or research direction
Write the missing item and attach it to the extension that already carries the concurrency dates rather than founding a new standard: a table of stage-specific treatment effects with standard errors for every randomized comparison within each fixed design, plus the stage membership and randomization probabilities of each arm, which is enough to rebuild the within-trial network and has already been illustrated in the platform-trial network literature. Add the mirror item on the review side, so that a network meta-analysis must state per contributing trial which contrasts were concurrently randomized and must report the network with and without the non-concurrent evidence. Quantify the exposure first, because it is the cheapest step and it decides how hard to push: audit published networks in oncology and critical care for platform and multi-arm multi-stage trials, and record how often the stage structure is recoverable from the trial report alone. Where it is not recoverable, compare the network built from digitized survival curves against the one built from the published aggregate effect, so that the error the missing item creates is measured rather than assumed.
Verification
Verdict. Partially addressed. One reader proposed this and the reviewer accepted it as a within-platform reporting problem distinct from covariate reporting, estimand mismatch and external-control bridging, on the ground that it concerns the stage structure needed to classify direct evidence at all. It is partially addressed because the adaptive-design reporting extension already mandates per-group recruitment and follow-up dates and an explicit concurrency statement; what no standard requires, and what the single methods paper behind this entry recommends, is the stage-specific effect estimates, and no review-side standard requires concurrency to be recorded.
Source
Derived from reading:nma-15 of the reviewed corpus.