No general NMA framework for non-parallel randomized trial designs
Statement
Network meta-analysis models assume every included trial contributes a randomized contrast between concurrently randomized arms followed from a common time zero. Re-randomization, randomized withdrawal, enrichment, response-adaptive allocation and dose titration break that assumption: their published contrasts are conditional on an interim response, on a run-in or on an adaptive allocation rule, and the arms the network needs were in part never randomized. No general framework states, for a given non-parallel design, which target contrast the trial identifies, which assumptions convert it to a parallel-design contrast, and how the conversion uncertainty enters the network likelihood; practice is a bespoke arithmetic reconstruction per trial, or more often silent use of the conditional subgroup as though it were the trial. The consequences are first order rather than second order. In the one worked example available, treating the re-randomized responder subgroup as the trial gave a placebo maintenance response of 31.2 percent against 15.5 percent after reconstruction and roughly halved the odds ratios against placebo, and a published network meta-analysis had already made that error. The canonical guidance does not close the gap: the Cochrane Handbook chapter on variants of randomized trials covers cluster, cross-over, stepped-wedge and factorial designs and says nothing about randomized-withdrawal, responder-enriched or re-randomization designs.
Why it is open
The reconstructions rest on assumptions no data can check: that a responder-to-non-responder efficacy ratio observed at one dose transports to another, that all protocol-mandated switchers would have dropped out, and, where an external prior fills a missing arm, that its arbitrary down-weighting is the right one; none of that uncertainty is separated from sampling uncertainty in the reported interval. The designs are also not one family with one fix, because an interim-response gate, an adaptive allocation rule and a titration schedule each destroy a different piece of the parallel-design structure, so a general framework must enumerate them rather than generalize over them, and the enumeration has not been written. The problem grows rather than shrinks: established class efficacy makes long placebo-controlled maintenance periods hard to justify, so more of the pivotal evidence entering networks comes from designs of this kind. The obvious conservative remedy is not obviously safe either, since work on trials stopped early for benefit concludes that excluding them from a synthesis underestimates effects and overestimates information while including them is close to unbiased, so exclusion of non-standard designs cannot be adopted as a default without its own analysis.
What has been tried
| Work | What it contributes |
|---|---|
| Druyts et al. 2014, Clinical Epidemiology | Names the challenge of incorporating alternative-design trials into network meta-analyses and indirect comparisons, and supplies the worked reconstruction of a re-randomized maintenance trial that shows how far the naive analysis is off. |
| Varu et al. 2019, Current Medical Research and Opinion | Treatment sequence network meta-analysis case study; a bespoke per-trial reconstruction presented as novel and explicitly unvalidated, with no simulation and no benchmark. |
| Higgins, Eldridge and Li, Cochrane Handbook for Systematic Reviews of Interventions, chapter 23 | The canonical guidance on including variants of randomized trials; covers cluster-randomized, cross-over, stepped-wedge and multi-arm or factorial designs, and does not address randomized-withdrawal, responder-enriched or re-randomization designs. |
| Elbourne et al. 2002, International Journal of Epidemiology | The template for successfully bringing one non-parallel design into meta-analysis, with the correlation structure and unit-of-analysis issues of cross-over trials handled explicitly rather than by conversion. |
| FDA 2019, Enrichment Strategies for Clinical Trials to Support Approval of Human Drugs and Biological Products | Regulatory guidance defining prognostic and predictive enrichment and randomized-withdrawal designs; establishes which population an enriched trial’s result belongs to, the fact a synthesis has to respect. |
| Barberio et al. 2025, Alimentary Pharmacology and Therapeutics | Runs design-stratified networks, separately for re-randomizing and treat-through maintenance trials, and reports different rankings in each; the conservative alternative to conversion, at the cost of two answers to one question. |
| Marschner and Schou 2022, Clinical Trials | Analyzes adaptive platform trials using a network approach, showing that some non-parallel randomization structures can be represented inside a network rather than converted into a parallel-design contrast. |
| Schou and Marschner 2013, Statistics in Medicine | Investigates bias in meta-analysis from trials stopped early and finds that excluding them underestimates effects and overestimates information while including them is close to unbiased; the counterweight to a default of dropping non-standard designs. |
| Marschner, Askie and Schou 2020, Research Synthesis Methods | Recommends sensitivity analyses for the impact of early stopping on a systematic review; a worked example of treating a design feature as a sensitivity dimension rather than an exclusion criterion. |
| Pallmann et al. 2018, BMC Medicine | Catalogues adaptive designs and their reporting requirements, supplying the enumeration of design features that a synthesis framework would have to work through. |
| Efthimiou et al. 2017, Statistics in Medicine | Combines randomized and non-randomized evidence in a network with explicit bias and down-weighting parameters; the closest existing template for admitting evidence of a different structure without pretending it has the same status. |
| Dias et al. 2013, Medical Decision Making | The generalized linear modeling framework on which network likelihoods are built, which takes concurrently randomized arms and a common time zero as given. |
Probable solution or research direction
Enumerate before generalizing: build a catalogue of non-parallel randomized designs and, for each, write the target contrast the trial identifies in potential-outcome terms, the assumptions mapping it to a parallel-design contrast, and whether those assumptions are checkable from what gets published. Represent the conversion as a model rather than as arithmetic, so its uncertainty is a parameter with a prior or a sensitivity range that enters the network likelihood alongside sampling error instead of disappearing into a point estimate. Where the conversion is not identified, make the design-stratified network the primary analysis and the converted network a sensitivity analysis, following the applied practice that already runs separate re-randomized and treat-through syntheses. Extend the network representation used for adaptive platform trials, which handles non-concurrent randomization without converting it away, to see how far interim-response gates and titration schedules can be modeled as extra nodes and time strata rather than reconstructed. Simulate networks mixing design families against known truth and report bias, coverage and rank reversals for naive pooling, stratification and conversion, since no benchmark exists for any of the three.
Verification
Verdict. Confirmed open. One reader proposed this from a single 2014 paper that names the challenge and supplies the worked reconstruction, and the reviewer accepted it as an identification problem not reducible to endpoint, population or treatment-switching heterogeneity. An independent search supports the absence: the conversion technique appears only in the methods sections of applied network meta-analyses without derivation or validation, and the Cochrane Handbook chapter on design variants does not cover these designs at all.
Source
Derived from reading:nma-42 of the reviewed corpus.