No common estimand or aggregate-data reconstruction for enriched and withdrawal trials
Statement
Maintenance and randomized-withdrawal trials randomize only patients who responded during an induction run-in, so each reports an effect conditional on its own responder population, defined by its own response criterion, its own assessment time and its own induction blinding. Pooling those arms in a network treats quantities conditioned on different populations as a single estimand, and the resulting comparison is driven by the enrichment rule as much as by the treatment: in one Crohn’s disease network the observed maintenance placebo response rates ran from 9 to 22 percent and remission rates from 10 to 44 percent, and higher placebo rates mechanically produced lower odds ratios. The dependence is systematic rather than incidental. Across 37 inflammatory bowel disease maintenance trials covering 12,075 patients, re-randomized designs gave remission in 41.9 percent on active drug against 20.3 percent on placebo, with a number needed to treat of 5, while treat-through designs gave 30.9 percent against 14.6 percent, with a number needed to treat of 7; the therapeutic gain itself is a function of the randomization structure. No standard method converts these arms to a common estimand from aggregate data. What is in use is an undocumented rescaling of re-randomized maintenance results to a treat-through equivalent through an induction response probability, appearing as one or two sentences in the methods sections of applied network meta-analyses with no derivation, no stated assumptions and no validation; the fuller version needs individual patient data from one trial to supply induction-to-maintenance transition probabilities stratified by depth of response, which are then assumed to transport unchanged to every other trial’s induction population.
Why it is open
The conditioning is a property of the design rather than of the reported numbers, so nothing published can undo it: converting a responder-randomized contrast to an all-randomized one requires the joint distribution of induction response and subsequent maintenance outcome, and trials publish only the marginal maintenance result within the enriched subgroup. The transition probabilities that would supply that joint distribution are estimable only from individual patient data, which in practice means only one trial’s sponsor can run the conversion, and once run they are assumed to transport to every other trial’s induction population with no way to check. Enrichment rules differ across trials in criterion, threshold and timing, so there is not one conversion but a family of them, each carrying its own untestable transport assumption. The canonical guidance is silent as well: the Cochrane Handbook chapter on variants of randomized trials covers cluster, cross-over, stepped-wedge and factorial designs and does not address randomized-withdrawal or responder-enriched designs, so a reviewer has no authoritative recommendation and pooling as though the conditioning did not exist remains the default.
What has been tried
| Work | What it contributes |
|---|---|
| Varu et al. 2019, Current Medical Research and Opinion | Treatment sequence network meta-analysis in Crohn’s disease, presented as a methodological case study that is explicitly novel and unvalidated; its central input, maintenance placebo response conditional on remitter versus responder-non-remitter status at the end of induction, is estimable only from individual patient data. |
| Barberio et al. 2024, Digestive and Liver Disease | Meta-analysis of 37 inflammatory bowel disease maintenance trials showing that remission rates and therapeutic gain over placebo differ systematically between re-randomizing and treat-through designs; the empirical demonstration that the design and not just the drug drives the pooled contrast. |
| Barberio et al. 2025, Alimentary Pharmacology and Therapeutics | Runs separate maintenance networks for re-randomizing and for treat-through trials rather than pooling them, and finds the rankings differ; the conservative alternative to conversion, at the cost of two answers to one question. |
| Jairath et al. 2016, Journal of Crohn’s and Colitis | Systematic review and meta-analysis of placebo rates in induction and maintenance trials of ulcerative colitis, quantifying how far the maintenance placebo rate depends on trial characteristics rather than on the disease. |
| Jairath et al. 2017, Alimentary Pharmacology and Therapeutics | The Crohn’s disease counterpart, establishing the same dependence of placebo rates on design and eligibility features across induction and maintenance phases. |
| Welty et al. 2020, Current Medical Research and Opinion | Applied network meta-analysis that recalculates maintenance data from re-randomized response designs to correspond to treat-through arms; the conversion appears as a methods-section statement with no derivation. |
| Disher et al. 2025, Advances in Therapy | Normalizes re-randomized trials to mimic a standard treat-through design inside an applied network, showing the same undocumented conversion is now routine in submissions. |
| Jairath et al. 2025, Journal of Comparative Effectiveness Research | Runs the conversion in the opposite direction, adjusting treat-through outcomes to align with responder re-randomized outcomes; the two directions have not been reconciled or compared. |
| Katz 2009, Clinical Journal of Pain | Reference treatment of enriched enrollment randomized withdrawal methodology and its interpretive limits, establishing that the estimand of such a trial is defined by the enrichment step. |
| Remiro-Azocar and Gorst-Rasmussen 2024, Research Synthesis Methods | Argues for importing estimand discipline into evidence synthesis; the framework in which a common target for enriched and parallel designs would have to be stated before any conversion is attempted. |
| ICH E9(R1) 2019, Addendum on Estimands and Sensitivity Analysis in Clinical Trials | Defines the population attribute of an estimand, which is precisely the attribute an enrichment rule changes; supplies the vocabulary but not the cross-trial reconstruction. |
Probable solution or research direction
Fix the target before the arithmetic: state whether the decision needs the effect among all patients starting induction or among induction responders, then express each trial’s reported contrast as a functional of the same underlying quantities so that any conversion is a derivation rather than a rescaling. Establish what is identified from aggregate data alone; with the induction response rate, maintenance arm sizes and maintenance rates published, the treat-through contrast is bounded by the range of possible dependence between induction response and maintenance outcome, and those bounds should be reported instead of a point estimate that silently assumes independence. Where individual data exist for one trial, estimate the transition probabilities from them and then test transport rather than assume it, by refitting with each candidate donor trial in turn and reporting the spread across donors. Simulate a network mixing treat-through and re-randomized trials whose enrichment rules differ in threshold and timing, and report bias, coverage and rank reversals for naive pooling, design-stratified networks and reconstruction. Standardize what a maintenance trial must report for any of this to be possible: the induction response definition, the assessment time, the induction response rate, and the maintenance outcome stratified by depth of induction response.
Verification
Verdict. Confirmed open. One reader proposed this from a single methodological case study whose approach is presented as novel and unvalidated, and the reviewer accepted it as more specific than the general estimand-mismatch entry EVB-03. Independent evidence beyond that paper supports it: a meta-analysis of 37 inflammatory bowel disease maintenance trials shows therapeutic gain differs by randomization structure, and the treat-through conversion now used in applied networks exists only as undocumented methods-section statements.
Source
Derived from reading:nma-30 of the reviewed corpus.