No estimand or correction for randomized-withdrawal anchors
Statement
In a randomized-withdrawal maintenance design the maintenance-phase placebo arm consists of patients who responded to the trial’s own active drug during induction and were then re-randomized off it, so its outcomes are a function of that drug’s off-treatment durability rather than of an untreated course. Anchoring two such trials on their withdrawal arms therefore violates the requirement that the common comparator be the same condition in both, and the violation is signed: the trial whose drug has the longer off-drug effect records the higher comparator response and hence the smaller active-minus-comparator contrast, so the anchored comparison penalizes durability even when on-treatment effects are identical. The magnitudes are not small; in one atopic dermatitis pair, maintenance of response at week 52 in the withdrawal arms was 56 percent against 30 percent for one endpoint and 40 percent against 14 percent for another (doi:10.1007/s13555-025-01622-9), and the risk ratios between the two withdrawal arms were reported as 1.995 and 3.724 (doi:10.1007/s13555-024-01107-1). No estimand definition states what a withdrawal-anchored contrast identifies, no sensitivity parameter indexes the residual index-treatment effect carried by the anchor arm, and no rule says when a withdrawal arm may be treated as placebo, so the dispute is settled by assertion: one side argues that the arms are demonstrably different, the other that pharmacokinetics leave no active drug at 36 weeks, and the raw between-arm difference is cited as decisive when it also contains the population imbalance that anchoring exists to absorb. Randomized-withdrawal maintenance designs are standard in atopic dermatitis, chronic pain and other chronic indications, so this is a recurring structural feature of the evidence base rather than one contested analysis, and the same argument has now run through two published exchanges over the same treatment pair.
Why it is open
The contamination is a property of the design rather than of the patients, so covariate balancing, weighting and outcome regression cannot reach it, and the quantity that would index it, the residual effect of the index treatment inside its own withdrawal arm, is never observed because no untreated control is carried through the maintenance phase. Separating that residue from ordinary between-trial population differences would require a third arm or an external reference that these designs do not supply, so the two candidate explanations for a between-trial difference in withdrawal-arm response are not identified apart. Pharmacological argument cannot settle it either, because the biasing quantity is a persisting biological effect rather than detectable drug, so an argument from half-life addresses a different quantity than the one that moves the contrast. The alternative on offer, an unanchored comparison, exchanges a signed and possibly modest bias for an unsigned and possibly large one, and nothing prices that exchange. The problem also has no owner, falling between trial design, where the withdrawal arm is a legitimate control for the trial’s own question, and evidence synthesis, where it is silently reused as a common comparator.
What has been tried
Probable solution or research direction
Define the estimand a withdrawal-anchored contrast identifies by writing the withdrawal-arm mean as an untreated maintenance outcome plus a residual index-treatment effect, so the bias of the Bucher contrast becomes an explicit function of the difference in residual effects between the two trials rather than a matter of opinion. Supply a sensitivity parameter for that difference, report the contrast across a range of it, and state the tipping point at which the conclusion changes; this is bias-parameter analysis of a familiar form and needs no new estimator. Elicit plausible ranges from relapse-time and pharmacodynamic evidence rather than from half-life, since the biasing quantity is persistence of effect and not persistence of drug. State a decision rule for the design: where the residual-effect difference cannot be bounded below the tipping point, report the anchored and unanchored analyses together rather than choosing between them. For trials still being planned, an untreated or long-washout maintenance arm, or reporting of relapse curves by induction treatment, would make the residual effect estimable and is the only route that removes the assumption instead of parameterizing it.
Verification
Verdict. Confirmed open. One reader proposed this from a published exchange of letters rather than from methodological work, and the reviewer accepted it as more specific than generic comparator-context heterogeneity while noting that a usable framework would have to distinguish pharmacologically negligible carryover from material residual durability. No methodological paper addresses withdrawal-arm anchors at all, so the evidence is applied correspondence in which the two sides reach opposite conclusions with no data-based adjudication available to either.
Source
Derived from reading:core-03 of the reviewed corpus.