Five Methods, One Problem: The Crossover-Adjustment Landscape Just Got Crowded

In hematologic oncology, post-randomization treatment switching is less an exception than a design feature. When effective subsequent therapies exist — and in AML, CLL, myeloma, and lymphoma they reliably do — control-arm patients switch, ITT OS estimates compress toward the null, and regulators and HTA bodies want to know what you plan to do about it. What has changed is the volume and diversity of published answers arriving simultaneously.

A simulation study in Statistics in Biopharmaceutical Research evaluates the hypothetical estimand for OS under varying subsequent-therapy scenarios, comparing RPSFT, IPCW, TSE, and MSMs head-to-head against real trial data — the kind of scenario-conditioned selection evidence that SAP authors actually need. A companion applied paper in the same journal grounds the comparison in real hematologic oncology data, moving the discussion from controlled simulation to the messier context of actual submissions.

Marginal structural models enter the RCT crossover conversation more formally via a Journal of Biopharmaceutical Statistics paper arguing that initial randomization makes MSM’s identifying assumptions more defensible than in pure observational work — a reasonable claim, with one caveat worth pressing: positivity strains badly post-progression when nearly all control patients switch, and whether the paper provides usable diagnostics for that scenario should be checked before adopting MSMs as the primary adjustment.

Multiple imputation arrives twice. One JBS paper frames the crossover problem explicitly within ICH E9(R1) estimand structure, imputing counterfactual survival times under a no-crossover assumption and using Rubin’s rules for uncertainty quantification — a practical advantage over deterministic RPSFT and IPCW approaches. A second JBS paper covers similar ground; whether the two implementations differ meaningfully in their post-switch time assumptions (clock-reset vs. clock-forward) is worth confirming before treating them as interchangeable options. A focused Clinical Trials note rounds out the cluster with a citable technical clarification on RPSFT for the standard-to-experimental-arm crossover scenario.

No consensus has emerged on which method belongs in the primary analysis slot and which in sensitivity. The international regulators convened at the ASA BIOP/FDA Project SignifiCanT forum in June 2025 — FDA, EMA, MHRA, HC, TGA, PEI, PMDA — confirmed that SOC-change scenarios structurally equivalent to crossover remain an active regulatory concern, with panelists noting that analytical adjustments carry significant assumptions and that restarting the trial is often the most defensible option. That is not an endorsement of any single method; it is a reminder that the burden of justification falls on the sponsor.

The implication is emerging, not immediate. Pre-specify your primary method with explicit scenario-dependent rationale and name the assumption that would falsify it; treat the remaining methods as a sensitivity hierarchy rather than a menu of equivalent choices, and plan time-varying covariate data structures for IPCW and MSMs at the protocol stage — not in a CDISC mapping exercise six months before database lock.

Protocol read: The crossover-adjustment menu finally has the shape of a methodology consensus — but only on the meta-rule that no single method is universally preferable, and that pre-specification with explicit rationale is the only defensible posture.

What to do now:

• Pre-specify the primary crossover-adjustment method in the SAP with explicit scenario-dependent rationale, and name the assumption whose violation would falsify it.
• Treat the remaining methods as an explicit sensitivity hierarchy, not as a menu of equivalent options — and document why each appears at its position in the list.
• Plan time-varying covariate data structures for IPCW and MSM at the protocol stage; retrofitting them six months before database lock is the recurring operational failure.