Post-Progression Treatment Regimes and the Limits of Cross-Trial Survival Transport

Models that simulate overall survival (OS) from progression-free survival (PFS) are fitted to published trial reports and used to transport treatment effects to new populations. We show that the quantity such transport requires depends on the post-progression treatment regime--the rates at which patients receive subsequent therapy, indexed by biomarker stratum and randomized arm--and that this object is systematically absent from the evidence base. Two families of correction are in routine use and neither reaches it. Crossover adjustment targets the counterfactual in which no patient receives subsequent therapy, a world no target population inhabits, and thereby eliminates precisely the information distinguishing source from target. Population adjustment matches baseline covariates, and a regime belongs to a trial rather than to a patient, so reweighting individuals cannot convert one regime into another. We establish the second point across four adjustment approaches spanning distinct function classes--- moment-matching MAIC, outcome-regression STC, augmented inverse probability weighting, and a gradient-boosted density ratio---which carry near-identical bias, showing the failure to be informational rather than methodological. We exhibit a regime that is exactly balanced across arms in the marginal, invisible to standard switching assessment, and nonetheless not transportable. Finally we compute the identified set for the transported effect over plausible regimes and mechanisms; it spans zero, so without knowledge of the target's post-progression landscape the sign of the treatment effect is undetermined. All results come from a single seeded script with a test suite, released with the camera-ready version.