The MAR Default Is Quietly Becoming Indefensible

A systematic review in Clinical Trials asks the question every SAP author has quietly wondered about: across real trials, does multiple imputation actually move the treatment effect away from what complete-case analysis would have given you? The review compiles the empirical record rather than relitigating the theory. That alone is useful — the methodologist consensus that MI is materially superior to CCA has rested more on simulation than on a head count of published trials where it changed the inferential conclusion.

Land that finding next to two near-simultaneous Journal of Clinical Epidemiology papers on patient-reported outcomes and the picture sharpens. A JCE commentary argues that PRO missingness is structurally MNAR — dropout is driven by toxicity, disease burden, deteriorating health — and that MAR-based MI applied without comment is therefore the wrong default, not a conservative one. A companion simulation study benchmarks MI, ML-based imputation, and CCA across mechanisms, supplying the empirical leg the commentary’s normative argument needs. Read together with the Clinical Trials review, the combined message is uncomfortable: the workhorse “MAR + MI + a token tipping-point” SAP paragraph is doing less work than its length suggests, particularly on PRO endpoints that are increasingly primary or key secondary in oncology, CNS, and rare disease submissions.

What regulators may actually accept

The constructive answers are arriving in the same publication cycle, which is convenient. A Journal of Biopharmaceutical Statistics paper develops a Bayesian treatment of the causal reference-based imputation model — situating jump-to-reference and copy-reference within potential outcomes, and letting delta-adjustment and tipping-point sensitivity analyses fall out of prior specification rather than be bolted on. This is the kind of framing ICH E9(R1) reviewers have been gesturing toward since the addendum; it gives the intercurrent-event strategy a coherent statistical home rather than a footnote. Separately, an SBR paper proposing MI-B, a B-spline multiple imputation method aimed at trial monitoring, addresses the narrower but real problem that parametric MI under multivariate normal assumptions misrepresents nonlinear trajectories — exactly the regime where interim decisions are most sensitive to imputation choice. A further JBS contribution pairs ML-based imputation with nonparametric multiple comparisons to address the uncongeniality problem when flexible imputation is followed by parametric inference. None of these is plug-and-play for a submission tomorrow, but together they describe what a defensible PRO missing-data section will look like in two years: a named mechanism, a model that matches it, and sensitivity analyses that vary the assumption rather than the seed.

The meta-analysis flank

The same cluster opens a second front on evidence synthesis, where sensitivity analysis is being rebuilt with less fanfare. A Statistics in Medicine paper extends Copas-Heckman publication-bias sensitivity from the normal-normal random-effects model to contrast-based GLMMs, eliminating the 0.5 continuity correction that has quietly distorted rare-event pooled estimates in safety meta-analyses for years. A Statistical Methods in Medical Research paper reasserts the three-way distinction between common-effect, random-effects, and fixed-effects (plural) meta-analysis — the last conditioning inference on the studies at hand without a superpopulation — which is directly relevant to scope-of-inference language in integrated summaries. And a provocatively titled SBR piece argues that mainstream meta-analytic estimators as applied to clinical trials are strongly inconsistent — a structural, not variance-driven, claim that warrants reading before it warrants repeating. Sweeping inconsistency claims usually rest on a specific model configuration; the scope of this one needs checking against the paper’s actual setup before it shows up in anyone’s HTA defense.

The common thread is that boilerplate is the failure mode. Whether the topic is PRO dropout, reference-based imputation, rare-event pooling, or the scope of a random-effects estimate, the new work consistently asks for a stated mechanism, a model that matches it, and a sensitivity analysis that actually stresses the assumption rather than redecorating it.

Protocol read: The MAR-default-plus-tipping-point template is no longer the safe choice it was — particularly for PRO-primary trials, where reviewers now have peer-reviewed ammunition to push back. Constructive alternatives exist; the work is in pre-specifying them before the SAP is locked, not appending them after.

What to do now:

• Audit your PRO SAP templates for the phrase “assumed missing at random” without clinical justification; either defend MAR against the JCE commentary’s argument or move to a mechanism-justified primary analysis.
• Treat the Bayesian causal reference-based framing as a candidate for the next confirmatory SAP iteration, with delta-adjustment priors replacing post-hoc tipping-point appendices.
• Re-examine rare-event safety meta-analyses still relying on 0.5 continuity corrections; the GLMM Copas-Heckman extension is the more defensible sensitivity tool.
• Read the “strongly inconsistent estimators” paper before citing it — verify the scope of the claim before letting it into a reviewer response.