Informative censoring and outcome-dependent missingness
Informative censoring cancels in an indirect RMST contrast only when the two mechanisms match. Unanchored, with differing mechanisms the bias was -2.3 to +1.4 months against a true difference of 0.62, reversing the comparison in up to 97% of analyses; equal mechanisms nearly cancelled. Anchored comparisons were protected (bias within 0.22) because both arms of each trial shared its mechanism. Adjusting only the source by inverse probability of censoring weighting helped partly where only the source was informative and broke the cancellation where both were. Read the study below.
Statement
Continuous outcomes go missing and time-to-event outcomes are censored for reasons related to the outcome itself, and the source and comparator trials can have different mechanisms, so the biases do not cancel in an indirect comparison. Dependence between the dropout process and the outcome after dropout is not identified from observed event and censoring data alone without auxiliary information or restrictions, though observable censoring patterns and some model implications can still be examined, so the untestability is specific rather than total. Baseline covariate adjustment cannot correct selection driven by unobserved post-baseline outcomes, but correction is not ruled out in principle: baseline covariates may suffice under conditional independent censoring, and time-varying inverse probability of censoring weighting, congenial imputation or joint models can identify effects under further assumptions. Externally controlled comparisons are where the asymmetry bites hardest, and censoring there is documented to be often informative because it reflects discontinuation for toxicity, patient preference or early switch to alternative therapies.
Why it is open
Ignorable missingness and non-informative censoring fail exactly where dropout is caused by toxicity or worsening disease, and the mechanisms can differ between the two trials, so the errors do not offset. Aggregate-data publications report at-risk tables rather than censoring reasons and give neither the missingness pattern nor the reasons in usable detail, so a missing-not-at-random sensitivity analysis has no anchor on the comparator side. Generic and calibrated inverse probability of censoring weighting adjust within a study rather than across the transport step, and the literature auditor found no method that makes such a sensitivity analysis compatible with a transport model. At commit 9d150e9, cpaic, the unreleased development package on GitHub that is not installable from CRAN, handles right, left and interval censoring and delayed entry but models the outcome conditional on the supplied rows and has no censoring-process, dropout or informative-censoring sensitivity model.
What has been tried
| Work | What it contributes |
|---|---|
| Informative censoring in externally controlled clinical trials: a potential source of bias, ESMO Open 2024 | asserts in the externally controlled comparison setting that censoring is often informative because it reflects discontinuation due to toxicity, preference or early switch to alternative therapies, and that this can bias progression-free survival comparisons; it does not supply a transport-compatible correction |
Probable solution or research direction
Build multiple imputation and inverse probability of censoring weighting that are congenial with the transport model rather than only with the within-study analysis, together with joint models of the outcome and the dropout or censoring process. Make missing-not-at-random and informative-censoring sensitivity analysis explicit, reporting how much outcome dependence would be needed to change the decision rather than assuming the mechanism away. State which features of the censoring process are being assumed and which are examinable, since observed censoring patterns can be inspected even though post-dropout dependence cannot.
Our study
Informative censoring on the two sides of an indirect comparison
The question put to it. When censoring depends on latent risk in the source and comparator trials to different degrees, does the bias cancel in the indirect RMST contrast, what does source-side IPCW achieve, and does a one-sided comparator sensitivity analysis cover the truth?
What it found. Informative censoring cancels in an indirect RMST contrast only when the two mechanisms match. Unanchored, with differing mechanisms the bias was -2.3 to +1.4 months against a true difference of 0.62, reversing the comparison in up to 97% of analyses; equal mechanisms nearly cancelled. Anchored comparisons were protected (bias within 0.22) because both arms of each trial shared its mechanism. Adjusting only the source by inverse probability of censoring weighting helped partly where only the source was informative and broke the cancellation where both were.
Results.
- Refuting sentence fails: unanchored bias -2.326 to 1.416 months when mechanisms differed; -0.080 to 0.021 when equal.
- Anchored: bias -0.202 to 0.219 months in all 12 scenarios.
- Source-side IPCW with a proxy correlated 0.7 with latent risk: 1.42 to 0.99 months where only the source was informative; -0.08 to -0.52 where both were equally informative.
- The comparator delta reproducing the truth lay outside 0.5 to 2 in up to 70% of analyses (median 2.4 with strong dependence at 50% censoring).
Design.
- Estimand
- RMST difference at 24 months, unanchored (A minus B) or anchored ((A minus C) minus (B minus C))
- Methods compared
- Kaplan-Meier RMST, source-side inverse probability of censoring weighting, one-sided comparator delta sensitivity analysis
- Data-generating mechanism
- Shared-frailty event and censoring processes with separate dependence in the source and comparator; 24 cells crossing the two dependences, censoring prevalence and anchored or unanchored design; naive Kaplan-Meier, source-side IPCW and a comparator delta-adjusted sensitivity analysis. 1000 replicates per cell.
- Replicates
- 1000 per scenario, 24,000 in total
What this does not answer. No covariate shift between trials; time-constant censoring hazards; one sensitivity parameterization; one baseline proxy. Peer review has not been done.
Peer review. Not yet done. The result stands on the study alone until it is.
Read it: PDF · Markdown · OpenDocument · Code and data: studies/OUT-07-censoring-mechanisms-that-do-not-cancel · Protocol: registered before the run
Verification
Verdict. Partially addressed. Auditors agreed the problem is open and no counterevidence surfaced. Reopened by the full-text reading: 3 findings from 3 papers survived both the reader and an independent second reviewer, so the verdict moves from confirmed-open to partially-addressed.
The source said. “Continuous outcomes go missing and time-to-event outcomes get censored for reasons related to the outcome itself, such as lack of efficacy or toxicity, yet population-adjusted analyses assume ignorable missingness and non-informative censoring.”
What was wrong with it. The claim that non-informative censoring is simply untestable was narrowed by the code auditor to dependence between censoring and the unobserved post-dropout outcome, and the claim that population adjustment cannot correct such selection was narrowed to baseline weighting alone, since time-varying inverse probability of censoring weighting, congenial imputation and joint models can identify effects under further assumptions. The literature auditor supplied a real source for the asymmetry in the indirect-comparison setting, an ESMO Open 2024 article on informative censoring in externally controlled trials, and voted the entry supported rather than caveated; the weaker caveated reading is used here and the disagreement stays in the audit record.
Dissent. literature dissented (supported): The core claim is structurally untestable and therefore cannot be closed by any paper: non-informative censoring and ignorable missingness are assumptions about the joint outcome-dropout distribution, so per the audit calibration this stays open regardless of how much sensitivity-analysis literature exists. I did retrieve a real source that asserts the problem specifically in the indirect-comparison setting rather than in trials generally: ‘Informative censoring in externally controlled clinical trials: a potential source of bias’ (ESMO Open 2024), which notes censoring is often informative because it reflects discontinuation due to toxicity, preference, or early switch to alternative therapies, and that this can bias progression-free survival comparisons. This supports the asymmetry claim, since external-control and unanchored ITC designs are exactly where the two sides have different censoring processes. I found no method that makes MNAR or informative-censoring sensitivity analysis compatible with a transport model, so the proposed direction is genuinely unbuilt. Generic IPCW and calibrated-IPCW sensitivity analyses exist but adjust within a study, not across the transport step.
Revised by the full-text reading. Applied 2026-07-26 full-text reading of the consolidated library (687 papers).
- Verdict changed
- confirmed-open -> partially-addressed 3 surviving findings from: Allowing for informative missingness in aggregate data meta-; Allowing for uncertainty due to missing continuous outcome d; Continuous(ly) missing outcome data in network meta-analysis. New citations: 10.1002/sim.6365, 10.1177/0962280220983544
- Literature and prior-art check
- open / supported. The core claim is structurally untestable and therefore cannot be closed by any paper: non-informative censoring and ignorable missingness are assumptions about the joint outcome-dropout distribution, so per the audit calibration this stays open regardless of how much sensitivity-analysis literature exists. I did retrieve a real source that asserts the problem specifically in the indirect-comparison setting rather than in trials generally: ‘Informative censoring in externally controlled clinical trials: a potential source of bias’ (ESMO Open 2024), which notes censoring is often informative because it reflects discontinuation due to toxicity, preference, or early switch to alternative therapies, and that this can bias progression-free survival comparisons. This supports the asymmetry claim, since external-control and unanchored ITC designs are exactly where the two sides have different censoring processes. I found no method that makes MNAR or informative-censoring sensitivity analysis compatible with a transport model, so the proposed direction is genuinely unbuilt. Generic IPCW and calibrated-IPCW sensitivity analyses exist but adjust within a study, not across the transport step. Cites https://www.esmoopen.com/article/S2059-7029(24)01864-7/fulltext: Provides a real, ITC-adjacent source asserting that censoring in externally controlled comparisons is often informative and can bias the comparison; converts the entry’s assertion into a sourced one without solving it..
- GPT-5.6 Sol, technical and source-code lens
- open / supported-with-caveat. C1, C2, C5, and C6 are technically sound. C3 needs a precise boundary: dependence between censoring or dropout and the unobserved outcome after dropout is not testable or identified from observed event and censoring data alone without auxiliary information or restrictions. Observable censoring patterns and some model implications can still be examined, so not every relevant feature is untestable. C4 is correct for baseline PAIC weighting alone when censoring depends on unobserved post-baseline outcomes. It is too broad if read as ruling out correction: baseline covariates may suffice under conditional independent censoring, and time-varying IPCW, compatible imputation, or joint models can identify effects under additional assumptions. At commit 9d150e9, cpaic handles right, left, and interval censoring and delayed entry, but models the outcome likelihood conditional on supplied rows and has no censoring-process, dropout, MNAR, or informative-censoring sensitivity model. The stated issue therefore remains open. Very high priority is proportionate when dropout is plausibly outcome-related and differs across trials, but is context-dependent rather than intrinsic to every PAIC.
- Decision path
-
R5:default-open → R11:caveated(n=1)
How the evidence falls in time
11 papers in the reviewed corpus bear on this problem, published 2014 to 2021. Progress and a fresh assertion of openness land in the same year.
It is also being closed in pieces: partial results come from more than one paper in more than one year, each covering a different part.
| Year | Says | Paper |
|---|---|---|
| 2014 | partly addresses it | Allowing for uncertainty due to missing continuous outcome data in pairwise and network meta-analysis |
| 2016 | confirms it is open | Identifying back pain subgroups: developing and applying approaches using individual patient data collected within clinical trials |
| 2016 | confirms it is open | Methods for network meta-analysis of continuous outcomes using individual patient data: a case study in acupuncture for chronic pain |
| 2018 | partly addresses it | Allowing for informative missingness in aggregate data meta-analysis with continuous or binary outcomes: Extensions to metamiss |
| 2018 | confirms it is open | A systematic survey shows that reporting and handling of missing outcome data in networks of interventions is poor |
| 2018 | confirms it is open | An innovative and comprehensive technique to evaluate different measures of medication adherence: The network meta-analysis |
| 2019 | confirms it is open | A comparative review of network meta-analysis models in longitudinal randomized controlled trial |
| 2019 | confirms it is open | Participants’ outcomes gone missing within a network of interventions: Bayesian modeling strategies |
| 2020 | confirms it is open | Comparison of exclusion, imputation and modelling of missing binary outcome data in frequentist network meta-analysis |
| 2021 | partly addresses it | Continuous(ly) missing outcome data in network meta-analysis: A one-stage pattern-mixture model approach |
| 2021 | confirms it is open | Pattern-mixture model in network meta-analysis of binary missing outcome data: one-stage or two-stage approach? |
Source
Derived from C:L1537-1546, C:L1578-1590, C:L250 of the reviewed corpus.