One major specification gap and two minor corrections remain. Administrative omissions do not determine this verdict.
Round 12 assessment
No C12 comments were raised. Changes responding to other reviewers are partly successful:
- Log-hazard Jacobian: resolved. Defining (g) explicitly on log hazards corrects covariance propagation. Numerical verification gives 0.016975 for round 11’s Kaplan-Meier example.
- F12-2 and F12-3: resolved. Extraction wording now separates published curves from requested covariance information and pairs author-supplied covariance with corresponding ordinate estimates.
- G12-M1 and F12-1 dependence correction: partly resolved. Joint modeling is appropriate, but replacement introduces an incompletely specified pooling model, detailed below.
- F12-note deferral: Existing fallback rules justify deferring practical availability questions about reported Aalen-Johansen estimates. Deferral of an older reference-risk prior issue does not specify covariance parameters introduced by this revision.
New findings, ordered by severity
C13-1. Major: multivariate pooling lacks common coordinates and a between-arm covariance specification.
Location: protocol/protocol.qmd:629, “Competing risks,” together with the union-grid rule at line 597 and protocol/priors.csv.
An arm reporting one interval from 0 to 104 weeks supplies two unique log hazards. An arm reporting intervals from 0 to 52 and 52 to 104 weeks supplies four. Saying that copies remain “one coordinate” does not specify how these differently indexed vectors enter one model.
Expanding both vectors onto the union grid requires explicit duplication maps and produces singular within-arm covariance for repeated coordinates. Moreover, preserving within-arm covariance does not specify between-arm covariance or guarantee equality of copied predictive coordinates. Multivariate random effects models require both covariance levels and sufficient identification or explicit regularization. Jackson, Riley, and White, Section 3.
This affects uncertainty in cumulative incidence, absolute differences, and magnitude-dependent certainty judgments.
Concrete conservative replacement: Replace the union-grid pooling instruction with:
Pooled piecewise integration requires identical reported interval boundaries across all covering reference arms through the analyzed horizon. Otherwise this route is unavailable, and the existing alternative-route hierarchy applies. Each arm contributes a vector (y_a) containing one log hazard per cause and reported interval, with within-arm covariance (V_a). Fit (y_a\sim N_K(\mu,V_a+T)), where (T=DRD), (D=\operatorname{diag}(\tau_1,\ldots,\tau_K)), independently (\mu_j\sim N(0,5^2)), (\tau_j\sim\operatorname{HalfNormal}(0,1^2)), and (R\sim\operatorname{LKJ}(2)). Hazards use years as the time unit. For each posterior draw, draw a new reference-arm vector from (N_K(\mu,T)) and integrate its hazards jointly. Numerical subdivision reuses each interval’s sampled hazard. With one covering arm, propagate its within-arm uncertainty without estimating between-arm heterogeneity.
Add these mean, heterogeneity, and correlation priors as PR15 to PR17 in
priors.csv, with respective sensitivity specifications (N(0,10^2)), (\operatorname{HalfNormal}(0,2^2)), and (\operatorname{LKJ}(1)).Retaining different reporting partitions is also possible, but requires a specified observation model and prediction rule instead of this conservative restriction.
C13-2. Minor: scenario uncertainty now targets a different quantity from the declared scenarios.
Location: protocol/protocol.qmd:636, compared with “Absolute effects” at line 974.
Scenarios remain empirical 25th and 75th percentiles of covering arms’ own cumulative incidences. New wording propagates their uncertainty using the pooled predictive distribution. Predictions for new reference arms do not represent uncertainty in those observed arms’ empirical quantiles.
Replacement text:
Joint predictive draws determine uncertainty for the pooled reference profile. Low and high scenario point estimates remain the empirical 25th and 75th percentiles of covering arms’ own cumulative incidences. For piecewise-route scenario uncertainty, draw each arm’s unique log-hazard vector from its working multivariate normal distribution with mean (y_a) and covariance (V_a), independently across arms. Reuse each sampled reported-interval hazard on every subdivision, integrate each arm’s cumulative incidence, and recompute the empirical quantiles in each draw. New-arm predictive draws do not replace these arm-specific draws.
C13-3. Minor: count-based uncertainty approximation is now stated as exact.
Location: protocol/protocol.qmd:615, “Competing risks.”
Revision replaces (\operatorname{Var}(\log h_k)\approx1/d_k) with equality and states independence without qualification. Inverse event counts provide an information-based large-sample approximation, not an exact finite-sample variance or independence result.
Replacement text:
For reported first-event counts with valid person time, use the working normal approximation for estimated log hazards from the piecewise exponential likelihood, with variance approximately (1/d_k) and zero cross-cause covariance under its standard independent-participant and noninformative-censoring assumptions. A zero count for either event makes this route unavailable.
All requested files inspected. Cross-table checks passed for 637 applicable outcome/population window pairs, 43 search-agent mappings, and all 861 registry-candidate rows. Search concepts and renderer are unchanged from version 0.12.0; renderer review was static, without live database recall testing. No files edited, created, or deleted.
VERDICT: MAJOR REVISION