Peer review round 14: LivingGLP1 protocol 0.14.0 (reviewer fable-5.1)
What I verified against the 0.13.0 snapshot: the revision changes protocol.qmd (subtitle; the round and outcome counts at lines 43 and 239; the Competing risks bullet, lines 587 to 659; line 985), priors.csv (PR15 to PR17), amendments.csv (A013) and the README status line. The eleven other authoritative tables, concepts.yaml (1.2.0) and R/search_strategy.R are byte identical to the reviewed snapshot, so the invariants of rounds 5 to 13 hold. comments.csv carries the eight round 13 rows that A013 counts, verdicts.csv the three round 13 verdicts, protocol.html the 0.14.0 subtitle and PR17. Numerical checks: PR15 at 1.96 SD spans hazards from 5.5 × 10⁻⁵ to 1.8 × 10⁴ per year, consistent with its rationale; the Aalen-Johansen inversion still returns 0.04 and 0.02 from constant hazards of 0.04 and 0.02 over one year; PR16 at τ = 1 allows a 2.7 fold proportional deviation of an arm's hazard function at one SD.
1. Round 13 comments
| Item | Status | Reason |
|---|---|---|
| F13-1 | Resolved | Lines 632 to 640 write the model I proposed: y_a ~ N(μ + Z u_a, V_a), one random effect per event shared across that event's intervals, u_ak ~ N(0, τ_k²), PR15 on μ, PR16 on τ_k, and a one arm rule. My observation equation across partitions was dropped because identical interval boundaries are now required (C13-1). The reason holds: the restriction is conservative and executable, the fallbacks at lines 648 to 651 and 1005 to 1013 are prespecified, and the length weighted equation stays logged as the amendment if unequal partitions must be accommodated. Two notes. The restriction will make the multi arm piecewise route rare, since trials publish numbers at risk on their own schedules; that is a loss of availability, not an error. PR16 takes Half-normal(0; 1²) where my row had 0.5², with no stated reason; with two to five arms this scale sets the predictive width, but 0.5² is in the sensitivity set, so the choice is defensible and testable. |
| F13-2 | Resolved | Lines 644 to 648: scenario values stay the empirical percentiles of the arms' own incidences, with uncertainty from N(y_a, V_a) per arm and the percentiles recomputed per draw. |
| F13-note | Resolved | Lines 642 to 644: the displayed reference cumulative incidence is the median of the integrated draws. The univariate case at line 986 is consistent, since the median is invariant to the back transform. |
| F12-note | Half settled | PR17 settles the prior half. The other half (reported Aalen-Johansen values that miss the analyzed horizon) remains deferred to the pre-analysis amendment, as agreed in round 12; I restate it so that it is not lost. |
2. New items in 0.14.0
No major items. None of the items below touches a model for relative effects, an estimand or a relative estimate; all are wording or cross reference fixes within the competing risk absolute effect routes.
Minor
One interval per arm belongs to one route, not two (lines 587 to 594 against 652 to 659). The exponential approximation is now "the same model with a single interval from 0 to the analyzed horizon", yet the piecewise bullet defines a count arm's reported interval as that of its reported events and person time, which is the arm's whole follow-up. Two cases are open. (a) Every covering arm reports one total count over the same fixed follow-up: the boundaries are identical, so the piecewise route qualifies with no constant hazard support judgment and no label, while the same data under the exponential bullet require both. (b) Arms report totals over different follow-up lengths that all pass the horizon: the piecewise bullet does not say whether these boundaries count as identical "through the analyzed horizon", and the exponential bullet's interval is not the data interval. Replace "is the same model with a single interval from 0 to the analyzed horizon" with: "uses the same pooling model with one coordinate per event: each arm's constant rate over its whole person time at risk is taken as its hazard on the single interval from 0 to the analyzed horizon, so the identical boundary requirement is met by construction. Whenever every covering arm has a single reported interval reaching the horizon, the route is this approximation and carries its label and support judgment, not the piecewise route." Both routes give the same F(t) for such data; only the label and the judgment differ.
PR17 is cited where PR16 applies (line 985). "(heterogeneity prior PR17)" sits on the sentence that governs all reference risk pooling, including competing event outcomes routed through piecewise or exponential integration, whose heterogeneity prior is PR16. Replace with "(heterogeneity prior PR17; PR16 for the competing risk routes of @sec-outcomes)".
The one arm distribution is implicit (lines 639 to 640). "The arm's within arm uncertainty is propagated" does not name a distribution, and the predictive sentence that follows has no u_new to draw. Replace with: "With one covering arm, τ_k is not estimated and the reference hazards are drawn from N(y_a, V_a) in place of μ + Z u_new."
Pooling scale for mixed inputs (PR17 against lines 1002 to 1005; pre-existing, made visible by PR17). For hazard based outcomes without a competing event, reference arms contribute Kaplan-Meier cumulative incidence when available and otherwise an event rate; PR17 names the cloglog and log rate scales but not which one carries a network holding both. Add after line 1005: "Both enter one model on the log cumulative hazard at the horizon: log{−log(1 − F)} for a Kaplan-Meier value, with variance by the delta method from its reported interval, and log(λt) for a rate, with variance approximately 1/d, under PR17 on the arm level heterogeneity." This predates 0.14.0 and does not affect the verdict.
3. Judgment on revisions made for the other reviewers
C13-1 and G13-M1 are resolved by the same restriction, with my random effect structure rather than GPT-6 Astra's T = DRD with an LKJ prior; the stated reason (a free between arm covariance over intervals and events is weakly identified with two to five arms, and a proportional deviation of the hazard function is the usual form of baseline heterogeneity) is sound, and the LKJ form would also have been executable. C13-2 and G13-1 coincide with my F13-2 and are adopted. C13-3 is adopted at lines 614 to 620: variance approximately 1/d_k and zero covariance under independent participants and noninformative censoring, which is the correct status of the piecewise exponential working approximation. Grok's displayed incidence rule is adopted at lines 642 to 644. The response overstates nothing.
Conclusion
All three of my round 13 items are adopted, the majors of the other two reviewers are resolved by a restriction whose stated reason I judge sound, and the authoritative tables other than priors.csv are unchanged. The four items above are wording and cross reference fixes within one branch of the absolute effect routes for competing risks; item 1 affects only which label the same F(t) carries, and item 4 predates this version. By the standard I applied in rounds 12 and 13, they are amendment level, to be settled in the next revision or in the pre-analysis amendment that already carries the first half of F12-note. The administrative items remain on the checklist as pre-registration conditions. I would approve the methods for registration once those conditions are completed.
VERDICT: ACCEPT