Peer review round 7: LivingGLP1 protocol 0.7.0 (reviewer fable-5.1)
What I verified on 2026-10-12 against the 0.6.0 commit: only protocol.qmd and amendments.csv (A006) changed; concepts.yaml (1.2.0), R/search_strategy.R and the twelve other authoritative tables are byte identical, so the invariants and search counts I reported in rounds 5 and 6 stand by invariance (re-read: 77 outcomes, 13 on log_hazard and 23 on log_hazard_cloglog, nrs_synthesis flags matching harms_nrs.csv, every node class resolving, 11 templates, and the four queries render). comments.csv has the 14 round 6 rows that A006 states, and protocol.html carries the 0.7.0 subtitle. I recomputed the new numbers and formulas: the two bias illustrations give 1.93 and 1.26; Q equals the exponentiated contrast of the complementary log log model with the log t offset (r = 0.10 and 0.14 with t = 1 and 1.5 years gives 0.954 both ways); the treated cumulative incidence with HR λ1 and the competing event ratio times λ2 matches numerical integration of overall survival times the cause specific hazard; and 1 - (1 - F_ref)^HR equals exp of minus HR times the cumulative subdistribution hazard.
1. Round 6 comments
| Item | Status | Reason |
|---|---|---|
| F6-1 | Resolved | The conclusions sentence is adopted verbatim (lines 514 to 516), and "a trial reporting inputs of both types contributes each to its own network" is in the competing risks paragraph. I still endorse keeping the cause specific network as the second step over grok-4.7's exclusion (G6-2): a connected network of reported Cox hazard ratios is a valid, labeled relative effect on the harm, whereas Q is descriptive. |
| F6-2 | Resolved | The adopted wording is gpt-6-astra's (C6-4), which counts competing deaths as mine did and also makes administrative censoring before t unknown and keeps competing deaths in the denominator; I prefer it to mine. |
| F6-3 | Resolved | Not applicable is anchored to the outcome, with the estimator rule for trials in which nobody died; O20 is all cause mortality, so the example is right (item 5 tightens "such as"). |
| F6-4 | Resolved | λ1 and λ2 are events per person year from the covering reference arms, pooled on the log scale, under a stated constant hazard; the reference risk bullet names Aalen-Johansen and Kaplan-Meier correctly. |
| F6-5 | Resolved | The Kaplan-Meier contrast is a cause specific input; the ban is limited to absolute and reference risks. |
| F6-6 | Resolved | The imprecision rule names the relative only rule and the Q range; the data items list estimator, hazard type, known status counts and observation period. The further items the new λ2 rule needs are item 3 below. |
2. New items in 0.7.0
No major items. Each item completes a rule the revision introduced; none changes a model, an estimand or a primary estimate.
Minor
The constant hazard fallback for reference risks is a net risk (Absolute effects, lines 863 to 868; Competing risks, lines 544 to 566). When no cumulative incidence is reported, the bullet estimates the reference risk from "their event rate under a constant hazard", which for an outcome with competing events is 1 - exp(-λ1 t), the parametric form of the net risk that the competing risks paragraph forbids as a reference risk (lines 540 to 542). The sentence at line 544 and the closing "Competing risks follow @sec-outcomes" override it, so this is a consistency edit; the fallback sentence predates 0.7.0, and the revision's definition of F_ref as the output of this procedure is what now points at it. Replacement for the bullet: "and otherwise their event rate under a constant hazard within their own follow-up; for outcomes with competing events the constant hazard estimate is the exponential competing risks expression computed from the pooled λ1 and λ2 of @sec-outcomes, and it serves as F_ref".
The time to event count rule has no downstream sentences (lines 465 to 469, 514 to 516, 680 to 683, 834 to 835). Counts for a time to event outcome now form a subdistribution network, but the conclusions ordering is scoped to "a binary harm", the likelihood list names the complementary log log model "for binary harms" only, the 95 percent sensitivity analysis sits inside the binary harms bullet and the sensitivity list says "harms counts". A time to event outcome whose Cox network fails the connectivity rule while its count network meets it has no rule. Replacements: "Conclusions about a binary harm or a time to event outcome use the primary hazard type network when it meets the connectivity rule; otherwise the other hazard type network, labeled with its hazard type; otherwise, for binary harms, the recorded harms analysis, labeled as such, and for time to event outcomes, descriptive reporting." Likelihoods: "binomial with complementary log log link and log follow-up offset for binary harms and for complete status counts of time to event outcomes". Sensitivity list: "counts from trials with at least 95 percent known event status added to the subdistribution network (binary harms and time to event outcomes)".
Data items for λ2 (lines 411 to 416; 549 to 553). The conversion needs competing deaths without a preceding target event by arm, person years by arm and any reported hazard ratio for the competing event; none is listed, and person years are rarely reported directly. Append to the outcome level items: "competing deaths without a preceding target event and person years by arm (person years as reported, otherwise the number randomized times the arm's mean follow-up, with median follow-up flagged); any reported hazard ratio for the competing event".
Two judgments without a criterion or a time (lines 535 to 540; 563 to 566). "Material nonproportionality" and "the constant hazard approximation cannot be supported" both change what is reported and are judged after data are seen. Replacement, appended to the competing risks paragraph: "Both judgments are made by two reviewers before relative effects are modeled and logged with reasons, with the alternative as a sensitivity analysis: constant hazard support from the reference arm cumulative incidence curves logged under @sec-absolute, and proportionality from reported tests and the published arm curves; nonproportionality is material when a reported test rejects proportionality at the 5 percent level or the arms' cumulative incidence curves cross after the first quarter of follow-up."
"Such as O20" leaves the not applicable set open (line 527). O20 is the only all cause mortality outcome; every other hazard scale outcome, including death by suicide (O35B), has competing deaths. Replacement: "only for outcomes with no competing event, at present only O20".
3. Judgment on revisions made for the other reviewers
C6-1 (constant hazards stated; integration when time varying baseline hazards are available) and G6-3 (constant rates for the reference node pooled on the log scale, λ1 replaced everywhere, F_ref defined) are correct; the closed form reproduces under numerical integration. C6-2 (proportional subdistribution hazards, with horizon specific contrasts and no cross horizon conversion when it fails) is the right assumption and the right fallback; item 4 only fixes who decides and when. C6-3 is correct: Q is exactly the model's contrast, and calling it descriptive is right because the nominal observation period is not person time. C6-5 and G6-1 are correctly hedged, with t defined and the numbers reproducing. G6-2's count rule for time to event outcomes is right (counts are subdistribution inputs and never pool with Cox estimates), and item 2 supplies the sentences it needs.
Conclusion
All six of my round 6 replacements were carried into 0.7.0, one by a better wording than mine, and the invariants hold by identity. The five items above are sentence edits that complete rules whose substance the owner has already chosen: which value serves as F_ref under constant hazards, which network carries conclusions for a time to event outcome, what is extracted for λ2, when the two judgments are made, and which outcome is not applicable. None changes a model, an estimand or a primary estimate. Applying the standard of rounds 4 to 6, I would approve the methods for registration once the administrative pre-registration conditions are completed, with the five edits made in the same pass.
VERDICT: ACCEPT