Peer review round 8: LivingGLP1 protocol 0.8.0 (reviewer fable-5.1)
What I verified on 2026-10-12 against the 0.7.0 commit: the revision changes only protocol.qmd, amendments.csv (A007) and the README; the twelve other authoritative tables, concepts.yaml (1.2.0) and R/search_strategy.R are byte identical, so the invariants of rounds 5 to 7 hold (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, the four queries rendering). comments.csv has the 13 round 7 rows that A007 states, protocol.html carries the 0.8.0 subtitle, and the list of log_hazard outcomes confirms that O20 is the only one whose event includes death from any cause, so "at present only O20" is right.
1. Round 7 comments
| Item | Status | Reason |
|---|---|---|
| F7-1 | Resolved | Lines 902 to 907: for an outcome with competing events a single rate is never used; the order is reported Aalen-Johansen, then the exponential expression from the pooled rates, which serves as F_ref, then the relative only rule. The added relative only step is right. |
| F7-2 | Resolved | All three sentences adopted: conclusions (522 to 526), likelihoods (713 to 716) and the sensitivity list (867 to 869), the last two extended for O20, correctly. |
| F7-3 | Resolved | Lines 416 to 419 list competing deaths without a preceding target event, person years with the stated fallback and flag, and any competing event hazard ratio. |
| F7-4 | Resolved | Lines 591 to 598 adopt the two reviewer timing, the logging, the sensitivity alternative and both criteria verbatim. |
| F7-5 | Resolved | Lines 540 to 541. |
2. New items in 0.8.0
No major items. Each is a consistency edit between sentences the revision added; none changes a model, an estimand or a primary estimate. For the one contradiction (item 1) an analyst would still build the right network, because line 474 is the more specific rule.
Minor
Primary network for O20 (line 545 against lines 472 to 474). "For time to event outcomes the cause specific network is primary" has no exception, while line 474 makes the not applicable network primary for O20. Replace with: "For time to event outcomes with a competing event the cause specific network is primary".
One input per trial (lines 472 to 474; also the subdistribution network). A trial reporting both a hazard ratio and complete status counts for O20, or a Fine and Gray ratio and counts for a harm, now has two eligible inputs to one network and no stated preference; line 468 implies hazard ratios first and the contribution rule (lines 966 to 970) gives an order, but neither is written as a rule for the network. Append after line 474: "A trial with more than one eligible input to a network contributes one, in this order: the reported hazard ratio, the cumulative incidence or Kaplan-Meier contrast, the arm counts; the alternatives are logged."
F_ref and the constant hazard requirement (lines 560 to 562, 585 to 588, 904 to 907). Line 560 defines F_ref as the Aalen-Johansen incidence; line 906 lets the exponential expression serve as F_ref; line 587 then says that subdistribution conversion does not require constant cause specific hazards, which is false on the exponential route. Replace lines 560 to 562 with: "where F_ref is the reference risk of @sec-absolute (a reported Aalen-Johansen cumulative incidence at that horizon, otherwise the exponential expression below)"; and lines 587 to 588 with: "constant cause specific hazards are not required when F_ref is a reported Aalen-Johansen incidence; otherwise the conversion inherits the constant hazard assumption and carries the exponential label".
Two baselines on the cause specific route (lines 562 to 583 against 902 to 907). The bullet makes the reference risk the reported Aalen-Johansen incidence whenever one exists, while the competing risks paragraph makes the reference cumulative incidence of a cause specific conversion the exponential expression F(t), or the integral, from the pooled hazards. The treated risk comes from the second, so the absolute difference subtracts a baseline from a different source; a pooled Aalen-Johansen figure and the closed form from pooled rates are not equal in general. The bullet also omits the integral as a route, which grok-4.7's G7-4 replacement named, and nothing says how the time varying hazards are pooled across several covering reference arms (lines 565 to 566 pool constant rates only). Replacement for the bullet's second sentence: "For an outcome with competing events a single event rate is never used. For a subdistribution hazard ratio the reference risk is the reported Aalen-Johansen cumulative incidence, otherwise the exponential expression of @sec-outcomes. For a cause specific hazard ratio the reference risk shown is the cumulative incidence of the conversion route at HR = 1: the time varying integral when time varying hazards are available, otherwise the exponential expression under a supported constant hazard. Otherwise the relative only rule applies." Append to the competing risks paragraph: "The time varying hazards are piecewise constant on the intervals of the digitized reference arm curves and are pooled per interval on the log scale with the reference risk model."
The Kaplan-Meier contrast in the contribution rule (lines 966 to 970; 533 to 537). Line 537 says the contrast enters the contribution network without continuity correction, but the contribution rule lists only the reported log hazard ratio, the cumulative incidence contrast and the count estimate. Replace "or the cumulative incidence contrast with its delta method variance" with "or the cumulative incidence or Kaplan-Meier contrast with its delta method variance". Line 536 should also name the source of the standard error, as line 490 does: "with a delta method standard error from the reported confidence intervals or the Greenwood variance".
3. Judgment on revisions made for the other reviewers
C7-1 (cause specific: relative only when neither integration nor a supported exponential approximation is possible; subdistribution: F_ref and proportional subdistribution hazards) is right; item 3 only reconciles it with the F7-1 fallback adopted in the same pass. C7-2 and G7-M1 (one not applicable network for O20, primary, with counts for outcomes with competing events going only to subdistribution networks) are right and close the contradiction grok-4.7 rated major; two parts of the C7-2 replacement were not carried, the exception at line 545 and "one input per trial" (items 1 and 2), although the response records it as adopted. C7-3 (offset log t) is right and Q reproduces. G7-1 (the hazard ratio inside the integral and inside all cause survival) is the correct treated cumulative incidence; I reproduced it numerically in round 7 and the wording now states it. G7-2's estimator is correct: log(-log S) is the log cumulative cause specific hazard, so the contrast is the log hazard ratio under proportional hazards. G7-3 is right, and my round 7 item 2 supplied the same sentences. G7-4 is only partly carried: its replacement named the time varying integral as a reference risk route and made that value the reference risk shown; the adopted bullet omits both (item 4). The response does not disclose either partial adoption.
Conclusion
All five of my round 7 items are resolved in the text, four by my wording and one (F7-1) by a wording with a better final step. The five items above are sentence edits between rules the revision introduced together: which network is primary for O20, which input a trial contributes, what F_ref is on the fallback route, which baseline a cause specific conversion subtracts from, and which inputs the contribution rule lists. None changes a model, an estimand or a primary estimate, which is the standard I applied in rounds 4 to 7. 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