Round 15: response to reviewers

Protocol version reviewed: 0.15.0. Revised version: 0.16.0.

Verdicts: GPT-6 Astra MINOR REVISION; Fable 5.1 ACCEPT; Grok 4.7 MINOR REVISION. Every comment is answered below and recorded in ../comments.csv.

Summary of main changes

• Reference arm meta-analysis: exact likelihood of each arm data (binomial logit; Poisson and binomial complementary log log on the log cumulative hazard at the horizon), so zero event arms are kept without correction; Kaplan-Meier values by normal approximation only with 0 < F < 1 and reported uncertainty; mean prior PR02; t in years.
• Exponential approximation rewritten as one passage: single interval piecewise fits and constructed single periods that reach or pass the horizon, both gated and labeled.

GPT-6 Astra (gpt-6-astra)

C15-1 (minor; Absolute effects). Undefined inputs in reference risk pooling: F = 0 or 1 and zero event rates make the normal model undefined.

Response (revised (alternative)). Agreed that the inputs needed a rule; resolved by an alternative that keeps the arms. Under your replacement a single zero event reference arm would disable absolute effects for the whole network, and zero event reference arms are routine for the many complementary log log harms outcomes (for example diabetic ketoacidosis, retinopathy complications, acute kidney injury). The reference arm meta-analysis now uses the likelihood of each arm own data: binomial with logit link for fixed time outcomes; for hazard based outcomes a Poisson likelihood for counts with person time and a binomial likelihood with complementary log log link and log follow-up offset for participants with an event, both on the log cumulative hazard at the horizon; a reported Kaplan-Meier value enters by the normal approximation only when 0 < F < 1 with a reported confidence interval or Greenwood variance. Zero counts need no correction and no arm is dropped for having none, which keeps the two requirements of your text (no omission because of the event count, no continuity correction). The mean prior is PR02 and the one arm case uses that arm likelihood under PR02. The competing risk routes keep the working normal likelihood, because curve inputs have no count likelihood; the protocol now states this.

Grok 4.7 (grok-4.7)

G15-1 (minor; Competing risks). The separate exponential route contradicted its own availability rule; gate and label only in the first sentence.

Response (revised). Agreed; the bullet is rewritten in one passage: a piecewise fit whose common interval is a single interval from 0 is the exponential approximation; when the piecewise route is unavailable, arms with counts and person time over one period from 0 that reaches or passes the horizon (finer intervals summed) give one coordinate per event, pooled with the same model; in both cases the constant hazard gate and the label apply. The clause that the constructed interval met the identical boundary requirement is removed. Periods that pass the horizon are admitted (fable-5.1 F15-1), matching the rule for outcomes without a competing event, since under the constant hazard gate a rate over a longer period estimates the same hazard.

G15-2 (minor; Absolute effects). Units of t and the variance input of the log cumulative hazard model unspecified.

Response (revised). Agreed; t is the analyzed horizon in years (weeks divided by 52.18), rates are per person year, and the Kaplan-Meier variance comes from the reported confidence interval or the Greenwood variance, requiring 0 < F < 1. Counts now enter by their exact likelihoods (C15-1).

Fable 5.1 (fable-5.1)

F15-1 (minor; Competing risks). Single reported intervals that pass the horizon.

Response (revised). Agreed; your second closure adopted in substance (see G15-1): single periods from 0 that reach or pass the horizon qualify, finer count intervals are summed, and the gate and label apply in every case; the piecewise bullet now states that a last count interval may end after the horizon, with its rates applied up to the horizon.

F15-2 (minor; Absolute effects). "Reported interval" ambiguous; Kaplan-Meier without uncertainty.

Response (revised). Agreed; "reported confidence interval or Greenwood variance". A Kaplan-Meier value without either enters through the arm counts; an arm with no qualifying input is listed with the reason. Because zero counts always qualify under the count likelihoods, such exclusion depends on reporting, not on the event count.

F15-3 (minor; Priors; absolute effects). No prior on the mean of the reference arm meta-analysis.

Response (revised). Agreed; PR02 now applies to the mean of the reference arm meta-analysis on the same scales, and the text cites it.

F15-note (note; Absolute effects). Reported Aalen-Johansen values that miss the analyzed horizon.

Response (deferred to amendment). Remains deferred to the pre-analysis amendment, as agreed in round 12.