Two minor specification issues remain. No new major finding.
Round 15 follow-up
C15-1: Resolved through an acceptable alternative. Binomial and Poisson likelihoods retain zero counts without continuity corrections or undefined transformed observations. Kaplan-Meier inputs now require (0<F<1) and reported uncertainty, addressing transformation boundaries documented in the survival documentation. Poisson mean (e^{\theta_a}E_a/t) and binomial offset (\log(f_a/t)) correctly express both inputs on the log cumulative hazard scale at (t).
Declining my proposal to disable pooling whenever any arm fails is reasonable. Valid zero counts carry information; excluding arms lacking usable reporting is a different decision from excluding arms because their event count is zero.
One correction to the response’s justification is needed. Round 15 response, C15-1 cites ketoacidosis, retinopathy, and acute kidney injury. However, the new hazard count likelihoods apply only to outcomes without competing events, currently O20. Competing-risk integration still rejects zero counts. Replace those examples with:
This alternative retains zero-event reference arms for O20 and boundary event counts for fixed-time binomial outcomes. It does not change the zero-count restrictions of the competing-risk integration routes.
Earlier closure judgments remain intact. Continued deferral of additional Aalen-Johansen horizon methods to a pre-analysis amendment remains acceptable under existing eligibility and relative-only rules.
New findings, all minor
C16-1. New reporting exclusions bypass the single-arm safeguard.
File: protocol/protocol.qmd, “Absolute effects”, particularly lines 1035 to 1049.
Arms without qualifying inputs now contribute nothing, but the single-arm rule still counts arms whose follow-up covers the horizon. With two covering arms and only one usable input, that exception does not apply. The model consequently retains heterogeneity estimation despite having only one informative arm. Scenario percentiles also need a clearly defined contributing set.
Replace the exclusion and single-arm provisions for outcomes without competing events with:
A usable reference arm both covers the analyzed horizon and supplies a qualifying input. Arms lacking qualifying inputs are listed with reasons and excluded from reference-risk estimation and scenario percentiles. With at least two usable arms, fit the specified random effects model. With one usable arm, use its likelihood with PR02, set between-arm heterogeneity to zero, and use that arm’s risk for both empirical percentile scenarios. With no usable arms, apply the relative-only rule. These provisions do not alter the separate competing-risk eligibility rules.
The binomial fallback should explicitly retain the existing known-status requirements:
Binomial inputs must satisfy the known-status rules in “Event outcomes and harms,” with the eligible follow-up (f_a) covering the analyzed horizon; the 95 percent rule remains restricted to its designated sensitivity analysis.
C16-2. PR02’s rationale does not distinguish annual hazard from cumulative hazard at the analyzed horizon.
Files: protocol/priors.csv, PR02 and protocol/protocol.qmd, “Absolute effects”.
PR02 now governs the reference-model mean, but its rationale still describes the baseline as a log cumulative hazard per year. The new model instead defines (\theta_a=\log\Lambda_a(t)). These parameterizations require different prior locations if they are intended to encode the same annual hazard prior. At two years, placing a zero-centered normal prior directly on log cumulative hazard gives prior median risk 0.632; placing it on log annual hazard gives 0.865.
Preserve the explicit protocol model by replacing PR02’s final rationale clause with:
In the primary trial complementary log log model, the offset is log follow-up in years and the baseline is the log annual hazard. In the noncompeting reference-arm model, the mean is on the log cumulative hazard scale at the analyzed horizon, with prior Normal(0; 5^2) and sensitivity prior Normal(0; 10^2). Fixed-time reference-arm models use the logit scale.
Verification and registration judgment
All requested documents and tables inspected. Read-only checks passed for 637 applicable outcome/population window pairs, 43 agent/search-term mappings, and identifiers and counts across 861 registry-candidate rows. Numerical checks confirmed count-likelihood scaling and finite zero-event likelihoods. Search renderer assessment was static; live database recall was not tested. Working tree remains clean; no files edited, created, or deleted.
Methods would be acceptable for registration after these minor clarifications and completion of administrative pre-registration conditions. Missing administrative information alone does not drive this verdict.
VERDICT: MINOR REVISION