One minor execution gap remains in newly specified reference-risk pooling. No new major findings.
Round 14 follow-up
My round 14 review raised no C14 comments and returned ACCEPT. No outstanding comments require resolution.
Changes addressing other reviewers are sound except for C15-1 below:
- G14-1, G14-2, F14-3: Resolved. Single-arm draws, exponentiation, count-based (y_a), and diagonal (V_a) are explicit and consistent with stated working approximation.
- G14-3, F14-1: Resolved. Single-interval integration now requires supported constant hazards and exponential-approximation labeling. Counts and person time must cover the specified horizon.
- F14-2: Resolved. PR16 now correctly governs competing-risk reference hazards.
- F14-4: Partly resolved. Kaplan-Meier and rate inputs now share the correct log cumulative-hazard scale, but boundary inputs lack an explicit handling rule.
- F14-note: PR16’s scale of 1, with sensitivity scales of 0.5 and 2, is defensible. Deferring additional Aalen-Johansen horizon methods to a pre-analysis amendment is acceptable while current eligibility and relative-only rules remain in force.
New finding
C15-1. Minor. Undefined inputs in reference-risk pooling.
File: protocol/protocol.qmd:1015, “Absolute effects,” “Hazard based outcomes.”
The new normal model requires finite transformed estimates and sampling variances. An eligible O20 reference arm can have no deaths even when its network meets connectivity requirements. Then (F=0) makes (\log{-\log(1-F)}=-\infty); a zero-event rate likewise gives (\log(\lambda t)=-\infty), with (1/d) undefined. (F=1) also fails. Official survival documentation confirms boundary problems for log-log transformations.
The general relative-only rule does not specify whether one invalid reference arm disables pooling, permits another input, or is omitted. Those choices can change baseline risk and resulting absolute effects.
Insert after “under PR17”:
This normal model requires a finite log cumulative hazard and a finite positive sampling variance from every covering reference arm. Kaplan-Meier inputs require (0<F<1) and valid reported uncertainty. Rate inputs require a positive event count and positive finite person time. If a Kaplan-Meier input fails these requirements, use that arm’s rate only when the stated constant-hazard assumptions and these requirements hold. If any covering arm still lacks a qualifying input, the reference-risk route is unavailable and the relative-only rule applies. Do not omit an arm because its input is invalid or add a continuity correction. These requirements also apply when only one reference arm covers the horizon.
This conservative replacement makes the selected approximation executable without introducing another statistical model.
Verification and registration judgment
All requested documents and tables inspected. Structural checks passed for 637 applicable outcome/population window pairs, all 43 agent/search-term mappings, and registration identifiers and counts across 861 registry-candidate rows. Tables other than amendments, search concepts, and renderer are unchanged from round 14.
Numerical checks confirmed identical constant-hazard results from single-interval and subdivided integration, and agreement between valid Kaplan-Meier-equivalent and rate inputs on the new pooling scale. Search renderer assessment was static; live database recall was not tested. No files edited, created, or deleted.
Methods would be acceptable for registration after C15-1 is resolved and administrative pre-registration conditions are completed. Missing administrative information alone does not drive this verdict.
VERDICT: MINOR REVISION