Round 16 comments resolved. Two minor specification defects remain in new wording. No new major finding.
Round 16 follow-up
- C16-0: Resolved. Correction in
protocol/peer-review/round-16/response.mdaccurately limits new count likelihoods to noncompeting hazard outcomes, currently O20, and fixed-time binomial outcomes. Competing-risk routes retain their zero-count restrictions. Keeping their working normal likelihood is an explicit modeling choice, not a mathematical necessity. Deferring extensions remains acceptable because eligibility restrictions and relative-only fallback are explicit. - C16-1: Resolved. Absolute effects now defines usable arms, excludes unusable arms from pooling and scenarios, distinguishes zero, one, and multiple usable arms, and restores known-status requirements. Independent arm posterior draws make scenario uncertainty executable.
- C16-2: Resolved. PR02, row 3 distinguishes annual trial baseline from mean log cumulative hazard at analyzed horizon and identifies logit scale for fixed-time reference risks. Retaining the same numerical prior on these explicitly different parameters is acceptable.
New findings, both minor
C17-1. Transformed confidence-limit rule admits undefined standard errors and assumes 95 percent coverage.
Location: protocol/protocol.qmd, “Absolute effects,” lines 1026 to 1031.
Requiring only (0<F<1) does not ensure finite transformed confidence limits. For example, (F=0.01) with interval ([0,0.04]) passes that requirement, but transforming its lower limit gives negative infinity. This can occur with bounded survival confidence intervals; multiple interval constructions and confidence levels are supported by standard survival software. R survival documentation.
Dividing by 1.96 also presumes 95 percent coverage. For limits 0.08 and 0.125, the specified standard error is 0.1201, whereas a 90 percent interval implies 0.1431.
Replace the uncertainty clause with:
Require (0<F<1) and a finite, positive standard error on (g(F)=\log{-\log(1-F)}). For a reported two-sided (100(1-\alpha)) percent confidence interval ((L,U)), require a known confidence level and (0<L<U<1), and use ([g(U)-g(L)]/[2z_{1-\alpha/2}]). Alternatively, derive the standard error from reported Greenwood variance by the delta method. If neither supplies qualifying uncertainty, proceed to the Poisson and then binomial input rules. An arm without any qualifying input is listed and excluded.
C17-2. New one-arm restatement extends log-hazard sampling to an Aalen-Johansen-only route.
Location: protocol/protocol.qmd, “Absolute effects,” lines 1063 to 1066.
The new sentence applies to every outcome with competing events. However, the preceding subdistribution route permits direct use of reported Aalen-Johansen cumulative incidence. That input need not identify both cause-specific hazards, so (y_a) and (V_a) may not exist. Restrict the restatement to routes that define those quantities.
Replace with:
For the piecewise integration and exponential approximation routes, usable arms and scenarios follow @sec-outcomes. With one covering arm, (\tau_k) is not estimated and each draw of the reference log hazards is from (\mathrm{N}(y_a,V_a)). This log-hazard sampling rule does not apply to direct use of a reported Aalen-Johansen cumulative incidence in the subdistribution route.
This correction preserves the previously accepted deferral of additional Aalen-Johansen horizon methods.
Verification and registration judgment
All requested files inspected. Read-only checks found no identifier or mapping errors across the authoritative tables, 637 applicable outcome/population window pairs, 43 agent/search-term mappings, or 861 registry-candidate rows. Only protocol.qmd, PR02, and amendments.csv changed among requested files since version 0.16.0. Count-likelihood scaling and revised interval-boundary matching passed numerical checks. Search renderer assessment was static; live database recall was not tested. Working tree remains clean.
These two narrow corrections would make methods acceptable for registration once administrative pre-registration conditions are completed. Missing administrative information alone does not drive this verdict.
VERDICT: MINOR REVISION