Methods require major revision because newly specified curve uncertainty can produce incorrect pooling weights and absolute effect intervals. Administrative pre-registration conditions do not determine this verdict.
Round 10 comments
- C10-1: Resolved. Revised inversion in protocol/protocol.qmd:602 correctly recovers constant cause specific hazards from compatible cumulative incidences. Numerical verification recovered hazards 0.4 and 0.2 and risk reduction 135.95 per 1,000 under target hazard ratio 0.5. Added Kaplan-Meier inversion is also appropriate under stated constancy. Newly introduced variance rule requires separate correction below.
- C10-2: Resolved. Competing risks, lines 590–595 explicitly excludes mortality curves containing deaths after nonfatal target events. Paired Kaplan-Meier curves with mutually exclusive first events constitute a defensible additional route for identifying cause specific hazards.
- C10-3: Resolved. Competing risks, lines 649–657 restricts failure to exponential approximation and preserves other eligible routes, including reported single-horizon Aalen-Johansen incidence.
No C10 proposal was declined or partly adopted. Responses and adopted boundaries are justified.
New findings, ordered by severity
C11-1. Major: curve-derived event counts and variances are not identified by specified inputs.
Location: protocol/protocol.qmd:609, “Competing risks,” piecewise integration, lines 609–616.
New rule reconstructs event counts from curve increments and interval-start numbers at risk, then assigns variance (1/d_k). Censoring within intervals prevents this reconstruction without additional information or assumptions.
Verified counterexample: both histories begin with 100 participants and end with 40:
- History A: 50 censored before events, then five target events and five competing events.
- History B: ten target events and ten competing events, then 40 censored.
With corresponding event times aligned, both produce identical paired Aalen-Johansen curves, ending at (F_1=F_2=0.10), and identical published endpoint risk counts. Actual event counts differ twofold. Specified inputs therefore cannot determine (d_k). Established reconstruction methods explicitly address censoring rather than deriving counts from interval increments alone. Guyot et al., Methods.
Moreover, even known event counts do not generally justify (1/d_k) for curve-inverted log hazards. For a Kaplan-Meier curve beginning with 100 participants and observing 90 target events before any censoring, Greenwood plus delta method gives [ \operatorname{Var}{\log(-\log \widehat S)} \approx \frac{90/(100\times10)}{(\log 0.10)^2} =0.01698, ] compared with (1/90=0.01111). Subsequent competing events can satisfy the positive competing-count requirement without changing this target-curve calculation.
Marginal variances also leave unspecified covariance needed by the existing “valid joint uncertainty” requirement. These defects can affect pooled reference hazards, absolute effect intervals, and certainty ratings.
Replace variance and associated eligibility sentences with:
For reported first-event counts and valid person-time, use the Poisson information approximation (\operatorname{Var}(\log h_k)\approx1/d_k). For curve inversion, require a reported or author-supplied joint covariance matrix (\Sigma) for all curve ordinates used across causes and interval endpoints. Let (g) map those ordinates to the vector of log hazards using the stated inversion formulas. Estimate its covariance as (J\Sigma J^\mathsf{T}), where (J) is the Jacobian of (g), and preserve this dependence during pooling and uncertainty propagation. Do not infer event counts from curve increments and interval-start risk counts. Subdivisions of an original interval reuse the same hazard draw. If any covering arm lacks qualifying inputs, positive finite hazards, or valid joint uncertainty, piecewise integration is unavailable; retain another eligible route.
This provides an executable conservative correction. Retaining reconstructed counts would instead require an explicit reconstruction method, censoring assumptions, and corresponding joint uncertainty calculation.
C11-2. Minor: revised scenario wording unnecessarily requires hazards for the reported-incidence route.
Location: protocol/protocol.qmd:950, “Absolute effects,” lines 950–953.
New parenthetical requires each covering arm’s cumulative incidence “from its own hazards.” However, lines 630–638 and 972–974 permit reported Aalen-Johansen incidence without recoverable cause specific hazards. Such trials support subdistribution conversion but cannot satisfy the new scenario instruction.
Replace parenthetical with:
For outcomes with competing events, scenario quantiles use each covering arm’s cumulative incidence at the analyzed horizon: its reported Aalen-Johansen estimate when that route is used, otherwise the incidence calculated from its own qualifying hazards by the route in use.
All requested files inspected. Programmatic checks found no identifier, applicable-window, reference-SD, Turner-mapping, nonrandomized-harms coverage, registry-count, or search-alias discrepancies across 637 outcome/population window pairs, 861 registry-candidate rows, and 43 agent mappings. Search renderer reviewed statically; R execution was blocked by temporary-file restrictions, and live search recall was not assessed. No files edited, created, or deleted.
VERDICT: MAJOR REVISION