Methods require major revision for two errors affecting absolute-risk calculations. Administrative pre-registration conditions do not determine this verdict.
| Round 7 comment | Status | Assessment |
|---|---|---|
| C7-1 | Partly resolved | Revised “Competing risks” fallback correctly distinguishes cause specific and subdistribution conversions. However, new “Absolute effects” fallback again excludes time varying integration. See C8-3. |
| C7-2 | Partly resolved | O20 counts and reported hazard ratios now correctly share a primary “not applicable” network. Blanket instruction making cause specific networks primary for all time to event outcomes remains. See C8-4. |
| C7-3 | Resolved | Recorded harms model now explicitly uses offset (\log(t)), consistent with (Q). |
All three proposals were accepted in responses. No declined proposal or alternative rationale requires adjudication.
Required corrections, ordered by severity:
C8-1. Major: new integration rule uses death survival instead of event-free survival. protocol/protocol.qmd:574, “Competing risks,” especially lines 577 and 580.
“All cause survival (any death an event)” is incorrect for nonfatal outcomes, including first heart failure hospitalization and incident diabetes. Integration requires survival free of both target event and competing event. People who experienced a nonfatal target event must leave this risk set. This follows directly from the competing-risk identity in Haque and Cortese, equation 6.
Illustrative calculation: constant nonfatal target hazard 0.2/year and independent death hazard 0.1/year give five-year cumulative incidence 0.518. Integrating target hazard against death-only survival gives 0.787.
Replace the integration specification with:
For arm (a), let (h_{1a}(u)) and (h_{2a}(u)) denote the cause specific hazards of the first target event and competing event. Compute [ S_a(u)=\exp!\left[-\int_0^u{h_{1a}(v)+h_{2a}(v)},dv\right], \qquad F_{1a}(t)=\int_0^t S_a(u)h_{1a}(u),du. ] Here (S_a) is survival free of both events. For the reference arm, use the reference hazards. For the treated arm, multiply the target hazard by its cause specific hazard ratio and the competing hazard by its corresponding hazard ratio when available; otherwise retain the stated assumption of no treatment effect on the competing hazard.
C8-2. Major: new person-time fallback permits invalid denominators for cause specific rates. protocol/protocol.qmd:416, “Data items,” applied at protocol/protocol.qmd:563, “Competing risks.”
Number randomized multiplied by general mean follow-up can include observation after a first nonfatal target event. Such time cannot enter either first-event cause specific rate denominator. Merely flagging median follow-up also does not establish valid person-time. Required denominator is time under observation while free of both events, as specified in SAVVY’s incidence-density definition.
Example: 100 participants followed for one year, with 50 first target events at six months, contribute 75 person-years at risk, although general follow-up totals 100 person-years.
Replace the person-years parenthesis with:
For cause specific rates, extract total person-time until the first target event, competing event, censoring, or analysis cutoff, whichever occurs first. Event counts and person-time must cover the same interval. Number randomized multiplied by mean follow-up is permitted only when that mean represents this exact time at risk across all randomized participants. General study follow-up, treatment exposure time, and median follow-up are not substitutes. If valid person-time cannot be obtained, do not estimate these rates by division; use another supported conversion or report relative effects only.
C8-3. Minor: new reference-risk hierarchy reinstates the conversion conflict addressed by C7-1. protocol/protocol.qmd:902, “Absolute effects.”
Revised hierarchy permits reported Aalen-Johansen incidence, then exponential approximation, then relative effects only. It omits integration from available time varying hazards, explicitly permitted in “Competing risks.” Consequently, identical available data lead to conflicting instructions.
Replace that hierarchy with:
For outcomes with competing events, obtain reference cumulative incidence at the analyzed horizon from reported Aalen-Johansen estimates or, when these are unavailable, by integrating the required time varying reference hazards using the event-free survival formula in @sec-outcomes. If neither is available, use the exponential competing risks approximation only when both required rates are available and constant hazards are supported. Otherwise apply the relative only rule. A single target-event rate is never sufficient. Cause specific treatment conversions additionally require the hazards specified in @sec-outcomes; reference cumulative incidence alone is insufficient.
C8-4. Minor: O20 exception remains absent from the general primary-network rule. protocol/protocol.qmd:545, “Competing risks,” compared with lines 472 to 474 and 522 to 526.
Specific O20 routing is now correct, but general rule still makes cause specific networks primary for every time to event outcome. Conclusion fallback also refers to an “other hazard type” that cannot exist for O20. This is the unresolved portion of C7-2.
Replace the primary-network sentence with:
For time to event outcomes with competing events, the cause specific network is primary. For O20, the not applicable network is primary; reported hazard ratios and eligible complete-status counts are combined under proportional hazards, with one input per trial. For binary harms analyzed on the hazard scale, the subdistribution network is primary.
Add to the conclusions rule:
For O20, use the not applicable network when the connectivity rule is met; otherwise report descriptively. No alternative hazard type network is constructed.
All requested files inspected. Read-only checks found no identifier, applicable-window, reference-SD, Turner-mapping, nonrandomized-harms coverage, search-alias, or registry-count inconsistencies. Tables and search files were unchanged from version 0.7.0 except the amendment log. Search renderer inspected statically; R execution was blocked by temporary-file creation restrictions, and live search recall was not assessed. No files edited, created, or deleted by this review.
VERDICT: MAJOR REVISION