Reproducibility checklist for censored gaze latency

Purpose

Use this checklist before preregistration, manuscript submission, or freezing a gaze-latency analysis. The aim is to make censoring, event construction, repeated-participant structure, diagnostics, sensitivity analysis, and software provenance auditable.

Design and event definition

Record the target event, target AOI, time origin, observation-window rule, event detector, minimum fixation-duration rule where relevant, and event-to-trial join key. Verify that event times are non-negative and never exceed the censoring limit. Review time-zero events explicitly rather than shifting them without a pre-specified measurement-resolution rule.

Censoring and gaze quality

Report observed events, right-censored trials, review-required rows, and censoring percentage overall and by condition. A valid censored trial requires a complete usable observation window. Incomplete or unresolved trials and gaze-quality failures are not ordinary censoring.

Repeated observations and estimand

State whether the reported effect is a marginal Cox hazard ratio, a conditional frailty-Cox hazard ratio, or an AFT time ratio. Name the participant identifier, repeated-observation structure, and AFT family explicitly. Do not use AIC/BIC to rank ordinary Cox partial likelihood, coxme penalized frailty likelihood, and AFT full likelihood as if they shared a likelihood basis.

Diagnostics

Inspect event sparsity, extreme censoring, convergence, non-finite estimates, and proportional-hazards diagnostics for Cox models. For Gaussian participant frailty, report a corresponding marginal Cox PH diagnostic instead of applying cox.zph() to the coxme object.

Sensitivity analysis

Pre-specify defensible alternatives for AOI geometry, detector, minimum fixation duration, time origin, quality threshold, clustered versus frailty Cox when scientifically relevant, Weibull versus log-normal AFT, and explicit review/exclusion rules. Rebuild the survival table whenever the event/AOI specification changes.

Reporting bundle

Archive participant/trial counts; event/censor/review counts; event/AOI/time-origin definitions; observation-window and quality rules; detector settings; model formula; repeated-observation structure; effect estimates and confidence intervals; diagnostics; sensitivity results; source/preprocessing provenance; and eyeprocess, survival, and coxme versions where relevant.

Final freeze questions

1. Is every non-event trial demonstrably censored rather than missing?
2. Is the event timestamp construction reproducible?
3. Is the estimand clearly identified as marginal, conditional/frailty, or parametric-time based?
4. Are AOI/detector/quality sensitivity branches documented?
5. Are review rows and failed diagnostics still visible in the audit trail?

API map

Preparation and validation: prepare_gaze_survival_data(), validate_gaze_survival_data(), summarise_gaze_censoring().

Models and diagnostics: fit_gaze_cox_model(), fit_gaze_mixed_cox_model(), fit_gaze_aft_model(), check_gaze_proportional_hazards(), compare_gaze_survival_models().

Sensitivity and reporting: compare_gaze_survival_specifications(), tidy_gaze_survival_model(), report_gaze_survival_model().