Skip to content

Reporting and limitations

Minimum reporting set

Report the participant count, trial count, observed-event count, censored-trial count and percentage, review-required rows, target event, target AOI, time origin, observation-window definition, event detector, fixation-duration rule where relevant, gaze-quality rule, repeated-participant structure, estimator family, effect estimate with 95% CI, diagnostics, and sensitivity branches.

report_gaze_survival_model() returns these fields together with model and software provenance. summarise_gaze_censoring() should normally be reported by experimental condition as well as overall.

Interpretation templates

For Cox models, an exponentiated coefficient is a hazard ratio. It describes the relative instantaneous target-event rate among trials still at risk; it is not a ratio of mean TTFF values.

For AFT models, an exponentiated coefficient is a time ratio under the fitted distribution. Values above one indicate multiplicatively longer event times and values below one indicate shorter event times, conditional on the model specification.

Example manuscript wording

Gaze latency was analysed as a right-censored time-to-event outcome. Trials in which the disclosure AOI was not fixated before the valid trial window ended were retained as censored observations; trials with unresolved gaze validity were not reclassified as censoring. We fitted a Cox proportional-hazards model with participant-clustered standard errors and reported hazard ratios with 95% confidence intervals. Proportional-hazards diagnostics and sensitivity analyses across pre-specified AOI/event definitions were examined. A Weibull AFT specification was reported as an alternative estimand and interpreted using time ratios.

Adapt this text to the actual detector, AOI, quality threshold, time origin, and repeated-effects structure used in the study.

Important limitations

  • A missing event is not evidence of censoring without a known usable risk window.
  • Censoring mechanisms should be scientifically defensible; informative dropout can bias standard survival estimators.
  • Very sparse events can make Cox and AFT estimates unstable even when software converges.
  • Time-zero events require a pre-specified rule grounded in sampling resolution and the definition of trial onset.
  • Python cluster-robust Cox and R latent frailty Cox are different estimands and uncertainty structures; numerical identity is not claimed.
  • Competing risks require a separate competing-risks framework rather than interpreting 1 - KM as a cause-specific cumulative incidence function.