Structural screening and event-aware analysis of official HYROX split data: a retrospective study of top-ranked results

Abstract Background HYROX combines eight 1-km running segments with eight functional work stations. Official split records enable observational analysis of repeated running patterns but may differ across events in timing architecture and may include repeated results from the same athlete. We evaluated split-derived pacing metrics while explicitly addressing structural comparability, event structure, and repeated-athlete dependence. Methods We retrospectively analyzed 2,000 publicly accessible top-ranked result records archived from four official HYROX ranking groups on March 29, 2026. Author-defined structural heuristics retained 1,909 records. The primary Model C used target-by-event fixed effects and two-way cluster-robust uncertainty by event cell and athlete. Running-split variability and late-race running fade were the two primary pacing estimates and were Holm-adjusted. Sensitivity analyses reproduced the best-ranked one-record-per-athlete Model D, applied an outcome-independent deterministic de-duplication rule, repeated random one-result-per-athlete selection 500 times, varied the handling of first-run anomaly event cells, and estimated event-adjusted within-athlete repeatability. A stratified audit independently compared archived result pages with the analytic data. Results Model C included 1,121 records from 682 target-specific athlete identifiers across 116 event cells; 212 athletes contributed repeated records. With event-cell and athlete two-way clustered uncertainty, neither running-split variability (β = 0.0105 per + 0.10, 95% confidence interval (CI) -0.0049 to 0.0260; Holm-adjusted p = 0.364) nor late-race running fade (β = 0.00449 per + 60 s, 95% CI -0.00271 to 0.01170; Holm-adjusted p = 0.364) was clearly associated with official time. The best-ranked-result rule yielded a larger positive fade estimate (β = 0.00929, 95% CI 0.00261 to 0.01597), whereas the outcome-independent deterministic rule yielded a smaller, imprecise positive estimate (β = 0.00355, 95% CI -0.00383 to 0.01093). Across 500 random selections, 99.8% of fade coefficients were positive, although only 9.0% of the corresponding 95% CIs excluded zero. Event-adjusted repeatability fits were at the non-negative boundary and are reported as approximately zero. At the page level, 48/48 archived pages agreed exactly with the analytic dataset (100%; exact 95% binomial CI 92.6%-100%); all 864 fields also agreed exactly and are reported descriptively because fields within pages are clustered. Conclusions The primary event-aware result-level analysis did not show clear associations between the two split-derived pacing metrics and official time. The larger fade estimate under the best-ranked-result rule contrasted with smaller, imprecise positive estimates under outcome-independent selection. Boundary estimates were approximately zero and did not permit strong inference regarding a stable athlete-specific component. The findings highlight the value of accounting for event structure, repeated athletes, and selection rules when analyzing official HYROX split data.