REF-1498 · Zhu et al., 2026 · Research Square — preprint, version 1, posted 11 June 2026; not peer reviewed · Paper
How much slower is the run after a HYROX station?
PREPRINT — not peer reviewed: in 13,618 mostly Pro records the next run slowed most after lunges (median +20 s) and sled push (+14 s); after the row, −4 s. Station and position mixed.
Main caveat: PREPRINT — not peer reviewed. Version 1 on Research Square; Crossref showed no published version on 2026-10-30.
Zhu and colleagues took public split results from 132 HYROX races in 2024-2025, reconstructed 13,618 athlete-event records into 95,326 station-to-run observations, and took the change in time from the run before a station to the run after it. Wall balls have no next run, so seven transitions were studied.
The largest median changes followed sandbag lunges (run 8, +20 s) and the sled push (run 3, +14 s); the row was the only one with a negative median (run 6, −4 s). Adjusted for previous run, percentile, sex, age group, division and event, lunges stayed highest (+25.4 s) and sled push next (+10.3 s).
A residual model that expected a straight-line slowing with race position then showed slower-than-expected running after lunges (+9.4 s) and sled push (+5.5 s) and faster-than-expected after the row (−9.3 s) and carry (−6.6 s). The directions held by sex, year and in Pro-only records. The authors read this as a task-specific pattern and do not claim station effects.
Our reading: a large, honest description with three things to keep in view. Each run is measured against the previous run, which itself followed a station, so the row's negative number may partly say that run 5 was slowed by burpees. Run 2 is compared with run 1, whose length can differ. And run 8 is also the last run.
It gives a first size for the slowdown after stations; it does not show which station costs what.
Protocol
Design: retrospective analysis of publicly available official split results; the source site is not named. Ethics approval stated as not required.
Sample: n = 13618 athlete-event result records from 132 races, 2024 (5,115, 37.6%) and 2025 (8,503, 62.4%). Men 7,106 (52.2%), women 6,512 (47.8%). Division: Pro 13,138 (96.5%), Open 397 (2.9%), Elite 83 (0.6%). How these records were drawn from the far larger field is not explained.
Repeated athletes: not handled. The unit is the record; repeated participations count separately and unique athletes are not reported. Each record gives seven observations: 95,326 in all.
Quality control: 13,942 parsed records; complete splits; every 1-km run within 120-900 s (13,740 left); missing grouping variables removed (13,618, 97.7%).
Outcome: Δrun = run after the station minus run before it (positive = slower); also as a percentage. Roxzone is not mentioned, and whether a run split includes transit is not stated.
Tiers: top 10%, 11-50%, 51-75%, bottom 25% of finish time within event, sex and division.
Adjusted model: linear model of Δrun on station-position pair, previous run, percentile, sex, age group, division and event fixed effects; sled pull as reference. Observations treated as nested but no clustered errors reported; p values not adjusted for multiple comparisons.
Residual model: run after the station predicted from previous run, linear position, percentile, sex, age group, division and event, with no station term; residual = observed minus expected.
Sensitivity: by sex and year, Pro only, relaxed 120-1200 s range, percentage scale, event-level ranks in 129 events with at least 30 records.
Funding and interests: no specific grant; no competing interests declared. Generative AI tools were used for language editing and manuscript preparation support, as printed. One author (KL) is named for data, analyses, figures and drafting. Overlapping authors with REF-1502. Derived summaries in supplementary files (not read); de-identified data and code on request.
Key Results
| Transition | Run before, median s | Run after, median s | Median Δrun, s (IQR) | Median Δ% | Adjusted mean Δrun, s | Contrast vs sled pull, s (95% CI) | Median residual, s (95% CI) |
|---|---|---|---|---|---|---|---|
| SkiErg → run 2 | 235.5 | 245 | +5 (77) | +1.93 | −0.142 | −3.99 (−4.8 to −3.1) | −4.01 (−4.4 to −3.6) |
| Sled push → run 3 | 245 | 261 | +14 (13) | +6 | +10.3 | +6.49 (5.9 to 7.1) | +5.52 (5.2 to 5.8) |
| Sled pull → run 4 | 261 | 262 | +2 (9) | +0.686 | +3.84 | reference | −2.2 (−2.5 to −1.9) |
| Burpee broad jump → run 5 | 262 | 267 | +5 (9) | +1.77 | +8.19 | +4.35 (3.9 to 4.8) | −0.533 (−0.7 to −0.3) |
| Row → run 6 | 267 | 263 | −4 (9) | −1.46 | +1.47 | −2.37 (−2.8 to −1.9) | −9.26 (−9.5 to −9.0) |
| Farmers carry → run 7 | 263 | 266 | +3 (9) | +1.2 | +6.97 | +3.12 (2.7 to 3.5) | −6.65 (−6.9 to −6.4) |
| Sandbag lunges → run 8 | 266 | 284 | +20 (27) | +7.59 | +25.4 | +21.6 (20.9 to 22.2) | +9.42 (9.0 to 9.9) |
Tiers (median Δrun, top 10% vs bottom 25%): sled push +11 vs +17 s; lunges +16 vs +24 s; SkiErg +2 vs +9 s; row −3 vs −4 s; burpees, sled pull and carry about 1 s apart.
Subgroups (sled push, row, lunges): men +14, −4, +20; women +15, −4, +20; 2024 +16, −4, +22; 2025 +14, −4, +19; Pro only +15, −4, +20; relaxed range +14.5, −4, +20.
Event level (129 events with at least 30 records): sled push in the top two in 78.3% (101/129); lunges the largest in 43.4% (56/129); row the lowest in 45.0% (58/129). So no ranking holds everywhere.
Dispersion: residual quartiles span zero for all but the row (Q1 −17.7, Q3 −0.951): for the lunges Q1 −5.02 and Q3 25.0. Many individuals run faster or slower than the median direction.
Mechanisms
Measured: split times only. No heart rate, lactate, muscle oxygenation, gait or RPE.
Authors' interpretation: sled push and lunges load the legs heavily, so residual fatigue and interference are plausible; the row is seated with a large upper-body part and the carry keeps a running-like rhythm, so they may disturb the next run less. They call these targets for mechanistic study, and state that a negative residual does not mean a station helps the next run.
Inference: the two analyses disagree on the carry. With position as categories the carry is above the sled pull by +3.12 s; with position as a straight line it is 6.65 s faster than expected. The row and carry residuals depend on that line, which the last run pulls upward. The straight-line expectation is a modelling choice.
Inference: Δrun measures against a run that followed its own station. Run 5 follows burpees and run 6 follows the row, so −4 s after the row may partly be recovery from a slowed run 5; the design cannot say.
Inference: the SkiErg transition's IQR is 77 s against 9 to 27 s for the others. The first run differs in length between venues (REF-1519: often about 800 m; REF-1496: first and last may differ), so that row is unreliable, and the last run's possibly different length may inflate the lunges number. The authors do not address run length.
Practical Application
How much slower is the run after a HYROX station? PREPRINT — not peer reviewed: in 13,618 mostly Pro records the next run slowed most after lunges (median +20 s) and sled push (+14 s); after the row, −4 s. Station and position mixed.
Reviewing a race: the authors propose the first run after each station as a checkpoint in race review. Comparing your own steps with these medians is, as our suggestion, reasonable as description for a Pro athlete, with the first run left out; a larger step does not diagnose why, and it does not fix a training method (the authors say the findings do not prescribe one).
Do not read it as "row helps" or "lunges are the station to train": neither follows. The row is at one fixed point, the lunges at another, and the top of the ranking changes by event.
Not from this paper; the base's own practice: the station-to-run pairs in SYNTH-71 stay practice. It fits sled push and lunges being among them but tests no training, so it does not support choosing them. It says nothing about wall balls, which have no next run.
Limitations
PREPRINT — not peer reviewed. Version 1 on Research Square; Crossref showed no published version on 2026-10-30. kb-watch checks monthly for one (relation is-preprint-of). Level capped at C.
Station and position confounded. Each station has one place; the authors say so and call results pair differences.
Records, not athletes. Repeated participations count separately; unique athletes are not given. Seven observations per record are not independent; the authors offer clustered errors only if a journal asks. p values unadjusted.
Selection. Finishers with complete splits; how 13,942 records were drawn from a far larger field is not explained. Nearly all Pro, so Open is untested: 397 records, and no Open-only analysis.
Model dependence. The expected-run model assumes a linear position effect; that the faster-than-expected finding on the row and carry rests on it is our reading, not the authors' statement. The abstract's wording that the row "remained lower" holds, but the adjusted model's lowest is the SkiErg transition (−0.142 s), not the row (+1.47 s).
No physiology, no course. Run length by venue is not addressed; no training data.
Supplementary tables not read.
How this entry was checked and changed
Why this level
read from the full text (Research Square PDF, CC BY 4.0). A large observational sample would earn B+ by design; the preprint rule sets one step below and caps at C. Records, not unique athletes, and the seven transitions in a record are not independent
Full text (checked 2026-10-30)
Main analytic sample: 13,618 athlete-event records, 132 races, 95,326 transition-level observations; repeated participations treated as separate records (full text)
Pro 13,138 records (96.5%), Open 397 (2.9%), Elite 83 (0.6%); men 7,106 (52.2%), women 6,512 (47.8%) (full text)
Median Δrun: lunges +20 s, sled push +14 s, burpees and SkiErg +5 s, carry +3 s, sled pull +2 s, row −4 s (full text)
Adjusted mean Δrun: lunges +25.4 s, sled push +10.3 s, burpees +8.19 s, carry +6.97 s, sled pull +3.84 s, row +1.47 s, SkiErg −0.142 s (full text)
Median residual: lunges +9.42 s, sled push +5.52 s, row −9.26 s, carry −6.65 s, SkiErg −4.01 s, sled pull −2.2 s, burpees −0.533 s (full text)
Event level, 129 events: sled push top two 78.3%, row lowest 45.0%, lunges largest 43.4% (full text)
No specific grant; no competing interests declared (full text)