A delivery rider can have an unusually long workday and a shorter night before the next shift. Following the same person across days lets researchers look for that sequence without confusing it with all the stable differences between riders. In this study, an extra working hour was linked with a steeper decline in sleep in the model’s longer-hours segment. The data left the location of any turning point uncertain.
- Fifty-five motorcycle delivery riders in South Korea wore wrist activity monitors and rated their fatigue before and after shifts.
- The within-person model estimated 3.59 and 11.52 fewer sleep minutes per additional working hour in its two segments.
- The authors’ breakpoint estimate was 13 hours, with an interval spanning 10 to 14 hours. Another model remained compatible with a smooth trend.
- Paper date
- 21 November 2026
- Participants
- 55 platform-based delivery riders
- Sleep measurement
- Wrist actigraphy
- Adjusted model
- 550 day-pairs
Original publication: 21 Nov 2026 · The date above refers to this brief.
How much shorter was sleep per additional working hour?
fewer minutes of sleep
The three measurements in a day-pair
Workday
Record work hours and workload.
Night after that workday
Estimate minutes of sleep from the wrist monitor.
Following workday
Collect the end-of-shift fatigue rating.
Start with the order of events
Work duration was recorded for one day. The wrist monitor estimated how long the rider slept afterwards, and the rider rated fatigue at the end of the next shift. The model compared longer and shorter days for the same rider and accounted for measured factors including rest time and delivery intensity. Workloads and circumstances can still change together from one day to the next, which leaves room for explanations beyond the modelled sequence.
Source 1 ↗An extra hour had a different estimate in each segment
With a bend placed at 12 working hours, the model estimated 3.59 minutes less sleep per additional hour in the shorter segment and 11.52 minutes less in the longer segment. The latter point estimate is 7.93 minutes steeper per working hour, or about 3.21 times the magnitude of the former. These are associations within the fitted segments. The analysis that tried to locate a bend estimated 13 hours, and its 95% interval spanned 10 to 14 hours. The paper also reports that a more flexible model could not rule out a linear association. These results leave several shapes consistent with the observations.
Source 1 ↗Which days went into each result?
The smoothed trend in Figure 3 uses 722 available day-pairs, while the adjusted mediation model uses 550. That is about 76.2% of the available pairs in the figure. Several pairs can come from one participant. When reading the fitted result, keep the 55 riders, the 722 available day-pairs and the 550 modelled day-pairs separate. The fatigue outcome is a self-rating at the end of the next shift.
Source 1 ↗A laboratory experiment answers a neighbouring question
Van Dongen and colleagues assigned different amounts of time in bed in a laboratory and tested participants’ cognitive performance over time. The delivery study instead observed naturally varying work hours and asked riders how fatigued they felt. Together the studies illustrate why a change in sleep, a change in task performance and a change in crash risk require their own measurements. A follow-up in riders could add objective performance and safety outcomes while keeping the day-by-day order.
Source 1 ↗Source 2 ↗Compare the estimated slopes and check who contributed
Publisher’s accepted full manuscript. Checked participant recruitment, daily measurements, Figure 3, the model sample and Table 2. The publisher may revise this version in production.
01Was sleep shorter after longer workdays?
- What was observed
- Both fitted slopes were negative. The estimate was steeper in the longer-hours segment.
- Where the conclusion stops
- The model used a chosen bend and the separate estimate for its location was imprecise.
Source 1 · Figure 3 and Table 2
02How many observations contributed to the adjusted results?
- What was observed
- The adjusted analysis used 550 pairs, compared with 722 in the descriptive trend.
- Where the conclusion stops
- A pair is a time-linked observation, and a rider can contribute multiple pairs.
Source 1 · Figure 3 caption
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| Riders | 55 people | Reported Participants contributing repeated observations. Source 1 · Methods and Table 1 |
| Day-pairs in descriptive figure | 722 day-pairs | Reported All available pairs for the smoothed trend. Source 1 · Figure 3A |
| Day-pairs in adjusted model | 550 day-pairs | Reported Analytic sample for the mediation model. Source 1 · Figure 3B |
| Adjusted pair count as share of descriptive pairs | 76.2 percent | Calculated Divide 550 by 722 and multiply by 100. Source 1 · Figure 3 caption |
| Shorter segment sleep slope | -3.59 sleep minutes per additional work hour | Reported Rounded from −3.589, 95% CI −6.097 to −1.081. Source 1 · Table 2, path a1 |
| Longer segment sleep slope | -11.52 sleep minutes per additional work hour | Reported 95% CI −17.791 to −5.248. Source 1 · Table 2, path a2 |
| Difference in magnitudes of fitted slopes | 7.93 sleep minutes per additional work hour | Calculated Subtract 3.59 from 11.52. Descriptive comparison of point estimates. Source 1 · Table 2 |
| Ratio of fitted slope magnitudes | 3.21 ratio | Calculated Divide 11.52 by 3.59 and round to two decimals. Source 1 · Table 2 |
| Bend in the main fitted model | 12 working hours | Reported Chosen location for the model spline. Source 1 · Methods, time-lagged modelling |
| Estimated breakpoint | 13 working hours | Reported Separate segmented regression, 95% CI 10 to 14 hours. Source 1 · Results, threshold estimation |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Rider study and sleep restriction experiment Source 2 · Van Dongen et al., Abstract | Working riders and healthy laboratory volunteers | Self-rated fatigue and tested performance | The laboratory observation concerns a different population, outcome and exposure. |
| Movement sensors and self-reports Source 1 · Methods, Measurements | Wrist monitor readings and rider ratings | Estimated sleep and perceived fatigue | Each measurement has a different meaning and source of error. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 22 Nov 2026
First publication. Any calculations in this brief use reported figures. The reading notes identify possible follow-up questions. AI source check; no human editorial or clinical review.
Read the estimates with their time window
| Comparison | Estimate | Statistical context |
|---|---|---|
| Sleep after a workday below 12 hours | 3.59 fewer minutes per additional work hour | 95% CI 1.08 to 6.10 fewer minutes. |
| Sleep after a workday in the longer segment | 11.52 fewer minutes per additional work hour | 95% CI 5.25 to 17.79 fewer minutes. |
| Estimated position of the bend | 13 working hours (95% CI 10 to 14) | Separate estimation step. The main model placed the bend at 12 hours. |
Rounded amounts from Table 2 and the breakpoint analysis.
Your questions, answered
Does the study establish a safe limit of 12 working hours?
Twelve hours is where the investigators placed the model’s bend. A separate estimate for the bend was uncertain, and the outcome was fatigue rather than measured crashes.
Was every available day-pair used in the adjusted model?
The descriptive curve used 722 available pairs. The adjusted mediation analysis used 550 pairs.
What does within-person mean here?
The model uses day-to-day differences for a given rider alongside the data from other riders. That helps account for stable differences between individuals.
Does the study measure how alert a rider actually was on the road?
The main fatigue endpoint came from riders’ own ratings. Laboratory tests can address performance directly in a different setting.
Limits of this interpretation
- The study observed working hours and sleep in one workforce. Day-specific factors could influence both.
- The model uses fewer day-pairs than the descriptive curve. Repeated observations from one rider are linked.
- Wrist actigraphy estimates sleep. The next-day fatigue endpoint was self-reported and the coefficients do not estimate crash risk.
Sources & transparency
- Min, Park and Kim (2026) — Sleep duration as a mediator between daily working hours and next-day fatigue among platform delivery riders
Publisher’s accepted full manuscript. Checked participant recruitment, daily measurements, Figure 3, the model sample and Table 2. The publisher may revise this version in production. · Accessed 22 Nov 2026
DOI: 10.1186/s12889-026-30229-8 - Van Dongen et al. (2003) — The cumulative cost of additional wakefulness
Published abstract of a laboratory experiment. Checked assigned time in bed and the outcome measures. · Accessed 22 Nov 2026
DOI: 10.1093/sleep/26.2.117
Prepared and source-checked with AI for Press-news Team. The source notes identify the material available during preparation. The article received no human editorial or clinical review. The source list supports the reported results and the calculations shown.
Source check: AI source check — primary research and stated source limits
Clinical review: Not applicable to this educational guide
Suggest a correction