What Do Two Weeks of Six-Hour Nights Do to Performance?

The abstract of Van Dongen et al. (2003) reports that chronic restriction to 6 hours or less per night produced cognitive deficits equivalent to up to 2 nights of total sleep deprivation. That sentence aggregates two conditions, and the results section separates them. After 14 nights, the 6-hour group reached the level of one night of total sleep loss on vigilance lapses and on the digit symbol substitution measure of working memory. Only the 4-hour group reached the two-night level, and even there only on those two tasks. Cognitive throughput on the serial addition and subtraction task reached the one-night level in the 4-hour condition.

So the defensible claim for a six-hour pattern is one night without sleep, not two. The dose-response itself was clear in the rate of change across the 14 days: vigilance lapses differed by condition (F2,30 = 3.67, P = 0.037), as did digit symbol substitution (F2,30 = 5.33, P = 0.010) and serial addition and subtraction (F2,30 = 6.19, P = 0.006). Lapses were defined as reaction times longer than 500 milliseconds, counted in a 10-minute test bout that was repeated every 2 hours through the waking day.

The 8-hour control condition is what makes the rest interpretable. Those participants showed only minor, non-significant increases in lapses over the two weeks, with an estimated mean slope not different from zero (t30 = 0.77, P = 0.45), and they produced normal practice curves on the other two tasks. The restricted groups were not getting worse at testing. They were getting worse at staying alert, and the curvature estimates indicate the impairment accumulated close to linearly across days rather than settling at a new level.

This is where the study disagrees with itself in the most operationally useful way. Subjective sleepiness did change across the restricted conditions (Stanford Sleepiness Scale F2,30 = 4.26, P = 0.024; Karolinska Sleepiness Scale F2,30 = 7.76, P = 0.002), but the 6-hour and 4-hour slopes were indistinguishable from each other on both scales (F1,30 = 0.10, P = 0.75 and F1,30 = 1.55, P = 0.22), even though their measured performance was not. People reported similar tiredness at two quite different levels of impairment.

The magnitudes make the point sharper. Three days of total sleep deprivation moved the Stanford scale by more than 2 units, while 14 nights at 6 or 4 hours moved it about 1 unit (Van Dongen et al., 2003). The shape differed too. Under total deprivation the sleepiness curve was close to linear, with a curvature estimate of 0.86 (standard error 0.14), while under chronic restriction it flattened toward a plateau at 0.24 (standard error 0.04). The Karolinska scale reproduced the same split, 0.81 (standard error 0.16) against 0.16 (standard error 0.03).

An all-nighter announces itself. A standing six-hour pattern does not, because the subjective signal saturates while the objective one keeps degrading. Any fatigue control that depends on a worker noticing and escalating their own tiredness is built on the one measure this experiment showed to stop tracking the deficit. The finding describes self-rated sleepiness under a controlled schedule, and it does not establish what workers would report under real workload, caffeine, and social pressure, none of which were present here.

Cumulative sleep loss across 3 days of total deprivation was 23.1 hours (standard deviation 2.6), significantly less than the cumulative loss built up over 14 nights at 4 hours (t20 = 10.58, P < 0.001). Yet the 4-hour group's deficits never exceeded those of the total deprivation group, which produced waking neurobehavioral and sleep delta power responses that were disproportionately large relative to how much sleep it lost. Total sleep debt therefore cannot explain the performance profiles on its own. Time spent awake past a limit tracks them better: lapses in behavioral alertness were near-linearly related to the cumulative duration of wakefulness in excess of a critical value of 15.84 hours (standard error 0.73).

That threshold is the number most often quoted from this paper. Read as the maximum daily wakefulness compatible with stable performance, it implies an average daily sleep need of 8.16 hours (standard error 0.73) to prevent cumulative deficits. The model explained 83.0% of the variance in the vigilance data, and each participant's estimated threshold closely matched their own habitual wake duration (difference 0.1 hours, standard error 0.5, t23 = 0.17, P = 0.86). This is a modeled population average, not a personal target: the between-subject standard deviation of the threshold was 3.58 hours (standard error 1.19).

Sleep physiology did not track the waking damage either. There was no significant progressive change in non-REM delta power across the 14 nights in any restriction condition (all absolute t15 values below 1.61, P above 0.13), while delta power in the first 4 hours of recovery sleep after 88 hours awake reached 172% of baseline (standard error 11%), more than the average under chronic 4-hour restriction (F1,18 = 5.40, P = 0.032). REM latency did move, falling about 2.2 minutes per day in the 4-hour condition (t16 = -7.55, P < 0.001). The practical reading is that the standard physiological marker of sleep pressure did not register what the performance tests were recording, so deeper sleep is not evidence that a restricted schedule has been absorbed.

One further result complicates the intuition that habitual short sleepers are the vulnerable ones. Controlling for condition, participants who had slept longer before the study tended to deteriorate faster on the vigilance task (r32 = 0.29, P = 0.048). That is a single modest correlation in a small sample, and it is a reason to measure individuals rather than a basis for screening anybody out.

Every number here comes from a controlled inpatient laboratory. Participants were healthy adults aged 21 to 38, screened free of medical, psychiatric, and sleep disorders, and kept off caffeine, alcohol, tobacco, and medication for 2 weeks before and throughout, verified by blood and urine screens. Wake time was enforced at a fixed hour and there was no work to do. Only 6 of the 48 participants were female, and the total deprivation group was entirely male. The authors state directly that they do not know whether the same cumulative impairment occurs in habitually short versus long sleepers, in people sleeping at other circadian times such as night workers, or in women compared with men. Carrying these figures onto a shift-working population is an inference beyond the paper.

The cells are small, at 9 to 13 people per condition, and the 8-hour control was the smallest at 9. Baseline differences were present: the 8-hour group was significantly younger than the 6-hour group (F3,44 = 3.16, P = 0.034, Bonferroni P = 0.022), and the 6-hour group performed slightly worse on digit symbol substitution at baseline (F3,44 = 2.83, P = 0.049). Habitual sleep before entry did not differ across conditions (F3,44 = 0.38, P = 0.77), which is the comparison that matters most for the dose-response claim. The authors also note that 8 hours in bed at baseline may itself have been insufficient, which would make the reported deficits underestimates rather than exaggerations.

Two limits deserve stating plainly. Recovery is not established here: only 3 recovery days followed the 14 restricted nights, and this paper does not report whether performance returned to baseline, so it cannot support any claim about catching up over a weekend. A correction to the paper was also published in the journal in 2004, so any single figure reproduced from it should be read as subject to that erratum.

None of that makes the core result shaky. The paper dates from 2003 and remains the field's landmark chronic-restriction experiment. A dose-response experiment run in another laboratory and published the same year in the Journal of Sleep Research put 66 volunteers on 3, 5, 7 or 9 hours of daily time in bed for 7 days, followed by 3 recovery days at 8 hours (Belenky et al., 2003). The two experiments agree on the part that matters most for a roster, which is that restriction leaves a performance deficit rather than one the body quietly absorbs, and Belenky et al. (2003) supplies the recovery evidence Van Dongen leaves open: the 7-hour and 5-hour groups had not returned to baseline by the end of their three recovery days.

Where the two part company is the shape of the decline, and that disagreement is unresolved. Van Dongen et al. (2003) has impairment accumulating across all 14 nights. Belenky et al. (2003) has vigilance speed in the 7-hour and 5-hour groups declining first and then appearing to stabilize at a reduced level, with lapses raised only in the 5-hour group and the 9-hour group staying at baseline, while only the 3-hour group showed speed declining and lapses rising steadily across the 7 days. Read onto a roster, one shape says a restricted pattern keeps getting worse and the other says it settles at a degraded level that recovery nights did not lift, and neither makes six hours look sustainable. What has moved in the two decades since is the literature on how long recovery takes and on why people differ, not the dose-response finding itself.