Sleep Opportunity and Decision Latency among Republic Fleet Watch Officers on Extended Patrol: A Records Study of 205 Officers, 22–21 BBY

1. Introduction

A watch officer's central task is to recognise what is on the display and commit to a classification before the situation moves on. Until 22 BBY the Republic kept only small patrol forces, and their watch rotations were built for short cruises with predictable rest (Harnell & Quill, 26 BBY). When the fleet became a fighting service in 22 BBY, patrols on station grew longer and the pool of qualified officers had to cover more watches. Medical ratings had already recorded fatigue as a routine complaint in the smaller forces (Carrowen, 23 BBY). Whether it also slows the decisions that officers are trained to make is a separate question, and one that the fleet had no data to answer.

Published work on sleep loss and performance comes almost entirely from civilian settings. Studies of crews on long cargo runs found lapses of vigilance that increased as accumulated sleep fell (Ordane, 28 BBY), and a study of night-watch keepers on orbital stations located most errors in the small hours, independent of how long the keeper had been awake (Dursk, 25 BBY). Neither measured the speed of a decision that requires judgement, and neither examined officers whose watches are bound to a patrol schedule they cannot alter. A further gap concerns the shape of the relation. If performance declines in proportion to each hour of lost sleep, a fixed adjustment to every rotation would recover it. If instead the cost is small until sleep falls below some level and then rises sharply, rotations should be planned around that level.

This paper is written from the vantage of 21 BBY. It draws on fleet records for 22 to 21 BBY, which cover only the first year or so of the fleet's service as a fighting force, and we make no claim about any other period. We used the Watch Decision Drill, an instrument of our own bureau (Coruscant Medical Academy, Bureau of Biological Sciences, 26 BBY), to ask two questions. First, are fewer hours of sleep opportunity in the preceding 72 standard hours (a standard hour being one twenty-fourth of a standard day, the convention of the logs and of this paper) followed by longer latency to a correct classification? Second, is that relation linear, or does it show a threshold below which it steepens? We expected a decline in performance with lost sleep, and we treated the existence and position of a threshold as open.

2. Methods

Sources and access. We used two record series. Republic Fleet Watch Logs supplied the drill entries, each stamped with the display onset and the time of the officer's first correct entry, together with the watch position and the berth-log entries for the patrol. Republic Fleet Medical Records supplied officers' years of service, recorded species and any note by the ship's medical rating that an officer's rest had been interrupted. The series were closed for this study at the end of the third quarter of 21 BBY, so the window runs from 22 BBY to that point. Access was granted by the Republic Health Service under its rules for the use of service records in research. Identifiers were replaced by coded numbers before analysis, and no individual officer, ship or patrol is named here.

Officers. Records for 214 officers standing routine watches on patrols of 40 to 90 standard days were screened. Nine were excluded: six because berth-log entries were missing for at least one drill, and three because fewer than three drills in the standard scenario had been entered. That left 205 officers, each contributing three drills within one patrol, for 615 drills; the first three qualifying drills, at least 10 standard days apart, were taken. Years of service ranged from 2 to 24 (median 7, interquartile range 5 to 9). The officers were predominantly human; nine (4.4%) were recorded as belonging to other species. They were kept in the analysis, but there were too few for separate estimates, no cross-species comparison was attempted, and no assumption was made about the sleep needs of any species other than as they appear in the logs.

Decision latency. The Watch Decision Drill presents a simulated contact on the tactical display with a fixed set of identifying cues, and the officer must classify it as friendly, neutral or hostile. The outcome is the time from display onset to the first correct classification. The drill has three difficulty tiers, set by how many cues are ambiguous. Before the window opened, the drill had been adopted into routine fleet watch training, so every officer had completed it many times. Watch position was night or day according to the ship's standard-time convention entered in the log for the drill's watch. Latency is right-skewed, with a long tail of slow responses, so we analysed its natural logarithm.

Sleep opportunity. The exposure was the number of standard hours in the 72 standard hours before display onset that the berth log showed the officer as off watch and berthed. This is opportunity to sleep, not sleep, and berth entries were made by the officer or the watch clerk (Republic Fleet Medical Service, 22 BBY). Elsewhere in service records the agreement between berth entries and sleep recorded by instrument has been reported as moderate and stronger for long than for short intervals (Ashvane, 24 BBY), a limit we return to below. The medical note of interrupted rest was used only in a sensitivity analysis.

Coding reliability. A second abstractor re-derived sleep opportunity for 120 drills from 40 officers, blind to the first abstraction; agreement was excellent (intraclass correlation, two-way random effects, single measure, 0.98). Drill timestamps were entered to the second, and watch position and years of service were transcribed from the logs and records without recoding, so no reliability statistic applies to them. Drill tier was re-coded for 60 drills from the scenario descriptions; the two codings agreed on 56 of 60 (93.3%; chance-corrected agreement, kappa, 0.90).

Analysis. Log latency was modelled by a linear mixed model with a random intercept for officer, fitted by restricted maximum likelihood, following the treatment of repeated performance tests in Wexley (27 BBY). Fixed effects were sleep opportunity, watch position (night or day), years of service centred at 6 and drill tier centred at 2. Sleep opportunity was first entered linearly and then as a segmented term with one knot. The knot was chosen by profile likelihood over half-hour steps from 10.0 to 24.0 hours under maximum likelihood, and then the segmented model was refitted by restricted maximum likelihood at the chosen knot. Because the knot is selected from the data, the test of a change in slope is conditional on it and is optimistic; problems of inference after selecting a breakpoint are treated in Toller (29 BBY). We report the range of knots within 1.92 log-likelihood units of the maximum, and compare the two models by likelihood ratio and by an information criterion (AIC) counting the knot as a parameter. Coefficients are on the log scale, and are converted to percentage changes in latency; conversions use unrounded coefficients and may differ in the last digit from a conversion of the rounded values shown. Confidence intervals are normal-approximation intervals; p-values use the normal approximation, which is adequate for 615 observations on 205 officers. Two sensitivity analyses were specified in advance: the linear model with sleep opportunity split into each officer's mean and the drill's deviation from it, and the segmented model refitted after removing drills whose medical note recorded interrupted rest.

3. Results

Of the 615 drills, 269 (43.7%) were held on a night watch, and 177 (28.8%), 243 (39.5%) and 195 (31.7%) were of tier 1, 2 and 3. Median latency was 8.9 s (interquartile range 6.8 to 11.7; range 3.1 to 31.9); the raw distribution was right-skewed (skewness 1.34) and the log distribution close to symmetric (0.09). Sleep opportunity averaged 20.4 hours in the preceding 72 standard hours (SD 5.1; range 6.0 to 34.0), or about 6.8 hours in each standard day. In 105 drills (17.1%) it was under 16 hours, and 80 officers (39.0%) had at least one such drill.

Latency rose as sleep opportunity fell, and the rise was concentrated at the short end (Table 1). Geometric mean latency was 13.54 s in the 33 drills with under 12 hours and 10.27 s in the 72 drills with 12 to under 16 hours, but only 8.99, 8.49 and 8.38 s in the three bands from 16 hours upward. The band figures are unadjusted; the models below adjust for watch position, service and tier.

In the linear mixed model each standard hour fewer of sleep opportunity was followed by a 1.90% longer latency (coefficient −0.0188 per hour of sleep, 95% CI −0.0244 to −0.0133; 1.34% to 2.47%; p < .001). In the segmented model reported below, the standard deviation of the officer random effect was 0.15 on the log scale and that of the residual 0.30, an intraclass correlation of 0.20, so a fifth of the variation in log latency lay between officers.

The segmented model fitted better (Table 2). Its knot was 17.0 hours. Below it, each hour fewer was followed by a 5.25% longer latency (coefficient −0.0512, 95% CI −0.0667 to −0.0356; 3.62% to 6.90%; p < .001). Above it the slope was much smaller, a 0.72% longer latency per hour fewer (coefficient −0.0072, −0.0147 to 0.0004; −0.04% to 1.49%; p = .064), and its interval includes zero. The two slopes differed by −0.0440 log units per hour (95% CI −0.0639 to −0.0241). The likelihood ratio against the linear model was 18.75 on one degree of freedom, and the segmented model's information criterion was 14.75 lower after counting the knot as a parameter. The associated p-value, below .001, is conditional on the selected knot and would be larger if the search were accounted for. Contrasted with 20 hours, an officer with 12 hours had a latency 1.32 times as long (95% CI 1.23 to 1.42) and an officer with 10 hours 1.46 times as long (1.32 to 1.62).

Where the knot lies was only weakly determined. Knots between 14.5 and 22.0 hours all lay within 1.92 log-likelihood units of the maximum, so the data support a bend somewhere in the range of about 15 to 22 hours and do not fix it at 17. Adjustment terms behaved as expected. A drill on the night watch was followed by a latency 15.95% longer (95% CI 10.14% to 22.08%; p < .001), each step up in tier by 20.90% longer (16.93% to 24.99%; p < .001), and each additional year of service by 0.98% shorter (−2.07% to +0.11%; p = .079).

Both prespecified sensitivity analyses gave similar results. Splitting sleep opportunity into an officer's mean and the drill's deviation from it gave 1.99% longer latency per hour fewer within officers (95% CI 1.22% to 2.75%) and 1.80% between officers (0.96% to 2.64%), so the association is not confined to differences between officers. The medical rating had noted interrupted rest before 46 drills (7.5%). With those removed, leaving 569 drills from 204 officers, the segmented slopes were 5.20% (3.50% to 6.91%) longer latency per hour fewer below the knot and 0.78% (0.00% to 1.57%) above it.

4. Discussion

Among these officers, short sleep opportunity was followed by slower correct decisions, and the association held within officers as well as between them. The size is not trivial. An officer who had been berthed for 12 of the preceding 72 standard hours took about a third longer to classify a contact than an otherwise similar officer with 20 hours, and the difference is larger in point estimate than the night-watch penalty (about 1.16 times), though the intervals are close and no formal comparison was made. In a task where the officer is trained to reach a correct classification quickly, a third longer is a delay in commitment that a commander would notice.

The shape of the relation matters more for planning than its average slope. The linear model describes the data adequately at a coarse level, but the segmented model fitted them better, though the test is conditional on the selected knot, and it attributes most of the decline to opportunity below the knot. Above it, the estimated slope is small and its interval reaches zero. The band medians agree: they fall by about 2.4 s between the shortest band and the next, and by only 0.07 s between the two longest. We do not claim to have found a threshold at 17 hours. The likelihood is nearly flat for knots between 14.5 and 22.0 hours, the knot was chosen from the same data on which it was tested, and a smooth curve that steepens gradually could not be distinguished from a sharp bend with these records. What the data do support is that the cost of each lost hour is larger when opportunity is already short.

One feature of the results argues against reading the association as a matter of who is short of sleep. The within-officer estimate was as large as the between-officer one, so an officer compared with his or her own drills on other days was slower after shorter opportunity. The estimate for years of service was small and its interval includes zero, so experience does not account for the pattern either. This supports the direction of the finding; it does not establish its cause, because a rotation that shortens berth time may also change workload, and neither is separated from the other in the logs.

Our results are compatible with the earlier civilian findings on vigilance (Ordane, 28 BBY) and add a measure of deliberate decision speed in a military setting. They also bear on the choice of an adjustment to rotations. If the steep segment lies where the records place it, a floor on 72-hour opportunity would be worth testing before any across-the-board increase in rest. We offer this as a hypothesis, to be tried by varying rotations in a planned comparison, and not as a recommendation drawn from records. The drill measures how quickly a correct classification is reached in a standard scenario. It does not measure the quality of a commander's judgement in an engagement, and these data say nothing about that.

5. Limitations

The exposure is opportunity, not sleep. Berth entries record when an officer was off watch and berthed, not whether the officer slept, and interrupted or poor sleep would make true sleep shorter than the entry suggests (Ashvane, 24 BBY). Removing drills flagged for interrupted rest did not move the estimates, but the flag depends on a medical rating noticing and recording it, so unflagged interruption remains likely. Error of this kind would generally dilute a linear association; its effect on the estimated knot and on the two slopes cannot be predicted.

Being observational, the design cannot separate cause from association, and both sleep opportunity and drill timing come from the same watch logs. Watch position is adjusted for, but night-watch officers have a different pattern of berth time by construction, so some of the shared structure between exposure and outcome may remain. Officers were drawn from patrols and ships that we did not model as a further level; clustering by ship would widen the intervals if ships differ in rotation practice and in how drills are run. The three drills per officer were adjusted for tier but not for their order, so a practice or fatigue trend across the patrol is not controlled.

Exploratory status applies to the threshold analysis. The knot was selected on the data, its test is conditional, and its location is uncertain across a span of about seven hours. The records run from 22 BBY to the end of the third quarter of 21 BBY only. Officers who were excused drills because of illness or duty are not in the sample, which could favour those who cope with short sleep. The officers were mostly human, and no cross-species inference is drawn. Finally, latency to a correct response in a standard scenario is the outcome; errors, incorrect first responses and the quality of a decision in a live contact were not analysed.