Sample and timeline
As individuals of interests were professional investigators (i.e., those who routinely conduct investigative interviews), existing professional contacts helped identify potential participants. Those who expressed interest received an enrollment packet that contained further instructions and a sleep-activity tracker. All participating officers worked as a part of a non-federal investigative or law-enforcement organization. The sample was drawn from seven institutions across Arizona, Iowa, Kansas, and Nevada. Overall, 79 officers enrolled in the study across 2019–2020 and provided at least some data, while 50 officers provided at least one day of joint actigraphic and investigative-diary data necessary for the analyses of interest. These key analyses thus involved 204 days across 50 officers, while analysis of actigraphic sleep–wake variables only (a larger set) involved 1442 days.
Out of the whole sample of 79 officers, 59 (74.7%) of the officers identified as male, and 20 (25.3%) as female. They ranged in age from 27 to 60 (mean = 42.06, SD = 8.06). Of 54 who did not decline to respond to the question on ethnicity, 95.8% identified as white and 4.2% as Hispanic. Twenty-two reported that they served in a detective role while three reported serving in a leadership role. All participants had investigative duties and conducted investigative interviews on a routine basis. Most respondents (79.6%) reported working typical morning to afternoon shifts during the study period, while 10.2% reported working from evening to late night, 8.2% working from late night to early morning, and 2% were scheduled to have multiple shifts during the study period.
First, participants received an envelope with instructions to enroll in the study, as well as a Fatigue Science Readiband® actigraph. Once they contacted the laboratory and registered the Readiband with the Fatigue Science platform, they completed an online background survey that inquired about demographics and sleep health (not discussed in detail within this report). Officers indicated their preferred time for daily surveys (to accommodate varying shifts), which were then e-mailed to them across 14 days (while they wore actigraphs). Following the two-week period, officers were debriefed over e-mail and their participation and data collection were terminated.
All research procedures were approved both by the Iowa State University Office for Responsible Research and the Federal Bureau of Investigation Internal Review Board. All methods were performed in accordance with relevant guidelines and regulations. Informed Consent was obtained from all subjects. The names of participating institutions and individual-level data are not disclosed due to confidentially assurances to participants during informed consent. Otherwise, all materials, analyses scripts, and results presented in this manuscript are available on Open Science Framework (https://osf.io/69nqt/).
Measures
Upon enrollment, participants completed a background survey which included measures of habitual sleep (see all materials on Open Science Framework). The survey also queried participants about overall well-being, exposure to traumatic events, well-being, and personality traits. Analyses of these individual differences are beyond the scope of the present report given the focus on daily fluctuations in alertness.
Actigraphy and alertness estimation
The Readiband® sleep-activity tracker uses three-dimensional accelerometer technology (sampled at 16 Hz) to measure movement and infer sleep–wake states. It relies on an automated proprietary algorithm to determine sleep-onset time In terms of sleep–wake variables, the Readiband outputs standard characteristics, including period-specific (e.g., nightly) sleep duration, sleep-onset latency, sleep-efficiency, and Wake-After-Sleep-Onset. The inter-device reliability of the Readiband in determining sleep–wake states is very high, estimated at 95% among healthy individuals in a recent study29. In terms of validity, Readiband is similar to other research-grade actigraphs when evaluated relative to polysomnography30,31.
Critically, the device and accompanying software utilize biomathematical modeling to continually estimate real-time fatigue. Specifically, the Readiband uses the extensively validated Sleep, Activity, Fatigue, Task, and Effectiveness Model (SAFTE) developed by the U.S. Army to estimate fatigue for each 30 s period given a minimum of three days of continuous data10,32. To do so, this model integrates actigraphically-recorded sleep duration (time spent asleep), sleep continuity (number and duration of interruptions), alongside time of day (circadian misalignment), and sleep consistency (regularity over time). Ultimately, the algorithm yields scores representing levels of alertness (vs. sleepiness or fatigue) between 0 and 100 for each 30 s epoch (“SAFTE Scores”). These scores indicate the person’s level of alertness (vs. sleepiness) relative to their own baseline and is only generated after 72 h of continuous recording.
More specifically, the SAFTE scores track the percentage of the person’s optimal baseline response speed, based on the observed level of sleep and circadian disruption. For example, a SAFTE score of 90 indicates that a person is around 11% slower than when at their normal. As a result, SAFTE scores can also be expressed as Blood Alcohol Content that would produce a similar level of impairment in response speed (https://www.fatiguescience.com/sleep-science-technology/). The SAFTE model has been extensively validated in laboratory contexts that assess reaction time across multiple days, as well as within real-world contexts involving railroad, aviation, and military operations that measure performance and accident risks33. Scores around 85 or higher are considered ideal, while scores below 80 indicate around 25% slower responses, and scores below 70 indicate dangerous fatigue impairment. For example, scores below 80 indicate threefold increase in likelihood of attentional lapses and approximate the effects of 0.05 Blood alcohol content34.
Daily survey
Each day during a time period they marked in the background survey, participants received a text-message with a link to an online daily survey. In this survey, officers first reported their subjective sleep quality of the prior rest period (“How well did you sleep last night?”). To capture global daily functioning, they also indicated that day’s stress (“How stressed do you feel today?”), as well as the amount of time spent on self-care (“How many hours did you spend on self-care today (exercising, relaxing, hobbies)?”) on 5-point scales. They also responded to “How many servings of Caffeine have you consumed since you have woken up?”. Finally, they indicated their working hours on that day, including any court time (“Please indicate the exact hours you have worked [including court] since the last time you completed this survey?”) these were used to accurately specify rest and active periods for data analysis that reflect varying shifts.
During each daily survey, participants were asked if they conducted an investigative interview, defined as a 10-min or longer conversation aimed at obtaining specific information. If so, they responded to the following questions regarding the interview (if a participant indicated multiple interviews that work period, they answered the same questions for each interview reported, up to 8). First, they indicated the general time of the interview (in three hour blocks starting at midnight and spanning twenty-four hours), the location of the interview (interrogation room, residence, vehicle, or other (e.g., street), and the duration of the interview (less than 30 min, 30–60 min, or 1 h or longer). Descriptive information for all the interviews appears in Supplemental Materials.
Critically, the investigators reported their assessment of each interview regarding their relationship with the subject (rapport and resistance), their own reactions (difficulties with focus and emotional composure), and the perceived usefulness of obtained information. First, established rapport was assessed by asking officers “How well was the rapport and co-operation established? (Please indicate the extent of the rapport, co-operation, and mutual respect you established with the interview subject). Second, subject resistance was assessed by “How difficult was it to obtain information?” (Please indicate how difficult was to obtain disclosure of desired information due to resistance from the subject?). Third, investigator composure was assessed with “How difficult was it to maintain your focus and emotional composure” (please indicate how difficult was it to sustain attention and control one’s emotional reactions). Fourth and final, the perceived information utility of obtained intelligence was assessed by “How useful was the information obtained?” (Please indicate the quantity and quality of information obtained during the interview). Responses to all these questions were made on “Not at All” (1) to “Extremely” (5) Likert-type scales.
Variables and analyses
First, in order to estimate average daily alertness during work periods when officers conducted interviews, each participant’s actogram (record of sleep–wake activity from the Readiband) was examined alongside reports of working hours to determine an active work period (time span during which officers conducted interviews and were awake, even if nighttime) and a rest period (time span during which officers slept and did not work, even if daytime). Cross-referencing ensured that dates of actigraphic rest-activity periods are appropriately paired with next-day diary reports. Then, the Readiband SAFTE alertness scores from each scored epoch across these active periods were aggregated to estimate average alertness for that officer during the respective work period (“day”). Note that SAFTE scores were generated only after 72 h of continuous recording, which results in a restricted set of days for these analyses (relative to other sleep variables which are generated every recorded rest period). Furthermore, all sleep periods are factored into the SAFTE model’s sleep algorithm and have a corresponding effect on individuals’ estimates. However, naps of 30 min or less may not be recorded by the Readiband so would not be reflected by the sleep duration estimates.
Second, in order to estimate daily interview–outcomes, we aggregated ratings for each dimension across all interviews reported that day. For example, if officers reported multiple interviews on a given day (29%), their responses about established rapport were averaged across all interviews to reflect overall rapport established across interviews conducted that day (commensurate with sleep–wake variables).
While SAFTE scores were available on a moment-to-moment basis, interview experiences were recalled only once daily without exact times, precluding a finer-grained analysis. As a result, for key analyses sleep characteristics (i.e., duration, continuity) of the prior rest episode (“prior night”) were utilized as predictors of waking function reported for the subsequent active period (e.g., interview outcomes the following day).
Statistical precision
For inferential tests of the predictive strengths of sleep and alertness for daily interview outcomes while taking account data-clustering within individuals and day-level co-variates, we estimated fixed coefficients within multi-level models with days (Level I) nested within participants (Level II). Models were implemented in R-Studio version 2021.09.1+372 using lmer package, with outputs are available on OSF.
Given high variability in the number of days with interviews across investigators (with multiple investigators reporting only one or two days with interviews), the analysis focused on the day-level associations across the whole sample. To this end, scores were grand-mean centered, such that daily alertness, sleep, and interview variables reflected deviations from the average day in the sample, regardless of the investigator. The estimated parameters thus represent day-level linear regression coefficients between sleep and alertness on one hand, and interview outcomes on the other, accounting for clustering within participants (partial in presence of day-level covariates).
According to simulations reported by Arend Schafer, exceeding 5 day-level observations and 50 person-level observations affords at least 80% power to detect small-to-moderate Level 1 direct effects35. To this end, the recruitment goal was 200 day-level observations or more (regardless of nesting), also approximating similar power to detect small-to-moderate correlations. Note that large day-level variance components are expected in sleep-tracking studies21, which contributes to power for identifying Level I effects even with few observations per Level I unit (i.e., number of individuals35).
Distributions and outliers
Distributions of key interview and sleep variables were inspected prior to the analyses to identify potential outliers (i.e., observations 3 or more standard deviations from the mean) and to identify anomalous distributions. For nearly all reported interview outcomes, ratings spanned the entire scale range (with the exception of ‘extremely’ ratings for investigator having composure difficulties). They were also normally distributed with minor skew. Only one daily interview data point was more than 3 SD below the mean (a rating of ‘not at all’ for establishing rapport). Given that the next-higher ratings regarding rapports were common and the distribution was continuous, we retained this data point. Inspection of alertness and sleep variables (duration, wake-after-sleep-onset, subjective quality) did not reveal any outliers (results appear in the Online Supplement).