Data came from the Work, Family, and Health Study.^9,24 The current study used data from a sample of employees working in teams within the IT division of a large Fortune 500 firm. Information technology employees working in the metropolitan areas with the 2 largest worksites were invited to participate in the study. The worksites were selected with the corporate leadership’s endorsement to facilitate the study.⁹ Trained field interviewers administered face-to-face computer-assisted personal interviews (CAPIs) to employees at the workplace, beginning in September 2009 and ending in September 2010, to get information on their demographic and work characteristics. Employees who completed a 60-minute worksite interview received a $20 incentive.

Immediately after the CAPI, interviewers introduced the actigraphy data collection process. If the participant agreed, the interviewer instructed them to wear a sleep monitor (Spectrum; Philips-Respironics, Murrysville, PA) on their nondominant wrist for a week except in situations in which the watch could be damaged (eg, excessive impact, extreme temperatures). Of the total 1171 employees who were invited, 823 employees completed the CAPI interview at baseline (70% response rate). Of these, 635 employees completed a weeklong actigraphy data collection (baseline). One year later, 569 employees (90% retention) completed a second week-long actigraphy data collection (12 months). Among 569 employees, 474 provided valid actigraphy for 3+ days at baseline and 12 months, a commonly used inclusion criterion in prior studies.⁸ In this study, we required a minimum of 5 days for reliable and valid estimation for day-to-day variability of sleep. Thus, the final sample for the current analyses was 396 employees who provided valid actigraphy data for 5 to 10 days at the 2 time points (waves, hereafter). On average, participants had 7 days of valid actigraphy across waves (SD, 0.87). Employees received additional $20 for providing their actigraphy sleep data at each wave. Therefore, the total compensation for our study sample was $60.

Participants had a mean (SD) age of 46.97 ( ± 8.45 years). Slightly more than half (59%) of the employees were men. Most (80%) were married or living with permanent romantic partner; 74% were white and 16% were Asian or Pacific Islander; 23% were providing care to their adult parents or relatives; 75% were college graduates and 22% had some college (1–3 years). The mean annual household income was in the range of 110,000–119,999. About half (48%) had children younger than 18 years living in the home. Mean (SD) work hours per week was 46.52 (7.33) hours. Most (79%) worked with standard dayshift (ie, Monday-Friday, 9–5) and the rest had variable shifts (that changes from day to day, but did not consist of regular evening/night shift workers, per se).

The Work, Family, and Health Study designed a workplace intervention to increase employee control over their work schedule and increase supervisor and coworker support.⁹ After the baseline CAPI interviews, a total of 56 study groups (aggregation of work teams that operated in the organization) were randomized to either the STAR intervention condition (n = 195) or the control condition (n = 201). The STAR intervention consisted of a 3-month structural and cultural change process, including 2 types of main activities: (1) facilitator-led sessions for both employees and managers to help transition them from rigid work schedule to give employees more control over when and where they work and (2) manager-specific training, goal setting, and behavioral self-monitoring focused on increasing family supportive supervisor behavior (for details on STAR procedures, see Kelly et al¹⁰ and Kossek et al²⁵). Employees who were assigned to the control condition continued business as usual (usual practice) without any entailed activities.

Actigraphy data came from each participant’s actiwatch using the Actiware Sleep Scoring Program (v5.71; Philips-Respironics). We used a standard algorithm recently validated²⁶ and at least 2 members of the scoring team determined the validity of each recording based on study-specific standard sleep criteria applied similarly to all recordings.⁸ Each sleep episode was identified if it began at the last epoch of high activity (>10 activity counts) preceding at least five 30-second epochs of <10 activity counts, indicating little to no movement. Possible sleep episodes were determined for low movement count periods by visual inspection and considering changes in light levels (eg, suddenly decreased light levels for that putative nap interval were confirmatory but not required for a nap designation), as well as the overall profile (eg, was the nap similar in low levels of activity to nighttime sleep?). A determination of the actual presence and amount of sleep during those rest intervals was calculated via algorithm (see below). Timing was also considered for categorization of each visually identified sleep periods as either a nighttime (main) sleep period or a daytime nap.

Among sleep episodes, an episode with the longest sleep period of the day was defined as the main, nighttime sleep episode (all other sleep episodes were defined as nap episodes). Daily nighttime sleep duration was computed for main nighttime sleep episodes as the amount of nighttime sleep attained on a day in minutes.

Daily WASO was computed as time spent “waking” after sleep onset and before sleep offset during nighttime sleep on a given day, in minutes, as previously validated vs polysomnography.²⁶

Sleep periods coded as nap episodes were summed at the daily level. We initially set the minimum nap rest interval duration at 30 minutes for discrimination of substantive naps from brief periods of no movement. Considering that naps may be rare among daytime employees, we also included naps in the range of 25.5–29.5 minutes in the analysis, which was 0.05% of total naps in our sample (53 naps from 15 employees). Daily nap duration for those who had any naps on a given day was calculated in minutes.

Employee age (in 1-year intervals) was tested as a moderator and also included as a covariate. Given that employees’ personal and family characteristics may denote family demands and responsibilities as well as their sleep behavior,^27,28 we controlled for sex (0 = female, 1 = male), marital/partner status (0 = unmarried, 1 = married or living with a partner), living with children younger than 18 years in household (0 = no, 1 = yes), and providing care to adult relatives (0 = no, 1 = yes) in all models. Sociodemographic factors, including race (0 = white, 1 = Asian/Pacific Islander/Other; 2 dummy variables were created) and range of current annual household income (1 = less than 49,999 to 12 = more than 150,000), were included in all models. Furthermore, to consider possible differences from work characteristics, total work hours assessed by asking “How many hours do you work at any job?” and work schedule (0 = standard dayshift, 1 = variable shifts) were controlled for. In addition, job titles were coded as core (=1; eg, software developers) vs noncore (=0; eg, technical support), following the practice in prior research.²¹ In the IT firm, there was an organizational level merger and about half of employees had not yet received the intervention when merger was announced, whereas others had already received it. To take into account potential differences from this, we included an indicator for the event sequence between merger and intervention (0 = Intervention Premerger, 1 = Intervention Postmerger) in our analysis. Lastly, we controlled for weekend vs weekday effect (0 = weekdays, 1 = weekends; 29% of the study days were weekends). All continuous variables were centered at sample means.

We fit multilevel models with SAS Proc Mixed²⁹ to take into account the clustered data structure: 5550 daily observations across the 2 waves clustered within 396 employees. There were relatively small variances at the study group level or wave level compared with variances at daily and individual levels, and hence, we used 2-level models to simplify model structure. Because of missing values in covariates, 10% of observations were not included in analyses; missing cases were not systematically different from nonmissing cases. The model for nap duration was estimated based only on 1448 nap days for employees (n = 163) who napped at both waves and had no missing responses on covariates. For the daytime napping models, we included only participants with naps at both waves to evaluate intervention-related changes on total daytime napping duration (but also conducted supplementary analyses with binary nap variable, see “Age differences in intervention effects on changes in daily amounts of nighttime sleep and daytime napping”). Examining changes in nap duration over time enables us to rule out a possibility that the likelihood of napping at one time point is random and nonsystematic.

To examine main intervention effects on changes in daily nighttime sleep duration, WASO, and daytime nap duration, we included a 2-way interaction between the intervention condition and a wave (0 = baseline, 1 = 12 months). This interaction parameter represents the difference in changes in a sleep outcome from baseline to 12 months between STAR employees relative to control employees.

To test whether intervention effects on changes in sleep differed by employee age, a 3-way interaction—intervention × wave × age—was included. The 3-way interaction parameter represents whether differences in changes in a sleep outcome from baseline to 12 months between STAR relative to control employees are further differentiated with increasing age. We conducted follow-up tests using estimate commands in Proc Mixed to assess differences between younger employees (38 years, or −1 SD) and older employees (56 years, or +1 SD) for each intervention condition.

Lastly, to examine intervention effects on changes in day-to-day variability of sleep, we used multilevel models with heterogeneous within-person variance. Daily variance in a sleep outcome was modeled in random effects and estimated by the exponential function (Exp) to use a linear prediction while eliminating the dependence of the variance on the mean level.³⁰ Positive coefficients indicate greater day-to-day variability in a sleep outcome. Of note, this approach can estimate more robust variability than the individual SD (iSD) or variance approach because it is model-based and takes into account temporal dependence of data (day effect) as well as the effects of covariates on the mean level.^21,31 In this approach, however, we cannot test the statistical significance of slopes for an interaction in random effects. Thus, in the case of a significant interaction, we conducted a follow-up test taking quantified individual-level variances across days (iSD²) as an outcome. By doing so, we may underestimate changes in employees’ day-to-day variability by the intervention and/or by age, because there is less power to detect effects when using quantified variance estimates compared with when modeling variability.³¹

Demographic and work characteristics of our sample of IT employees, with actigraphic sleep measures at baseline, are presented in Table 1. We also present results from difference tests (analysis of variance, χ² tests) by employee age, divided into 3 age groups based on mean age ± 1 SD at baseline: younger, 24 to 38 years (n = 72); middle-aged, 39 to 55 years (n = 257); or older employees, 56 to 70 years (n = 67). On average, employees had 424 minutes (7 hours) of nighttime sleep duration (SD, 54.8) and 41 minutes of WASO (SD, 16.6), per day, with no statistical difference by age groups.

In terms of daytime napping, 253 employees (64%) had any naps at baseline. Among those with any naps at baseline (Table 1), 22% of their sleep episodes were naps, and they had 2.2 naps, on average, across the study days (SD, 1.6) and 1.2 naps on days when they napped (median, 1 nap; about 19% of nap days included 2 naps on average for that day). For those who had any naps, the mean (SD) daytime napping across total study days was 23 (21.7) minutes and the mean (SD) daytime napping across only days when they nap was 67 (38.2) minutes (median, 57.5 minutes) at baseline. Thus, two-thirds of our sample napped for less than 30 minutes across all days in a week of collection, and for about an hour on a nap day (that happened on average twice a week). The model for nap duration included 1448 nap days in participants who napped at both waves and had no missing responses on covariates (n = 163 employees; 41% of the sample). These nap days occupied about 30% of total study days used in the analyses (Table 2).

Although age was not related to having naps, younger employees had longer naps than did middle-aged and older employees. To provide proxy estimates for day-to-day variability in sleep measures, we included individual SDs (iSDs) of each sleep variable. The degrees of daily variation were 64, 14, and 32 minutes around person means for nighttime sleep, WASO, and nap duration, respectively. There were no age differences in these iSDs of sleep measures. Greater iSD of nighttime sleep duration was associated with longer nap duration (r = 0.18, P < .01) and greater iSD of WASO (r = 0.27, P < .001).

Table 2 shows results from multilevel models separately predicting daily nighttime sleep duration, WASO, and nap duration. In Model 1, there was a significant interaction between wave and intervention predicting nighttime sleep duration: intervention employees increased nighttime sleep duration (3 minutes more per day) but control employees decreased (6 minutes less per day), resulting in a significant 9-minute difference in daily nighttime sleep duration between the 2 groups (Fig. 1, panel 1). This result is consistent with previous intervention effects for 24-hour total sleep duration.⁸ There were no significant interactions between wave and intervention predicting changes in the amount of WASO or nap duration.

In Model 2 of Table 2, we tested whether intervention effects differed by employee age. There was a significant 3-way interaction between age, wave, and intervention predicting nap duration (but no effects on nighttime sleep duration or WASO). Panel 2 in Fig. 1 shows that the difference in nap duration between older employees (+1 SD corresponds to 56 years) in the intervention condition and those in the control was significant with a difference of 17 minutes less per day. Specifically, control older employees increased nap duration, but intervention older employees did not. Younger employees (−1 SD corresponds to 38 years) in the intervention condition did not significantly change nap duration compared with control younger employees. We also conducted supplementary analyses with binary nap variable (napping vs no napping) that includes those who napped at only one wave. Neither the main effect of the intervention nor age differences in the intervention effect on changes in the likelihood of napping were significant.

Table 3 shows results from heterogeneous within-person variance models predicting changes in day-to-day variability in nighttime sleep duration, WASO, and nap duration. There was no significant interaction between age, wave, and intervention predicting daily variance of nighttime sleep duration. However, significant 3-way interaction effects emerged for daily variance of WASO and nap duration. To interpret the interaction effect on daily variability in WASO, we conducted a follow-up analysis predicting individual variance of WASO with the same set of predictors (individual mean levels of WASO were controlled). The 3-way interaction also predicted individual variance of WASO at a trend level (B = −8.56, SE = 4.56, P = .062). The post-hoc test indicated that older employees in the intervention condition significantly decreased day-to-day variability in WASO at 12 months compared with their baseline—90 minutes² less variance or 9.5 minutes less SD around person means (Fig. 1, panel 3). Older employees in the control condition did not significantly change daily variability of WASO and the difference in variability of WASO between the 2 groups did not reach statistical significance. Lastly, with regard to daily variability in nap duration, younger employees in the intervention condition exhibited an increase in daily variability of nap duration at 12 months relative to their baseline.