IntroductionBackground

Sleep is a crucial component of daily life, closely interconnected with all aspects of our routines and overall well-being, including mental health [,], physical health [], and work performance [,]. The COVID-19 pandemic profoundly impacted various aspects of daily life, with sleep patterns being a particularly significant area of concern. However, the effects on sleep were often indirect, resulting from changes in daily routines and lifestyle adjustments rather than being a direct consequence of the virus.

In response to the pandemic, outdoor restrictions limited our exposure to natural daylight, a crucial element for regulating circadian rhythms and sleep patterns []. Similarly, mobility restrictions altered daily physical activity (PA) patterns. Additionally, workplace restrictions led to work-from-home policies, which resulted in reduced mobility and flexible working hours. While these changes led to more relaxed work schedules, they also blurred the boundaries between professional and personal life. Notably, factors such as daylight exposure, PA, and work routines—each significantly affected by the pandemic—are well-established influences on sleep health [,].

Traditional sleep measurements often rely on self-reported methods, such as the Karolinska [] or Pittsburgh sleep diary []. While these methods are effective for tracking day-to-day sleep over short periods, conducting diary studies over longer intervals is generally not feasible due to the cognitive burden on participants. Nonintrusive measurements using smartphones and fitness trackers have recently emerged as a more viable alternative for capturing sleep data over extended periods. While consumer-grade devices may not precisely detect sleep stages, they have shown promising results for measuring sleep onset, duration, and wake-up time. Assessing sleep with these devices has the advantage of capturing data in people’s natural living environments, unlike sleep laboratories. Additionally, this method is not subject to memory biases that can occur with survey responses and sleep diaries.

The evolution of mobile health (mHealth) technologies has significantly enhanced traditional sleep monitoring methods, particularly through the use of wearable devices. These devices offer a more accessible and less invasive way to monitor sleep patterns, while also deepening our understanding of sleep-related phenomena. For instance, wearable devices have been used to determine individuals’ chronotypes and track their sleep and activity rhythms over extended periods [,]. They have also been used to measure sleep alignment between coworkers [], examine the relationship between sleep and burnout [], and assess sleep patterns in various populations, including patients with mental disorders []. Several studies have confirmed the validity and reliability of wearable devices, demonstrating notable sensitivity compared with the gold-standard polysomnography (PSG). For example, a review of 7 consumer sleep-tracking devices [] highlighted their high effectiveness in detecting sleep relative to PSG. Similarly, a study [] evaluated 6 consumer wearable devices and validated their accuracy in assessing sleep timing and duration compared with PSG.

Prior research comparing sleep patterns before and during the pandemic has revealed notable differences. Studies found that following the pandemic’s onset, individuals tended to go to bed later [], slept for longer durations [], exhibited reduced variability between weekday and weekend sleep [,], and experienced increased sleep disturbances or diminished sleep quality []. Various factors have been identified as contributing to these disruptions in sleep routines, including decreased PA [], social isolation [], increased use of electronic devices [], and the shift to working from home [].

While previous studies have focused on the immediate consequences of lockdowns and restriction policies, less attention has been paid to the long-term effects, particularly during the late stages of the pandemic when restrictions began to relax. This phase is crucial for understanding the residual effects of the pandemic on sleep patterns and how quickly individuals revert to their prepandemic sleep habits. The transition to working from home as the default mode has resulted in a less constrained work-life routine, leading to more flexible sleep-wake schedules. Certain demographics may benefit more from these transitions, such as individuals with more flexible routines (eg, research personnel) or those who tend to snooze their alarms after waking, referred to as “snoozers.” Additionally, occupation is a known factor influencing sleep patterns, with a classic example being the contrast between shift workers and nonshift workers [,]. However, less is understood about the differences between various roles within academia, such as researchers with deadline-driven roles and administrative personnel typically following a 9-to-5 schedule. Therefore, a comprehensive, longitudinal analysis of sleep patterns that includes these variables and extends into the late stages of the pandemic is important.

Objectives

Our study aims to provide a holistic view of how the pandemic has influenced sleep patterns. We evaluate the long-term relationships between sleep patterns, including average and variability in total sleep duration and sleep timing, alongside individuals’ characteristics (demographics, occupation, and PA) and external factors (stringency of restriction policies and seasonal variations). Our research utilizes longitudinal data from fitness trackers and questionnaire responses collected from working adults at a Finnish university. This extensive data set enables us to examine shifts in sleep behavior during the later stages of the COVID-19 pandemic, from June 2021 to June 2022. The study’s timeframe covers a full annual seasonal cycle, which is crucial for analyzing sleep patterns in Finland, where significant seasonal changes and daylight variations occur due to its northern latitude.

MethodsStudy Data

This work used data from the cor:ona (comparison of rhythms: old vs. new) study [] as part of a 1-year multimodal data set of working adults.

Ethics Approval

The study was approved by the Aalto University Research Ethics Committee (approval number D/536/03.04/2021_COR_ONA).

Participants and Procedures

The cor:ona study recruited 128 full-time employees from a university in Finland for a 1-year investigation into how their daily activities changed during different stages of the COVID-19 pandemic. Throughout the study, participants wore a Polar Ignite fitness tracker (Polar Electro Oy), enabling us to unobtrusively collect various measures related to sleep and PA. In addition, participants completed an initial baseline questionnaire, an exit questionnaire, and a shorter version of the baseline questionnaire each month. The monthly questionnaires asked for information about their daily routines, work, and sleep quality over the past month. The detailed recruitment procedure and participants’ demographics were described in a previous study [].

Fitness Tracker DataSleep Measures

The fitness trackers measured bedtime (defined as the recorded time when a person fell asleep), waketime (defined as the recorded time when a person woke up), and interruption duration (defined as the total time in seconds spent awake between sleep start and end times) for each day. A sleep period was defined as the longest sleep episode for each day. Sleep patterns were measured using 4 metrics: (1) total sleep time (TST), which measured the time a person spent asleep, calculated as the duration from bedtime minus the interruption duration; (2) midsleep (MS), the midpoint between bedtime and waketime, which was used to measure sleep timing and computed as (bedtime + TST)/2. Additionally, we proposed 2 other metrics to measure sleep regularity: (3) TST variability, computed as the SD of TST during weekdays (Sunday night to Thursday night); and (4) MS variability, computed as the SD of MS during weekdays. We focused exclusively on weekdays due to the expected differences between weekday and weekend sleep patterns. The Niimpy behavioral data analysis toolbox [] was used for extracting sleep measurements.

Physical Activity Measures

The fitness tracker recorded the number of steps taken each hour, which were then summed to provide a daily step count. To comprehensively account for daily PA patterns, including their timing and distribution, we introduced 2 additional metrics: midstep and intradaily variability (IV) []. These metrics are designed to capture the timing and dispersion of PAs throughout the day. Specifically, midstep represents the hour of the day when half of the total number of steps is achieved, analogous to MS in the context of PA. By contrast, IV quantifies the fragmentation of the activity-rest rhythm and is measured as follows:

where N=24 is the total number of samples within each day; Xi is the i measurement sampled at P=60-minute interval; and is the average value of all samples in a day. Low IV indicates less fragmented activity-rest rhythm, whereas high IV could imply daytime naps or nighttime awakenings.

External Data

Seasonal data were collected from the World Weather Online developer application programming interface []. Given the significant variation in day length in Finland during the study (up to 13 hours), day length was used as a proxy for seasonal variables. The choice of day length as a proxy was motivated by Friborg et al []. The study compared 2 geographically distinct locations with substantial differences in day length variability: Ghana and Norway. Although no noticeable seasonal effects of day length were observed in Ghanaians, Norwegians showed a delay in both bedtime and waketime during summer weekdays, though sleep duration remained relatively unaffected.

We also utilized the Stringency Index (SI) [], a composite measure ranging from 0 to 100, to assess daily COVID-19 restriction policies. Higher values on this index indicate more stringent COVID-19 restrictions, including measures such as school and workplace closures, the cancellation of public events, and the enforcement of stay-at-home orders. This index allows for standardized comparisons of policy responses across different countries or regions, as well as changes within the same region over time.

Questionnaire Data

Upon entering the study, participants completed a baseline questionnaire that collected basic background information, including age, gender, chronotype, occupation, and origin, among others. Chronotype was assessed using the reduced Morningness-Eveningness Questionnaire (MEQ) [], with higher scores indicating a morning type and lower scores indicating an evening type. For the origin-related question, participants chose from 3 options: Finland, Europe (excluding Finland), or outside of Europe. Participants indicating they were from Finland were classified as Finnish, while those selecting other options were described as having a “migrant background.” Regarding occupation, participants specified whether they were academic or service staff. The term “academic staff” refers to individuals involved in academic and research activities within the organization, while “service staff” includes those in roles such as human resources and other administrative or support functions. Participants were determined as a snoozer if they answered “yes” to the following question: “Snoozing can be considered as choosing to go back to sleep after an alarm has awakened you intending to wake up later; setting the alarm earlier than when you intend to wake up; or setting multiple alarms with the intent to not wake up on the first alarm. Do you currently consider yourself a snoozer using this definition?,” as adapted from [].

For the analysis of snoozer characteristics, we used the 2-item Patient Health Questionnaire (PHQ-2) [] and the short form of the Pittsburgh Sleep Quality Index (PSQI) [], averaging the values collected from the monthly questionnaires. Additionally, the short form of the Positive and Negative Affect Schedule (PANAS-SF) [] was used in the initial baseline questionnaire.

Data Exclusion and Preprocessing

Sleep data were restricted to the period from July 1, 2021, to May 31, 2022. Because of our rolling recruitment process, which started in mid-June 2021 and ended in June 2022, we excluded data from June of both years. This exclusion was necessary because we lacked complete data for these months, and including partial data could have introduced bias. A standard filter, adopted from [], was applied to remove outliers in TST (TST<3 hours and TST>13 hours). Participants with fewer than 30 recorded nights due to dropout or technical issues were excluded. For gender-related analysis, nonbinary participants (n=1) were excluded to preserve their privacy. To maintain the interpretability of the relationships between sleep patterns and the examined variables, we chose not to normalize the dependent and independent variables.

Statistical Analysis

We used a logistic regression model to examine factors predicting snoozing behavior. Using snoozing behavior as the dependent variable, and to replicate the findings from [], we included the same set of independent variables: age, gender, step count, TST, BIG-5 personality traits (openness, conscientiousness, extraversion, agreeableness, and neuroticism), PANAS-SF, PHQ-2, PSQI, and MEQ. To further investigate the potential confounding effects of chronotype (measured by MEQ) on the relationship between personality traits and snoozing behavior, we conducted a Baron and Kenny [] mediation analysis.

Given the nature of our data set, which included repeated sleep measurements for each participant, we used mixed effects linear models [] to analyze how sleep patterns and their regularity evolve over time. The models included TST, MS, and the variability of TST and MS as dependent variables. For models with variability of TST and MS as dependent variables, the numerical independent variables were averaged across weekdays. We adopted a sequential modeling strategy, building 3 distinct models for each dependent variable. Model 1 included basic characteristics such as chronotype, age, gender, origin, occupation, and parenting cohabitation status (number of children in the household). Model 2 extended model 1 by adjusting for external factors such as the stringency of restrictions and day length. Finally, model 3 built on model 2 by incorporating PA metrics, including step count, midstep, and IV. This approach allows for the exploration of the unique contributions of each new set of variables beyond those accounted for in the previous model. All models included hierarchical random effects for the study participants to account for repeated measurements. The models are formulated as follows:

Model 1: Yij = β0 + β1xij1 + β2xij2 +···+ β2xij7 + β8xij8 + uj + ϵijModel 2: Yij = β0 + β1xij1 + β2xij2 +···+ β9xij9 + β10xij10 + uj + ϵijModel 3: Yij = β0 + β1xij1 + β2xij2 +···+ β11xij11 + β12xij12 + β13xij13 + uj + ϵij

where the independent variables are xij1=age, xij2=gender, xij3=number of children, xij4=origin, xij5=occupation, xij6=MEQ, xij7=snoozer, xij8=free day, xij9=Stringency Index, xij10=day length, xij11=steps(×1000), xij12=midstep, and xij13=step entropy.

95% CIs were reported using bootstrapping. The performance of the model was compared using the likelihood ratio test (LRT) to ensure model parsimony. All statistical analyses were performed using R software (version 3.6.1; R Foundation) []. Linear mixed models were tested using the lme4 package [], and P values for these models were calculated using the lmerTest package [].

ResultsData Summary

In total, 112 users and 27,350 nights were included in the TST and MS analyses. The models for the variability of TST and MS used the weekday SD of both measures, which included 3682 observations. The average age of participants was 39.5 (SD 9.9) years. Of these 112 participants, 49 were academic staff and 63 were service staff. presents the average values of the 4 sleep metrics—TST, MS, and their corresponding SDs—for each participant included in the analysis. illustrates the sleep patterns over time for 2 participants: 1 with low variability and 1 with high variability in their sleep patterns.

‎Figure 1. TST, MS, and their SDs of participants included in the analysis. Each dot represents the participant's mean value for the corresponding metrics. MS: midsleep; TST: total sleep time. ‎Figure 2. Sleep data over time from two participants. Participant 1 (red line) demonstrates shorter, later, and more variable sleep compared to Participant 2 (blue line). Total Sleep Time

We begin by investigating the factors that influence TST using the 3 linear mixed models described earlier. presents the results of these models predicting TST. For improved interpretability, the rate of change in TST is expressed as the estimate of the predictors multiplied by 60 minutes. In the full model (model 3), an increase in age by 1 year was associated with a 1.2-minute decrease in TST (95% CI –1.8 to –0.6; P=.008). Regarding gender, males were found to sleep 20.4 minutes less than females (95% CI –33.0 to –7.8; P<.001). Comparing occupations, service staff were found to sleep 22.2 minutes more than academic staff (95% CI 8.4-36.6; P=.004). A detailed monthly breakdown of the variations in sleep pattern measurements across different occupations is provided in . After adjusting for day length and the SI, an additional hour of day length was associated with a 0.60-minute decrease in TST (95% CI –0.72 to –0.36; P<.001). Conversely, a 1-point increase in the SI offset this decrease by 0.18 minutes (95% CI 0.06-0.30; P<.001). In the full model, which included PA, a 1-unit increase in IV was associated with a 15.6-minute decrease in TST (95% CI –17.5 to –13.8; P<.001). Moreover, an additional hour in midstep was associated with a 1.8-minute increase in TST (95% CI 1.2-2.4; P<.001).

Table 1. Estimates of fixed effects from the linear mixed effects model predicting TSTa.PredictorsModel 1bModel 2cModel 3d
EstimatedCIP valueeEstimatedCIP valueeEstimatedCIP valueeAge–0.02–0.03 to –0.01.002f–0.02–0.03 to –0.01.002f–0.02–0.03 to –0.01.008fGender (male)–0.34–0.55 to –0.14<.001g–0.34–0.55 to –0.14<.001g–0.34–0.55 to –0.13<.001gNumber of children0.04–0.07 to 0.15.510.04–0.07 to 0.15.510.03–0.07 to 0.14.56Origin (migrant background)0.01–0.27 to 0.27.970.01–0.27 to 0.27.970.01–0.24 to 0.24.93Occupation (service)0.360.11 to 0.60.01h0.360.11 to 0.60.01h0.370.14 to 0.61.004fMEQ–0.01–0.04 to 0.02.59–0.01–0.04 to 0.02.600.30–0.20 to 0.22.79Snoozer (Yes)–0.2–0.44 to 0.05.11–0.2–0.44 to 0.05.11–0.18–0.44 to 0.05.14Free day (Yes)0.100.07 to 0.13<.001g0.100.07 to 0.13<.001g0.080.06 to 0.11<.001gStringency Index—i——0.0050.003 to 0.007<.001g0.0030.001 to 0.005<.001gDay length———–0.01–0.01 to –0.01<.001g–0.01–0.012 to –0.006<.001gSteps (×1000)——————–0.01–0.01 to 0.01<.001gMidsteps——————0.030.02 to 0.04<.001gIntradaily variability——————–0.26–0.29 to –0.23<.001g

aThe σ2 values for models 1-3 were 1.13, 1.13, and 1.1, respectively. The intraclass correlation coefficient values for models 1-3 were 0.19, 0.19, and 0.19, respectively. The marginal R2/conditional R2 values for models 1-3 were 0.055/0.234, 0.057/0.235, and 0.069/0.245, respectively. The Akaike information criterion values for models 1-3 were 81,208.16, 81,154.98, and 80,783.26, respectively.

bIncludes demographic and occupational variables.

cIncludes model 1 + restriction and seasonal factors.

dIncludes model 2 + physical activity influences.

eItalicized values denote significance.

fP<.01.

gP<.001.

hP<.05.

iNot available.

The marginal R2 values represent the proportion of variance explained by the fixed effects, while the conditional R2 values indicate the proportion of variance accounted for by both fixed and random effects. The increase in both R2 values suggests that more complex models, particularly model 3, explained a greater proportion of the variance in the dependent variable. The LRT between models 1 and 2 indicated that model 2 was a significantly better fit (χ22=57.17; P<.001). Additionally, the LRT between models 2 and 3 showed that model 3 provided a significantly improved fit (χ23=377.72; P<.001). The performance of the full model (model 3) was further supported by the Akaike information criterion (AIC), which was lowest for model 3 (AIC 80,783.26), indicating that it offered the most optimal fit for the data.

Midsleep

Using the same approach, we developed 3 linear mixed models to assess the associations between the same set of predictors and MS. The results are presented in . To enhance interpretability, the rate of change in MS is measured as the estimate of the predictors multiplied by 60 minutes. Across all 3 models, chronotype (MEQ) (P<.001) and sleep on a free day (P<.001) consistently emerged as significant factors. In the full model (model 3), a 1-point increase in the MEQ was associated with an 8.4-minute decrease in MS (95% CI –10.8 to –5.4; P<.001). Sleep on a free day occurred 11.4 minutes later (95% CI 9.6-12.6; P<.001) compared with a workday. After adjusting for season and restriction policies, MS was delayed by 0.6 minutes (95% CI 0.6-1.2; P<.001) for each additional hour of day length. A 1-point increase in the SI was associated with a 1.2-minute increase in MS (95% CI 1.2-1.8; P<.001). In the full model, which included PA variables, a 1-unit increase in IV was linked to a 17.4-minute earlier MS (95% CI –19.8 to –15.6; P<.001). Similarly, an increase in step count was associated with a 0.6-minute earlier MS (95% CI –1.2 to 0.0; P=.04).

The LRT between models 1 and 2 indicated that model 2 was a better fit (χ22=443.70; P<.001). Additionally, the LRT between models 2 and 3 showed that model 3 provided a significantly improved fit (χ23=291.63; P<.001). The AIC value for model 3 was also the lowest (AIC 86,315.145), indicating that it provided the best fit for the data.

Table 2. Estimates of fixed effects from the linear mixed effects models predicting MSa.PredictorsModel 1bModel 2cModel 3d
EstimatedCIP valueeEstimatedCIP valueeEstimatedCIP valueeAge–0.01–0.02 to 0.01.38–0.01–0.02 to 0.01.46–0.006–0.02 to 0.01.56Gender (male)0.13–0.16 to 0.41.390.12–0.17 to 0.40.430.11–0.19 to 0.38.49Number of children–0.1f–0.33 to –0.02.02f–0.18–0.33 to –0.02.02f–0.14–0.30 to 0.02.08Origin (migrant background)0.19–0.18 to 0.55.320.18–0.20 to 0.53.340.09–0.29 to 0.02.63Occupation (service)–0.17–0.51 to 0.17.33–0.18–0.52 to 0.15.29–0.21–0.55 to 0.14.23MEQ–0.14–0.18 to –0.09<.001g–0.14–0.18 to –0.09<.001g–0.14–0.18 to –0.09<.001gSnoozer (Yes)0.27–0.06 to 0.61.090.29–0.04 to 0.63.080.32–0.05 to 0.66.08Free day (Yes)0.210.18 to 0.24<.001g0.210.18 to 0.24<.001g0.190.16 to 0.21<.001gStringency Index—h——0.020.02 to 0.03<.001g0.020.02 to 0.03<.001gDay length———0.000.00 to 0.01.048f0.010.00 to 0.01.002iSteps (×1000)——————–0.01–0.02 to –0.01<.001gMidsteps——————0.00–0.00 to 0.01.43Intradaily variability——————–0.29–0.33 to –0.26<.001f

aThe σ2 values for models 1-3 were 1.38, 1.36, and 1.36, respectively. The intraclass correlation coefficient values for models 1-3 were 0.26, 0.26, and 0.27, respectively. The marginal R2/conditional R2 values for models 1-3 were 0.168/0.389, 0.178/0.400, and 0.179/0.400, respectively. The Akaike information criterion values for models 1-3 were 86,990.188, 86,573.060, and 86,315.145, respectively.

bIncludes demographic and occupational variables.

cIncludes model 1 + restriction and seasonal factors.

dIncludes model 2 + physical activity influences.

eItalicized values denote significance.

fP<.05.

gP<.001.

hNot available.

iP<.01.

Total Sleep Time Variability

presents the factors predicting the variability in TST. Across the 3 models, age (P=.01), number of children (P=.03), occupation (P<.001), and snoozing behavior (P=.006) emerged as significant factors. In the final model (model 3), each additional year of age was associated with a 0.01-unit increase in TST variability (95% CI 0.00-0.01; P=.01). Notably, participants with snoozing habits exhibited higher TST variability, increasing by 0.15 units (95% CI 0.05-0.27; P=.006). Each additional child was associated with a 0.06-unit reduction in TST variability (95% CI –0.11 to –0.00; P=.03). Service staff also demonstrated lower TST variability, with a reduction of 0.15 units compared with academic staff (95% CI –0.27 to –0.05; P<.001). When accounting for PA, a decrease of 1 hour in midsteps was correlated with a 0.01-unit increase in TST variability (95% CI –0.02 to –0.00; P=.03), while a 1-unit increase in IV was associated with a 0.16-unit decrease in TST variability (95% CI –0.23 to –0.09; P=.03). The LRT indicated that model 2 did not provide an improvement over the baseline model (χ22=4.78; P=.09). However, model 3 demonstrated better performance compared with the baseline model (χ25=31.95; P<.001).

Table 3. Estimates of fixed effects from the linear mixed effects model predicting TST variabilitya.PredictorsModel 1bModel 2cModel 3d
EstimatedCIP valueeEstimatedCIP valueeEstimatedCIP valuee
Age0.010.00 to 0.01.01f0.010.00 to 0.01.01f0.010.00 to 0.01.01f
Gender (male)0.110.01 to 0.21.038g0.110.01 to 0.21.03g0.100.00 to 0.21.06
Number of children–0.05–0.11 to 0.00.056–0.05–0.11 to 0.00.052–0.06–0.11 to –0.00.01g
Origin (migrant background)–0.03–0.15 to 0.10.56–0.03–0.15 to 0.09.54–0.03–0.15 to 0.08.56
Occupation (service)–0.17–0.28 to –0.05.004g–0.17–0.28 to –0.05.004g–0.15–0.27 to –0.05<.001h
MEQ0–0.01 to 0.02.550–0.01 to 0.02.570–0.01 to 0.02.60
Snoozer (yes)0.180.07 to 0.30.002g0.180.07 to 0.30.002f0.150.05 to 0.27.006f
Daylength—i——0–0.00 to 0.01.160–0.00 to 0.01.08
Stringency Index———0–0.00 to 0.00.100–0.00 to 0.01.17
Steps (×1000)——————–0.01–0.01 to 0.00.007g
Midsteps——————–0.01–0.02 to –0.00.02f
Intradaily variability——————–0.16–0.23 to –0.09.001h

aThe σ2 values for models 1-3 were 0.24, 0.24, and 0.24, respectively. The intraclass correlation coefficient values for models 1-3 were 0.14, 0.14, and 0.14, respectively. The marginal R2/conditional R2 values for models 1-3 were 0.059/0.195, 0.060/0.194, and 0.068/0.200, respectively. The Akaike information criterion values for models 1-3 were 5458.745, 5457.957, and 5436.793, respectively.

bIncludes demographic and occupational variables.

cIncludes model 1 + restriction and seasonal factors.

dIncludes model 2 + physical activity influences.

fP<.05.

gP<.01.

hP<.001.

iNot available.

eItalicized values denote significance.

Midsleep Variability

presents the factors predicting the variability of MS. Across the 3 models, the number of children (P=.004), snoozing behavior (P=.01), midsteps (P=.008), and IV (P=.001) emerged as significant factors. For each additional child, MS variability was reduced by 0.10 units (95% CI –0.16 to –0.03; P=.004). In all models, being a snoozer correlated with increased MS variability. Specifically, snoozers experienced a 0.17-unit increase in MS variability compared with nonsnoozers (95% CI 0.03-0.31; P=.01). To better understand the characteristics of snoozers, we conducted an analysis based on Mattingly et al’s study []. Interestingly, our results revealed that age (P=.02) and chronotype (P=.002) were significant factors in predicting snoozing behavior. The full results are detailed in .

In the full model, including PA variables, midsteps also became significant. Each hour increase in midsteps was associated with a 0.02-unit decrease in MS variability (95% CI –0.04 to –0.00; P=.008). However, the more complex models did not show a significant improvement over the baseline model, as indicated by the LRT (model 2: χ22=1.00; P=.60/model 3: χ25=10.17; P=.07).

Table 4. Estimates of fixed effects from the linear mixed effects model predicting MS variabilitya.PredictorsModel 1bModel 2cModel 3d
EstimatedCIP valueeEstimatedCIP valueeEstimatedCIP valuee
Age0.00–0.00 to 0.01.410.00–0.00 to 0.01.410.00–0.00 to 0.01.55
Gender (male)0.10–0.03 to 0.23.120.10–0.03 to 0.23.120.09–0.04 to 0.22.17
Number of children–0.09–0.16 to –0.02.01f–0.09f–0.16 to –0.02.01f–0.10–0.16 to –0.03.004g
Origin (migrant background)–0.05–0.19 to 0.11.49–0.05–0.19 to 0.11.49–0.05–0.20 to 0.09.45
Occupation (service)–0.12–0.26 to 0.02.11–0.12–0.25 to 0.02.11–0.11–0.25 to 0.02.10
MEQ0.00–0.02 to 0.02.960.00–0.02 to 0.02.980.00–0.02 to 0.02.94
Snoozer (yes)0.200.06 to 0.35.006g0.200.06 to 0.35.006g0.170.03 to 0.31.01f
Daylength—h——0.00–0.00 to 0.01.340.00–0.00 to 0.01.26
Stringency Index———0.00–0.00 to 0.00.940.00–0.00 to 0.00.89
Steps (×1000)——————0.00–0.01 to 0.01.41
Midsteps——————–0.02–0.03 to –0.00.008g
Intradaily variability——————–0.09–0.21 to –0.0209

aThe σ2 values for models 1-3 were 0.59, 0.59, and 0.59, respectively. The intraclass correlation coefficient values for models 1-3 were 0.09, 0.09, and 0.09, respectively. The marginal R2/conditional R2 values for models 1-3 were 0.034/0.120, 0.034/0.120, and 0.038/0.122, respectively. The Akaike information criterion values for models 1-3 were 8679.371, 8682.369, and 8679.197, respectively.

bIncludes demographic and occupational variables.

cIncludes model 1 + restriction and seasonal factors.

dIncludes model 2 + physical activity influences.

eItalicized values denote significance.

fP<.05.

gP<.01.

hNot available.

DiscussionPrincipal Findings

In this study, we used a year-long longitudinal data set from 112 working adults and identified several significant relationships between changes in sleep over time and various factors, including restriction policies, seasonal changes, PA, and sociodemographics. We found that more stringent restrictions were associated with increased TST and delayed MS. Additionally, seasonal factors played a notable role: increased day length was linked to reduced TST and delayed MS. Changes in work arrangements, particularly the shift to remote work, directly impacted individuals based on their occupations and sleep patterns. Academic personnel, with more flexible schedules, slept less and exhibited greater variability in TST compared with service personnel, who had more structured work schedules. Additionally, individuals identified as “snoozers” had more flexible sleep schedules with greater variability in both TST and MS compared with nonsnoozers. Moreover, activity patterns played a significant role: exercising later in the day was associated with longer TST and reduced variability in both TST and MS. To contextualize our findings within the broader scope of sleep during the pandemic, the following section details our results and compares them with previous studies.

Demographic Factors

Previous research has highlighted several epidemiological factors affecting sleep patterns, notably, age, gender, and chronotype. Consistent with previous studies, we found that older individuals tend to sleep less [,]. However, our findings reveal a correlation between older age and increased TST variability, which contrasts with prior results []. The variance in the observed correlations may be due to our study using objective sleep measures, while [] relied on self-reported data. Additionally, we found no significant association between MS variability and age. Regarding gender differences, our study shows that males tend to have shorter and less consistent TST compared with females. While the shorter TST among males is well-documented [,], evidence regarding gender disparity in TST variability is inconsistent. For instance, an actigraphy study on a middle-aged cohort found that females exhibited greater TST variability than males []. Conversely, a survey-based study on university students [] reported no gender differences in TST variability. Additionally, our study observes that parental duties significantly impact sleep patterns. Parents typically exhibited earlier sleep times and more consistent TST and MS than nonparents. The underlying reasons for these observations remain uncertain, but one hypothesis is that parents’ sleep/wake schedules are more stable due to the need to synchronize their sleep patterns with those of their children. While the specific relationship between parenting and sleep pattern variability has not been extensively studied, research on cohabitation suggests that living with others can influence sleep patterns by reducing variability in sleep timing and duration [,]. This context highlights how factors related to shared living arrangements, such as parenting, can contribute to greater sleep pattern regularity.

Snoozing Behavior

We observed higher variability in TST and MS among individuals identified as “snoozers.” Interestingly, younger individuals and those with an evening chronotype are more likely to be “snoozers,” suggesting an interplay between age, chronotype, and snoozing habits. The natural sleep-wake patterns associated with an individual’s chronotype may influence their tendency to snooze alarms. Morning types, who wake up earlier, might not feel the need to snooze as much because their schedules align better with societal norms, in contrast to evening types.

Clinically, snoozing can be linked to prolonged sleep inertia, a state of reduced alertness upon waking []. Morning types (with high MEQ scores) may be less pr