To analyze sleep and alcohol use patterns before and during federal holidays and during DST transitions, data collected from adult subscribers to the wearable biometric device platform WHOOP, Inc. (Boston, Massachusetts, United States) were analyzed retrospectively. During the 1-year study interval (1 May 2020 to 1 May 2021), WHOOP users recorded sleep-wake data via continuous monitoring from the WHOOP device, with optional daily surveys on health behaviors including alcohol use. With WHOOP membership growing rapidly, the study interval was chosen to optimize the balance of a growing participant pool with sufficient historical data (i.e., a calendar year). The age range chosen represents working-age individuals, who were most likely to have their work schedules impacted by public holidays (United States Census Bureau, 2021). Participants were chosen from only the US to avoid confounding the data with international differences in holiday observances. Participants without sleep-wake data on any of the public holidays were excluded from the analysis. Additionally, individual sleep episodes for which complete data were not available (for example, due to loss of device charge) were not counted towards participants’ required minimum of 70% compliance for inclusion.

Inclusion criteria for the primary analytic sample (sleep data) were: 1) being a US-based member of the WHOOP platform who collected data for 255 (70%) or more days out of the year-long study interval from 1 May 2020 through 1 May 2021; and 2) being between the ages of 18 and 65 years for the entirety of the data collection interval. From the primary analytic sample, a subsample of participants who responded to optional daily surveys about alcohol consumption for 255 (70%) or more days and all holidays during the data collection interval were included in the alcohol use subsample.

Holidays analyzed included all 10 US public holidays as observed by the Federal Reserve System (Federal Reserve Bank of Chicago, 2021): Memorial Day (25 May 2020), Independence Day (also known as Fourth of July; 4 July 2020), Labor Day (7 September 2020), Columbus Day (also known as Indigenous Peoples’ Day or Indigenous Peoples Day; 12 October 2020), Veterans Day (11 November 2020), Thanksgiving Day (26 November 2020), Christmas Day (25 December 2020), New Year’s Day (1 January 2021), Martin Luther King Jr Day (also known as Martin Luther King Junior’s Birthday; 18 January 2021), and Presidents’ Day (also known as Washington’s Birthday; 15 February 2021). DST transitions included setting clocks earlier 1 h in the fall (“Fall Back”; 1 November 2020) and setting clocks later by 1 h in the spring (“Spring Ahead”; 14 March 2021).

The Monash University Human Research Ethics Committee reviewed and approved the study protocol (#26304). All participants in the study consented at registration with WHOOP to the use of their data for the purposes of scientific research and could withdraw consent at any time.

Sleep data were derived from data captured using a wrist-worn multi-sensor (tri-axial accelerometer, optical heart rate sensor, capacitive touch sensor and ambient temperature sensor) on the WHOOP device (Berryhill et al., 2020). Sleep variables included in this analysis were sleep duration, sleep onset, sleep offset, and the consistency of sleep timing. Sleep consistency was measure using a proprietary method adapted from the previously validated Sleep Regularity Index (SRI) (Phillips et al., 2017), which is the percentage of concordance of individuals being in the same state (asleep versus awake) at different timepoints. Whereas the SRI compares two total timepoints 24 h apart, the WHOOP sleep consistency compares 24-h timepoints over a 4-day interval (Czeisler et al., 2022). Additional timepoints were included to incorporate additional recent historical sleep-wake data into the WHOOP sleep consistency score, with the intervals assigned progressively lower weights as they become further apart. Four sleep outcomes were included given increased recognition of the multidimensionality of sleep (Buysse, 2014).

The WHOOP device has undergone multiple performance assessment in young, healthy individuals (Berryhill et al., 2020; Miller et al., 2020, Miller et al., 2021) and has demonstrated acceptable two-stage (wake or sleep) categorization to automatically detect sleep, with 86%–89% agreement with polysomnography (Miller et al., 2020; Miller et al., 2021).

For the alcohol use variable, subscribers were prompted in optional daily surveys delivered through the WHOOP mobile application, which included a question on alcohol use during the previous day, reported as a binary (yes/no).

First, for each holiday, the day preceding the holiday, and DST transition, the percent change to each sleep outcome relative to a baseline was calculated on the primary analytic sample, as was the percent change in the point prevalence of alcohol use within the subsample of eligible participants. To determine a baseline for holiday and DST transition comparisons, sleep and alcohol data from the same day of the week for 4 weeks preceding and 4 weeks following the date of each public holiday were averaged. For example, data for Thanksgiving Day (Thursday, 26 November 2020) were compared to data from the four calendar Thursdays preceding and four following Thanksgiving Day. The baselines were comprised of the same day of the week to account for population-level variance in sleep behaviors based on the day of the week (Capodilupo and Miller, 2021). A local rather than a complete-year baseline was selected to avoid the confounding effect of seasonal variation in sleep and alcohol use patterns and the effect of fluctuating severity of the coronavirus disease 2019 (COVID-19) pandemic and its associated mitigation policies (Capodilupo and Miller, 2021; Czeisler et al., 2022).

Days preceding each holiday (herein, eves) were analyzed separately to account for the potential impacts on both sleep and behavior leading into as well as occurring on a holiday. For example, we hypothesized that in anticipation of Labor Day, which occurs on a Monday, the preceding Sunday night might have higher prevalence of alcohol consumption, shifted sleep timing, and overall extended sleep relative to a typical Sunday night, but Labor Day itself may resemble a non-holiday Monday night in which sleep and alcohol patterns are indifferent from typical pre-work night patterns. Christmas Eve and Day and New Year’s Eve and Day are exactly 1 week apart and therefore a necessary exception to the described baseline methodology. These days were excluded from each others’ baseline calculations to avoid confounding the baseline with another holiday.

Daily sleep duration, consistency (scored on a 0 to 100 scale), and timing were calculated for participants over the year-long study interval. An average by day was taken across participants for each sleep metric to calculate the variance in metrics over the duration of the study. Sleep episodes were assigned to the calendar date in the local time zone in which the sleep ended. Both going to bed before and after midnight are common. Having used the date of wake avoided misattributing sleep episodes to different days based on the episodes beginning before or after midnight. For example, sleep episodes concluding the morning of 2 January 2020 were treated as if the sleep episode were initiated on date 1 January 2020. If the sleep episode initiated in the early morning hours of 2 January 2020, the episode was still assigned to 1 January 2020. The average values for each participant were compared to their individual baseline values and used to calculate percent changes for sleep duration and sleep consistency for both the public holidays and their eves. Sample estimates for percent change were created by averaging participant estimates. The 95% confidence intervals based on between-subjects variability were estimated as well.

The alcohol data were analyzed using the daily percentage of subscribers who reported alcohol consumption in the optional survey out of the total subscribers who responded to the question. Individuals included in this subset of analysis who did not respond on a given day were omitted from the sample for that day. Similar to the method of analyzing sleep data, a baseline using the same day of week from 4 weeks prior to and 4 weeks after each holiday was created by taking an average of the 8 days. The binomial proportion confidence interval was calculated for both the holidays and their baselines. The percent change from baseline was calculated for both the public holidays and their eves. The propagation of error for the percent change was computed, accounting for the correlation of measurement error and correlation of uncertainties in the group averages. Two-sided p-values were obtained from confidence intervals for a difference (Altman and Bland, 2011). The changes in individual measures (sleep duration, sleep consistency, sleep onset, sleep offset, and alcohol use) for each holiday and eve were then ranked among the holidays and eves in each measure from most positive to most negative change compared to baseline.

Second, to evaluate whether there was a relationship between alcohol use and changes in sleep patterns, for each holiday, eve, and DST transition, analyses of paired percent changes in alcohol use and 1) sleep duration and 2) sleep consistency were conducted in the subsample of eligible participants.

All analyses were conducted using the Python programming language version 3.6.2 (Python Software Foundation). The threshold for statistical significance was set at 0.05.

The mean absolute change (average percent change in absolute value) from baseline sleep duration during US public holidays during the study interval was 1.9%. Including holiday eves, the absolute change was 2.1%. Overall, 15 of the 20 (75%) holidays and eves were associated with significant increases in average sleep duration, with an average increase of 2.5% (Table 1). There were significant decreases in sleep duration on two of the 20 (10%) holidays and eves, with an average decrease of 1.4%. Increases and decreases in average sleep duration were observed following the Fall Back and Spring Ahead DST transitions, respectively. The largest average increases from baseline sleep duration occurred on Thanksgiving Day [5.2% ± 0.3 (7h31 m vs. 7h10 m)], Christmas Day [5.1% ± 0.3 (7h38 m vs. 7h18 m)], the Fall Back DST transition [4.9% ± 0.3 (7h39 m vs. 7h19 m)], and New Year’s Day [4.8% ± 0.3 (7h37 m vs. 7h17 m)]. The largest average decreases from baseline sleep duration occurred on New Year’s Eve [−2.5% ± 0.3 (6h58 m vs. 7h10 m)] and the Spring Ahead DST transition [−0.8% ± 0.3 (7h07 m vs. 7h12 m)], with Memorial Day [−0.3% ± 0.2 (7h03 m vs. 7h05 m)] as the only other public holiday with a significant decrease in sleep duration. Changes to sleep duration on Christmas Eve, Columbus Day, and Veterans Day were not significant (p > 0.05) (Tables 1, 2).

The mean absolute change from baseline sleep consistency during US public holidays was 2.1%. Including eves, the absolute change was 2.8%. The eight holidays with significant increases in sleep consistency had an average increase of 1.9%. Nine of 20 (45%) holidays and eves were associated with significant decreases in sleep consistency, with an average decrease of 4.2%. New Year’s Eve had the largest change [−13.8% ± 0.3 (65.0 vs. 75.8 out of 100)] (Table 1). The next largest changes were observed on New Year’s Day [−6.3% ± 0.3 (68.3 vs. 73.2)], the Fall Back DST transition [−6.1% ± 0.3 (67.3 vs. 71.9)], and Christmas Eve [−5.0% ± 0.3 (71.5 vs. 75.7)]. The night before Labor Day had the highest average increase in sleep consistency from baseline [2.6% (75.3 vs. 73.1)]. Sleep consistency differences spanned a larger range (20.2%) when compared to sleep duration differences (16.5%). The percent changes in sleep consistency on Columbus Day, the night before Martin Luther King Jr Day, and the Spring Ahead DST transition were not significant (p > 0.05) (Tables 1, 2).

The mean absolute change from baseline in the timing of sleep onset was 11.8 min on public holidays. Including their eves, sleep onset changed on average 16.2 min. Most changes in sleep onset timing [12 of 20 (60%)] were delays in sleep onset, indicating a population-level bedtime that was, on average, 22.1 min later on those holidays (Table 1). The largest average delay in sleep onset was on New Year’s Eve, on which the average change was an 88.9-min (±3.2) delay in sleep onset time, from a baseline average of 22:33 to 00:01 on New Year’s Eve (Table 2). Later sleep onset times were also observed with Spring Ahead and Fall Back DST transitions. Seven of the 20 (35%) holidays and eves were associated with significantly earlier sleep onset times compared to baseline—on average 8.3 min earlier (Table 1). Each of the holidays that had an earlier average time of sleep onset, indicating an earlier average bedtime, were on Monday holidays.

The mean absolute change from baseline in the timing of sleep offset on public holidays was 21.5 min. This was almost twice the average change of sleep onset timing. Including the holiday eves, the average change was 24.5 min 13 of 20 (65%) holidays and eves were associated with a later sleep offset—on average 32.8 min later. Both Spring Ahead and Fall Back DST transitions were associated with delayed sleep offset as well, even following the Fall Back DST transition when an extra hour was added to the night. On New Year’s Eve (the morning of New Year’s Day), offset timing delayed 78.1 min (±3.1) later—from a 06:39 baseline to 07:56 on the morning after New Year’s Eve—the largest change in sleep offset (Table 1). We observed earlier sleep offsets (on average 8.9 min earlier) on seven of the 20 (35%) holidays and eves, each of which were Monday holidays (Table 1).

The mean absolute change from the baseline alcohol use on holidays was 32.0%. Including their eves, the mean absolute change was 36.8%. We observed a significantly increased prevalence of alcohol consumption on 14 of 20 (70%) holidays and eves, with an average increase of 46.7% (Table 1). The majority of significant changes in the number of alcohol consumers were increases, with one exception of New Year’s Day, when alcohol consumption averaged 9.9% lower than surrounding same days of the week. More than twice as many users reported alcohol consumption on New Year’s Eve [138.0% ± 6.7 (74.2% vs. 31.2%)] and Christmas Eve [111.6% ± 4.7 (66.7% vs. 31.6%)] compared to the baseline (Tables 1 and 2). Regarding DST transitions, the point prevalence of alcohol use increased by 9.5% ± 0.85 during the Fall Back DST transition (63.2% vs. 57.7%), and did not differ significantly during the Spring Ahead DST transition (56.7% vs. 56.9%).

The analysis of paired alcohol use and each of sleep duration and sleep consistency revealed a non-significant correlation (r = −0.211; p = 0.34) between changes in alcohol use and changes in sleep duration (Figure 3A), yet a strong, negative correlation (r = −0.635; p = 0.001) between alcohol use and sleep consistency, with an increased percentage of alcohol use associated with decreased sleep consistency across the sample (Figure 3B).

Strengths of this study include the large sample size and the utilization of sleep-wake data that were 1) high-resolution (i.e., continuous 1 Hz recordings to generate 30-s epoch resolution sleep stage estimates) 2) largely complete (i.e., minimal missingness, with data for an average of 87.9% of the study interval), and 3) comprehensive (i.e., capturing multiple dimensions of sleep patterns, including the duration and timing of sleep), collected from a sleep wearable device validated against polysomnography (Berryhill et al., 2020). Additionally, in the alcohol use subsample, daily assessment of the variable minimized potential recall bias.

Limitations of this study may include selection bias potentially amplified by the Big Data Paradox, self-reported alcohol data, and potential unmeasured effects of the COVID-19 pandemic. Regarding selection bias, the data were retrospectively analyzed from subscribers to the WHOOP platform. Given that WHOOP is a personalized digital fitness and health platform, participants may be more informed about sleep and its impact on performance compared to the general population. Moreover, demographic characteristics outside of sex and age were not collected, so beyond the overrepresentation of men in the study sample, the extent to which the sample is representative of the US adult population is unknown. Sample under-representativeness can be especially concerning with large datasets such as the one used for this analysis, as potentially incorrect point estimates of biased samples can be compounded by substantially underestimated widths of confidence intervals afforded by large sample sizes—the Big Data Paradox (Meng, 2018; Bradley et al., 2021). Regarding self-reported alcohol data, surveys could be subject to social desirability and other biases. For alcohol data in particular, as survey responses were opt-in for subscribers; subscribers who rarely drink alcohol may have been less likely to track their alcohol use. The data shown should be interpreted as relative drinking prevalence among those who drink, which may overrepresent the true prevalence of population-level drinking. On the other hand, if the findings in our sample were generalizable, the increased prevalence on holidays may underestimate the population-level increase. Finally, the study interval analyzed occurred entirely during the COVID-19 pandemic, and the associated stay-at-home orders and social distancing mandates may have impacted the behavioral changes ordinarily associated with holiday observance. However, the selection of the study interval as beginning on 1 May 2020 was partially to avoid the interval during which pandemic effects on population movement would be strongest. Mobility data reveal that population movement increased rapidly following the lifting of the first stay-at-home order, which occurred prior to the start of the study interval (24 April 2020), suggesting that the peak impact of mobility restrictions due to the pandemic had passed (Moreland et al., 2020). Nevertheless, further research following the complete lifting of social distancing mandates will be required to understand how these data were impacted by the present mandates.

The results of these analyses should be considered with the understanding that a large data set with nonrandom recruitment was used. As such, due to the nature of WHOOP platform membership at the time of the study, most participants were male. The big data paradox is minimized in this analysis given the use of real-world data to assess discrete hypotheses and an analysis plan that was determined before beginning research and carried through to completion. Although estimating expected systematic differences in outcomes with a more representative sample is extremely difficult, we found no significant differences to report when stratifying the presented analyses by age and sex groups. Additionally, results could be somewhat confounded by behavior changes of participants due to COVID-19 pandemic as the time interval analyzed coincided with an active interval of COVID-19. This could lead to an underestimation of the impact of holidays as there were later bed times and wake times and longer sleep duration overall during this interval (Robbins et al., 2021). Sleep consistency had also increased during the pandemic (Czeisler et al., 2022) potentially leading to an overestimation in the changes in sleep consistency associated with holidays.