Participants.

Twenty-four healthy adults, ages 18 to 26 years (mean ± SE: 20.6 ± 0.35 yr, 54% female) completed a counterbalanced, randomized crossover design (described below). Participants abstained from any psychoactive drugs (including caffeine and alcohol) for 24 h before each study session. Participants’ habitual sleep–wake rhythm was monitored for 3 nights prior to study participation verified by sleep logs and actigraphy (a wristwatch movement sensor, sensitive to wake and sleep states). Exclusion criteria, assessed using a prescreening questionnaire, included a history of sleep disorders, neurologic disorders, closed-head injury, Axis 1 psychiatric disorders, history of drug abuse, and current use of antidepressant or hypnotic medication. Participants were also excluded from entering the study if they reported: sleeping less than 7 h per night, traveling across time zones in the past month, doing shift work in the past year, having their bedtime and/or wake time change by more than 2 h more than 3 times a week or consuming 3 or more daily caffeine-containing drinks. The study was approved by the local human studies committee of the University of California Berkeley, with all participants providing written informed consent.

Experimental design.

Following successful completion of screening, participants entered a randomized crossover study design involving two sessions, conducted in a counterbalanced order: one after a rested night of sleep and one after 24 h of sleep deprivation. Participants were randomly assigned to start with either a sleep-deprived or a sleep-rested session followed by the crossover session, separated by at least 7 days. Upon entering the study, and prior to any sleep manipulation, participants completed the Interpersonal Reactivity Index (IRI [82]) in order to measure interindividual differences in empathy.

Within each session, helping behavior was assessed in the morning (between 9 to 11 AM), using the helping behavior questionnaire (details below). In addition, mood states were measured twice in each session using the Positive and Negative Affect Schedule (PANAS [83]), first in the evening prior to any sleep manipulation (between 8 to 10 PM) and again in the morning following both sleep sessions (between 8 to 9 AM).

In the sleep-deprived session, participants arrived at the laboratory at 9:30 PM and were continuously monitored throughout the enforced waking period by trained personnel. During the sleep deprivation period, participants engaged in a limited set of activities such as studying, being online, reading, or watching movies. The following morning at approximately 10:00 AM (±45 min), participants performed a social judgment fMRI paradigm inside the scanner (details below). In the sleep-rested session, participants arrived at the lab at 7:00 PM and were prepared for an ambulatory electroencephalography (EEG) polysomnography recording, after which they were sent home to sleep allowing for more naturalistic measurement. The next morning, participants returned to the laboratory and had the electrodes removed. Participants then performed the same activities as those described above in the sleep deprivation condition, starting at the same matched circadian time.

Helping behavior assessment.

Helping behavior was assessed using a 40-item questionnaire sourced from the Self-Report Altruism Scale [84], a scale that is also part of the Prosocial Personality Battery [85,86]. Both are frequently used in studies assessing prosocial and altruistic helping behaviors [71,86–91]. Each item described a social situation requiring various types of help (e.g., “If I was in a hurry to get to work and someone stopped me to ask for directions I would…” or “I would help a stranger struggling with her grocery bags to carry them,” for a full list of the items see Table B in S1 Text). For each statement, participants were asked to indicate how they would respond to the social situation at this moment in time using a 5-box vertical scale ranging from “I would definitely help” or “I would stop to help” (wording was tailored to each scenario, see Table B in S1 Text) to “I would not help” or “I would ignore them.” The reliability of this scale in this sample was high (Cronbach’s alpha = 0.82 in the sleep-rested session and 0.87 in the sleep-deprived session).

The helping requests in the assessment were equally divided between strangers and familiar others to control for the known impact of familiarity on helping behavior [2] (see Note A in S1 Text for an analysis of familiarity effects). For example, “I would offer my seat on a crowded bus to a 60-year-old woman” reflects a social scenario involving a stranger while “If a coworker who lives near me asked me to give him/her a ride home I would…” involves a familiar other.

In addition to these social situations, 10 nonsocial control items that targeted factual questions were included in the questionnaire (e.g., “Is Denmark larger than Sweden?”), requiring a yes/no binary reply. Social items only were subsequently reversed scored (such that higher scores denote greater helping, range 1 to 5) and averaged for each participant. One participant was excluded from the analysis due to partial completion of the questionnaire (35% missing values). The helping behavior questionnaire included 2 separate balanced versions (each with different 30 social and nonsocial statements), such that participants never replied to the same questionnaire twice. Versions were counterbalanced across participants and sessions. Notably, no significant order effects were found when comparing the participants who completed their sleep deprivation session first, relative to those who completed it second (mean difference = −0.05, 95% CI = [−0.68, 0.58], t = −0.18, P = 0.9) and similarly for the order of the sleep-rested session (mean difference = −0.32, 95% CI = [−0.82, 0.17], t = −1.37, P = 0.2).

Effortful behavior control task.

Sleep loss can impact the willingness to exert effort [57] and could therefore indirectly influence the desire to help others. To explore this factor of effort, participants in Study 1 performed an incentivized effort task in both the sleep-rested and sleep deprivation sessions (similar to a grip effort task [92]). During each session, participants were required to use their right index finger to press the “m” key at variable speeds for a certain amount of timed effort intervals.

The effort task involved 3 blocks. The first block lasted 3 min and asked participants to press the “m” key at a speed faster than 4 presses per second. The second block required participants to choose between hard trials (4 to 6 key presses per second) or easy trials (2 presses per second) for a total of 20 trials. Here, participants were offered monetary rewards based on their choice: 10 cents for an easy trial and 10 to 80 cents for hard trials. In the third and final block, participants were asked to press “m” at a constant speed of 4 strokes or faster per second for as long as they wished, up to a maximum of 8 min. Each minute rewarded participants with $1 for their effort, such that the maximum reward available was $8. Participants were informed they could stop at any time during this last block and would be rewarded based on the time they had already spent. This “time on task” variable indicated the level of sustained effort participants were voluntarily willing to exert and was therefore used to assess participants’ volitional effort levels, subsequently factored into the main analyses examining helping. Results of the time on task analysis indicated that participants chose to exert less effort following sleep deprivation (time on task duration in minutes (means ± SE: SR = 5.82 ± 0.68, SD = 4.03 ± 0.62, a 31% reduction; P = 0.015, d = −0.57), yet as described in the Results, differences in helping behavior remained significant after statistically accounting for changes in effort.

fMRI social judgment paradigm.

The relationship between neural activity within the social cognition network and prosocial behavior was examined using a well-documented mentalizing task, wherein participants were asked to think about personal attributes of social targets [17,45–49]. This allowed for an ecological assessment of social cognition brain activity without any overt biases of knowing the study’s primary motivation which was that of helping choices. Such a design offers an orthogonal approach to examining the neural correlates of prosocial behavior triggered by empathy [20,21] or incentivized donation choices [22,23], by targeting a key foundational process of prosocial behavior: the comprehension and understanding of another’s mind [37,39].

During fMRI scanning, participants viewed 48 experimentally controlled information cards depicting various adults based in the US (including name and profession, represented by a silhouette black and white image to avoid visual face biases, see example in S4 Fig). Participants were asked to make explicit personality trait judgments for each individual presented in each trial, a process that evokes a robust engagement of the social cognition brain network [46,47,49]. For each information card trial, participants made social assessment ratings regarding the level of perceived competence or warmth of each individual based on the details provided to the participant on that trial (ranging from 0 to 6, warmth and competence ratings randomized across runs). There were no significant differences between the sleep deprivation and sleep-rested conditions for scores of either competence (SR = 3.28 ± 0.09, SD = 3.27 ± 0.09, P = 0.9) or warmth (SR = 2.95 ± 0.11, SD = 2.84 ± 0.08, P = 0.6) ratings.

In each trial, the information card was presented for 3 s, followed by the social judgment response screen for 4 s and an inter-trial fixation period (jittered, 2 to 6 s for optimal event-related fMRI design). For the nonsocial control trials, a total of 32 additional information cards of equivalent size were used that depicted various objects using a black and white image (e.g., vacuum cleaner, guitar). In these object trials, participants were asked to rate how old they believed the objects to be, using a similar scale of 0 (very old) to 6 (very new). As with previous such fMRI paradigms [93,94], these nonsocial cards provided an on-task comparison with the human information cards (i.e., human > object judgment), allowing for the discrimination of brain activity that is unique to the processing of social stimuli above and beyond the presentation of a visual object or other nonspecific task demands.

The in-scanner social judgment paradigm contained 2 repeatable versions, each including a different set of 80 information cards (48 humans and 32 objects, though see Note B in S1 Text for a control fMRI analysis in which human and object trials are equally balanced). The version used was counterbalanced across participants, such that each version was viewed in a sleep-rested session for half of the participants and in a sleep-deprived session for the others. In each session, the information cards were presented in 2 runs, with human and object cards presented in a randomized order within each run. The start of each run contained a 10-s fixation block, allowing for a steady-state equilibrium of the BOLD fMRI sequence. Stimulus presentation and response collection were controlled by PsychoPy [95].

Similar to prior studies, helping behavior was assessed outside the scanner [20,21] (see Helping behavior assessment above for full details) to enable the use of real and multidimensional scenarios of helping others [6], rather than focusing on a predetermined domain of prosocial behavior (e.g., providing financial support [22,56]). This assessment further provided an ecological measure of helping choices, allowing for time-dependent processes (e.g., personal values, attitudes, dispositions) to come into play [96], unconstrained by typical event-related fMRI limitations.

fMRI acquisition and analysis.

Blood oxygenation level-dependent contrast functional images were acquired with echo-planar T2*-weighted (EPI) imaging using a Siemens 3 Tesla MRI scanner with a 12-channel head coil. Each image volume consisted of 37 descending 3.5 mm slices (96 × 96 matrix; TR = 2,000 ms; TE = 22 ms; voxel size 3.5 × 3.5 × 3.2 mm, flip angle = 50°, 0.3 mm interslice gap). One high-resolution, T1-weighted structural scan was acquired at the end of the sleep-rested session (256 × 256 matrix, TR = 1,900; TE = 2.52; flip angle = 9°; FOV 256 mm; 1 × 1 × 1 mm voxels).

Preprocessing and data analysis were performed using fMRIprep v1.25 [97] and Statistical Parametric Mapping software implemented in Matlab (SPM12; Wellcome Department of Cognitive Neurology, London, UK). Using fMRIprep, T1-weighted (T1w) images were corrected for intensity non-uniformity, skull-stripped, and spatially normalized to the ICBM 152 Nonlinear Asymmetrical template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter, and gray matter were then performed on the brain-extracted T1w. Functional image preprocessing included coregistration with the T1w structural scan, slice-time correction, resampling to MNI152NLin2009cAsym standard space (voxel size 2 × 2 × 2 mm), and spatial smoothing (6-mm Gaussian kernel), using the fMRIprep pipeline. To control for head motion and physiological artifacts, 13 nuisance regressors were calculated as well, including 6 rotation and translation parameters, framewise displacement (FD), and the first 6 principal components of the anatomical CompCor pipeline (taking into account CSF and white matter signals).

Following preprocessing, a standard general linear model (GLM) was specified using SPM for each participant to investigate the effects of interest. Contrasts were created at the first level focusing on human < > object judgments to elucidate regions sensitive to social cognition. The resulting contrasts were then taken through to a second level, random-effects analysis to assess group-level effects, examined using a paired t test (Sleep Rested < > Sleep Deprived). Analyses focused a priori on activity in a set of brain regions comprising the social cognition network, regions that have been implicated in studies of social cognition and helping behavior [20–23]. These regions include: bilateral TPJ [MNI coordinates: −48; −60; 32 left, 56; −60; 22 right], dorsomedial PFC [2; 56; 22], ventromedial PFC [2; 44; −18], precuneus [2; −52; 32], bilateral middle temporal sulcus [−52; 2; −26 left, 56; −2; −24 right], and right inferior frontal gyrus [52; 30; −8].

Regions of interest (ROIs) were derived using the NeuroSynth framework [98], a large-scale automated meta-analysis tool for neuroimaging data. An activation map was calculated from 80 social cognition studies that were corrected for multiple comparisons (FDR < 0.01). A minimal size of 150 voxels per cluster was further applied to the activation map, resulting in a mask of 8 binarized ROIs (see Table C in S1 Text for a complete list of ROIs). Analyses focused on averaged activity across all ROIs in order to examine task-evoked network-level activation and avoid multiple comparisons across separate ROIs [99]. Using the binarized mask, condition-specific (human > object) activity was extracted and then compared between sleep conditions. Beyond these network-level ROIs, non-a priori whole-brain results are provided in Table D in S1 Text for results-reporting completeness but are not discussed further.

Finally, to examine the issue of specificity, two post hoc analyses were conducted. The first explored activity within six additional standard brain networks using the same analysis steps described above. The networks were derived from a validated seven-network parcellation [100] and included the limbic, frontoparietal, ventral/dorsal attention, somatomotor, and visual networks (absent the default mode network, given its substantial overlap with regions of the social cognition network [101,102]). None of the activation patterns in these networks showed a significant main effect of sleep deprivation during the social judgment task, and similarly, sleep loss changes in activity within each of these networks did not predict the degree of impairment in helping behavior (all P > 0.1, corrected for multiple comparisons, see Table A in S1 Text).

The second analysis examined the possible influence of attention. Reaction time is often used as a proxy for attentiveness and attention [103–106] and is commonly included as a cofactor variable in typical fMRI analyses to control for attention [107–109]. Using this same approach, single-trial reaction times during the fMRI scanning task were extracted per individual and per session (sleep rested, sleep deprived). These reaction times were then included as a “nuisance regressor” in the single-subject level of the fMRI analysis, further controlling for measures of movement and physiological confounds as in the main analysis. When controlling for trial-specific reaction times, the results of the main findings remained similarly significant, evincing a significant and selective reduction in social cognition network activity following sleep deprivation, relative to the sleep-rested condition (mean change = −0.46 ± 0.17, P = 0.02, d = 0.7). Moreover, activity in the social cognition brain network remained significantly associated with helping behavior when reaction times were added as a cofactor in the analysis model (R = 0.43, P = 0.04).

Sleep recordings.

Sleep was recorded using standard polysomnography including EEG, electromyography (EMG), and electrooculography (EOG) recordings. EEG was recorded from 11 scalp electrodes (Fp1, Fp2, F3, F4, F7, F8, C3, C4, P3, P4, and Oz; International 10–20 System), referenced to left and right mastoid (A1, A2). EEG signals were sampled at 200 Hz. Polysomnographic recordings were scored according to standard criteria [110]. Sleep statistics are provided in Table E in S1 Text and conform to population norms for this age range [111].

Participants.

Study 2 tested whether more modest night-to-night variability in self-reported sleep efficiency and sleep duration predicted day-to-day changes in helping behavior the next day. Unlike the experimental sleep manipulation of Study 1, Study 2 examined sleep variations under free-living conditions. A total of 171 participants (age 36.96 ± 0.73 yr, 41.2% female) signed up for this 4-day study using Amazon Mechanical Turk (MTurk)—a platform where individuals can perform online tasks for a specified reimbursement (here, $4.25 to 5.75 depending on the final number of daily surveys). Enrollment was restricted to those with IP addresses in the US and a prior online MTurk approval rating of 95% or higher. Additional exclusion criteria included a current diagnosis of an Axis 1 psychiatric disorder and/or the confirmed diagnosis of a sleep disorder. Participants were also excluded from further analysis if they completed only 1 daily survey, which would otherwise have prevented sufficient variability in assessing within-person effects (and see the robustness assessment regarding the main regression using a minimum of 3 or 4 nights of data in Note C in S1 Text). Furthermore, participants whose sleep logs reported extreme sleep duration values were similarly excluded (less than 3 h or more than 12 h). The final sample, therefore, included 136 participants (mean age ± SE = 37.83 ± 0.87 yr, 41.9% female), yielding a total of 441 valid observations across the study period.

Study design.

Following recruitment, participants were asked to complete validated daily sleep diaries (see Table F in S1 Text), quantifying their sleep across four consecutive nights. The next day, participants completed an assessment of helping behavior using a shorter form of the helping questionnaire applied in Study 1 and described above. The short-form version included 10 social items depicting requests for help, presented in random order, and counterbalanced for familiarity such that requests from strangers and familiar others were equally represented in each daily survey. Each survey day included a different version of the short-form questionnaire depicting different social scenarios. Similar to Study 1, reliability measures for this scale were also strong (Cronbach’s alpha = 0.87 for version 1, 0.85 for version 2, 0.88 for version 3, and 0.9 for version 4). The longitudinal nature of Study 2 further allowed for an examination of test-retest reliability in this sample, which was 0.79 for the first consecutive days of the survey, 0.78 for days 2 to 3, and 0.72 for days 3 to 4.

To measure helping behavior with respect to prior sleep, the survey was only available online during a specific time window in the morning (until 1 PM local time), and participants were requested to complete the survey as close as possible to their wake-up time. In addition to the key outcome variable of helping behavior, measures of mood were also collected in each daily survey using the short PANAS questionnaire [112] described above. Finally, trait empathy was assessed upon entry to the study using the IRI [82], as in Study 1.

Data preprocessing and analysis.

Analyses focused a priori on sleep efficiency and sleep duration, given previous work linking both sleep parameters to social and interpersonal functioning [28,30]. Sleep efficiency was measured using participants’ daily sleep diaries, based on the percent of time asleep out of total time in bed (i.e., total time in bed minus sleep latency and time spent awake after sleep onset, mean ± SD = 90.02 ± 9.36%). Sleep duration was calculated as the total time elapsed from sleep onset to wake time minus sleep latency and time spent awake after sleep onset (mean ± SD = 429 ± 68.8 min).

A linear mixed-effects model was calculated to test whether night-to-night variability in sleep efficiency and sleep duration, within participants, predicted day-to-day changes in helping behavior the next day. The predictors of sleep efficiency and duration were calculated for both between- and within-person effects using person mean centering [113] (see Statistical analyses below and S1 Fig). The between-person effect refers to a person’s average over the study period (e.g., mean sleep duration/efficiency across 4 days), while the within-person effect refers to that person’s deviation from their average on a particular day (mean deviation in sleep efficiency = ± 2.03%, mean deviation in sleep duration = ±35.48 min, or 8.2% of total sleep time). All assessment days were weekdays to avoid concerns of weekend changes in sleep patterns. All linear mixed-effects models were adjusted for age, sex, and survey version.

While the main model focused on sleep duration and efficiency, it is important to note that circadian rhythms and circadian disruption also influence emotional and mood states [114–117]. We therefore, sought to further explore the contribution of circadian influence to helping behavior. For each participant, the measure of mid-sleep was calculated as the halfway point between sleep onset and sleep offset time during the study period, a marker of habitual circadian phase [118,119]. Similar to the main analysis, mid-sleep was assessed for both between- and within-person effects. Adding these parameters to the main model, there was no significant effect of circadian phase (mid-sleep) on helping behavior (within-person effect, β = −0.04 ± 0.03; between-person effect, β = 0.03 ± 0.04, both P > 0.25).

Donation data and analysis.

Study 3 tested the prediction that the loss of 1 h of sleep opportunity (due to DST) resulted in a real-world, large-scale decrease in helping behavior. Data were obtained from an online database of donations made between the years 2001 to 2016 in the US, via the DonorsChoose website, a platform that helps raise funds for school projects in the US (e.g., buy books, get supplies for a science project).

A total of 6,211,956 donations were available for analysis, including information about donor location, the timestamp of each donation, and the project each donation was intended to fund. Donations were excluded from further analyses if they did not include information on date/time or on donor location or were intended for projects that were not eventually funded (e.g., projects that expired before meeting their funding goal or were still not funded at the moment of download). Donations for projects that lasted less than a day were also excluded to allow for more stable predictors of donation behavior over time (including possible effects of sleep).

The main model focused on nationwide donations coming from states that observe DST (i.e., excluding Hawaii and Arizona), totaling 3,871,500 eligible donations (average daily donation amount $82.27 ± 0.14). For each donation, the following information was calculated and used in the statistical analyses: the day of the week/month/year of the donation and the time of day the donation was made. In accordance with prior reports [41,43,63,120,121], analysis focused on the weekdays following the transition, as both the ST and DST transitions result in sleep consequences lasting up to 5 days before sleep onset and offset times revert, and habitual sleep patterns return [59,122]. Analyses therefore focused on a robust window that spanned multiple days of assessment (see Note D in S1 Text for a secondary analysis focusing on post-transition Monday alone).

Using timestamp data of each donation, analyses tested the hypothesis that during the week of the transition to DST (the weekdays following the second Sunday of March since 2007 or the first Sunday of April before that), the corresponding loss of 1 h of sleep opportunity would significantly decrease altruistic helping behavior reflected in lower donation amounts. The statistical models examined the 4 weeks before and after the DST transition to avoid annual seasonal effects on donation amounts (e.g., donation amounts are lowest during the summer vacation when school is out, see S2 Fig). Donations were then aggregated to a daily average amount across the examined period (2001 to 2016, total number of days = 18,454) and subsequently log-transformed before being implemented in a multiple regression model (see Statistical analysis below). Donation data were filtered to exclude extreme outliers (above or below 3 standard deviations from the mean), most of which from a small number of donors giving very large amounts of money (e.g., more than $100,000 in a single donation) or from days that included extremely high numbers of recorded donations. Following these criteria, a total of 3,420,996 donations were aggregated across 18,034 days and utilized in the analyses.

Time of year control analysis.

In order to test for nonspecific effects of time of year on donation amounts, 3 additional models were constructed. The first model examined whether the mere switching of the clock might impact donation behavior irrespective of sleep changes by focusing on the months surrounding the transition back to ST (the weekdays following the first Sunday of November since 2007 or the last one in October before that). Analyses for this model similarly focused on the 4 weeks before and after the transition to ST as the key predictor of interest and controlling for the same covariates as the main model.

The second model examined possible time of year effects on donation behavior, probing whether the months of March/April when DST transition takes place, might impact donation behavior irrespective of changes in sleep. This model focused on the same timeframe as the main model (i.e., the weekdays following the second Sunday of March since 2007 or the first Sunday of April before that), but was now applied to donations coming from Arizona and Hawaii, states that do not observe DST and therefore are unlikely to experience sleep loss during the transition week. Data from Arizona and Hawaii included 76,276 donations made between the years 2001 to 2016 (average donation amount $70.23 ± 1.11).

Finally, the last model accounted for time availability effects. Since the day of DST transition is technically a 23-h day, changes in the available time to make donations, and thus a reduced number of donations, could impact donation amounts irrespective of changes in sleep. To account for such effects, an additional model was analyzed that controls for the number of daily donations made during the study period, given that the act of donation gifting is indeed subject to time availability irrespective of donation amount. Analyses for this model similarly focused on the 4 weeks before and after the transition to DST/ST as the key predictor of interest, controlling for the same covariates as the main model.

Helping behavior.

The key outcome measures of helping behavior used in Studies 1 to 3 were each designed to offer different but complementary measures of helping, thus allowing us to demonstrate a broader phenotype of prosocial behavior across all studies. The outcome measures taken in Studies 1 and 2 assessed participants’ desire to help, using the helping behavior questionnaire, assessing numerous forms of helping deeds and acts common in everyday social life [6,71]. Adding to this, Study 3 examined consequential helping, measuring real-world donation behavior using an online donation database and thus reflect a decisive action that directly resulted in the financial giving of help: an altruistic one, since it importantly did not result in (nor depend on) any reciprocal direct financial gain to the donor (a core construct of altruism) [123,124]. As such, and fitting prior studies [20,21,125,126], the assessment of prosocial helping did not incentivize the choice to help others, since acts of altruism typically do not involve reciprocal material exchanges, but instead, are benevolent and unidirectional [123,124].

Thus, Studies 1 and 2 measure the motivational desire to act altruistically, with the study design choice of a helping behavior questionnaire being motivated by the need to assess a wide selection of prosocial deeds and acts common in everyday life. Broadening the aperture of helping assessment, Study 3 evaluated an altruistic, consequential helping choice, which was the decision to give away money without any incentivized reciprocal benefit.

Behavioral results.

To test the hypothesis of reduced helping behavior following sleep loss in Study 1, a repeated measure ANOVA was calculated, taking into account familiarity (stranger versus familiar other) across the different sleep conditions (sleep rested, sleep deprived). In case of significance, post hoc tests were computed using paired 2-sided t tests corrected for multiple comparisons using the Bonferroni correction. To control for changes in mood and effort following sleep loss, a multiple linear regression was applied using the difference scores of each variable (sleep deprivation, sleep rested). In this model, the intercept coefficient reflected the sleep loss effect on helping (i.e., zero denotes no effect), while the mood and effort variables were added as covariates. Associations between social cognition network activity and helping behavior were tested using Pearson’s correlation, with mean activity from the entire a priori network of interest (i.e., not only limited to the activated cluster) to avoid spurious fMRI-behavior correlations. All statistical analyses were conducted using JASP (JASP Team 2021) and the pingouin library implemented in Python [127].

Multilevel modeling.

Linear mixed-effects models were used to determine associations between sleep and helping behavior across days in Study 2. All multilevel models were adjusted for age, sex, and survey day, with subject identifier defined as a random effect. A person-mean centering was applied to the predictor variables in all models to disaggregate the between-person and within-person effects [113]. The time-invariant person average and time-varying deviation from an individual’s average were then both included as fixed effects in the multilevel model (between- and within-person components, respectively). To control for outliers in the 2 key predictors of interest, sleep efficiency and sleep duration values that were ±3 STD from the mean were filtered prior to the final analysis. In addition to the main model, 2 additional models were calculated controlling for changes in (i) mood states (within-person change across days) and trait empathy (between person factor); and (ii) prior helping behavior (a daily variable). These models were identical to the main model in terms of the key predictors and covariates. All multilevel analyses were performed in R [128] using the “lme4,” “lmerTest,” and “sjPlot” packages [129–131]. Goodness-of-fit was evaluated using the conditional R² [132].

Multiple regression.

Study 3 implemented a multiple regression model to evaluate the impact of DST transition on donation behavior. The log-transformed daily donation amount in US dollars was set as the outcome variable, while the key predictor of interest was “week type” (with levels Monday to Friday before DST, DST weekdays, weekdays after DST, and any other day of the year). All models were adjusted for additional predictors influencing donation likelihood including: time of day (using 4 equal 6-h bins: morning: 6 AM to 11:59 AM, afternoon: noon to 5:59 PM, evening: 6 PM to 11:59 PM, and night: midnight to 5:59 AM), day of the week, month, year, weekend day (Saturday or Sunday), and holidays (as a binary True/False variable). For the control analysis of available time to donate, the predictor of number of donations (log-transformed) was also added to the main model. If DST transition lowered donation amount in the 5 days after switching to DST, a significant coefficient for the level “DST week” was expected compared to all other Monday to Friday in the variable “week type.” Similarly, a nonsignificant coefficient for “week type” was expected in the 2 control models exploring the transition back to ST as the key predictor or when focusing on donations from states that do not observe DST (i.e., Arizona and Hawaii). All tests of statistical significance were two-sided, and p values less than 0.05 were considered statistically significant. All analyses were performed in R [128] using the “lme4” and “lmerTest” packages.