Delaying high school start times impacts depressed mood among students: evidence from a natural experiment

Abstract

Purpose

Delaying high school start times prolongs weekday sleep. However, it is not clear if longer sleep reduces depression symptoms and if the impact of such policy change is the same across groups of adolescents.

Methods

We examined how gains in weekday sleep impact depression symptoms in 2,134 high school students (mean age 15.16 ± 0.35 years) from the Minneapolis metropolitan area. Leveraging a natural experiment design, we used the policy change to delay school start times as an instrument to estimate the effect of a sustained gain in weekday sleep on repeatedly measured Kandel-Davies depression symptoms. We also evaluated whether allocating the policy change to subgroups with expected benefit could improve the impact of the policy.

Results

Over 2 years, a sustained half-hour gain in weekday sleep expected as a result of the policy change to delay start times decreased depression symptoms by 0.78 points, 95%CI (−1.32,−0.28), or 15.6% of a standard deviation. The benefit was driven by a decrease in fatigue and sleep-related symptoms. While symptoms of low mood, hopelessness, and worry were not affected by the policy on average, older students with greater daily screen use and higher BMI experienced greater improvements in mood symptoms than would be expected on average, signaling heterogeneity. Nevertheless, universal implementation outperformed prescriptive strategies.

Conclusion

High school start time delays are likely to universally decrease fatigue and overall depression symptoms in adolescents. Students who benefit most with respect to mood are older, spend more time on screens and have higher BMI.

Introduction

The prevalence of mood disorders, including major depression, increases sharply from 1%–17–25% as children progress through adolescence, with the fastest incidence rate evident in ages 15 to 18 [1, 2]. Numerous underlying processes make adolescence a period of vulnerability to mood dysregulation. Adolescence is marked by rapid physiological changes, development of social networks, increased academic pressures, a greater propensity for risk-taking behaviors, and heightened reactivity to emotional stimuli [3–5]. Puberty – which demarcates adolescence from childhood – also brings about greater sensitivity to gonadal and adrenal hormones, and morphological changes in brain regions responsible for emotion regulation [6–9].

Moreover, extensive changes in sleep physiology occur over the course of puberty [10]. Overall melatonin levels decrease and the timing of melatonin peak shifts later as children undergo puberty, making it difficult to fall asleep before 11PM [11–13]. The homeostatic accumulation of sleep pressure slackens, making it easier for adolescents to forego sleep and accumulate sleep debt [14, 15]. In addition, adolescents are confronted with earlier school start times as they progress from elementary to middle and high school. More than 36% of U.S. public high schools start before 8 a.m. and approximately 70% of U.S. high school students sleep less than the recommended 8–10 h per night [16–19]. The confluence of biological and structural shifts that arise during adolescence and results in short and ill-timed weeknight sleep has been termed a “the perfect storm” [20, 21].

Sleep deprivation in adolescence can disrupt normative development of the prefrontal cortex and multiple brain networks involved in cognitive control and emotional regulation [22]. Short sleep duration has been implicated in greater odds of depressed mood in adolescents [23]. However, several questions remain unanswered. First, if a structural intervention such as a policy to delay school start times is implemented and successfully increases weekday sleep duration, is there a sustained protective effect on depression symptoms? Second, would such a structural intervention affect all adolescents homogeneously, or would identifiable subgroups of adolescents benefit more than others? Lastly, changes in sleep patterns and reductions in motivation due to fatigue are two of the minimally five diagnostic criteria required for a diagnosis of a major depressive episode [24]. As such, does increasing sleep prospectively predict lower depression symptoms overall or does it simply ameliorate the dysregulated sleep component of the syndrome? There is a lack of clarity regarding these questions in epidemiologic findings to date.

We strive to address these questions with data from a natural experiment conducted in the Minneapolis metropolitan area – the START study [25]. Prior work using this data demonstrates a sustained gain in expected sleep duration following a delay in start times [26]. We will assess the impact of the sustained gain in sleep attributable to delaying school start times on depression symptoms, and examine the subgroups – determined by baseline socioeconomic and lifestyle covariates – that benefit most from the implementation of a school start time delay policy.

Method

Study & setting

START collected three waves of data on students from five high schools, two of which delayed their start time. Baseline data was collected from 9th -graders in the spring of 2016, while all the schools were on an early start schedule. In the fall of 2016, two schools shifted their start time from 7:30AM and 7:45AM to 8:20AM and 8:50AM, respectively, while three comparison schools, similar on demographic, cultural, and geographic characteristics to the delay schools, continued to start at 7:30AM. Follow-up assessments were conducted in the spring of 2017 and 2018, when the enrolled students were in 10th and 11th grade. Detailed study procedures have been previously described [25]. The conduct of this analysis was approved by institutional review boards of the University of Minnesota and Harvard University. This study followed the STrengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.

Measurement and variables

Weekday sleep duration was self-reported. The students were asked when they usually go to bed on school days, how long it takes them to fall asleep after going to bed (less than 10 min, 10–20 min, more than 20 min, or unknown), and when they usually wake up on school days. Bedtimes and wake-up times were collected using multiple-choice questions with answers binned into 15-minute intervals (bedtimes between 9PM and 2AM and wake-up times between 5AM and 8:30AM). Weekday sleep duration was calculated as the hours between going to bed and waking up, minus the approximate amount of time it takes to fall asleep.

Depression symptoms were measured at all three time points with the six-item Kandel-Davies questionnaire, which is validated in adolescents aged 12–18 years [27]. Participants were asked how often they have been bothered or troubled by the following during the last two weeks: (a) feeling too tired to do things, (b) having trouble going to sleep or staying asleep, (c) feeling unhappy, sad, or depressed, (d) feeling hopeless about the future, (e) feeling nervous or tense, and (f) worrying too much about things. Likert-scale responses of “not at all”, “somewhat”, and “very much” conferred 1, 2 and 3 points, respectively. The primary outcome was overall depression symptoms operationalized as the average of the individual item scores, multiplied by 10 for a final score ranging from 10 (no depression symptoms) to 30 (most severe symptoms). As secondary outcomes, we analyzed average scores on fatigue symptoms (items a and b) as well as items on low mood (item c), hopelessness (item d), and worry symptoms (items e and f) – together termed “mood” symptoms henceforth. These subdomains have been operationalized in prior work utilizing the Kandel-Davies questionnaire [28].

We obtained a comprehensive set of covariates to control for possible confounding and to study the heterogeneity of the policy’s effect: participants’ biological sex (male or female), Hispanic ethnicity, race (American Indian or Alaska Native; Asian; Black, African, or African American; Native Hawaiian or other Pacific Islander; or White), an indicator of a one-parent household, the highest level of parental education (less than high school, high school or GED, some college, college degree, or advanced degree), time in minutes that it takes to get to school, body-mass index (BMI), an indicator of a goal to lose weight, an exertion-weighted total of hours of mild (weight of 1), moderate (weight of 2) and strenuous (weight of 3) exercise in the past week, a count of days with at least 60 min of physical activity in the past week, severity of uncontrollable eating behavior (reverse-coded average of Likert-scale responses of “Hardly ever”(= 1), “Sometimes”(= 2), “Most of the time”(= 3), “Almost always”(= 4) to “How often you stop eating when you feel full?” and “How often you trust your body to tell you how much to eat?”), an ordinal measure of feeling hungry at school (“Hardly ever”(= 1), “Sometimes”(= 2), “Most of the time”(= 3), “Almost always”(= 4)), an indicator of having no meals with family in the past week, a count of caffeinated beverages consumed in the past month (including energy drinks, coffee and tea), an indicator of morning obligations (school activities, homework, chores, or work for pay), an indicator of working for pay, eligibility for free or reduced-price lunch, the number of hours of screen time dedicated to things other than school work on an average weekday, and the count of nights per week using a device with a screen in bed. Asian and Native Hawaiian and Other Pacific Islander race categories, and parental education of less than high school and high school or GED were combined for analysis due to small cell counts.

Missing data procedures

Of the 2,134 students enrolled in the study, 2,054 fully completed the baseline survey. Attrition over the 2-year follow-up was modest; 1,917 participants contributed data at either the first or second follow-up. Item-level missing data at baseline was imputed using hot-deck imputation, an approach well suited for the mostly ordinal and categorical data collected in START [29]. Non-response at follow-up was addressed using stabilized inverse probability of censoring weights, which involved conducting logistic regressions for the numerator and denominator to predict probabilities of remaining in the study at the first and second follow-up times [30]. At each time point, the probability of remaining in the study was modeled separately for those in the delay and early start groups, with numerator models reflecting the group-wise probabilities (intercept-only models). The denominator models accounted for the school, eligibility for free or reduced-price lunch, parental education, one-parent household status, not having any family meals in the past week, uncontrollable eating behavior, caffeine intake, BMI, depression symptoms, and weekday sleep duration at baseline. Weights were computed as inverted quotients of the numerator and denominator predicted probabilities and cumulated over time.

Statistical analyses

We compared the distributions of baseline covariates between groups of students from schools that implemented the delay policy (the policy change group) and schools that remained on an early schedule (the comparison group) over the duration of the study using studentized t-tests for continuous variables and Chi-square tests for categorical variables. Using hierarchical targeted maximum likelihood estimation (htmle), we estimated the effects of the policy change on weekday sleep duration, overall depression symptoms, and fatigue and mood symptom clusters at 1- and 2-years post-baseline. Htmle is an efficient, doubly robust algorithm that combines inverse probability of treatment weighting and outcome regression to estimate average treatment effects in data where the intervention is assigned to clusters rather than individuals [31, 32].

We next examined whether the policy change had a heterogeneous effect on depression symptom outcomes. Our aim was to learn which subgroups of students experienced the greatest reduction in symptoms following the school start time delay. The data was split into 70% training and 30% testing sets. We used an X-learner algorithm to estimate the conditional average treatment effect for each individual in the training set [33]. Estimates were calculated by fitting SuperLearner models for the outcome within strata of the delay variable and cross-applying the predictions. The difference in the predictions was relearned with another SuperLearner, adjusting for non-random allocation of the delay policy and potentially informative censoring using the product of inverse probability of treatment and censoring weights. This step directly estimated the conditional average treatment effect, eliminating the reliance on the correct specification of the model for the outcome under no intervention.

Trained models for the conditional average treatment effect were applied to the testing sample to identify those for whom the delay policy is prescribed: those with negative predicted conditional average treatment effect values. Htmle was used to compute the contrast in the expected depression outcomes had everyone been assigned to their prescribed school start time regime (delayed or early) versus (a) had the delay policy been assigned universally or (b) had the policy been applied to a random 50% of the population. Improvement given prescribed rather than universal assignment indicates the existence of identifiable subgroups determined by combinations of baseline prescriptive covariates for whom the policy is both more beneficial and more harmful than expected on average. Improvement relative to at-random allocation, on the other hand, only signified subgroups with greater than average benefit expected from the policy change. Contrast with either universal or at-random implementation indicated heterogeneity of the delay policy’s effect. Where heterogeneity was found, an xgboost variable importance analysis identified student characteristics most predictive of the contrast in the testing sample using the gain and frequency metrics [34]. Section 1 of the online supplement provides a detailed overview of the computational algorithms used to detect heterogeneity.

Lastly, leveraging the policy change as an instrumental variable, we homed in on the effect of a sustained sleep gain that is expected had everyone transitioned to a delayed start schedule versus stayed on an early schedule. To do so, we utilized g-estimation of the parameters of a structural nested mean model (SNMM) comprising two equations for the two follow-up periods [35]. These models were run for outcomes which were homogeneously impacted by the delay policy. Standard errors for the estimates were bootstrapped with 1000 samples.

Initial data wrangling and imputation were conducted using SAS 9.4 (SAS Institute Inc., Cary, NC, USA) and subsequent analysis of the data was conducted using R version 4.2.1 (R Core Team (2022)).

Results

Of the 1,917 students who contributed data over the 2-year follow-up, 1,157 (60.4%) experienced the policy change to delay start times in the fall of their sophomore year. At baseline, students who were about to undergo the policy change were less likely to experience hunger during the school day, reported less difficulty with uncontrollable eating, lower intake of caffeinated beverages, slightly longer weekday sleep duration (7.2 (SD = 1.1) versus 7.1 (SD = 1.2) hours on average) and lower baseline depression scores (15.9 (SD = 4.8) versus 17.0 (SD = 5.2) on average) relative to students in comparison schools. A comprehensive comparison of baseline student characteristics – measured while all students were still on an early schedule – is outlined in Table 1.

Table 1.

Distributions of Baseline Covariates Compared by Policy Change Among Participants with Follow-Up Data from the START Study, Set in the Minneapolis Metropolitan Area, 2016 (N = 1,917)

	Policy changea	Comparisonb	Overall
n (%)	1157 (60.4)	760 (39.6)	1,917	P c
Age in years: mean(SD)	15.1 (0.3)	15.2 (0.4)	15.2 (0.4)	< 0.01
Female biological sex: n(%)	573 (49.5)	368 (48.4)	941 (49.1)	0.67
Race/ethnicity: n(%) d
Hispanic ethnicity	39 (3.4)	18 (2.4)	57 (3.0)	0.26
American Indian or Alaska Native	33 (2.9)	30 (3.9)	63 (3.3)	0.24
Asian	150 (13.0)	28 (3.7)	178 (9.3)	< 0.01
Native Hawaiian or Other Pacific Islander	4 (0.3)	8 (1.1)	12 (0.6)	0.10
Black, African American	78 (6.7)	23 (3.0)	101 (5.3)	< 0.01
White	968 (83.7)	711 (93.6)	1679 (87.6)	< 0.01
One parent household: n(%)	38 (3.3)	24 (3.2)	62 (3.2)	0.98
Highest parental education: n(%)				< 0.01
Less than high school	13 (1.1)	7 (0.9)	20 (1.0)
High school degree	57 (6.0)	65 (9.7)	122 (7.4)
Some college	113 (9.8)	118 (15.5)	231 (12.1)
College degree	460 (39.8)	337 (44.3)	797 (41.6)
Advanced degree	514 (44.4)	233 (30.7)	747 (39.0)
Time (hours) to get to school: mean(SD) : n(%)	0.8 (0.3)	0.8 (0.4)	0.8 (0.3)	< 0.01
Student has work, chores, activities, or schoolwork obligations in the morning before school: n(%)	295 (25.5)	221 (29.1)	516 (26.9)	0.09
Student works for pay: n(%)	172 (14.9)	112 (14.7)	284 (14.8)	0.99
Eligible for free or reduced-price lunch: n(%)	164 (14.2)	112 (14.7)	276 (14.4)	0.78
Body mass index: mean(SD)	21.5 (4.2)	21.7 (3.9)	21.6 (4.1)	0.30
Intention to lose weight: n(%)	312 (27.0)	214 (28.2)	526 (27.4)	0.60
Weighted total of physical activity in the past week: mean(SD)	18.5 (9.1)	19.8 (9.7)	19.0 (9.3)	< 0.01
Days with at least 60 min of physical activity in the past week: mean(SD)	4.9 (1.8)	4.9 (1.9)	4.9 (1.8)	0.82
Difficulty controlling eating: mean(SD)	1.8 (0.7)	2.0 (0.7)	1.9 (0.7)	< 0.01
Hungry at school: n(%)				< 0.01
Hardly ever	104 (9.0)	35 (4.6)	139 (7.3)
Sometimes	560 (48.4)	277 (36.4)	837 (43.7)
Much of the time	325 (28.1)	214 (28.2)	539 (28.1)
Almost always	168 (14.5)	234 (30.8)	402 (21.0)
No meals with family in the past week: n(%)	85 (7.3)	53 (7.0)	138 (7.2)	0.83
# Caffeinated beverages in the past month: mean(SD)	9.0 (20.4)	11.0 (23.9)	9.8 (21.9)	0.05
Hours of screen time on average weekday: mean(SD)	2.4 (1.4)	2.7 (1.4)	2.5 (1.4)	< 0.01
Nights per week using device in bed: mean(SD)	3.6 (1.9)	4.0 (1.7)	3.8 (1.8)	< 0.01
Weekday sleep (hours) : mean(SD)	7.1 (1.1)	7.1 (1.2)	7.1 (1.1)	0.21
Overall depression symptoms: mean(SD)	15.9 (4.8)	17.0 (5.1)	16.3 (5.0)	< 0.01

Note. SD = standard deviation

^aThe policy change group reflects schools that delayed their start times from 7:30AM and 7:45AM to 8:20AM and 8:50AM, respectively

^bThe comparison school group reflects the schools that remained on an early schedule (start time of 7:30AM)

^cDouble-sided p-values from studentized t-tests for continuous variables and Chi-sq tests for categorical variables

^dPercent do not add to 100, multiple categories allowed as choices

The comparison of baseline characteristics between those who contributed data at follow-up and those who were censored post baseline is shown in Table S1 in the online supplement. Remaining on an early school start schedule, eligibility for free or reduced-price school lunch, having higher BMI, exercising less, experiencing shorter weekday sleep, and higher baseline depression symptoms all signaled a greater likelihood of dropping out of the study after baseline.

In Fig. 1, we compare counterfactual sleep and depression symptom outcomes across follow up under the policy change and comparison conditions. The policy change condition reflects a delay in school start times from 7:30AM and 7:45AM to 8:20AM and 8:50AM, respectively, whereas the comparison condition reflects a school start time of 7:30AM. Weeknight sleep duration would be expected to increase from baseline if the policy change was universally implemented, but gradually decrease if all had remained on an early schedule. Depression symptoms – overall, mood and fatigue symptoms – are expected to increase under both counterfactual conditions, but to a lesser degree for overall and fatigue symptoms if school start times were delayed. Average effects of the policy change, estimated as differences in counterfactual outcomes, are outlined in Table 2. Given a delayed versus early schedule, overall depression symptoms are expected to be on average 0.13 points lower, 95% CI (−0.17,−0.08)) and 0.26 points lower, 95% CI (−0.50,−0.02) at 1- and 2-years post baseline, respectively, had everyone transitioned to a later start time. Larger effects are expected for fatigue symptoms if the delay policy is implemented (−0.54 points, 95% CI (−0.73,−0.35) and − 0.91 points, 95% CI (−1.45,−0.38) at 1- and 2-years post-baseline, respectively). There is no average effect of the policy change on mood symptoms. Baseline differences in weekday sleep and depression symptom outcomes were minimized by using inverse probability weighting, as demonstrated in Fig. 1, Table S2 and Table S3 in the online supplement, and are not expected to influence expected changes in the outcomes across follow-up.

Fig. 1.

Counterfactual sleep and depression symptom outcomes across follow-up (a. weeknight sleep duration; b. overall depression symptom score; c. mood symptom score; d. fatigue and sleep-related symptom score)

Table 2.

Average Effects of the Policy Change on Sleep and Depression Outcomes Over Follow-up among Participants of the START Study, Set in the Minneapolis Metropolitan Area, 2017–2018 (N = 1,917)

	1-year post-baseline	2-years post-baseline
	Est (95% CI)	Est (95% CI)
Weekday sleep duration	0.53(0.34,0.72)	0.45(0.23,0.66)
Depression symptoms
Overall	−0.13(−0.17,−0.08)	−0.26(−0.50,−0.02)
Mood	−0.04(−0.46,0.38)	−0.03(−0.09,0.02)
Fatigue	−0.54(−0.73,−0.35)	−0.91(−1.45,−0.38)

Note. Est = point estimate of the average effect of the delay estimated using htmle; CI = confidence interval

Counterfactual conditions reflect the policy change where schools delay their start time from 7:30AM and 7:45AM to 8:20AM and 8:50AM, respectively, and a comparison condition where schools continue to start at 7:30AM.

We next addressed whether the impact of the delay was heterogeneous with respect to depression symptoms over time. In Table 3, we compared population outcomes if everyone followed their prescribed assignment versus (a) if the delay policy was implemented universally or (b) at-random. Prescriptive allocation of the delay did not yield benefit relative to universal implementation. However, a 0.76-point decrease, 95% CI (−1.43,−0.08) in mood symptoms 1-year post-baseline is expected if the delay is implemented in subgroups with expected benefit of the policy rather than at-random in the population. Overall depression symptoms and fatigue symptoms did not exhibit heterogeneous response to the policy. The putative benefits of universal relative to prescribed and prescribed relative to at-random implementation did not survive multiple testing correction using the Benjamini-Hochberg method. However, multiple testing correction may be an overly conservative strategy in this case, given that the outcomes are all based on the same depression symptom scale.

Table 3.

Comparison of Prescribed vs. Universal or At-Random Policy Change Implementation on Counterfaetual Depression Symptom Outcomes among Participants of the START Study, Set in the Minneapolis Metropolitan Area, 2017–2018 (N = 1,917)

	Universal Implementation		Random implementation
	1-year post-baseline	2-years post-baseline	1-year post-baseline	2-years post-baseline
Depression symptoms	Est (95% CI)	Est (95% CI)	Est (95% CI)	Est(95% CI)
Overall	0.03(−1.24,1.29)	0.77(−0.46,1.99)	−0.03(−1.30,1.23)	0.21(−1.03,1.44)
Mood	0.85(0.19,1.51)	0.85(−0.35,2.04)	−0.76(−1.43,−0.08)	0.49(−0.72,1.70)
Fatigue	−0.13(−1.44,1.18)	1.32(−0.12,2.75)	−0.24(−1.55,1.07)	0.01(−1.44,1.45)

Note. Est = point estimate of the average effect of prescribing the delay contrasted with a counterfactual strategy (universal or at-random implementation) estimated using htmle; CI = confidence interval

Table 4 summarizes the variable importance statistics (gain and frequency) for the conditional average effect on mood symptoms and outlines the top 10 variables that in combination contribute most strongly to the accuracy of its prediction. Student age, daily school-unrelated screen-time, BMI, physical activity, and the intension to lose weight are statistically different for those for whom the delay is prescribed (46.7%). The delay is expected to improve low mood for students who are older (15.28 (SD = 0.32) vs. 15.05 (SD = 0.35)), have higher BMI (21.94 (SD = 4.02) vs. 21.20 (SD = 3.62)), spend more time on screens on an average day (3.12 h (SD = 1.42) vs. 2.02 h (SD = 1.31)), exercise less often (4.74 days (SD = 1.95) vs. 5.12 days (SD = 1.16) per week), are more likely to want to lose weight (37% vs. 19%), and consume more caffeine in a given month (10.71 cups (SD = 22.56) vs. 7.59 cups (SD = 19.06)).

Table 4.

Variable Importance Analysis Outlining the Top 10 Characteristics Most Predictive of the Conditional Average Treatment Effect for the 1-year Mood Symptom Outcome; Testing Sample of START Study Participants from the Minneapolis Metropolitan Area, 2017 (N = 534)

Characteristic	Gaina	Frequencyb	Delay prescribed	Delay not prescribed	P c
			n = 240	n = 274
			Mean(SD)	Mean(SD)
Age (years)	0.21	0.15	15.28 (0.32)	15.05 (0.35)	< 0.01
Hours of screen time on average weekday	0.12	0.04	3.12 (1.42)	2.02 (1.31)	< 0.01
Weekday sleep duration (hours)	0.10	0.17	7.16 (1.13)	7.13 (1.07)	0.73
Weighted hours of physical activity (past week)	0.10	0.10	19.37 (9.80)	19.40 (9.00)	0.97
Body mass index	0.09	0.14	21.94 (4.02)	21.20 (3.62)	0.03
Difficulty controlling eating	0.06	0.04	2.04 (0.75)	1.69 (0.61)	< 0.01
# Caffeinated beverages in the past month	0.05	0.06	10.71 (22.56)	7.59 (19.06)	0.09
Days with 60 + min of physical activity (past week)	0.04	0.04	4.74 (1.95)	5.12 (1.61)	0.02
Intension to lose weight (proportion)	0.04	0.01	0.37 (0.48)	0.19 (0.40)	< 0.01
Time it takes to get to school (hours)	0.03	0.04	0.79 (0.37)	0.78 (0.31)	0.73

Note.SD = standard deviation

^aGain statistic quantifies each characteristic’s contribution to the accuracy of the counterfactual contrast prediction

^bFrequency quantifies the relative number of times a feature appears in regression trees

^cDouble-sided p-values from studentized t-tests

We used the policy change as an instrumental variable to estimate the impact of a sustained 30-minute sleep gain – as approximately expected following the school start time delay – for overall and fatigue symptoms, which were homogeneously impacted by the delay. We estimated a reduction in overall depression symptoms by 0.78, 95%CI (−1.32,−0.28) points over 2 years and a more pronounced impact on fatigue (−1.36 points, 95% CI (−2.19−0.69)), which likely drives the reductions in overall symptoms.

Discussion

This study examined how a structural intervention to delay school start times impacts depression symptoms in a population of high school students. We found that start time delay policies have the potential to reduce population-level depression symptom burden in high school students, particularly improving self-reported difficulty due to fatigue. While the delay does not confer an average impact on mood symptoms, students who are older, have higher BMI, express the intension to lose weight, spend more time on screens, exercise less, and consume more caffeine may experience improvements in mood symptoms following a shift to later school start times.

Insufficient sleep is a prevalent issue among adolescents. According to a 2018 report of nationally representative data from the Youth Risk Behavior Surveillance System, 72.7% of high school students in the U.S. get less than the American Academy of Sleep Medicine’s recommended 8–10 h of sleep per night [17, 18]. Nearly half of 16-year-olds in the U.S. regularly sleep less than 7 h per night [36]. Sustained sleep deprivation increases the risk of developing chronic health conditions including major depressive disorder [37]. Recent work theorizes that sleep deficits in adolescence may increase internalizing symptoms by promoting functional imbalance in the default brain network, promoting repetitive negative thoughts and rumination, further compromising sleep [38].

School-based interventions have been shown to increase sleep duration in adolescents. A successful reduction of insufficient sleep (< 7 h/night) was evidenced by an experimental school-based sleep education curriculum, but the intervention group reported no change in perceived stress over 1 year of follow-up [39]. Interventions targeting individual behaviors to improve sleep may be insufficient to safeguard downstream mental health outcomes when chronobiological changes in adolescents collide with structural morning-side schedule demands. The START study demonstrates that a school start delay results in a meaningful and sustained average gain in weekday sleep duration – self-reported and objectively-measured with actigraphy in a sub-sample [26]. Our findings confirm that a structural change towards later school start times reduces depression symptoms in high school-aged youth [40]. We expect students who maintain a 30-minute gain in weekday sleep over two years to have a lower burden of depression symptoms longitudinally. Overall depression symptoms are expected to be lower than if students remained on an early schedule, largely driven by a pronounced reduction in fatigue.

This study is novel in demonstrating that delaying high school start times may be particularly protective for students who are older for their cohort, who have higher BMI and indicate struggling with excess weight, who are more sedentary than their peers, and spend more time on screens. Struggling with excess weight and spending more time on screens have been cited as independent risk factors for dysregulated mood [41–43]. A large study found significant increases in adolescent depression symptoms over time associated with greater time spent on social media and watching television, revealing reinforcing interactions between the within-person and between-person impacts of media use on deteriorating self-esteem [44]. Overweight and obesity have a strong bidirectional connection with major depression, particularly among female adolescents and young adults [45, 46]. There is a potential link between excess adipose tissue and adolescent-onset major depression via inflammatory mechanisms that are mediated by increases in sex hormones during puberty [47]. In adolescence, being heavier considerably limits exercise and promotes sedentary activities – including time spent on screens [48]. Notably, children growing up in disadvantaged socioeconomic circumstances are more likely to fit the profile of characteristics identifying those with greatest benefits from the school start time delay intervention [49–51]. Our findings suggest that delaying start times in underserved school districts – as part of a universal implementation strategy – may lower population-level depression symptom burden and potentially reduce observed disparities in depression.

Our finding that prescribed assignment is better than at-random but not universal allocation of delayed start times indicates that the subgroups with empirically identified harm are not well-characterized by the set of prescriptive covariates examined. The impact of the delay policy is likely heterogeneous with respect to the mood symptom cluster, but the information we have only characterizes subgroups who have better-than-average mood symptoms as a result of the policy change. This work may help inform the choice of school if alternatives are available to a family with a high school student. Otherwise, schools may consider scheduling study hall or free periods at the start of the school days such that students who need extra sleep can opt out of early morning obligations. In addition, future research should assess if school bus routes can be optimized to provide students who benefit most from later school start times with extra time in the mornings. Such studies would need to connect individual characteristics to geocoded addresses of participating students.

There are notable limitations and strengths to this work. First, the school start time delay was not randomly assigned; however, we used inverse probability of treatment weighting to control for imbalances in measured student characteristics that may affect depressed mood over follow-up. Second, while overall retention was good, loss to follow-up depended on baseline student characteristics. We ameliorated this potential source of bias by using stabilized inverse probability weights. Third, we recognize that unmeasured school-district level characteristics such as average household income were not well-accounted for and may have been drivers in the school districts’ decisions to delay start times, study attrition, and levels of depression symptoms in the study population. The geographic proximity of the participating schools and efforts made to select schools similar on economic and demographic characteristics gives us hope that the delayed and early-start schools were comparable on unmeasured cluster-level characteristics. On the other hand, the selection of schools from a proximal geographic area limits generalizability. For estimation and inference of overall and heterogeneous effects, we utilized hierarchical targeted learning, a doubly robust algorithm that accounts for clustering of the delay implementation, but with only 5 clusters, our precision may have been reduced. As such, evidence for heterogeneity of treatment response was limited after correction for multiple testing. Lastly, we were not able to assess gender identity as a potential prescriptive factor – only biological sex (male or female) was collected in this study. Despite limitations, the START study gave us an opportunity to leverage the delay policy as an instrumental variable and thus control for unmeasured confounding of the relationship between individual weekday sleep duration and self-reported mood.

Our findings stand in favor of universal high school start time delays, which result in meaningful reductions in fatigue, driving modest reductions in overall depression symptoms over two years of follow-up. Delays are expected to have greater than average impact on low mood in older students who have higher average BMI, indicate a goal to lose weight, spend more time on screens, exercise less, and consume more caffeine. Universal delays of high school start times have the potential to not only sustainably prolong weekday sleep among high school students but improve depression symptom outcomes.

Data availability

No datasets were generated or analysed during the current study.