School start time delays and high school educational outcomes: Evidence from the START/LEARN study

Abstract

Introduction:

Delaying high school start times extends adolescents’ nightly sleep, but it is less clear how it affects educational outcomes. We expect links between school start time delays and academic performance because getting enough sleep is a key input to the cognitive, health, and behavioral factors necessary for educational success. Thus, we evaluated how educational outcomes changed in the two years following a school start time delay.

Methods:

We analyzed 2,153 adolescents (51% male, 49% female; mean age 15 at baseline) from START/LEARN, a cohort study of high school students in the Minneapolis – St. Paul, Minnesota, USA metropolitan area. Adolescents experienced either a school start time delay (“policy change schools”) or consistently early school start times (“comparison schools”). We compared patterns of late arrivals, absences, behavior referrals, and grade point average (GPA) one year before (baseline, 2015-2016) and two years after (follow-up 1, 2016-2017 and follow-up 2, 2017-2018) the policy change using a difference-in-differences analysis.

Results:

A school start time delay of 50 to 65 minutes led to 3 fewer late arrivals, 1 fewer absence, a 14% lower probability of behavior referral, and 0.07-0.17 higher GPA in policy change schools versus comparison schools. Effects were larger in the second year of follow-up than in the first year of follow-up, and differences in absences and GPA emerged in the second year of follow-up only.

Conclusions:

Delaying high school start times is a promising policy intervention not only for improving sleep and health but for improving adolescents’ performance in school.

Introduction

Adolescents must get an adequate amount of high-quality sleep to function at their best (Beebe, 2011; Shochat et al., 2014). However, 73% of adolescents in the United States do not get enough sleep (Wheaton et al., 2018). High rates of inadequate sleep in adolescence are produced by a “perfect storm” of biological, psychosocial, and contextual factors that intersect in adolescence (Carskadon, 2011; Crowley et al., 2018). Biologically, circadian delays (Crowley et al., 2007) and delayed sleep pressure accumulation (Jenni et al., 2005) push adolescents to feel ready for sleep at later hours than younger children or adults. At the same time, adolescents are gaining autonomy over their own schedules, face increasing educational demands, and have more social interactions than younger children (Crowley et al., 2018). These pressures toward later bedtimes combine with early secondary school start times to curtail adolescents’ nightly sleep for much of the calendar year. Thus, delaying school start times has been identified as an effective policy change that can extend adolescent sleep and improve overall sleep health (ADOLESCENT SLEEP WORKING GROUP et al., 2014; Bowers and Moyer, 2017; Minges and Redeker, 2016; Morgenthaler et al., 2016; Owens et al., 2014; Wheaton et al., 2016). Many organizations that promote child wellbeing have published position statements supporting later school start times for adolescents, including the American Academy of Pediatrics, Centers for Disease Control & Prevention, and the U.S. Surgeon General (Start School Later, 2022).

Educational performance depends on both behavioral and cognitive factors, each of which require sufficient sleep. Adolescents who do not get enough sleep have higher rates of behavior regulation problems, depression, anxiety, and risky behaviors than adolescents with healthy sleep durations (Beebe, 2011; Shochat et al., 2014), because sleep deprivation changes risk-taking, sensation-seeking, and impulsivity and is linked to poorer emotional processing (Krause et al., 2017). Sleep deprivation also negatively affects attention, working memory, and learning (Alhola and Polo-Kantola, 2007; Krause et al., 2017; Walker and Stickgold, 2006). Correspondingly, adolescents who extend their nightly sleep improve their cognitive performance (Dewald-Kaufmann et al., 2013). For example, one experimental study compared adolescents’ outcomes after 5 nights of short sleep durations (6.5 hours in bed per night for 5 nights) with 5 nights of healthy sleep durations (10 hours in bed per night for 5 nights). In a simulated school day with learning tasks and assessments, sleep-deprived adolescents had lower test scores and paid poorer attention (Beebe et al., 2010). Therefore, we expect school start time delays to improve adolescents’ educational outcomes such as attendance, behavior, and grades.

Despite evidence that school timing matters for student outcomes (Carrell et al., 2011; Diette and Raghav, 2017; Edwards, 2012; Heissel and Norris, 2018), prior research has produced inconsistent evidence on the effects of school start time policies on educational outcomes, as reviewed by Biller et al. (2022). Though most research on high school start time changes in the United States finds that later school start times are associated with fewer late arrivals (Dunster et al., 2018; Lenard et al., 2020; Owens et al., 2010; Thacher and Onyper, 2016; Wahlstrom et al., 2014), fewer absences (Lenard et al., 2020; McKeever and Clark, 2017; Wahlstrom et al., 2014), and fewer disciplinary problems (Thacher and Onyper, 2016), some studies show null or opposite findings (Dunster et al., 2018; Fuller and Bastian, 2022). Evidence on the effects of later school start times on academic performance among high school students in the United States is mixed and may be driven by differences in study methodologies with different degrees of causal rigor: Studies that compare average grades at high schools in the United States before and after they changed their start times find higher letter grades (Wahlstrom et al., 1997) and higher first period GPA (Wahlstrom et al., 2014) after school start time delays. More rigorous studies that compare individual students’ academic performance before and after high school start time changes typically find no change in grades (Boergers et al., 2014; Owens et al., 2010; Thacher and Onyper, 2016; Wahlstrom, 2002), but do identify improvements in others (Dunster et al., 2018; Fuller and Bastian, 2022).

In the present study, we add to this literature by using a rigorous difference-in-differences study design to evaluate the effect of school start time delays on academic outcomes among students in the United States. It is important to understand the effect of school start time policies on educational outcomes because academic performance is of great importance to both students and to the local policymakers (typically school boards) who determine school start times in the United States. Additionally, policy changes that improve educational outcomes can promote health equity by addressing a fundamental cause of health disparities (Link and Phelan, 1995).

Contribution

Despite evidence that high school start time changes should improve students’ educational performance, prior research on this topic in the United States has identified mixed effects that vary both by specific educational indicator and study methodology. This study used a more rigorous difference-in-differences approach to evaluate the educational outcomes of students who did and did not experience a school start time delay, following both groups one year before and two years after the policy change. We consider a school start time delay of 50 to 65 minutes that occurred in the fall of 2016 at two high schools in the Minneapolis-St. Paul metropolitan area, with three comparison schools that maintained early start times. This policy change extended students’ objectively-measured nightly sleep, measured using wrist actigraphy, by 40 minutes per night on school nights at follow-ups that occurred both one and two years later (Widome et al., 2020a). We consider four outcomes related to school engagement and performance – late arrivals, absences, behavioral referrals, and grades – using high-quality data drawn from administrative records provided by school districts.

Methods

Data

The START/LEARN cohort study included approximately 2,500 adolescents who attended five high schools in three school districts in the Minneapolis – St. Paul metropolitan area. All participants were in 9^th grade in the 2015-2016 academic year and were followed through 10^th grade (2016-2017, follow-up 1) and 11^th grade (2017-2018, follow-up 2). Additional details on study recruitment and sampling are available in Widome et al. (2020b). The START/LEARN study was approved by the University of Minnesota Institutional Review Board and participating school districts’ research review panels.

Of the 2,466 adolescents who participated in START/LEARN, our analyses include 2,153 adolescents (87% of participants) for whom we could determine each student’s policy condition, specific school, educational outcomes, and demographic covariates. Appendix A shows the sample selection process, and Table 1 describes the analytic sample.

Table 1.

Analytic sample description; mean (SE) or percent

	Full sample N = 2,153		Policy change N = 1,269		No change N = 884
Policy condition
No change	41%		0%		100%
Policy change (SST delay)	59%		100%		0%
District
Wayzata	36%		62%		0%
Buffalo	22%		38%		0%
Elk River	41%		0%		100%
Number of schools	5		2		3
Outcomes
Late arrivals
Baseline (2015-2016)	2.1 (0.1)		2.3 (0.1)		1.8 (0.1)
Follow-up 1 (2016-2017)	3.1 (0.1)		2.5 (0.4)		4.0 (0.2)
Follow-up 2 (2017-2018)	4.6 (0.2)		3.5 (0.2)		6.1 (0.3)
Absences
Baseline (2015-2016)	6.2 (0.1)		6.2 (0.2)		6.3 (0.2)
Follow-up 1 (2016-2017)	8.3 (0.2)		8.0 (0.2)		8.7 (0.2)
Follow-up 2 (2017-2018)	9.4 (0.2)		8.9 (0.2)		10.1 (0.3)
Any behavior referral
Baseline (2015-2016)	7%		5%		11%
Follow-up 1 (2016-2017)	8%		4%		15%
Follow-up 2 (2017-2018)	11%		3%		23%
GPA	All schools, Wayzata weighted	All schools, Wayzata unweighted	All schools, Wayzata weighted	All schools, Wayzata unweighted
Baseline (2015-2016)	3.19 (0.02)	3.19 (0.02)	3.30 (0.02)	3.30 (0.02)	3.02 (0.03)
Follow-up 1 (2016-2017)	3.13 (0.02)	3.10 (0.02)	3.26 (0.02)	3.21 (0.02)	2.94 (0.03)
Follow-up 2 (2017-2018)	3.09 (0.02)	3.04 (0.02)	3.29 (0.02)	3.19 (0.02)	2.81 (0.03)
GPA	All schools, Wayzata weighted	All schools, Wayzata unweighted	All schools, Wayzata weighted	All schools, Wayzata unweighted
Pre-baseline (Fall 2015)	3.23 (0.02)	3.22 (0.02)	3.35 (0.02)	3.34 (0.02)	3.05 (0.03)
Baseline (Spring 2016)	3.15 (0.02)	3.14 (0.02)	3.25 (0.02)	3.25 (0.02)	2.98 (0.03)
Follow-up 1a (Fall 2016)	3.14 (0.02)	3.10 (0.02)	3.27 (0.02)	3.22 (0.02)	2.95 (0.03)
Follow-up 1b (Spring 2017)	3.11 (0.02)	3.09 (0.02)	3.23 (0.02)	3.20 (0.02)	2.93 (0.03)
Follow-up 2a (Fall 2017)	3.11 (0.02)	3.05 (0.02)	3.31 (0.02)	3.20 (0.02)	2.84 (0.04)
Follow-up 2b (Spring 2018)	3.06 (0.02)	3.01 (0.02)	3.25 (0.02)	3.17 (0.02)	2.79 (0.04)
ACT score1	23.0		24.0		21.3
Covariates
Male	51%		50%		52%
Non-white	20%		25%		13%
Free-or-reduced-price lunch	18%		19%		16%
Parent graduated college	82%		85%		77%

¹ACT score is provided for purposes of characterizing school districts only. It is not used in our analyses. This measure is missing for 221 cases that are otherwise included in the analytic sample; these values are the mean among non-missing data.

Measures

Educational outcomes

All data on educational outcomes were provided by school districts and were drawn from administrative records.

Late arrivals, absences, and behavior referrals.

For each student, we measured late arrivals, absences, and behavioral referrals for each of the three academic years. We dichotomized this measure into an indicator of whether a student had any behavior referrals for a given school year.

Grade point average.

Grade point average (GPA) averaged a student’s grades in all courses that a student took in a given semester. School districts had slightly different methods of calculating GPA. Two of the three districts calculated a “weighted” GPA for high grades in more challenging courses (e.g., Advanced Placement courses), such that a student in an advanced course would receive more GPA points than a student with the same grade in a regular course. The third district calculated unweighted GPA. Only one of the districts that weighted GPA reported both weighted and unweighted GPA for its students. These differing GPA calculation practices make comparing GPA trajectories across districts more complex. Thus, we show two sets of results, one using weighted GPA for the district that provided both weighted and unweighted GPA and the other using unweighted GPA for that district.

• Policy change school district 1: Wayzata Public Schools (one high school) reported both weighted and unweighted GPA. In the district’s weighting scheme, students received one additional grade point for grades of C-or higher in Advanced Placement Courses. Wayzata offered approximately 32 advanced courses during the years of this study.

• Policy change school district 2: Buffalo-Hanover-Montrose Schools (one high school) reported unweighted GPA. The district offered five Advanced Placement courses and 37 concurrent college courses during this study.

• Comparison school district: Independent School District 728 (two high schools and one combined middle/high school; we include only high school students at the combined school) provided weighted GPA measures only. In the district’s weighting scheme, students received 0.34 more grade points for grades of C-or higher in Advanced Placement courses or concurrent college courses. In the years of this study, the district offered about 16 Advanced Placement courses and 10 concurrent college courses.

Exposure

School start time.

When students were in ninth grade, all three districts had early high school start times. Beginning in the fall of 2016 (fall of students’ tenth grade year), two of the three districts delayed start times by approximately one hour.

• Policy change school district 1: Wayzata Public Schools (one high school) delayed its start time from 7:30am to 8:20am (a 50-minute delay) in fall 2016.

• Policy change school district 2: Buffalo-Hanover-Montrose Schools (one high school) shifted from a 7:45am start time to an 8:50am start time (a 65-minute delay) in fall 2016.

• Comparison school district: Independent School District 728 (two high schools and one middle/high school; we included only high school students at the combined school) maintained a 7:30am start time throughout the study period.

Demographic covariates

Students reported their biological sex (male, female, or prefer not to answer), ethnoracial group, and parental educational attainment. We collapsed ethnoracial group into two categories, non-Hispanic white and nonwhite (including American Indian, Alaskan Native, Asian, Black, Hispanic, Native Hawaiian or Pacific Islander, multiracial, or any other group) due to small numbers of nonwhite respondents in this sample. Students also indicated the educational attainment of up to two parents or guardians. We created a dichotomous measure indicating whether or not the student had at least one parent/guardian who completed college; 80% of students had a parent or guardian who completed college. Both students and schools provided information on whether the student received free or reduced-price lunch (FRPL) over the three years of this study. We coded students for whom there was any indication that they ever received FRPL (either the student’s survey report of receiving FRPL or a school report that the student ever received FRPL) as ever having received FRPL. We included these covariates to attempt to account for factors that may have contributed to both which school districts chose to delay their start times and the educational outcomes of students in these districts. (We were not able to account for these factors at the school level, e.g., by including the share of nonwhite students as a school-level covariate, because of the small number of schools in our study.)

Analytic approach

We tested whether patterns of late arrivals, absences, behavior referrals, and GPA differed over time between policy conditions (early vs. late start times) using a difference-in-differences approach. Specifically, we estimate multilevel mixed effects linear regressions of late arrivals, absences, and behavior referrals across three school years and GPA across six semesters. The statistical model is:

$$\text{L}1:{\text{Y}}_{\text{ij}}={\beta }_{0\text{j}}+{\beta }_{1\text{j}}\left({\text{PolicyCondition}}_{\text{i}}\right)+{\beta }_{2\text{j}}\left({\text{Time}}_{\text{j}}\right)+{\beta }_{3\text{j}}\left({\text{PolicyCondition}}_{\text{i}}{\text{Time}}_{\text{j}}\right)+{\beta }_{4\text{j}}\left({\text{Covariate}}_{\text{i}}\right)+{\beta }_{5\text{j}}\left({\text{Covariate}}_{\text{i}}{\text{Time}}_{\text{j}}\right)+{\text{e}}_{\text{ij}}$$

$$\text{L}2:{\beta }_{0\text{j}}={\gamma }_{00}+{\text{u}}_{0\text{j}}$$

The betas are the regression coefficients associated with each predictor, eij is the individual (L1) error, γ₀₀ is the average intercept across schools (L2s), and u_oj is the school (L2) error. Our models included both school and student random effects (u_0j is a random intercept for school) and adjusted for students’ sex, ethnoracial group, free or reduced-price lunch receipt, and parental college completion. (Due to extremely small cell sizes, it was not feasible to adjust for race in models of behavior referrals.) These covariates adjusted for any confounding between policy condition and educational outcomes. We allowed effects of covariates to vary across year/semester by using interaction terms as noted in the equation above. The baseline time period in analyses of late arrivals, absences, and behavior referrals was the first and only observed school year before the policy change took place (2015-2016). For analyses of GPA, the baseline time period was the semester immediately preceding the policy change (spring 2016, which was the second semester in which we observed GPA). We then used these models to predict levels of the outcomes variables for each outcome by policy condition and time, as presented in Figures 1 and 2.

Figure 1.

Predicted number of late arrivals and absences and probability of behavior referral by policy condition

Note: Models of late arrivals and absences are adjusted for sex, race, free-or-reduced-price lunch receipt, and parental college completion. Models of the probability of any behavior referral are adjusted for sex, free-or-reduced-price lunch receipt, and parental college completion. Error bars indicate 95% confidence intervals. Asterisks indicate statistical significance of difference-in-differences, *** p < 0.001, ** p < 0.01, * p < 0.05.

Figure 2.

Predicted GPA by policy condition

Note: Adjusted for sex, race, free-or-reduced-price lunch receipt, and parental college completion. Error bars indicate 95% confidence intervals. Asterisks indicate statistical significance of difference-in-differences, *** p < 0.001, ** p < 0.01, * p < 0.05.

For each educational outcome, the target estimand was the average difference in the educational outcome with later school start time versus a continued early school start time. We used the difference-in-difference estimator to estimate this effect. Specifically, we used the difference in average educational outcome in the delay schools before and after school start time change minus the difference in average educational outcome in the comparison schools over the same time periods. In the equation above, the policy*time effect is our difference-in-differences estimate.

Results

Sample description

The analytic sample included 2,153 adolescents, of whom 51% were male, 20% were non-white, 18% received free-or-reduced-price lunch, and 82% had a parent who graduated college. Students who attended policy change schools were 50% male, 25% non-white, 19% free-or-reduced-price lunch recipients, and 85% students with a college graduate parent. Students in comparison schools were 52% male, 13% non-white, 16% free-or-reduced price lunch recipients, and 77% students with a college-graduate parent.

School start times and educational outcomes

Students who experienced a school start time delay had different trajectories of late arrivals, absences, behavior referrals, and GPA than students who did not experience a school start time delay. Figures 1 and 2 display the predicted values for each of these outcomes, and Tables 2 and 3 show the corresponding difference-in-differences analyses.

Table 2.

Predicted marginal effects and difference-in-differences (95% confidence intervals in parentheses) from adjusted models of educational behavior

	Policy change schools			Comparison schools			Difference-in-differences
	Baseline (2015-2016)	Follow-up 1 (2016-2017)	Follow-up 2 (2017-2018)	Baseline (2015-2016)	Follow-up 1 (2016-2017)	Follow-up 2 (2017-2018)	Baseline to follow-up 1	Baseline to follow-up 2
Number of late arrivals	2.0 (1.3, 2.8)	2.2 (1.5, 3.0)	3.2 (2.4, 4.1)	2.0 (1.3, 2.6)	4.2 (3.5, 4.9)	6.3 (5.5, 7.1)	−2.0*** (−2.5, −1.4)	−3.1*** (−3.8, −2.4)
Number of absences	6.8 (4.4, 9.2)	8.7 (6.4, 11.1)	9.8 (7.4, 12.2)	6.1 (4.2, 8.1)	8.5 (6.6, 10.5)	9.9 (7.8, 11.9)	−0.5 (−1.0, 0.09)	−0.8** (−1.5, −0.2)
Probability of any behavior referral	0.05 (−0.03, 0.13)	0.04 (−0.03, 0.11)	0.03 (−0.03, 0.10)	0.17 (0.02, 0.32)	0.22 (0.05, 0.39)	0.29 (0.11, 0.47)	−0.06** (−0.10, −0.02)	−0.14*** (−0.20, −0.08)

Notes:

^***p < 0.001,

^**p < 0.01,

^*p < 0.05.

Models of late arrivals and absences are adjusted for sex, race, free-or-reduced-price lunch receipt, and parental college completion. Models of the probability of any behavior referral are adjusted for sex, free-or-reduced-price lunch receipt, and parental college completion.

Table 3.

Predicted marginal effects and difference-in-differences (95% confidence intervals in parentheses) from adjusted models of Grade Point Average

	Policy change schools
	Pre-baseline (Fall 2015)	Baseline (Spring 2016)	Follow-up 1a (Fall 2016)	Follow-up 1b (Spring 2017)	Follow-up 2a (Fall 2017)	Follow-up 2b (Spring 2018)
GPA, Wayzata weighted	3.32 (3.23, 3.40)	3.23 (3.14, 3.31)	3.23 (3.14, 3.31)	3.20 (3.11, 3.28)	3.27 (3.18, 3.36)	3.20 (3.11, 3.29)
GPA, Wayzata unweighted	3.32 (3.24, 3.39)	3.23 (3.15, 3.30)	3.18 (3.11, 3.26)	3.16 (3.09, 3.24)	3.17 (3.09, 3.25)	3.13 (3.05, 3.21)
	Comparison schools
	Pre-baseline (Fall 2015)	Baseline (Spring 2016)	Follow-up 1a (Fall 2016)	Follow-up 1b (Spring 2017)	Follow-up 2a (Fall 2017)	Follow-up 2b (Spring 2018)
GPA, Wayzata weighted	3.03 (2.95, 3.11)	2.96 (2.88, 3.04)	2.95 (2.87, 3.03)	2.93 (2.85, 3.01)	2.84 (2.75, 2.92)	2.79 (2.71, 2.88)
GPA, Wayzata unweighted	3.03 (2.96, 3.10)	2.96 (2.89, 3.03)	2.94 (2.87, 3.02)	2.93 (2.85, 3.00)	2.83 (2.75, 2.90)	2.79 (2.71, 2.87)

	Difference-in-differences
	Pre-baseline to baseline	Baseline to follow-up 1a	Baseline to follow-up 1b	Baseline to follow-up 2a	Baseline to follow-up 2b
GPA, Wayzata weighted	0.02 (−0.03, 0.07)	0.01 (−0.03, 0.06)	−0.0005 (−0.05, 0.05)	0.17*** (0.11, 0.23)	0.14*** (0.07, 0.21)
GPA, Wayzata unweighted	0.02 (−0.03, 0.06)	−0.03 (−0.08, 0.02)	−0.03 (−0.08, 0.02)	0.08* (0.02, 0.13)	0.07* (0.006, 0.14)

Notes:

^***p < 0.001,

^**p < 0.01,

^*p < 0.05.

Adjusted for sex, race, free-or-reduced-price lunch receipt, and parental college completion.

Late arrivals

Late arrivals increase over time for all students, but this increase was more modest in policy change schools than comparison schools. Students who experienced a school start time delay had a similar number of late arrivals to those in comparison schools at baseline (top panel of Figure 1) but fewer late arrivals at follow-ups after the policy change (Table 2 and top panel of Figure 1). Specifically, students in both policy change and comparison schools had an average of two late arrivals at baseline. The school start time delay led to about two fewer late arrivals per year at follow-up 1 and three fewer late arrivals per year at follow-up 2 (p < 0.001) at policy change schools relative to comparison schools (Table 2).

Absences

Patterns of absences are shown in Table 2 and the middle panel of Figure 1; absences increased over time for all students. Students in both policy change and comparison schools had similar numbers of absences before the school start time delay. The start time delay did not have an immediate effect on absences, with students in both policy conditions having about 9 absences per year at follow-up 1. However, at follow-up 2, students at policy change schools were absent about one fewer day per year than students in comparison schools (p < 0.01, Table 2).

Behavior referrals

Students in comparison schools that did not experience a start time delay had a higher probability of experiencing a behavior referral than students in policy changes schools both before and after the school start time policy change, likely reflecting differences in schools’ disciplinary practices (Table 2 and bottom panel of Figure 1). The policy change exacerbated this difference, however. Both before and after the policy change, 3-5% of students who experienced a school start time delay had behavior referrals. For students in comparison schools that maintained early start times, the probability of a behavior referral increased from about 17% at baseline to 22% at the first follow-up and 29% at the second follow-up. This led to a difference-in-differences of 6% (p < 0.01) at follow-up 1 and 14% (p < 0.001) at follow-up 2.

Grade point average

As described in the Measures section, we analyzed GPA calculated in two ways to address differences in GPA weighting schemes across school districts. Figure 2 shows predicted GPA values. The left panel uses a weighted GPA measure from Wayzata high school and the right panel uses an unweighted GPA measure from Wayzata high school. For both measures, trajectories of GPA were parallel before baseline and through follow-ups 1a and 1b. These trajectories diverged at follow-ups 2a and 2b, however. Students who experienced a school start time delay had less steeply declining GPA trajectories than students who did not experience a school start time delay, though GPA decreased over time in all schools. When using the GPA measure in which Wayzata is weighted, these differences were 0.17 GPA points at follow-up 2a (p < 0.001) and 0.14 GPA points at follow-up 2b (p < 0.001). When Wayzata is unweighted, the respective differences were 0.08 (p < 0.05) and 0.07 (p < 0.05).

Supplementary analyses: Demographic moderation

We tested whether these effects differed by student demographic characteristics for late arrivals, absences, and GPA by adding three-way interaction terms for covariate*time*policy to our models. (We were not able to test moderation for behavior referrals due to small sample sizes.) Most associations were similar across most groups, and we did not identify any moderation of effects on absences or by race. The effect of the school start time delay on late arrivals was larger for males than females at follow-up 2 (3.9 vs 2.4, p<0.05). Effects of the policy change on GPA were larger for males than females at follow-up 1a only (the semester immediately after the policy change) (0.06 vs −0.04 when weighted, p<0.05; 0.02 vs −0.08 unweighted, p<0.05). Socioeconomically more advantaged students’ GPAs benefitted more from the school start time delay than did less advantaged students. Students who did not receive free-or-reduced-price lunch had higher weighted GPAs at follow-ups 2a (0.21 vs −0.02, p<0.01) and 2b (0.1 vs −0.05) and higher unweighted GPAs at follow-up 2b (0.11 versus −0.1, p<0.05) than students who received free-or-reduced-price lunch. Similarly, students who had a parent who graduated college had higher weighted GPAs at follow-ups 2a (0.21 vs −0.02, p<0.01) and 2b (0.19 vs −0.06) and higher unweighted GPAs at follow-up 2b (0.11 versus −0.06, p<0.05) than students whose parents did not graduate from college.

Conclusions

We found that a high school start time delay of 50 to 65 minutes led to 3 fewer late arrivals, 1 fewer absence, a 14% lower probability of behavior referral, and 0.07-0.17 higher GPA in the two years after the start time delay compared to maintaining an early school start time. These effects were more pronounced (or, for absences and GPA, limited to) the follow-up that occurred two years after the implementation of the policy change than in the first year of follow-up. Though differences were modest for any specific outcome (with late arrivals having the largest improvement), small gains across several dimensions are collectively promising for student engagement and performance.

These findings strengthen existing evidence that later high school start times improved on-time arrivals, absences, and disciplinary problems in the United States (Lenard et al., 2020; Owens et al., 2010; Thacher and Onyper, 2016; Wahlstrom et al., 2014). The role of school start time delays on academic performance has varied in prior work, with positive effects in some cases (Wahlstrom et al., 2014, 1997) but null effects in others (Boergers et al., 2014; Owens et al., 2010; Thacher and Onyper, 2016; Wahlstrom, 2002). None of those studies used a difference-in-differences design to evaluate a school start time delay, though Lenard et al. (2020) considered a school start time advance in which high school start times were changed to be 40 minutes earlier (the opposite direction of recommended policy change) and found that late arrivals, absences, and dropouts increased while test scores were unaffected. Additionally, Fuller and Bastian (2022) used a comparative interrupted time series approach that is similar to a difference-in-differences framework and found that a 90-minute start time delay improved grades but not test scores, with inconsistent effects on late arrivals and suspensions.

With our causally rigorous approach, we find that GPA is affected by school start time, but only two years after a policy change.

In addition to providing new evidence on the effect of high school start time delays, our findings are aligned with a broader literature on the importance of school timing for student outcomes in the United States. For example, prior research in the United States has found that earlier course timing decreases grades among college students (Carrell et al., 2011; Diette and Raghav, 2017); later start times increase math and reading test scores among middle school students (Edwards, 2012); and start times that are later relative to sunrise increase math and reading scores among middle and high school students, with larger effects for older students (Heissel and Norris, 2018). Biller et al. (2022) provide a complete review of extant studies of the effects of school start times on grades and test scores among adolescents worldwide.

We note that the two-year follow-up period used in this study is also conflated with our cohort study design. The START/LEARN cohort is a group of ninth grade students followed prospectively. Thus, over the study period these students experienced several simultaneous exposures: the passage of time since the policy change occurred, growing older, and progressing into tenth and eleventh grades. We cannot disentangle how these factors may have combined to produce delayed effects for some outcomes but not others. First, time since policy change could have affected outcomes differently. However, since the mechanism presumed to link the policy change and student outcomes – nightly sleep duration – increased immediately after the policy change in this cohort and remained stable over two years (Widome et al., 2020a), this explanation is less likely to explain differences between outcomes. Second, effects could be determined by the age of the student. Older students, who are more advanced in puberty, have greater pressure to sleep at later hours than younger adolescents and also have increasing autonomy over their own schedules and transportation to school. This may be one reason that absences increase in the second follow-up only. Finally, there may be grade-level-specific differences. For example, eligibility for advanced courses may be limited to students in higher grades. Grades in these more rigorous courses could be differently affected by sleep than grades in less demanding courses. The most similar prior study to the present approach used a comparable longitudinal cohort approach to examine the consequences of a school start time delay but unlike the present study did not have a control group. Thacher and Onyper (2016) compared GPA trajectories in successive cohorts of high school students at a single school who experienced a school start time delay after their ninth, tenth, and eleventh grade school years. They did not identify any effects of school start time delay on GPA, though all students’ grades rose in twelfth grade regardless of start time. This difference from our results may be an artifact of their lack of a comparison group that did not experience a school start time delay.

Though this study is strengthened by its difference-in-differences design, it also has several limitations. First, this is a natural experiment. Schools/districts were not randomly assigned to policy conditions by the research team, meaning that there could be confounding by other student or school characteristics. The difference-in-differences approach used here assumes that trends would be otherwise parallel between the treatment and control group in the absence of policy change, meaning our findings rest on the assumption that our comparison group is a good proxy for counterfactual change in the treated group. We attempted to correct for these potential biases by accounting for student characteristics in our models, but we could not adjust for school-level confounders due to small number of schools in our study. Additional pre-treatment data would have strengthened the parallel trends assumption, as we have only one pre-treatment data point (grade 9). However, four of our five schools served students in grades 9-12 only. Even if it had been possible to get data from students’ prior schools, it is not clear that performance in a different set of middle schools would have represented an appropriate pre-treatment comparison. Second, data come from adolescents living in the Minneapolis – St. Paul metropolitan area, and findings may not generalize to students in other geographic areas. This is a relatively advantaged sample of students and schools, and it is not clear whether the effect of a school start time delay on educational outcomes would be similar for students or schools with fewer resources. Third, comparing student grades across schools is complex because course grades for the same level of performance vary between schools. In our study, these differences may be exacerbated by differing GPA weighting schemes. We attempt to address this by presenting analyses of both weighted and unweighted GPA measures for the school that provided both such measures, but we cannot know the true effect of measurement differences on our results. Despite these complexities, high school GPA is a stronger and more consistent predictor of college graduation than standardized tests such as ACT scores (Allensworth and Clark, 2020), making it an important target for intervention.

Based on the findings and limitations of this study, we offer three directions for future research on school start times and educational outcomes. First, we need to better understand why the effect of a school start time delay on educational outcomes that we identified here manifested gradually over time. This pattern is particularly perplexing because a prior study using the same policy change found immediate and sustained effects of the school start time delay on sleep that remained stable in the two years after the policy change (Widome et al., 2020a). Future research should elucidate in which ways these differences are produced by the combination of time-since-policy-change, student age, and grade-level-specific factors discussed above. Second, we recommend that subsequent studies consider longer follow-up periods. In particular, understanding academic performance in the final year of high school and following students through high school graduation is important for assessing the long-term educational attainment consequences of school start time delay policies; one study has found that school start time delays increased high school graduation rates (McKeever and Clark, 2017). Additionally, it is unknown whether improvements to nightly sleep brought on by a school start time delay might modify life course sleep trajectories in a lasting manner beyond adolescence. Third, we suggest that future research consider the timing of events within the school day. Prior research has indicated that school start times have a larger effect on absences, late arrivals, and grades in first period courses (Bastian and Fuller, 2018; Dunster et al., 2018), though these studies used a less rigorous design than that presented here. We were unable to obtain data on specific course grades for this study. Fourth, future studies should assess whether school start time effects on grades vary across subject areas (e.g., math versus language arts). Though we were not able to obtain information on grades in specific courses, prior work suggests that effects may differ by subject: Groen and Pabilonia (2019) found that a high school start time delay improved reading scores (for girls only) but did not change math scores.

Nonetheless, this study provides crucial information for understanding the effects of school start time changes on adolescents’ educational performance – a domain of great importance to both adolescents and to the policymakers who make school start time policies. We found that students who experience a school start time delay are in class on time and for more days per school year, have fewer behavior problems, and get better grades than their peers who have earlier start times. Thus, delaying high school start times is a promising policy intervention that not only improves adolescents’ sleep, health, and behavior (Minges and Redeker, 2016) but can also strengthens their educational performance. Educational attainment is in turn a fundamental cause of health, suggesting yet another pathway through which delaying school start times for adolescents can promote health equity.

Appendix A. Analytic sample selection

Data availability statement:

Because this study was funded by the National Institutes of Health (NIH), data from this study will be made available within one year as specified NIH open access policy. At present, a de-identified dataset that includes the variables used in this analysis can be requested by submitting a data analysis proposal and plan for review to the corresponding author.