School Start Times, Sleep, and Youth Outcomes: A Meta-analysis

Literature searches were conducted Scopus, ScienceDirect, JSTOR, Pubmed, PsychInfo, ERIC, Proquest, EBSCO, and Google Scholar, using a combination of keywords associated with school start time (“school start time” or “school starting times”), sleep (“sleep”, “sleep duration”, or “sleep pattern”), and developmental outcomes (“achievement”, “academic performance”, “mental health”, “well-being”, or “health”) across samples with different ages and stages of development (ie, elementary, middle, or high school). The searches included studies through the end of 2019 and produced 1263 records. One hundred and nineteen duplicate records were removed, and the remaining 1144 abstracts were reviewed and screened. Of these, only articles meeting all of the following inclusion criteria were retained: (1) measured school start time; (2) measured sleep or other developmental outcomes; (3) published in English; and (4) could not be excluded based on available information in the abstract. The research team requested unpublished data from authors who have published on pediatric sleep, school start time, sleep, and developmental outcomes via e-mail and ResearchGate (n = 10); however, no records were added through this approach. Taken together, the literature searches and screening resulted in 80 reports retained for further examination.

The full texts of these 80 reports were screened for the following exclusion criteria: (1) did not include school start time (n = 9); (2) did not include empirical data (n = 21, including 17 meta-analysis and review articles); (3) did not include appropriate statistics (eg, qualitative research, n = 1); and (4) duplicate reports (n = 3); as a result, 34 were excluded (46 reports were retained). The research team also reviewed the reference lists of 17 meta-analysis and systematic review articles, producing 125 records. After screening and cross-checking, 114 reports were excluded and 11 articles were retained, yielding a total of 57 (46 + 11) reports. Approximately half of these 57 articles (n = 28) were coded by 2 to 3 coders. Coding discrepancies were discussed and resolved to achieve intercoder agreement (interrater reliability rs = 0.99–1.00 for continuous variables and κs = 0.83−1.00 for categorical or string variables). Once reliability was achieved, the remaining 29 articles were independently coded by 1 of the 3 trained coders.

The 57 reports were coded, and the primary investigator contacted authors (n = 53) to obtain missing information (eg, group comparison statistics, sample descriptives, and demographic information) for effect size computation (eg, correlations between school start time and sleep) or moderation analysis (eg, participants’ race or ethnicity). Only 21 requests were met; 13 authors indicated they were not able to provide requested statistics or raw data, and 19 authors did not respond despite several attempts. The coders coded 39 reports with complete data, whereas 29 reports had sufficient statistical information for inclusion (see Supplemental Table 9 for additional information about included studies). Two reports analyzed the same research data; therefore, 28 studies were included in the final analyses (see Fig 1 PRISMA flowchart²¹ ): 27 peer-reviewed articles and 1 research report. A detailed data-extraction manual can be found in the Supplemental Information. Additional information about studies for which there was insufficient information for inclusion is shown in Supplemental Table 10.

Many studies (n = 18) measured specific school start times (eg, 7:45 am, “What time do you need to arrive at school?”¹⁷ ), whereas others (n = 10) used school start time ranges (eg, 7:30–7:45 am²² ). Approximately 36% of studies (n = 10) included data pre- and postdata examining developmental outcomes before and after a change in school start time (eg, delaying start time from 7:50 to 8:45 am in the Seattle School District²³ ), whereas remaining studies (n = 18) compared SSTs across countries (Australia versus United States¹⁷ ), districts (eg, Altoona versus Chippewa Falls in Wisconsin²⁴ ), or schools (eg, “early-starting schools” versus “late-starting schools”²⁵ ).

The following terminology is used to synthesize across analytic approaches: “initial SST” refers to the SST before the implementation of an SST delay, whereas “new SST” refers to the SST after the implementation of an SST delay. Two additional terms allow for the inclusion of studies that did not implement a change in SST but analyzed schools with different start times. These terms include “earlier SSTs”, which are discussed in contrast to “later SSTs” as a descriptive comparison of SST timing irrespective of any changes implemented in the SSTs.

First, the current study investigated the direct effects of school start times (ie, new, later school start times) on outcomes without considering initial or later start time. Next, initial and later SST were considered as moderators of the direct effect. To conduct the moderation analyses, all SSTs were coded into 5 groups: (1) before 7:30 am, (2) between 7:30 and 7:59 am, (3) between 8:00 and 8:29 am, (4) between 8:30 and 8:59 am, and (5) after 9:00 am to account for differences in how SSTs were reported; for example, some studies reported SSTs as specific times whereas others reported a range of times. However, although SSTs were coded categorically in the moderation analyses, they were treated as a continuous variable when estimating direct effects on developmental outcomes.

To investigate the broad impact of SSTs, this study focused on a range of developmental outcomes: (1) sleep (n = 25), (2) socioemotional health (both positive and negative; n = 8), (3) academics (eg, grades; n = 9), (4) cognitive development (eg, sustained attention; n = 3), (5) behavioral health (eg, substance use; n = 3), and (6) physical health (n = 3). Sleep indicators consisted of duration (n = 23), quality (n = 10), bedtime (n = 17), wake time (n = 15), sleepiness (n = 13), chronotype (n = 7), hygiene (n = 2), and social/jet lag and naps (n = 3). Sleep indicators were assessed using self-reports (n = 18), actigraphy (n = 2), a combination of self-reports and actigraphy (n = 3), a combination of actigraphy and laboratory assessment (n = 1), and a combination of self-reports and other method (ie, “school records”; n = 1). Other outcomes were assessed using self-reports (n = 6), school records (n = 4), parent reports (n = 1), cognitive tests (n = 1), biomarkers (n = 1), a combination of self-reports and parent reports (n = 1), a combination of self-reports and school records (n = 1), a combination of self-reports, teacher reports, and cognitive tests (n = 1), a combination of self-reports, teacher reports, and school records (n = 1), and a combination of self-reports and health center records (n = 1).

To standardize interpretation, all effect sizes were coded such that positive effect sizes represent later SSTs having a stronger association with better youth outcomes. For example, if a study reported a negative correlation between SST and depressive symptoms, the correlation coefficient was reverse coded to capture a positive association between later SST and better socioemotional well-being. As another example, if a study reported a mean difference in sleep duration between 2 groups with different SSTs, Cohen’s d was computed to compare sleep duration in the latter SST group to the earlier SST group.

Risk of bias was evaluated for all synthesized studies using an adapted version of the Newcastle-Ottawa Scale for assessing the quality of nonrandomized studies in meta-analyses.²⁹  The third and fourth authors coded the representativeness of the sample (r = 0.93), inclusion of intervention cohorts (r = 0.93), assessment of school start times (r = 0.98), whether analyses controlled for baseline measure of outcome (r = 1.00), control confounding variables (r = 0.91), assessment of outcome (r = 0.98), and adequacy of follow-up of cohorts (r = 0.90). A detailed coding manual is reported in Supplemental Table 11. The rating resulted in a sum score per study (possible ranges = 0–21, with higher scores indicating better quality). Next, the quality of each study was coded as high (sum scores of 17–21), moderate (sum scores of 10–16), low (sum scores of 5–9), or poor (sum scores of 0–4). The strength of evidence was coded for each outcome domain following prior meta-analyses.³⁰  The first and second authors coded risk of bias (r = 0.94), imprecision (r = 1.00), inconsistency (r = 1.00), indirectness (r = 1.00), publication/reporting bias (r = 0.88). A detailed coding manual is reported in Supplemental Table 12. The coding resulted in a sum score for each study that ranged between 5 and 12. The sum scores were averaged within each outcome domain. Next, strength of evidence for each outcome was coded where scores of 5 to 6 were coded as high, 7 to 10 as moderate, and 11 to 12 as low.

All analyses were conducted in a meta-regression framework using the robumeta R package.³¹  Random-effects models were used to allow true effect sizes to vary among studies.²⁶  Robust variance estimations³¹  were used to handle nonindependence in effect sizes (eg, effect sizes drawn from the same study or project). Correlated effects weights were used based on the most prevalent source of dependence in the data (ie, multiple measures from the same participants³² ). The average correlation between dependent effect sizes (ρ; ranging between 0 and 1) was specified at the default value (0.80).³¹  Previous research demonstrates robustness in estimates of effect sizes across reasonable values of ρ.³³  Finally, small sample adjustments were implemented to provide unbiased estimates with small numbers of studies (eg, <40) and skewed covariates.³⁴  Simulation research shows that after incorporating the small sample adjustments, the RVEs of effect sizes are robust when the degree of freedom is >4.³⁴  As such, estimates that had <4 degrees of freedom were excluded from analysis.

Intercept-only meta-regression models were conducted to examine the effect of SSTs on youth outcomes overall and for each domain separately. The estimated mean effect sizes were then converted to correlations, and Cohen’s³⁵  criteria were used to evaluate the effect sizes. The significance of SST effects was determined by the 95% confidence intervals (CIs) and P values. Heterogeneity was evaluated using 2 indices: the between-study variance in study-average effect sizes (τ²), and the ratio of true heterogeneity to total variance across the observed effect sizes (I²).

Next, univariate meta-regression models examined how the association between SST and youth developmental outcomes was moderated by school start time, youth characteristics (age, sex, and race or ethnicity), and school characteristics (school level, sector, size, SES, racial or ethnic composition, urbanicity, region, and country). Categorical moderators were dummy coded, and the reference group was rotated to obtain all possible pairwise comparisons. Next, subgroup analyses were conducted for each categorical moderator to obtain the mean effect size for each group. Meta-regression analyses investigate whether there are statistically significant differences by groups (ie, are the groups statistically different from each other) based on the moderating variable. In contrast, subgroup analyses complement the moderation analyses by providing estimates for the individual effect size by subgroup (ie, the specific subgroup effect size estimate).

Multiple sensitivity analyses were conducted to examine robustness. First, the overall association between SST and youth developmental outcomes was investigated by (1) controlling for participant age and (2) focusing on studies that reported youth-level data. Next, sensitivity analyses examined the moderating effects of SST separately at each school level (ie, elementary, middle, or high school). Two sets of sensitivity analyses were investigated for the moderating effects of youth and school characteristics: (1) controlling for age in all meta-regression analyses and (2) addressing potential heterogeneity across countries by focusing on US studies only for youth race or ethnicity, school sector, and urbanicity.

Finally, publication biases were investigated using funnel plots with Egger’s tests³⁶  and trim-and-fill analyses.³⁷  Because RVE is not available for assessing publication bias,³⁸  effect sizes were aggregated within each project, and publication bias using traditional meta-analytic methods (assuming effect sizes are independent) in random effects models were examined using the Metafor R package.³⁹ 

First, the average effect of SST was estimated across SST, youth characteristics, and school characteristics. This was done in 2 steps: (1) estimating the effect between SST and an aggregate indicator of overall development and (2) estimating the effect of SST for each indicator (ie, sleep, socioemotional, academic, behavioral, physical, and cognitive) separately. For socioemotional well-being, additional analyses examined positive and negative outcomes separately. Each sleep indicator (duration, quality, bedtime, wake time, hygiene, chronotype, social/jet lag and naps, and sleepiness) was analyzed separately. Table 1 presents the effect size, confidence interval, and heterogeneity estimates. Figure 2 presents a forest plot for all effect sizes. Supplemental Figs 4–13 present forest plots for each developmental indicator.

For aggregated developmental outcomes, there was a small SST effect (Table 1, top row) such that later SSTs were associated with better outcomes (mean effect size = 0.061; SE = 0.022, 95% CI = 0.015–0.107; P = .01). For sleep outcomes (Table 1, second row), there was no significant effect for aggregated sleep or for individual sleep indicators with 1 exception: later SSTs were associated with longer sleep duration, with a small effect size (mean effect size = 0.109; SE = 0.043; 95% CI = 0.019–0.199; P = .02). Studies on sleep hygiene and social/jet lag and naps were not sufficiently powered to produce reliable results (degrees of freedom [df] < 4). For socioemotional outcomes, positive and negative socioemotional well-being were investigated separately, and there was a small effect between later SSTs and lower levels of negative socioemotional well-being (higher scores indicate lower levels of negative socioemotional well-being; mean effect size = 0.060; SE = 0.023; 95% CI = 0.001–0.119; P = .04). There were no significant associations between SST and academic outcomes. Because of limited degrees of freedom (df < 4), there were not sufficient data to examine the association between SST and positive socioemotional well-being, and other developmental domains (ie, behavioral, physical, or cognitive).

Two sets of sensitivity analyses were conducted. The first set controlled for participant age and centered age at 15 years (ie, the mean age across studies; range was 9–21 years old), so that the intercept of each meta-regression model would estimate the mean effect size across studies. Results are shown in the top portion of Supplemental Table 13. Covarying age necessitated removing 2 studies that did not report participant age; this resulted in insufficient data to examine negative socioemotional outcomes, academic outcomes, or sleep chronotype. For outcomes that had sufficient degrees of freedom for analysis, an identical pattern of significance was observed as the primary findings reported above. The second set of sensitivity analyses examined the overall effects of SST on developmental outcomes, including only studies that report youth-level data. Results are shown in the bottom portion of Supplemental Table 13. Again, an identical pattern of significance was observed, supporting the robustness of the primary findings reported above.

To investigate whether associations between SST and outcomes were dependent upon initial SSTs, and to identify which start time categories may benefit most from later SSTs, moderation analyses were conducted on initial SST. Because previous research suggests potential nonlinear effects of SST,¹²  the 5 SST categories were dummy coded and tested as moderators in a meta-regression. The reference groups were rotated to obtain all possible comparisons. Subgroup analyses provided estimates of the effect sizes for the association between SST and youth outcomes for each SST category.

Several significant moderating effects emerged for initial SST, and the meta-regression results are shown in Table 2. For the aggregated developmental outcome indicator, later SST was associated with better outcomes for youth whose initial SST was 7:30 to 7:59 am (compared with youth with SSTs before 7:30 am). For sleep outcomes, later SST was associated with better overall sleep for youth whose initial SST was 7:30 to 7:59 am (compared with youth with SSTs before 7:30 am). No significant moderating effects were observed for other outcomes. Subgroup analyses (Table 4) observed that later SSTs were significantly associated with better aggregated outcomes, better aggregated sleep indicators, and longer sleep duration for youth whose initial SST was between 7:30 and 7:59 am, but not youth in other SST categories.

Sensitivity analyses examined the moderating effects of SST separately at each school level (elementary, middle, and high schools). Because of limited study numbers, however, analyses were only sufficiently powered for high school studies. Among high school studies, later SST was associated with better youth outcomes when initial SST was 7:30 to 7:59 am. Focusing on the moderating effects of initial SST (Supplemental Table 14), we observed 2 significant findings. Later SSTs were associated with better overall sleep and earlier bedtime for youth whose initial SST was 7:30 to 7:59 am (compared with youth with initial SSTs of 8:30–8:59 am).

To investigate whether associations between SST and outcomes were dependent upon new SSTs, moderation analyses investigated the impact of the new SST after the implementation of a later, delayed SST. The 5 SST categories were dummy coded and tested as moderators in a meta-regression. The reference groups were rotated to obtain all possible comparisons. Subgroup analyses provided the effect size estimates for the association between SST and youth outcomes for each SST category.

Significant moderating effects emerged for sleep outcomes and estimates from the meta-reg ression results are presented in Table 3. Later SSTs were associated with better sleep quality for youth whose new SST was 8:30 to 8:59 am (compared with those whose new SST was 8:00–8:29 am). No significant moderating effects were observed for other outcomes. Subgroup analyses showed that later SSTs were significantly associated with better overall outcomes, better overall sleep, and longer sleep duration among youth whose new SST was 8:30 to 8:59 am but not for other new SSTs (Table 5).

Sensitivity analyses examined the moderating effects of SST separately at each school level (elementary, middle, and high schools). Because of limited study numbers, however, analyses were only sufficiently powered for high school studies. Among high school studies, later SST was associated with better youth outcomes only when new SST was 8:30 to 8:59 am.

The final set of analyses examined the extent to which the association between SST and youth outcomes varied by youth (age, sex, and race or ethnicity) and school characteristics (school level, sector, size, socioeconomic status [SES], racial or ethnic composition, urbanicity, region, and country). Meta-regression analyses were conducted for each moderator separately. Each moderator (school level, sector, urbanicity, region, and country) was dummy coded, and the reference group was rotated to obtain all possible comparisons. To obtain effect size estimates, subgroup analyses examined the association between SST and youth outcomes within each moderator category.

No significant moderating effect emerged for youth age. Unfortunately, there were insufficient degrees of freedom to produce reliable estimates for moderation by sex or race or ethnicity (Table 6).

Significant moderating effects emerged for school characteristics, including school level, sector, and country (Table 7). Specifically, a significant difference in effect sizes emerged between middle and high school youth in which later SSTs were more strongly associated with lower levels of sleepiness for high school (compared with middle school) youth. No significant differences emerged for other outcomes. All subgroup analyses and coefficient estimates can be found in Table 8. For school sector (Table 7), significant differences emerged; for youth in private (compared with public) schools, later SSTs were more strongly associated with better overall sleep and later wake times. For geographic differences, effect sizes among youth in North America, Europe, and Asia were compared (Table 7). A significant pattern emerged; for youth in North America (compared with Europe), later SSTs were more strongly associated with earlier bedtimes.