Habitual Sleep Duration and Risk of Childhood Obesity: Systematic Review and Dose-response Meta-analysis of Prospective Cohort Studies

A meta-analysis of cross-sectional studies found that shorter-time sleep was correlated with increased risk of obesity in children. However, findings from prospective cohort studies were inconsistent. PubMed and other data resources were searched through May 2015. Twenty-five eligible studies were identified including 56,584 children and adolescents with an average 3.4-year follow-up. Compared with children having the longest sleep duration (~12.2 hours), kids with the shortest sleep duration (~10.0 hours) were 76% more likely to be overweight/obese (pooled odds ratio [OR]: 1.76; 95% confidence interval [CI]: 1.39, 2.23); and had relatively larger annual BMI gain (pooled β coefficient: 0.13; 95% CI: 0.01, 0.25 kg/m2). With every 1 hour/day increment in sleep duration, the risk of overweight/obesity was reduced by 21% (OR: 0.79; 95% CI: 0.70, 0.89); and the annual BMI gain declined by 0.05 kg/m2 (β = −0.05; 95% CI: −0.09, −0.01). The observed associations were not appreciably modified by region, baseline age or the length of follow-up. Accumulated literature indicates a modest inverse association between sleep duration and the risk of childhood overweight/obesity. Further research is needed to determine the age and gender specified optimal hours of sleep and ideal sleep pattern with respect to obesity prevention in children.

Statistical Analysis. According to the recommendation of the World Health Organization (WHO), overweight was defined as an age and gender specific BMI between the 85 th and 95 th percentile and obesity was defined as a BMI above the 95 th percentile 9 . The average follow-up time was calculated as the sum of person-years divided by the total number of participants.
To estimate the overall association between sleep duration and risk of overweight/obesity, we used the inverse of variance as the weight to calculate the pooled ORs and 95% CIs comparing the shortest to the longest category of sleep duration. Standard errors (SEs) were derived from the fully adjusted ORs and 95% CIs in the primary studies, which were transformed to natural logarithms (ln). To estimate the association between sleep duration and the continuous outcomes (e.g. annual BMI gain), we pooled the β regression coefficients weighted by the inverse of their variances considering that both exposure (i.e. sleep duration) and outcomes (e. g. annual BMI gain) were measured similarly in the primary studies 10 . If the information on linear association was not available in the primary study, it would be derived from generalized least-squares for trend test if the number of data points was ≥ 3 11 , or calculated directly under a linear assumption if the number of data points was < 3. If the extreme sleep duration category was open-ended (e.g. ≥ 12 hours/day), its lower/upper limit was estimated by assuming the range equivalent to its adjacent close-ended category.
To assess heterogeneity among the original studies, we inspected forest plots and conducted a Cochran's Q test with a P ≤ 0.10 considered as significant heterogeneity. We also computed the I 2 statistic to measure the magnitude of heterogeneity. The low, moderate, and high levels of heterogeneity were defined as < 30%, 30-50%, and > 50%, respectively. Sources of heterogeneity were explored using meta-regression and subgroup analyses with pre-defined factors including age (< 3, 3-< 5, ≥ 5 years), gender, follow-up time (above or below median) and study region (USA vs. non-USA).
Small-study effects or publication bias was assessed by funnel plot asymmetry followed by Egger's regression asymmetry test (when the number of studies was ≥ 3) or Begg's adjusted rank correlation test (when the number of studies was < 3). The Duval and Tweedie nonparametric "trim and fill" method was used to adjust for publication bias, if needed 12 .
Results from a random-effects model were presented as our main findings because we found that there were moderate/high heterogeneities in most of the pooled analyses and publication bias existed in some analyses 13 . Sensitivity analyses were performed to evaluate the robustness of the findings. Specifically, we determined the effects of a single study on the pooled results by removing one study at a time in the meta-analysis. Also, we explored the possible changes if replacing a random-effects model with a fixed-effects model.
All analyses were performed using STATA statistical software (Version 13.0; STATA Corporation LP, College Station, Texas, USA). A two-sided P value ≤ 0.05 was considered significant if not specified.
Scientific RepoRts | 5:16160 | DOi: 10.1038/srep16160 Results Literature search. As shown in Figure 1, 168 relevant articles were retrieved from PubMed. Of them, 145 were excluded for one of the following reasons: 1) not in English (n = 5); 2) not original studies (e.g., letter to editor) (n = 22); 3) not prospective cohort studies (n = 26); 4) not in children/adolescents (n = 26); 5) not on sleep duration or no outcome of interest (n = 72); 6) no sufficient information on measures of the association of interest (n = 1). In addition, 10 studies were identified from EMBASE, Google Scholar, or the reference lists of the relevant articles, for a total of 25 eligible studies in the meta-analysis 14-38 . Characteristics of included studies. Table 1 presents characteristics of the 25 included studies. Eleven studies were conducted in the USA [15][16][17]19,20,25,29,31,33,37,38 , four in Australia 23,28,32,35 , two in Canada 18,24 , two in Denmark 30,34 , and one in each of the following countries: UK 14 19,23,27,30 or three different age groups 38 , or separately by genders 29 . We treated them as independent cohorts. In the final dataset, 56,584 children and adolescents from 25 studies (32 independent cohorts), with an average 3.4-year follow-up, were included.

BMI
Weekly household income, breastfed or not, the age the child stopped being breastfed completely, birth weight, mother/father education, and gender.
Shorter sleep duration are primarily associated with BMI in children with early onset obesity(a subgroup): Sleep duration at age 6 to 7 was inversely associated with BMI at age 8 to 9 years (β = − 0.68, P = 0.017); and sleep duration at age 8 to 9 was inversely associated with BMI at age 10 to 11 years (β = − 1.21, P = 0.003).

BMI
Gender, sleep problems, household income, maternal education, and maternal weight status.
Short sleep duration at age 4 to 5 years was significantly associated with higher BMI at age 8 to 9 years (β = − 0.07, P = 0.044), which was slightly attenuated by television viewing at age 6 (β = − 0.06, P = 0.076). Information on sleep duration and the risk of overweight/obesity were available in 11 studies (14 coh orts) [14][15][16][17][18][19][20][21]24,25,38 , including 31,185 participants and 4,473 cases with a follow-up range from 21 months to 5 years. Data on sleep duration and annual BMI gain were available in six studies (8 cohorts) 22,25,27,28,31,33 including 5,341 participants with a follow-up range from 2 to 5 years. Figure 2 shows the association between sleep duration and the risk of overweight/obesity. By combining data from seven studies (10 independent cohorts), the pooled OR (95% CIs) was 1.76 (95% CI: 1.39, 2.23) for participants in the shortest sleep duration group as compared with those who were in the longest sleep duration group. The heterogeneity was high among studies (I 2 = 70.5%, P < 0.01). No strong evidence of publication bias was observed (Egger's test, P = 0.10).   For the dose-response relationship, after pooling available data from 8 studies (9 independent cohorts), the risk of overweight/obesity was 21% lower with every 1 hour/day sleep duration increment (combined OR: 0.79; 95% CI: 0.70, 0.89). A high heterogeneity among studies was observed (I 2 = 81.2%, P < 0.01). Because Egger's test indicated a significant publication bias (P = 0.01), the "trim and fill" method was used to adjust for the publication bias. With this adjustment, the pooled linear association was somewhat attenuated and became statistically non-significant (pooled OR: 0.92; 95% CI: 0.81, 1.05). Figure 3 demonstrates the relation between sleep duration and annual BMI gain. Four studies (5 independent cohorts) provided data on the multivariable-adjusted β coefficient (95% CIs) comparing children with the shortest to the longest sleep duration. The pooled results indicated that participants with the shortest sleep duration had significantly more annual BMI gain (β coefficient: 0.13; 95% CI: 0.01, 0.25). High heterogeneity among studies existed (I 2 = 67.5%, P = 0.02). The test for publication bias approached significance (Egger's test, P = 0.054).
Subgroup and sensitivity analysis. According to the available data, we examined potential effect modification by study region (USA vs. Non-USA), age (< 3, 3 < 5, ≥ 5 year), or follow-up duration (above or below the median). None of these factors materially modified the observed associations (Table 2). A few studies 26,29,31 examined gender difference between sleep duration and BMI z-score or PBF. Although the number of studies is not sufficient for us to pool the results, none of these studies found a significant gender difference.
In addition, the results were not appreciably changed when replacing random-effects models with fixed-effects models (data not shown). No single study substantially influenced the pooled results on sleep duration and the risk of overweight/obesity when omitting one study each time in the meta-analysis (Supplemental Table). However, the association between sleep duration and annual BMI gain was substantially attenuated when removing one of three independent cohorts (Silva, 201125 , Lee, 2012 , or O'Dea, 2012 28 ) at a time (Supplemental Table). Table 3 shows the association between sleep duration and secondary outcomes. Sleep duration was inversely associated with BMI, BMI z-score and WC, and positively related to FMI.

Discussion
This meta-analysis of prospective cohort studies provides accumulated evidence supporting the hypothesis that sleep duration in children and/or adolescents is inversely associated with the risk of overweight/ obesity in a dose-response manner. This inverse association is not appreciably modified by study region, baseline age or the length of follow-up.

Strengths and limitations.
This meta-analysis was based on up-to-date literature with the largest synthesis of prospective cohort studies from various populations. The overall sample size was large (n = 56,584) and the average follow-up period was relatively long (~3.4 years) among children and adolescents. Data for the pooled analysis were derived from fully adjusted models in the primary studies, which should reduce the likelihood of confounding. Also, our conclusions were strengthened by the consistent findings in the binary associations (the shortest vs. the longest sleep duration), the dose-response relationships, and the results of other related anthropometric measures such as annual BMI gain and WC.
A few limitations of this meta-analysis should also be considered. First, because sleep duration was parent-or self-reported, misclassification was possible, though it was likely to be non-differential and may attenuate the observed associations. Second, incident cases of overweight/obesity in the primary studies could not be determined. However, this was a common methodological issue in the studies of childhood obesity. Third, sleep durations were classified differently across the original studies, especially the reference groups were somewhat different. This inconsistency could affect our pooled results. Fourth, the findings could be biased by residual confounding or unmeasured confounders in the primary studies. Fifth, results from some sub-group analyses might not be very robust because of the high heterogeneity across the primary studies. Finally, publication bias was found when assessing linear association between sleep duration and the risk of overweight/obesity. However, the conclusion generally remained after controlling for the publication bias using Duval and Tweedie's "trim and fill" method.
Comparison with previous reviews or meta-analyses. A few previous narrative or systematic reviews found an inverse association between sleep duration and risk of obesity or weight gains among children and adolescents 6,[39][40][41][42] . However, these reviews did not conduct any quantitative analysis. A meta-analysis published in 2008 found that children or adolescents with shorter sleep duration had a 58% higher risk of overweight/obesity 43 . Of note, nine out of eleven studies included in that meta-analysis were cross-sectional, and only 2 were longitudinal studies. Similar results were reported in the same year in another meta-analysis, which included 12 cross-sectional studies 5 . A recently published meta-analysis focused on longitudinal studies in children and adolescents 44 , but this analysis did not   48 , and social or cultural factors 49,50 are major contributors. Although the exact mechanism remains mysterious, there are several possible pathways that could explain the inverse association between sleep duration and the risk of obesity/overweight. First, sleep has a beneficial effect on the sympathetic nervous system and hypothalamic hormones 6 . For example, sleep deprivation is associated with low levels of leptin and high levels of ghrelin, which suggests an important role in appetite regulation 51,52 . Sleep restriction is also associated with an increase in cortisol levels, which promotes increased food intake and the accumulation of visceral fat 6 . Second, sleep restriction or deprivation may lead to reduced energy expenditure by reducing thyroid-stimulating hormone (TSH), and therefore reducing basal metabolic rate, or by breaking the balance of neural circuits that maintain body weight. Third, sleep restriction can contribute to obesity by promoting waking behavior that causes weight gain. For instance, increased waking hours may provide more opportunities for an individual to eat more energy dense foods 53,54 . Also, lack of sleep is associated with sedentary activities such as television watching that might lead to weight gain due to increased snacking 35 . In addition, people with sleep restriction often experience fatigue and sleepiness throughout the day, which may cause a decrease in physical activity 53 . It is also possible that people with short sleep duration might drink/eat more high-energy drinks/foods to conquer fatigue 53 .

Summary.
In conclusion, this meta-analysis of prospective cohort studies, based on up-to-date literature, provides moderate evidence that sleep duration may be inversely and longitudinally associated with the risk of overweight/obesity in children and adolescents. To prevent childhood obesity and, thereafter, chronic diseases, a certain amount of sleep time for children and adolescents should be recommended.  Table 3. The association of sleep duration with secondary outcomes BMI: body mass index; FFMI: fat free mass index; FM: fat mass; FMI: fat mass index; PBF: Percentage body fat; SSF: sum of the skin-fold; WC: waist circumference. * P = 0.053 Since a randomized clinical trial on sleep duration may not be feasible due to ethical and practical considerations, further prospective cohort studies with longer follow-up periods, valid objective measures of sleep duration and dynamic sleep quality, as well as multiple measures of body composition, are warranted to further explore age and gender specified optimal sleep duration, and ideal sleep pattern (e.g., sleep timing and chronotype) in terms of weight maintenance.