We used a non-clinical convenience sample (N=278; 168 girls) of typically developing native Portuguese-speakers from the city of São Paulo, Brazil, aged 9 to 15 years old, an age span that includes the full range of pubertal development stages (37, 38). The exclusion criteria (based on reports given by guardians) were: 1) the use of chronic medication, to rule out the presence of clinical disorders and/or use of drugs that could impact development, sleep and/or cardiometabolic profiles; 2) having been held back at school for a year or more or being a student with special needs, which would be indicative of the presence of possible clinical problems. See below for information on sample size estimation.

This cross-sectional observational study received ethical approval from the Ethics Committee of the Universidade Federal de São Paulo (CAAE 56284216.7.0000.5505). After obtaining permission from a number of schools in various areas of the city, arrangements were made with the staff to invite potential participants and their families to take part in the study. Those who expressed an interest in being involved in the study took home written instructions, consent forms and questionnaires (regarding demographics and medical history, among other factors) to be completed by one of their guardians. Participants were only included if they met the eligibility criteria, assented in taking part and returned signed informed consent forms from their guardians. Participants were tested at their school in two sessions. The first was an individual session in which they completed questionnaires (including reports on sleep patterns and on self-rated pubertal status). The second was a group session, during which a trained team obtained anthropometric measurements and capillary blood from the participants for the determination of their lipid profiles and glycated hemoglobin levels. We could not fully control for circadian influences regarding cardiometabolic biomarkers because Brazilian students study either in the morning or in the afternoon, as the schools use a “shift” system. Nonetheless, to minimize these circadian effects, we collected the biological data towards the end of the morning and the beginning of the afternoon.

After data collection, the participants were awarded a “Science Partner” certificate and reimbursed for their commuting costs on the days of the experiment. The families received an individualized report of the findings. Referrals were provided when results indicated cardiometabolic problems in compliance with local clinical standards. Other information was obtained from the participants and their guardians but was not used in the current study.

There was no imputation of missing data in any of the analyses. Descriptive statistics for all variables were determined using the statistics package SPSS version 26.0 (62). The same software was used to compare SJL between sexes (Mann-Whitney-U-Test was conducted due to the lack of the normality assumptions for t-test), and we also provided effect sizes.

To allow comparisons with prior literature and to decide which cardiometabolic markers would compose the tested cardiometabolic latent factor (see below) we determined the Pearson inter-correlations between variables using the same statistical package. The other analyses (CFA and path modeling) were conducted using Mplus version 8.2 computer package (63). Figure 1 illustrates the tested path models.

Because we failed to find good fitting path models when we controlled for age and PDS scores in the path analyses (see Results), we conducted other, exploratory post hoc analyses using SPSS v.26.0 to determine whether these developmental markers were associated with SJL and cardiometabolic indicators that were found to form the cardiometabolic latent factor. The first exploratory analysis were Pearson correlation matrices controlled for multiple comparisons to describe linear inter-associations between the individual cardiometabolic status indicators, SJL, and developmental markers (age and PDS), by sex.

The next exploratory analyses were multivariate analyses of covariance (MANCOVA) in which all four metabolic indicators used to build the latent factors (BMI, waist/height ratio, systolic blood pressure, and total cholesterol/HDL) were entered as multivariate dependent variables, and SJL and PDS (or age in separate models) as continuous predictors, including the correlation of SJL and the developmental measures (PDS or age), per sex. The level of significance was 5%. When Pillai’s Trace values were significant, follow-up univariate analyses for each cardiometabolic indicator were carried out. If interactions of SJL and developmental markers was significant, Johnson-Neyman plots (77) were use to interpret effects. Effect sizes were determined with partial eta squared (η_p ²): values between 0.059 and 0.138 are considered medium effect sizes and those higher than 0.138, large effect sizes (73).

All descriptive statistics are shown in Table 1 , including sample size per variable as we did not impute missing data. The result of Mann-Whitney-U-Test showed that there is no difference between girls and boys in terms of SJL (Z= -1.037; p=0.300; R^{2 =} 0.004).

We started with 8 cardiometabolic variables [BMI, percentage of body fat, waist/height ratio, systolic and diastolic blood pressure, glycated hemoglobin, triglycerides, and total cholesterol/HDL (ratio)] to find the single factor structure. The best fitting unidimensional cardiometabolic latent factor was comprised of the following variables (see details of the method to select the indicators in the Supplementary File 1 ): anthropometric measurements (BMI, waist/height ratio), systolic blood pressure and total cholesterol/HDL (ratio). In data of girls, PPP was 0.262, showing adequate fit. The same model for boys converged, but reached a poor fit: PPP equal to 0.041.

PPP and 95% confidence interval for the difference between the observed and replicated Chi-square χ2 values of all tested models are shown in Supplementary File 1 ; Table 2S . PSR of each of these models are shown in the Supplementary File 1 ; Table 3S . The standardized factor loadings and their 95% credible interval of the selected models per sex are shown in Figure 2 .

First, we analyzed the association of SJL on the cardiometabolic latent factor, separately by sex, to investigate if our data confirmed this association for adolescents without considering their individual ages and pubertal status. PPP and 95% confidence interval for the difference between the observed and replicated Chi-square χ2 values are shown in the Supplementary File 1 ; Table 2S . The convergence criterion are shown in Supplementary File 1 ; Table 3S . In girls, the PPP for this model was 0.613, which is a good fit. The standardized factor loadings and their 95% credible interval are shown in Figure 3 . The standardized regression coefficient was β = 0.268, CI [0.080 to 0.411], showing that with every increase of one standard deviation in SJL, the cardiometabolic latent trait rose by 0.268 (note, however, that factor loading of systolic blood pressure was less than 0.4); R squared was 0.072. The comparable model for boys did not converge.

Next, we controlled for PDS scores in the cardiometabolic latent factor and, in separate models, for age instead of PDS scores only using data of girls because the path model for boys including SJL did not show adequate fit to go forward to run the full path model. The model with data from girls controlled for both PDS and age did not converge (see Tables 2S , 3S in the Supplementary File 1 ).

Because the path analyses of data controlled for age and PDS did not fit the data well, we could not determine whether developmental changes exerted an effect on the association of SJL and cardiometabolic status. We therefore carried out additional post hoc analyses to explore the extent to which other types of statistical analyses were able to capture whether age and pubertal status had effects on SJL and cardiometabolic indicators, per sex.

Correlations corrected for multiple comparisons between SJL, cardiometabolic and developmental markers for both sexes analyzed separately showed that SJL was higher at older ages and at more mature pubertal stages ( Table 2 ). BMI was also higher in older and more sexually matured individuals, but only in girls. Differently, for boys, the only cardiometabolic change across development was a higher systolic blood pressure in older individuals. All these significant associations were small (did not exceed r=0.40). Importantly, correlations of SJL with cardiometabolic indicators did not survive correction for multiple comparisons, which is not usually controlled for in the literature.

Regarding the exploratory MANCOVA with data from girls, the model including SJL and PDS scores had Pillai’s Trace result that showed that there was only a main effect of PDS scores (F_(1,124)=7.881, p<0.001, η_p ² = 0.203), while SJL (F_(1,124)=1.565, p=0.188) and their interaction (F_(1,124)=2.122, p=0.082) were not significant. Follow-up univariate analyses of between-participant effects showed that this effect was driven by the association of PDS scores with body mass index (BMI) and total cholesterol/HDL, but did not relate to SJL. - For BMI, multivariate result values were F_(1,127)= 4.887, p=0.029, η_p ² = 0.037 (small effect size); unstandardized regression coefficients (B) was 1.903, indicating that for every unit increase in PDS scores there was an increase of 1.903 in BMI. For total cholesterol/HDL, results were F_(1,127)= 11.085, p=0.001, η_p ² = 0.080 (medium effect size); the B was -0.485, indicating that for every unit increase in PDS scores there was a decrease of 0.485 in total cholesterol/HDL (that is, less cardiometabolic risk in more sexually mature participants). Adjusted R² of BMI and total cholesterol/HDL were 0.090 and 0.091, respectively, indicating that about 9% of the variance of BMI and cholesterol/HDL was explained by the whole model (small effect sizes). All other factors failed to reach significant effects (p>0.05).

In the equivalent MANCOVA model with data of girls, including age instead of the PDS scores, Pillai’s Trace result showed that there was a main effect of SJL (F_(1,124)=2.741, p= 0.032, η_p ² = 0.081), age (F_(1,124)= 6.802, p<0.001, η_p ² = 0.180), and their interaction (F_(1,124)= 2.948, p=0.023, η_p ² = 0.087). This interaction was driven only by association of age with systolic blood pressure (F_(1,127)=5.938, p= 0.016, η_p ² = 0.045, small effect size; B=0.067; R² = 0.023). This model explained only 2.3% of the variance of systolic blood pressure so is a negligible effect. Additionally, Johnson-Neyman plots of this interaction ( Figure 4 ) showed that it was significant only below around the age of 120 months (10 years of age) and above 180 months (15 years), while in between (which comprises the great majority of our sample) the interaction was not significant because confidence intervals of the regression crossed zero on the y axis in this age interval.

The pattern of exploratory MANCOVA results for boys was different from that of girls, showing no significant associations among the variables, which might explain why the data for this sex did not fit the path analyses models. More specifically, in the model including pubertal development there was no main effect of SJL (F_(1,78)=1.027, p=0.399), PDS (F_(1,78)=1.017, p=0.404) or their interaction (F_(1,78)=1.431, p=0.231). In the model with data of boys using age instead of PDS scores there were also no significant main effect of SJL (F_(1,78)=1.568, p= 0.191), age (F_(1,78)=1.940, p= 0.112) or their interaction (F_(1,78)=1.801, p=0.137).

Our unidimensional cardiometabolic latent factor fitted the data, mirroring results of various previous studies in adolescents (10, 49, 69), adults and elderly (71, 72) from other populations, despite the fact that we used non-adjusted raw scores and a different combination of indicators to form this latent factor (see Supplementary File 1 for a detailed comparison of our and their latent factors). This is unlikely to be a limitation of our study because we started off with similar biomarkers that were used in these publications, all of which are regarded as related to cardiometabolic status (47). Crucially, a property of latent factors such as those modeled here is that if they are found to have adequate fit indices this means that the latent construct of interest, which was measured indirectly through the shared variance of multiple cardiometabolic biomarkers (also called indicators), actually exists, independently of the indicators used to measure it (46). Therefore, indicators that theoretically reflect the same construct can be used interchangeably across studies and still measure the same construct. Stated differently, our findings show that cardiometabolic risk can be determined as a single latent factor, even in non-clinical samples of adolescents, enabling replications of findings which are not usually possible when different studies use different biomarkers. This has a great advantage because this latent factor can be established even if different cardiometabolic indicators are used across studies. This finding in itself, regardless of the effects of this latent factor on SJL, is a contribution to the literature because it mirrors prior results (10, 49, 69) in a sample from a developing nation with different characteristic from that of other populations of adolescents assessed in the literature. Nonetheless, usually at least three indicators (46) must be used to build a reflective unifactorial latent structure (here we used four).

The association of SJL and cardiometabolic latent risk in girls found here has only previously been consistently shown in respect of increases in adiposity in adolescence (17–23) and was not identified when using a non-latent composite measures of cardiometabolic risk in a large sample of non-clinical early adolescents (21). Although Cespedes Feliciano et al. (21) also assessed cardiometabolic status as a continuous variable (averaging across many standardized indicators), thereby characterizing the participants on a continuum from a better to a worse cardiometabolic profile, differently from their approach our latent factor allowed us to show that the higher the SJL, the worse the shared variance of cardiometabolic measures beyond adiposity. This was clear for females only (without considering age or pubertal status), for which acceptable fit indices were found. In other words, our statistical approach was more sensitive to this effect and picked up this association in girls either because they are more prone to this relation as hypothesized (8, 18, 41, 43–45) or because our sample had more girls. This corroborates findings that components of cardiometabolic risk and adverse endocrine profiles other than adiposity are associated to SJL in pediatric populations (27) and other sleep disturbances (e.g., short sleep duration and bad quality sleep (31–34). The reason for this higher sensitivity of a latent factor is due to the fact that a cardiometabolic latent factor represents the shared variance of different continuous parameters that express an underlying theoretical cardiometabolic construct and partials out the specific effects of individual indicators (46). This makes this analytic approach advantageous to assess its association with SJL compared to: 1) averaging across (21, 35) or combining various indicators in multivariate analyses (27); 2) using a metabolic syndrome index (28, 30, 31, 34), for which there is no consensus in terms of which indicators should be used (49), their cutoff scores or for which adjustments (e.g., for age, sex and pubertal status) should be made (51–54); and 3) using multiple analyses, one for each cardiometabolic-related variables, which still seems to be the norm in the sleep literature. In point of fact, our post hoc exploratory correlations of SJL with each cardiometabolic indicator did not survive corrections for multiple comparison in either sex. This is so because these composites or use of many separate correlation analyses per indicator do not account for the fact that cardiometabolic variables are intercorrelated and that carrying out multiple analyses per variable can lead to false positive results, distorting statistical findings. Therefore, determining a cardiometabolic latent factor may better allow future replications of the associations of SJL and cardiometabolic risk compared to the prevailing approaches used so far to establish this relation. Although these effects were found here to be present in healthy girls, this approach may be more important when studying clinical samples of adolescent and may also be more adequate to analyze the effects of sleep disturbances other than SJL in adolescents on cardiometabolic risk. The reason is that the abovementioned statistical inadequacies possibly led to the inconsistency in the individual markers/components of cardiometabolic dysregulation found to be associated with sleep problems (31–34, 80), a relation that has not been found when using non-latent cardiometabolic composites [metabolic syndrome index: (31, 34) the exception being (35)].

Regarding the contribution of each indicator to the latent factor, we stress that BMI and waist/hight ratio made by far the highest contribution in all models, which can explain why SJL has so often been found to relate to adiposity (17–23) (for different results, in possibly underpowered studies, see 24, 26). Although our non-clinical sample had a general healthy cardiometabolic profile, it should be noted that a worse latent cardiometabolic score, even if mostly apparent through a higher BMI, may exert carryover effects that can culminate in the development of the metabolic syndrome (80, 81), which doubles in older compared to younger adolescents (39).

Having obtained an adequate latent cardiometabolic score and finding this to be related with SJL in girls posed the following question: does this latent construct alter SJL or is the inverse true? Although there is no consensus in the literature regarding the mechanisms through which SJL and cardiometabolic risk are interrelated, it has been assumed that disruptions of central and/or peripheral circadian rhythms, beyond the effects of sleep duration/quality, lead to chronic stress and, through this (12–14) and other pathways, dysregulate energy homeostasis (3, 11, 27). Indeed, it is notoriously difficult to differentiate the physiological effects of chronic stress (allostatic load) from those of the metabolic syndrome (15) even at the latent trait level (16). Adolescents who present high SJL have also been found to have worse eating habits, a possible consequences of SJL-induced changes in appetite (22, 82), which potentiates cardiometabolic problems associated with SJL (in animal models: (83). Additionally, a meta-analysis found a positive dose–response relationship of the metabolic syndrome in adults with the duration of exposure to shiftwork (inherently linked to SJL) (81). Prospective studies in pediatric samples also show that short sleep duration, which is often associated with higher SJL, can lead to the development of higher BMI/obesity (26, 84). However, the specific relation between SJL and cardiometabolic markers was not determined in these studies. Hence, it seems that SJL may lead to cardiometabolic changes and not vice-versa, although this was not determined here because we used a cross-sectional design.

The full path models controlling for age and PDS scores did not have adequate fit. Follow up exploratory correlation and multivariate analyses showed that cardiometabolism (considering indicators individually with correlations and as a multivariate) was not clearly related to SJL when age and pubertal status were accounted for, so these developmental variables were likely inconsistent with the tested full path models, leading to inadequate fits. In other words, it could be that age/PDS scores may have little to do with the association of SJL with metabolism, the same being true when considering age. The only exception was a very small relation of SJL with systolic blood by age in girls, but this correlation reached a very low effect size and was limited to the extreme ages of our sample. This interaction was not modeled in the full path analyses so it cannot be excluded that it interfered with model fit. Admittedly, the lower sample size considering boys probably influenced this lack of fit.

Regarding the exploratory correlations (corrected for multiple comparisons), for which our sample size can be considered large enough, SJL increased in more mature individuals (older and more sexually developed adolescents), as expected (4, 36). Additionally, BMI was positively related to both chronological and sexual maturity, but only in girls, which also confirms prior findings of higher accumulation of adiposity in this sex (40–43). In the case of boys, the only developmental association with cardiometabolic variables was that of age and systolic blood pressure, which has also been shown to have higher weight in metabolic latent factor in this sex in some ethnicities (10). Despite this, SJL was not significantly correlated with any of the cardiometabolic indicators once corrections for multiple comparisons was carried out, suggesting that some of the reported associations found in prior studies were inflated by type I errors, indicating a questionable link of chronological misalignment (SJL) and worse cardiometabolic profiles when cardiometabolic indicators are analyses separately in either sex. Inflated type I errors are avoided using the latent factor approach, which showed here a small positive association of SJL and a latent cardiometabolic risk score, which was only clear in adolescent girls, confirming two studies that found associations of SJL and adiposity only in this sex (18–21).

Prior studies did not investigate the effects of developmental stage on SJL, having usually averaged effects across different ages and/or sexes, or used age/sex adjusted measures in some or all variables. Therefore, although we cannot corroborate our hypotheses of a possible role of chronological and/or pubertal development across the first half of adolescence on the association between SJL and cardiometabolic status, our results show the clear need for further investigations on this issue, especially in clinical samples with either cardiometabolic problems and/or developmental issues. We nonetheless show that age and pubertal status potential influence on the association of SJL and worse cardiometabolic profiles must be analyzed separately in girls and boys as they mature and grow older because this varies widely across sex. Merely statistically controlling for age, pubertal status and/or sex is probably not adequate because these variables likely differently influence SJL and cardiometabolic status, which have not yet been confirmed as invariant (i.e., that they measure the same constructs) across these variables (85).

We conclude that in early adolescence SJL is associated not only with higher adiposity (BMI and waist/height ratio), but also with its shared variance (continuous latent factor) with other indicators of adverse cardiometabolic profiles (higher total cholesterol/HDL ratio and systolic blood pressure) in typically developing girls, but results were not clear in boys.

Our findings showed that a worse cardiometabolic profile was already present in a non-clinical sample of early adolescent girls who had higher SJL. Given the increasing prevalence of metabolic syndrome in adolescents worldwide, our results highlight the need for interventions to help reduce the global burden of cardiometabolic ill health (36), and should serve as a warning that measures must be taken to improve adolescent sleep schedules, particularly for girls who tend to exhibit higher SJL in most studies (8, 18, 41, 43–45). Reducing SJL and improving cardiometabolic health can be attained by: 1) delaying school start times so that adolescents can get more sleep during school nights and therefore decrease the need to compensate for lost sleep on days off (e.g. 86-88); 2) using experimentally successful behavioral strategies in adolescents to extend sleep, including help in organizing schedules to allow earlier bed time, having written instructions and parental support, reducing caffeine intake (89) and, importantly, decreasing electronic screen time at night because this disrupts sleep patterns, particularly in adolescence (88). Improving adolescent eating habits, increasing their levels of physical activity (88) and avoiding exposure to endocrine disrupters (90) are also means of decreasing cardiometabolic risk that is not directly associated with sleep problems. Establishing the mediating and/or moderating contribution of these factors to the association of cardiometabolic risk and SJL were not the aim of the present study, but should be explored in coming experiments. Furthermore, future studies should explore how age and pubertal status contribute to the relation of SJL and cardiometabolic profiles during the pubertal transition, separately in boys and girls, and whether this increases in older adolescence, especially because: 1) children and adolescents seem to be particularly vulnerable to develop obesity due to short sleep duration (26, 84, 91), especially girls (91); 2) rates of metabolic syndrome double in older compared to younger individuals (39); and 3) SJL increases across adolescence (3–5).

This study has some limitations that should be noted. First, the cross-sectional design does not enable the determination of whether SJL impairs cardiometabolic status or vice-versa. Second, the path model analyses could have had a better fit if we had tested a larger sample, especially a larger sample of boys, who were under represented compared to girls and likely led to a poor but convergent cardiometabolic latent factor. Third, the SJL and pubertal status measures were obtained here through self-reports, which only partially correspondent to objective measures of these variables (20, 21, 37). For instance, lower SJL (of around one hour) has been found in studies that used actigraphy compared to self-report (20, 21). However, although the presence of SJL can be determined objectively using actigraphy, this is usually only assessed over a limited period of time. Therefore, these results might not represent usual disrupted sleep-wake patterns evaluated with a subjective rating about general sleep habits, as was done in the current study (29), which may be more highly associated, in the long-term, with cardiometabolic dysregulation. Furthermore, there are cultural differences in sleep timing (29), which might explain why Tamura et al. (8) reported only one hour of self-reported SJL, while here, and in other publications (5, 19, 25), SJL was closer to two hours. Lastly, our statistical models did not control for sleep durations, sleep quality, nutritional status or physical activity, which have been shown to be related with the association of SJL and adiposity (18, 21, 26, 90), nor for exposure to endocrine disrupters which affect metabolism and pubertal timing (90). We also took the decision not to control for ethnicity to avoid fallacious racial stereotypes, despite suggestions that race influences cardiometabolic profiles (48) and pubertal status (37). The reasons were that Brazil has a highly admixed population with unclear skin color boundaries (92) and that race labels have a much lower association with genetic ancestry than most of the literature recognizes (93, 94).