Chronic Insufficient Sleep in Women Impairs Insulin Sensitivity Independent of Adiposity Changes: Results of a Randomized Trial

More than one in three adults sleep less than the recommended 7 h/night (1). This is concerning given robust associations of short sleep with cardiometabolic diseases (2). Data from observational studies demonstrating that shorter sleep predicts poorer metabolic outcomes (3–6) are supported by short-term sleep restriction (SR) trials showing that total and partial sleep deprivation adversely affect glucose metabolism (7). However, the acute, extreme SR (to <5 h/night) imposed in previous studies does not resemble real-world sleep patterns. In a representative sample of U.S. adults, the median sleep duration of adults with short sleep (<7 h/night) was ∼6.0 h/night (8). Clinical trials designed to assess the impact of prolonged exposure to widely prevalent patterns of insufficient sleep are needed to augment our understanding of the impact of chronic short sleep on metabolic health.

To date, the only study to assess an impact of milder sleep curtailment on insulin resistance (IR) was conducted among healthy young men randomized to either maintained adequate sleep (AS; ∼7.6 h/night) or short sleep (∼6.2 h/night) for 3 weeks (9). Results showed an initial reduction in insulin sensitivity after 1 week of SR that returned to baseline by the final week. Although that SR protocol better represented population patterns of short sleep than did previous trials, the 3-week duration may not have been sufficient to accumulate the sleep debt experienced with chronic insufficient sleep. Beyond this, the study did not include women, for whom suboptimal sleep is particularly pervasive; women report sleep disturbances more frequently than men across all stages of adulthood (1). It is particularly important to study the metabolic consequences of sleep deficiencies in women because associations of suboptimal sleep with adverse metabolic outcomes can be worse in women than men (10).

Further highlighting the importance of elucidating the impact of prevalent patterns of sleep on glucose metabolism in women are the more pronounced aging-related declines in sleep duration compared with men (11). Indeed, the transition to menopause is accompanied by worsened sleep (12), and lower levels of endogenous estrogen, a protective factor against IR (13), may increase susceptibility to adverse metabolic consequences of short sleep among postmenopausal women. Thus, a goal of the present study was to evaluate the causal effects of chronic insufficient sleep on glucose metabolism in pre- and postmenopausal women using a mild SR protocol implemented over 6 weeks. We hypothesized that insulin sensitivity, glucose tolerance, and β-cell function would be worsened by SR compared with maintenance of AS, with effects exacerbated in postmenopausal women.

This study was part of the American Heart Association Go Red for Women Strategically Focused Research Network at the Columbia University Irving Medical Center (CUIMC) and was designed to evaluate the causal impact of prolonged, mild SR on cardiometabolic risk factors in women (ClinicalTrials.gov identifier NCT02835261). The trial comprised two 6-week phases in a randomized crossover design: AS (7–9 h/night) and SR (reduction of 1.5 h/night). Study phases were separated by a 6-week washout period. All procedures were conducted in accordance with the standards of the Declaration of Helsinki and approved by the CUIMC institutional review board. All participants provided informed consent before participation.

Study recruitment took place between August 2016 and February 2021 in the New York City area. Metabolically healthy women aged 20–75 years who were at elevated risk for cardiometabolic disease due to family history were recruited. Elevated risk was defined as having overweight or class I obesity (BMI 25–33.5 kg/m²) or BMI 20–24.9 kg/m² with at least one parent with type 2 diabetes (T2D), hyperlipidemia, or cardiovascular disease. Initial eligibility criteria were assessed via phone interview. Prospective participants meeting all preliminary eligibility criteria were invited to the laboratory for an in-person screening visit, during which BMI was measured and information on sex (as a biological variable), age, health history (of both the individual and immediately family members), and sleep habits was collected. Individuals were enrolled in the study if they self-identified as a biological woman and met the definition for elevated cardiometabolic risk (verified from BMI and health history questionnaire), reported adequate habitual sleep duration (≥7 h/night) and low risk for obstructive sleep apnea (no more than one positive component on Berlin Questionnaire [14] and Epworth Sleepiness Scale [15] score <10), and did not meet any of the exclusion criteria: current or recent (<1 year) pregnancy; hormone-based medication use; sleep, mood, psychiatric, or cardiometabolic disease or disorder; smoking; history of drug or alcohol abuse; excessive caffeine intake (>300 mg/day); significant delayed or advanced sleep phase (Morningness-Eveningness Scale [16] composite score 16–30 or 70–86); travel across time zones; and work involving nonday shifts, operation of heavy machinery, or long-distance driving. In addition to providing written informed consent, participants were required to provide signed verification that they would not operate a motor vehicle during SR.

Demographic information, including self-reported age, race, ethnicity, highest level of education, and employment status, was collected from enrolled participants who then underwent an intensive sleep screening period. Habitual patterns of sleep were assessed in the free-living setting over 2 weeks via wrist actigraphy (Actigraph GT3×+; ActiLife LLC, Pensacola, FL) and nightly sleep diaries. This method of sleep assessment is endorsed by the American Academy of Sleep Medicine for measuring free-living sleep (17), and wrist-actigraphy provides valid estimates of total sleep time (TST) when compared with polysomnography (18). Data were used to determine eligibility based on habitual sleep (i.e., average TST of 7–9 h/night with ≥70% of the nights reaching at least 7 h).

Sleep assignments were revealed at the first baseline visit. The order of sleep conditions was randomized by the study statistician, who was blinded to the intervention, using a computer-based random sequence generator, and provided to the research assistant before participant accrual. For each condition, personalized bedtime and wake time targets were developed in collaboration with the participant on the basis of sleep patterns measured during screening. During AS, participants maintained their typical bedtimes and wake times, determined from the actigraphy-based sleep screening period, to achieve TST of 7–9 h/night. During SR, participants delayed their bedtime by 1.5 h/night, relative to screening (habitual actigraphy-derived bedtime) while maintaining their typical wake time, with the goal of achieving a nightly reduction in TST of ∼1.5 h. Each sleep condition was maintained for 6 weeks to induce a meaningful degree of cumulative sleep debt during SR, more than twice that of our previous shorter-term trial with extreme SR (19), while not causing undue burden and/or risk.

At baseline and end point of each study phase, participants came to CUIMC in the morning after a ≥12-h fast for assessments. Participants underwent MRI scans for body composition assessment followed by an oral glucose tolerance test (OGTT) at the Irving Institute for Clinical and Translational Research. Blood samples were collected from an antecubital vein before and 15, 30, 60, 90, and 120 min after a 75-g glucose load (Trutol Beverage; ThermoFisher, Middletown, VA). Fasted blood samples were also collected at week 3 and, from a subset, at week 4 (n = 20) study visits during each phase.

Free-living sleep was assessed using actigraphy throughout in study phases. Participants were instructed to wear the actigraphy device, which collects data on movement across three planes as well as ambient light, on their nondominant wrist at all times aside from bathing. Actigraphy devices were collected at each in-person study visit, scheduled every 1–2 weeks, for data download and device charging. Raw data were scored using the Cole–Kripke algorithm in ActiLife version 6 (Actigraph LLC), with sleep diary information used to enhance accuracy of bedtime and wake time estimates from the algorithm. Actigraphy-derived sleep data, including bedtimes and wake times and TST, were reviewed with participants at study visits, at which time discrepancies between actigraphy-derived bedtimes and wake times and sleep diaries, as well as between actual and target sleep times, were discussed and resolved. If needed, adjustments were made to the bedtimes to ensure compliance with the target weekly TST goals.

Whole blood was processed after collection, and resultant plasma was aspirated, aliquoted, and stored at −80°C until analysis of levels of glucose (mg/dL) and insulin (µU/mL) by the Biomarkers Core Laboratory of CUIMC. Plasma insulin levels were assessed using radioimmunoassay, and both glucose and insulin levels were measured in duplicate at each time point. The HOMA-IR was calculated as fasting plasma glucose (FPG; converted to mmol/L) × fasting plasma insulin (FPI) ÷ 22.5 (20); fasting measures of glucose metabolism were available at baseline, week 3, week 4, and at the end point of each phase. From the OGTT at baseline and end point of study phases, total area under the curve for glucose (AUCg) and insulin (AUCi) were calculated using the trapezoidal rule (21); Matsuda and oral disposition indices were calculated using the web-based tool from DeFronzo and Matsuda (https://mmatsuda.diabetes-smc.jp/EnglishHPMM.html) (22).

At baseline and end point of each phase, whole-body MRI was carried out using a 1.5T General Electric system (6× Horizon; Milwaukee, WI) (23). MRI-derived outcomes of interest for our investigation were volumes (L) of total adipose tissue (TAT) and visceral adipose tissue (VAT), given their roles in the pathophysiology of T2D (24,25).

This trial aimed to accrue 80 participants to achieve a complete sample size of 40. Ultimately, 50 participants were accrued to achieve 40 participants with end-point data from both phases (n = 38 for the primary outcome of glucose metabolism). Using data from a previous trial of the effects of SR on energy balance (26), we estimated a between-condition (SR versus AS) difference for within-subject change in body weight, the closest available correlate for the co-secondary outcome of adiposity change, of 0.91 (SD, ±0.65 kg). On the basis of this estimated effect size, a sample of 40 participants would provide >90% power to detect main effects of condition at α = 0.05 using paired two-tailed t tests and sufficient power to test for differences between pre- and postmenopausal women (assuming equal numbers per group).

Descriptive characteristics of the sample are shown as mean ± SD for continuous variables and n (%) for categorical variables. An independent samples t test and χ² test for proportions were used to compare continuous and categorical characteristics, respectively, between pre- and postmenopausal women. The primary outcomes for this investigation were markers of glucose metabolism in the fasted state (i.e., FPG, FPI, and HOMA-IR) and in response to the OGTT (AUCg, AUCi, Matsuda index, and disposition index). Linear mixed models with repeated measures were used to test the effect of sleep condition on post-baseline outcome measures (fasted measures at weeks 3, 4, and 6; OGTT and MRI measures at week 6), which were adjusted for baseline values. Study condition, week, phase, and menopausal status were fixed effects. A condition by menopause interaction was included in initial models to evaluate whether menopausal status influenced condition effects; the interaction was dropped when it was nonsignificant. However, models were repeated and stratified by menopausal status for all outcomes because a CUIMC goal was to characterize cardiometabolic risk as a function of menopausal status. Cohen’s f² values were calculated from the mixed-effects models as a standardized measure of effect size (27). Main analyses were conducted in SAS 9.4 (SAS Institute, Cary, NC).

Post hoc analyses were conducted to determine whether effects of SR on markers of glucose metabolism were modulated by, or independent of, changes in TAT or VAT. If SR affected an adiposity measure at P ≤ 0.10, its change from baseline was evaluated as a mediator of the effect sleep condition on primary outcomes using the Causal Mediation package in R Studio. Where mediation was nonsignificant, thereby limiting risk for collider bias, linear mixed models similar to those used in the primary analysis were conducted with change in TAT and VAT added, separately, as covariates to determine if SR effects on glucose metabolism were, in part, dependent on change in adiposity. Models were repeated in the full sample and in pre- and postmenopausal women separately.

All results of primary and post hoc analyses are presented as β ± SEM and considered significant at P < 0.05. Values for glucose and insulin were converted to SI units for presentation of results (28).

Data generated from this study and statistical code are available upon written request to the corresponding author (M.-P.S-O.).

Fifty women were accrued and 40 completed the study. Primary outcome data were available for 38 women, 11 of whom were postmenopausal (see the Consolidated Standards of Reporting Trials [CONSORT] diagram in Supplementary Fig. 1). Descriptive characteristics of the analytic sample are provided in Table 1. Among participants with outcome data, >60% self-identified as a racial minority (i.e., non-White) and >25% self-identified as Hispanic. On average, participants had normal FPG levels (4.86 ± 0.50 mmol/L [87.6 ± 9.0 mg/dL]) and 50% had BMI >25 kg/m² (25.5 ± 3.3 kg/m²).

Thirty-four women (89%) completed the study prior to the COVID-19 pandemic, and their screening TST did not differ from those completing after the pandemic interruption (457.7 ± 3.8 vs. 455.7 ± 7.1 min/night; P = 0.86). Compliance with the outpatient sleep protocol was excellent: across the 6 weeks of SR, 78.9% of women reduced TST by ≥75 min/night from screening, 50% of whom reduced TST by ≥90 min/night (average TST reduction, 80.7 ± 2.3 min/night; P < 0.0001; SR: 370.8 ± 2.9 vs. AS: 451.5 ± 2.9 min/night) (Supplementary Fig. 2).

In the full sample, SR increased HOMA-IR (β = 0.30 ± 0.12; P = 0.016) and FPI (β = 6.8 ± 2.8 pmol/L; P = 0.016) compared with AS (Fig. 1) with no effect on FPG (β = 0.09 ± 0.08 mmol/L; P = 0.26).

Menopausal status influenced the effect of SR on HOMA-IR (P for interaction = 0.042), which increased to a greater extent among postmenopausal women (β = 0.45 ± 0.25) relative to premenopausal women (β = 0.27 ± 0.13) (Fig. 1). Results of a priori analyses stratified by menopausal status showed a significant effect of SR on FPI in premenopausal women (β = 8.2 ± 3.3 pmol/L; P = 0.016). Although of similar magnitude to that of premenopausal women, the increase in FPI during SR versus AS among postmenopausal women was not statistically significant (β = 7.8 ± 5.3 pmol/L; P = 0.15). No effect of SR on FPG was observed in either premenopausal (β = 0.02 ± 0.10 mmol/L; P = 0.84) or postmenopausal women (β = 0.24 ± 0.15 mmol/L; P = 0.13).

In the full sample, SR did not significantly affect any of the OGTT-derived measures of glucose metabolism (all P > 0.10; Table 2), and no interactions between sleep condition and menopausal status were observed. Results of analyses stratified by menopausal status revealed that Matsuda index, a marker of whole-body insulin sensitivity, tended to decrease in SR relative to AS in postmenopausal (β = −1.29 ± 0.60; P = 0.069), but not premenopausal (β = 0.13 ± 0.96; P = 0.89), women.

No impact of SR on VAT and TAT was observed in analyses conducted with all women (VAT: β = 0.01 ± 0.04 L, P = 0.81; TAT: β = −0.07 ± 0.26 L, P = 0.80) and in postmenopausal women only (β = −0.09 ± 0.12 L, P = 0.49; TAT: β = −0.35 ± 0.72 L, P = 0.64). In premenopausal women, VAT tended to increase in SR compared with AS (β = 0.04 ± 0.02 L; P = 0.082). However, subsequent mediation analyses demonstrated that VAT did not explain a significant proportion of the observed effects of SR on primary outcomes in the full sample or in premenopausal or postmenopausal women only (for proportion mediated, all P > 0.20).

Given the lack of mediation by VAT, we evaluated whether changes in primary outcomes in response to sleep condition were independent of changes in adiposity. Among the full sample, including change in VAT as a fixed effect in linear models did not alter the significance of any of the reported main effects of SR, or interactions of sleep condition with menopausal status, on measures of insulin sensitivity (Table 3). The same was true for TAT.

In analyses stratified by menopausal status, controlling for change in VAT attenuated the effects of SR on HOMA-IR and FPI in premenopausal women, whereas results remained borderline significant in postmenopausal women (Table 3). Similar findings were observed when change in TAT was included as a covariate. Controlling for change in VAT revealed a significant effect of SR on FPG in postmenopausal women (β = 0.29 ± 0.08 mmol/L; P = 0.023) that was not observed in the primary analyses. For the outcome of Matsuda index, adding change in VAT to the model did not modify our primary findings of a postmenopausal-specific tendency in response to SR (β = −1.35 ± 0.57; P = 0.076), whereas adding change in TAT strengthened the effects (β = −1.80 ± 0.60; P = 0.039).

Within this model of chronic insufficient sleep, representing, to our knowledge, the longest free-living trial of continuous SR to date, we showed that curtailing TST to ∼6 h/night for 6 weeks increased key risk factors for T2D (29). Small but significant reductions in insulin sensitivity were observed among generally healthy women in response to prolonged SR, with postmenopausal women exhibiting more deleterious responses for measures of both fasted and whole-body insulin sensitivity. That mild SR has a causal adverse effect on glucose metabolism in women with family history of cardiometabolic risk has important public health relevance because IR plays a key role in the pathogenesis of T2D (30). Our results highlight sleep as a modifiable lifestyle behavior to consider for the prevention of T2D, particularly in postmenopausal women, who are most vulnerable to experiencing sleep deficiencies (31).

Adverse effects of SR on insulin sensitivity in the present trial align with previous observational (2) and short-term SR studies (7) and provide important new insights into the role of sleep in the development of T2D. This study enhances the ecological validity of previous studies (7) by showing that adverse effects of SR on key aspects of glucose metabolism extend to longer-term, milder SR within a free-living context. We found that incremental accumulation of sleep debt, independent of inherent changes in timing of sleep, over 6 weeks impairs fasting insulin sensitivity in pre- and postmenopausal women and increases insulin levels, particularly among premenopausal women. Although the magnitudes of change in these markers of glucose metabolism in response to SR are modest, relatively small increases in these outcomes could have meaningful clinical implications if sustained over time. For example, a prospective analysis of data from the Coronary Artery Risk Development in Young Adults study demonstrated that each 1 µU/mL increase in fasting insulin predicted a 3% higher risk for hypertension (32). Another epidemiologic study reported 48% higher odds for the metabolic syndrome per 1.4 µU/mL increase in fasting insulin level (33). In comparison, we found that SR increases FPI by >1.3 µU/mL (7.8 pmol/L) relative to AS. In addition, linear relationships have been observed between insulin sensitivity, assessed through HOMA-IR, and risk for diabetes (29) and cardiovascular disease (34), suggesting that even incremental increases in HOMA-IR can be clinically significant. Hence, even the average person with short sleep could be at elevated risk for the development of T2D.

Although markers of IR were elevated in response to mild, chronic SR, we did not observe significant differences between conditions in glucose levels from fasted blood samples or the OGTT in primary analyses. This is contrary to findings of a seminal study conducted with healthy young men, in which Spiegel et al. (35) observed a significantly lower rate of glucose clearance after six nights of curtailing time in bed to 4 h/night. Sex of study participants likely accounts for the discordant findings. Many previous SR studies have included only male participants (7). Of note, in trials comprising both male and female participants, no effects of acute SR on fasting glucose levels (19,36) or glucose tolerance (36) are observed. Sex differences in the influence of insufficient sleep on glucose metabolism deserve further investigation.

Another aim of this trial was to explore whether the metabolic effects of prolonged SR differed between pre- and postmenopausal women. Aligning with our hypothesis, deleterious effects of SR on insulin sensitivity were exacerbated among postmenopausal women. Our work corroborates prior findings of reduced insulin sensitivity, measured using a hyperinsulinemic-euglycemic clamp, in postmenopausal women undergoing SR but no influence on glucose and insulin in response to a meal tolerance test (37). Interestingly, data from the fasted state appear to suggest differential phenotypes of SR-induced impairment of insulin sensitivity between the groups of women: among premenopausal women, increases in fasting insulin in response to SR were responsible for higher HOMA-IR, whereas both fasting insulin and glucose levels tended to be elevated in SR versus AS among postmenopausal women, particularly when adjusting for abdominal adiposity. These potential menopause-based differences in the response to chronic SR warrant further investigation.

This study has several important strengths. First, to our knowledge, it is the longest crossover SR intervention to date. In addition, mild SR, which closely resembles the chronic insufficient sleep profile adopted by a large portion of adults, and the outpatient setting lend strong ecological validity. These lifelike chronic sleep conditions were achieved without sacrificing rigor. Indeed, compliance with the study protocol was excellent and attrition was low. In addition, we gained insight into the potential direct effects of sleep on glucose metabolism by incorporating gold-standard measures of adiposity into analyses, which revealed that the observed effects of SR on insulin sensitivity were largely independent of change in adiposity. Another unique aspect of this study is the focus on sleep in women across all stages of adulthood and racial and ethnic backgrounds.

Our findings also highlight important avenues for future research. A crucial next step to improve our understanding of the role of insufficient sleep in glycemic dysregulation will be to extend the current SR paradigm to different population groups. First and foremost, the study must be repeated in men to evaluate potential gender and sex differences. Beyond this, replication is needed in a larger sample of postmenopausal women because power to detect effects in stratified analyses was limited in the present study. Results also warrant investigating chronic short sleep in individuals with existing metabolic conditions, such as prediabetes. Glucose tolerance often remains normal in the early stages of T2D, as insulin secretion is initially increased to maintain homeostasis (30). Given that we observe elevated insulin levels in response to SR among women without existing impairments to insulin sensitivity, it is possible that prolonged exposure to insufficient sleep among individuals with prediabetes could accelerate the progression to T2D. A notable limitation of the present study is the lack of gold-standard measures of insulin sensitivity from hyperinsulinemic-euglycemic clamp. Although there are modest to strong correlations between measures of glucose metabolism derived from hyperinsulinemic-euglycemic clamp with those of fasted samples and OGTT (38), application of clamping techniques (hyperinsulinemic-euglycemic and hyperglycemic) would allow for more precise and detailed assessment of changes in insulin sensitivity in response to prolonged SR and provide a direct measure of insulin secretion. Along with the need for replication with these gold-standard measures of glucose metabolism, interrogation of the mechanisms underlying effects of chronic insufficient sleep on glucose metabolism is warranted.

Despite the need for additional work to uncover the negative consequences of insufficient sleep and mechanistic underpinnings, it is also imperative to evaluate whether augmenting sleep can improve glucose metabolism among individuals with insufficient sleep. At present, few studies have been designed to test this question, and results of existing studies are inconsistent. One parallel-arm study applying a single, personalized sleep-hygiene consultation did not observe differences in 4-week changes in insulin sensitivity between participants in the treatment and control groups; however, sleep was only increased by ∼21 min/night in the treatment group (39). Other studies have observed improvements in glycemic control and IR with greater degrees of sleep extension (40,41), perhaps especially if sleep is extended to beyond 6 h/night (41). To our knowledge, there is one fully powered randomized controlled trial investigating changes in glycemic parameters in response to a longer-term sleep extension intervention (i.e., >4 weeks), conducted exclusively with male participants (42). In that study, men with overweight or obesity and habitual TST ≤6.5 h/night randomly assigned to the sleep-extension arm increased their sleep duration by >70 min/night, relative to the control group, over the 6-week intervention. Within the sleep-extension arm, average reductions in fasting insulin (11.03 pmol/L) and HOMA-IR (0.51) were comparable to the increases we observed in these measures in response to SR. Results of longer-term sleep-extension studies and our SR study in women support a causal impact of sleep on glycemic control. However, nuanced results related to the effects of sleep extension, along with limited data and interstudy heterogeneity, highlight the need for further investigation into whether behavioral interventions to enhance home-based sleep duration can improve glucose metabolism, assessed using gold-standard measures, among larger cohorts of men and women.

In summary, the causal impact of chronic insufficient sleep on reduced insulin sensitivity established in this study indicates that optimizing sleep can be an important strategy to prevent T2D in women. Achieving healthy sleep may be particularly important during the transition to menopause, a critical window for preventing cardiometabolic disease (43), given declines in sleep and T2D-protective hormones. This deserves further attention from the scientific community. In the meantime, clinicians should educate patients on the crucial role of sleep for health and discuss strategies to increase sleep duration, which may protect against IR and the development of T2D, and promote healthy life span in women.

Acknowledgments. The authors thank all participants of the study as well as research staff and members of the core laboratories at CUIMC.

Funding. This clinical trial was supported by the American Heart Association (AHA; grant 16SFRN27950012 to M.-P.S.-O.), a National Institutes of Health (NIH) Clinical and Translational Science Award to Columbia University (grant UL1TR001873), and the NY Nutrition Obesity Research Center (grant P30 DK026687-39). M.-P.S.-O. also receives support from the NIH National Heart, Lung, and Blood Institute (NHLBI; grants R01HL128226 and R35HL155670) and the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK; grant R01DK128154). F.M.Z. is a Berrie Fellow in Diabetes and Obesity Research and is supported by NIH NHLBI grant T32HL007343 and NIDDK grant R01DK128154. B.L. is supported by NIH National Institute of Aging grant R01AG065569 and NIDDK grants P30DK063608 and R01DK128154. S.J. is supported by AHA grant 16SFRN29050000 and NIH NHLBI grants R01HL106041 and R01HL137234. B.A. is supported by AHA grant AHA811531.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

The funding sources were not involved in study development, design, implementation, or dissemination.

Author Contributions. F.M.Z. and M.-P.S.-O. designed the research. B.A. and S.J. contributed to the study conceptualization. F.M.Z., B.L., and S.E.S. conducted the research. B.C. and Z.C. analyzed data. F.M.Z. wrote the manuscript. B.L., S.E.S., B.A., S.J., and M.-P.S.-O. reviewed and edited the manuscript. M.-P.S.-O. was responsible for final content. M.-P.S.-O. is the guarantor for this work and, as such, had full access to all of the data and takes responsibility for the integrity of the data and accuracy of analysis.