Cross‐sectional relation of long‐term glucocorticoids in hair with anthropometric measurements and their possible determinants: A systematic review and meta‐analysis

Summary Background Long‐term glucocorticoids (HairGC) measured in scalp hair have been associated with body mass index (BMI), waist circumference (WC), and waist‐hip‐ratio (WHR) in several cross‐sectional studies. We aimed to investigate the magnitude, strength, and clinical relevance of these relations across all ages. Methods We performed a systematic review and meta‐analysis (PROSPERO registration CRD42020205187) searching for articles relating HairGC to measures of obesity. Main outcomes were bivariate correlation coefficients and unadjusted simple linear regression coefficients relating hair cortisol (HairF) and hair cortisone (HairE) to BMI, WC, and WHR. Results We included k = 146 cohorts (n = 34,342 individuals). HairGC were positively related to all anthropometric measurements. The strongest correlation and largest effect size were seen for HairE‐WC: pooled correlation 0.18 (95%CI 0.11–0.24; k = 7; n = 3,158; I 2 = 45.7%) and pooled regression coefficient 11.0 cm increase in WC per point increase in 10‐log‐transformed HairE (pg/mg) on liquid‐chromatography‐(tandem) mass spectrometry (LC–MS) (95%CI 10.1–11.9 cm; k = 6; n = 3,102). Pooled correlation for HairF‐BMI was 0.10 (95%CI 0.08–0.13; k = 122; n = 26,527; I 2 = 51.2%) and pooled regression coefficient 0.049 kg/m2 per point increase in 10‐log‐transformed HairF (pg/mg) on LC–MS (95%CI 0.045–0.054 kg/m2; k = 26; n = 11,635). Discussion There is a consistent positive association between HairGC and BMI, WC, and WHR, most prominently and clinically relevant for HairE‐WC. These findings overall suggest an altered setpoint of the hypothalamic–pituitary–adrenal axis with increasing central adiposity.


| BACKGROUND
The prevalence of obesity, defined in adults as a body mass index (BMI; weight in kg divided by height in meters squared) ≥ 30 kg/m 2 , has increased dramatically worldwide over the past decades. 1 An imbalance between energy intake and expenditure is regarded as the major cause of obesity. Numerous distinct characteristics and conditions can contribute to obesity within an individual. 2 One important contributing factor may be chronic exposure to the stress hormone cortisol, the major end-product of the hypothalamic-pituitary-adrenal (HPA) axis. In healthy individuals, cortisol secretion and metabolism are closely linked and tightly regulated. Cortisol is converted by 11-beta-hydroxysteroid dehydrogenase type 2 (11β-HSD-2) to the biologically inactive cortisone in end-organ tissues, but can be converted back to cortisol by 11-beta-hydroxysteroid dehydrogenase type 1 (11β-HSD-1) on tissue-level. 3 Exposure to very high levels of endogenous or exogenous glucocorticoids (GC), such as in Cushing's syndrome, leads to a phenotype characterized by abdominal obesity and other features of the metabolic syndrome. 4,5 It is hypothesized that even a chronic mild increase of GC, that is, in the high-physiological range, can contribute to overweight and obesity in the general population. 2 Despite many efforts over the last decades to explore this relation in different matrices such as blood, saliva and urine, conflicting results were found. 6 This may be due to cortisol's circadian rhythm, its pulsatile secretion, and the daily variation following changing circumstances such as acute stress. Hence, measurements that reflect a shorter term (minutes or hours for serum and saliva, days for urine) seem less suitable to investigate this association in the general population. 7 In the past decennium, a relatively novel technique has allowed researchers to study long-term levels of GC by measuring cortisol and cortisone levels in scalp hair (HairF and HairE, respectively). Every centimeter of scalp hair is believed to represent the cumulative GC exposure of one month. 8 HairGC measurements are now considered an easily applicable, noninvasive and reproducible method for assessing longterm GC exposure. 8  to differences in mean BMI of the study populations. 9 Gray et al. and Ling et al. also reported that BMI and BMI standard deviation score (SDS), that is, BMI z-scores adjusted for age and sex that are most often used in pediatric studies, 10 were important determinants of HairF levels in children. 11,12 However, in the last years, many new large-scale studies in various age categories have been published that have investigated the relation between HairGC and anthropometric features. Some of these studies showed a positive relation, 13,14 while other studies showed no relation between HairGC and anthropometric measurements. 15,16 It is unclear whether these conflicting results can be explained by differing population characteristics such as mean age, sex, and prevalence of obesity, use of corticosteroids, handling of outliers, or the various laboratory methods that were used.
Moreover, other anthropometric measurements than BMI are considered equally or even more relevant to cardiometabolic health, such as waist circumference (WC) and waist-hip-ratio (WHR), which both are markers of central adiposity. 17 These deserve specific attention as GC are known to particularly induce abdominal obesity. 18 Likewise, there are suggestions that hair cortisone might correlate stronger to obesity than hair cortisol itself. 19 However, a metaanalysis that summarizes all evidence considering different anthropometric parameters in association with both HairF and HairE as well as relevant moderators of these relationships is missing.
Therefore, the aim of the current systematic review and metaanalysis was to investigate the cross-sectional relations between HairGC levels (HairF and HairE) and anthropometric measurements (BMI, BMI SDS, WC, and WHR) and to explore the possible influence of relevant characteristics of the population and laboratory methods.

| METHODS
We performed this systematic review and meta-analysis in concordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement and Meta-analysis of Observational Studies in Epidemiology (MOOSE) checklist. 20,21 This systematic review was registered at the PROSPERO database (Registration number CRD42020205187, December 7, 2020). 22

| Search strategy and selection criteria
A university health sciences librarian designed a comprehensive search to identify studies and conference abstracts concerning hair cortisol and/or hair cortisone and measurements of obesity.
To avoid missing potentially relevant papers we designed a broad search strategy combining the elements "hair," "cortisol/cortisone," and "BMI/WC/WHR/anthropometrics", including their synonyms without any restrictions other than "studies in humans". The search was conducted in the following databases from inception up to November 16, 2020: Medline (Ovid), Embase, Cochrane, Web of Science, Scopus, Cinahl, PsycInfo, and Google Scholar. The complete search strategy is provided in the supporting information Appendix S1. Search results were exported to reference management software (EndNote version X9, Clarivate Analytics), and duplicates were removed prior to screening.
All identified studies were independently screened in two stages by two physicians (EV, OA, or MM) with a background in adult (EV and MM) and pediatric (OA) endocrinology. All studies that reported original HairGC data in humans were included in the title/ abstract screening stage and were subsequently assessed full text.
Disagreements were solved by discussion among the first authors (EV, OA, and MM), and the senior author (EvR) until consensus was reached. Additionally, reference lists of all included studies and relevant reviews were screened systematically for potentially relevant articles. 23 We included studies that reported cross-sectional associations between HairGC and measurements of obesity. We excluded case reports, animal studies, review articles, non-English or nonpeer reviewed studies, and studies in which hair sampling and weight measurements were not performed simultaneously ( Figure 1). Pediatric studies that only included children younger than age 2 years were also excluded because BMI-based definitions of obesity are not available for this age group. 10 We contacted all corresponding authors of articles that reported both HairGC and anthropometric data but did not report an association between these two outcomes to ask if they could provide us with an association measure. Of articles that also included patients with mental or physical diseases that are known to influence the relation between GCs and obesity, we only included the separate analyses of healthy controls if available. When data of the same participants were reported in several studies, we included the study that reported a bivariate association (correlation coefficient or unstandardized simple linear regression coefficient) between HairGC and measurements of obesity. If more than one article reported a bivariate association, we included the study with the largest sample size.

| Data extraction
Descriptive, methodological, and outcome data were extracted from all included studies by two researchers independently (EV, OA, or MM) using a predesigned standardized data extraction sheet. Discrepancies were resolved by discussion among the first authors (EV, OA, and MM) and the senior author (EvR). The following descriptive data were extracted: study population characteristics (sample size and cohort characteristics: age, sex, prevalence of obesity, mean levels of and measurements of obesity, that is, BMI, BMI SDS, WC, and WHR. In studies presenting multiple data points of the same participants (e.g., before and after an intervention), only baseline associations were extracted. When insufficient data were reported for meta-analysis, corresponding authors were contacted twice in a 2-week time frame.
In case of nonresponse, data were extracted from previous metaanalyses where possible. 9,12

| Risk of bias assessment
Risk of bias was assessed by two researchers independently (EV, OA, or MM) using the Quality In Prognostic Studies (QUIPS) tool. 24 In short, the QUIPS tool aids in the assessment of potential bias sources F I G U R E 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram. HairGC, hair glucocorticoids from the following study domains: study participation, study attrition, prognostic factor measurement, outcome measurement, confounding measurement, and statistical analysis. The subdomains on which risk of bias was assessed were the following: population selection criteria (QUIPS 1; study participation), the used laboratory methods (QUIPS 3; prognostic factor measurement), whether or not anthropometric measurements were objectively measured (QUIPS 4; outcome measurement), whether or not corticosteroid use was taken into account and whether any consideration was given to handling outliers in HairGC values (QUIPS 5; study confounding), and reporting of relevant statistics (QUIPS 6; statistical analysis and reporting). All subdomains were scored as 'low', 'moderate', or 'high' risk of bias on individual cohort level. We omitted the study attrition domain of the QUIPS tool (QUIPS 2) since it was not applicable to our cross-sectional research question. Discrepancies between the researchers were solved by discussion among the first authors (EV, OA, and MM) and the senior author (EvR).

| Qualitative synthesis
For the qualitative synthesis, we summarized all authors' conclusions regarding cross-sectional associations between HairGC levels and obesity measurements, that is, correlation coefficients, regression coefficients, or comparison of HairGC levels and obesity measurements across categories.

| Statistical analysis
All meta-analyses were conducted in R version 3.6.3 with an α of 0.05. 25 For all descriptive data, medians and (interquartile) ranges were converted to means and standard deviations prior to analyses. 26 Furthermore, subgroup means from individual studies as well as the pooled means across all studies were pooled. 27 When not originally reported, standard errors were calculated based on reported confidence intervals or p-values and degrees of freedom using the Tdistribution.

| Meta-analysis of correlation coefficients
For all studies reporting bivariate correlations (correlation coefficients), Fisher's r-to-z transformation was applied to transform individual correlations stratified on all combinations of HairGC (HairF and HairE) and obesity measurements (BMI, BMI SDS, WC, and WHR). As several studies reported correlations within distinct subgroups, we calculated the pooled correlation coefficients, 95% confidence intervals (CIs) and prediction intervals (PIs) using multilevel random effects models. 28,29 One study was excluded for all meta-analyses, as the reported correlation coefficient for BMI versus HairF of the total cohort was 0.91. We assume this is a typographic error, as the authors state that they only found a statistically significant correlation in the highest tertile of the polygenic susceptibility score (which was reported to be 0.269, making a correlation of 0.91 for the total cohort impossible). 30 These authors did not respond to our contact attempts.
The I 2 statistic and Cochrane's Q test were used for the assessment of between-study heterogeneity, with I 2 > 25% and p-value for Cochrane's Q test <0.05 indicating heterogeneity. For all metaanalyses with data from at least 10 cohorts, exploratory moderator analyses were performed using mixed-effect models for categorical parameters (e.g., used laboratory method) and random effects models for continuous parameters (e.g., mean age of the study participants).
Publication bias was assessed using contour-enhanced funnel plots.

| Meta-analysis of unstandardized simple linear regression coefficients
For all studies reporting unstandardized simple linear regression coefficients between 10-log transformed HairGC (HairF or HairE) in pg/mg as independent variable and untransformed obesity measurements (BMI, BMI SDS, WC, and WHR) as dependent variable, pooled regression coefficients and 95% CIs were calculated using the statistical approach described by Bini et al. and Becker & Wu. 31,32 In short, this approach allows pooling of linear regression coefficients using weighted least squares provided that the independent and dependent variables have been measured in the same manner across all studies. Therefore, we calculated pooled regression coefficients of 10-log transformed HairGC on untransformed obesity measurements, stratified on laboratory method. Between-study heterogeneity was assessed using the Q w -statistic described by Bini et al. 31

| RESULTS
The literature search identified 1017 unique citation titles of which a total of 120 studies 5,13,14,16,19,30, comprising 146 separate cohorts were included ( Figure 1). This corresponds to a total of 34,342 included participants of which 15,698 (46%) were sampled from general population-based studies ( Table 1). The remaining 18,644 (54%) participants were sampled from studies where study inclusion was based on medical criteria (e.g., individuals with obesity), occupational characteristics (e.g., health-care workers), or socioeconomic characteristics (e.g., children from low-income parents).

| Risk of bias
Risks of bias assessments on cohort level are presented in Table 1.

| Qualitative synthesis
An overview of all outcomes reporting any relation between HairGC and obesity measurements is shown in the supporting information  (Table 4 and supporting information Figure S9). Mean age and mean HairF concentration of the study population did not moderate the correlations between HairGC and obesity measurements (all p-values >0.05, Table 4). In contrast, higher mean HairE was associated with stronger positive correlations (estimated slope 0.0046 per point increase in mean HairE on study level, 95% CI 0.0025-0.0068, p < 0.0001).
Visual inspection of the funnel plots showed no evidence for publication bias; that is, no systematic trends were found between standard error (as proxy for study sample size) and magnitude and direction of the reported correlation coefficients (supporting information Figures S10-S15).

| Meta-analysis of regression coefficients
The pooled regression coefficients stratified on analysis method are presented in  and BMI/BMI SDS. 9,12 Evidently, there is a relation between measures of obesity and long-term glucocorticoid levels, a relation that has been controversial for measurement of GC levels in other matrices that reflect shorter time periods. 6 As GC are known to contribute to central adiposity, for example, in Cushing's syndrome, it might be possible that in the study of a gradually developing disease such as obesity, long-term GC measurements offer a different and perhaps more appropriate perspective to the role of the HPA-axis.
The current study indicates that this relation is strongest (i.e., the highest correlation coefficient and the largest effect size) for cortisone, the inactive form of cortisol, and WC. Although the pooled HairGC levels, although this is highly speculative. Other possible areas of bias, for example, the selection of participants (whether or not the participant selection was population-based or based on medical, occupational or socio-economic characteristics), the consideration of possible confounders (outliers of HairGC measurements and corticosteroid use), and the statistical reporting all did not affect the outcomes.
As expected, given the large number of included studies, we observed a relatively high between-study heterogeneity in our metaanalyses of correlation coefficients, up to an I 2 of 68% for HairF versus WC. Although some of our studied moderators could explain part of this heterogeneity, the majority is still unexplained. Hence, there may be a role for other factors that are known to influence HairGC levels and/or obesity that we did not account for in the current report. For example, a recent meta-analysis demonstrated that adversity also relates to long-term GC levels, although this relation is complex and depends on the type and timing of adversity and on the studied population. 154 Adversity and stressful conditions can have similar complex relations to obesity. 155 We did not include these factors as possible moderators in our analyses due to a lack of universally accepted definitions that we could apply to all studies. However, we do not suspect a major influence of stressful conditions on our results as sensitivity analyses focusing on population-based cohorts were comparable with the analyses based on all data.
A major strength of the current study was our comprehensive search in which we included all studies that reported any association between measures of adiposity and HairGC levels, including studies that did not primarily aim to investigate these associations. To minimize the risk of publication bias due to incomplete reporting of results based on statistical significance, we contacted corresponding authors of all included studies for additional information. In addition, we contacted all corresponding authors of studies that reported anthropometric measurements and HairGC but not an association. This yielded additional information for 70 cohorts (48%). This limits the risk of publication bias, which was also confirmed by our funnel plots (supporting information Figures S10-S15). Moreover, an important addition of our work compared with the two systematic reviews and metaanalyses that have already been published on this topic was that we studied both the active form cortisol and the inactive form cortisone, their relations to different measures of adiposity, and also investigated effect sizes complementary to correlations. This has yielded the valuable conclusion that both the strongest correlation as well as the strongest, clinically relevant effect size are actually seen for HairE versus WC, instead of the most commonly studied association HairF versus BMI. Another strength of our study is that we focused on studies that did not include participants with severe diseases affecting GC levels, which have therefore not disturbed our findings.
A limitation of our study was that we obtained data that are related to full cohorts instead of individual person-data. This restricts our conclusions to comparisons across cohorts instead of across individuals. However, by pooling regression coefficients, we could provide an effect size that is applicable on individual level. Other limitations relate to the lack of standardization of HairGC analysis methods and the usefulness of HairGC itself, as there are still numerous issues unsolved. For example, the ubiquitously reported growth speed of scalp hair, 1 cm per month, may vary considerably by ethnicity and season. 8 Other issues represent the high prevalence of overall CS use (which may influence basal cortisol levels and were found to be used by 11% of the Dutch population, a number that may be even higher in other countries 140,156 ), hair characteristics such as color, treatment and washing frequency, 157 and the unresolved issue of how to handle HairGC outliers. 158,159 These characteristics were often not reported in the included studies, which prevented comparison across studies.
Then again, the results of our analyses in the subgroup of studies that accounted for outliers and corticosteroid use, the two issues that are most likely related to obesity, did not differ significantly from the results in the subgroup of studies that did not account for outliers, corticosteroid use, or neither. It should however be noted that we only assessed whether studies handled outliers at all and that the exact manner of handling outliers in (psycho)endocrine research is still a separate topic of discussion. 159 Lastly, this review only included cross-sectional associations while any conclusion on the prognostic or predictive value of HairGC for future obesity should come from studies investigating longitudinal relations, which have however until now only been performed scarcely. 15,134 Altogether, we confirmed a consistent positive association between anthropometric measurements and hair glucocorticoids. This relation was most often studied for hair cortisol and BMI but showed the strongest correlation and largest effect size for hair cortisone and WC. These relations were not influenced by mean age, mean BMI, or mean HairGC levels nor by the used laboratory methods of the studies. However, the percentage of males, the percentage of participants with obesity, and objective measurement of weight instead of selfreported weight represented important features to take into account when assessing hair glucocorticoids in cohorts. Although causality is not yet proven, our results suggest that higher long-term glucocorti-