Childhood Adiposity and Adolescent Sex Steroids in the Exploring Perinatal Outcomes Among Children Study

Catherine Kim; Kylie K. Harrall; Deborah H. Glueck; Daniel Shumer; Dana Dabelea


Clin Endocrinol. 2019;91(4):525-533. 

In This Article

Materials and Methods


The design, methods and baseline characteristics of EPOCH participants have been previously described.[7] EPOCH is an observational prospective study that recruited healthy 6- to 13-year-old children who were offspring of singleton pregnancies, born at a single hospital in Denver between 1992 and 2002, whose biological mothers were members of Kaiser Permanente of Colorado (KPCO). The study population was sampled to reflect similar racial and ethnic distributions of Colorado. The EPOCH study included an oversampling of offspring of mothers who had experienced gestational diabetes in pregnancy, and average BMI in EPOCH is slightly higher than the national standard.[9] Children and their mothers were invited to participate in two research visits at average ages of 10.5 (SD = 1.5) and 16.7 (SD = 1.2). While a total of 604 children attended the 1st research visit, this analysis focused on the 418 children who attended both visits. All participants provided informed consent, and youths provided written assent. The study was approved by the Colorado Multiple Institutional Review Board.

Measures of Adiposity

Childhood height and weight were measured in light clothing and without shoes. Weight was measured to the nearest 0.1 kg using an electronic scale. Height was measured to the nearest 0.1 cm using a portable stadiometer. BMI was calculated as kg/m2. Waist circumference was measured according to the National Health and Nutrition Examination Survey protocol as previously described.[10] At both visits, MRI of the abdominal region was used to quantify visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) with a 3 T HDx Imager (General Electric) by a trained technician. Each study participant was placed supine and a series of T1-weighted coronal images were taken to locate the L4/L5 plane. One axial, 10 mm, T1-weighted image, at the umbilicus or L4/L5 vertebrae, was analysed to determine SAT and VAT content. Images were analysed by a single reader, blinded to exposure status. With measures of VAT and SAT, we generated VAT/SAT ratios. Previous reports in the EPOCH cohort have examined the correlation between BMI and specific markers of visceral adiposity. BMI has high correlation with SAT in boys (Pearson's r = .94) and girls (r = .91), but the correlation with VAT is moderate (r = .65 in boys, r = .68 in girls); correlation between visceral fat mass and waist circumference was similar.[10] Hepatic imaging was performed at the 2nd research visit using a magnitude based, 6-echo, spoiled gradient-recalled echo sequence. Hepatic fat fraction (HFF) was calculated from the mean pixel signal intensity data, for each echo acquisition using an open source Osirix algorithm. This fraction was then multiplied by 100 such that a value of 1 is equivalent to 1% HFF.

Dehydroepiandrosterone, Testosterone, Oestradiol Measurements

A fasting blood draw after an overnight fast occurred for all consenting children at the second visit. Sera were refrigerated and analysed within 24 hours of collection. All laboratory measurements were performed at the Colorado Clinical Translational Science Institute Core Laboratories. Serum oestradiol was measured by using a Beckman Coulter chemiluminescent with a limit of detection of 10.0 pg/mL. Serum total testosterone was measured by using a Beckman Coulter 1-step competitive with a limit of detection of 17 ng/dL. Serum DHEA was measured by using a Beckman Coulter chemiluminescent with a limit of detection of 2 μg/dL.

Covariate Measurements

The KPCO Perinatal database, an electronic database linking the neonatal and perinatal medical record, was used to collect birthweight. Race/ethnicity was self-reported using 2000 US Census-based questions and categorized as Hispanic (any race), non-Hispanic white, non-Hispanic African American and non-Hispanic other. Pubertal development was self-assessed using a diagrammatic representation of Tanner staging adapted from Marshall and Tanner; sexual maturation self-assessment was recently shown to be in excellent agreement with physician-assessed Tanner stage,[11] although lack of direct testicular measurement could potentially over-estimate pubertal stage. For the purpose of the analysis, youth were categorized as prepubertal (Tanner <2) and pubertal (Tanner 2–5), and as white or nonwhite. Maternal level of education (high school or less vs more than high school) and total household income (<$50 000 vs more than $50 000) were self-reported at the study visit. Fasting insulin was measured by a radioimmunoassay method. Plasma leptin concentration was measured by using a Millipore radioimmunoassay with a sensitivity of 0.5 ng/mL. Serum LH was determined by using a Beckman Coulter chemiluminescent assay with a sensitivity of 0.12 mIU/mL.

Statistical Analysis

Girls and boys were examined separately. Baseline characteristics were described using sample size and percentages for categorical variables and means and standard deviations for numeric variables (Table 1). When considering sex steroid profiles, given the young age of study participants, we needed to consider the substantial number of values that fell below the level of detection.[9–12] Many authors choose to replace values below the limit of detection with randomly sampled values, or with a value halfway between zero and the detection limit. These approaches have been critiqued for introducing bias or reducing statistical power.[12–15] We chose instead to employ the reverse-scale Cox proportional hazards approach,[16] an unbiased approach which maintains power by including both values above and below the limit of detection. The approach does not rely on either estimation or imputation.[16]

In classical survival analysis, the data are right-censored. By contrast, sex hormone data are left-censored at the level of detection. It is important to realize that data below the level of detection still conveys information. Intuitively, if a person has a sex hormone level below the limit of detection, the level must be very low. This leaves the analyst with a thorny problem. How do we apply methods for right-censored data to left-censored data?

To apply classical survival methods to sex hormone data, the scale of the data must be reversed, through a transformation. The scale of each sex hormone is reversed by subtracting observed sex hormone levels from the maximum observed value for that sex hormone in the data set. This effectively converts left-censored to right-censored data. The resulting model predicts the risk of sex steroid levels above the limit of detection. Thus, a hazard ratio less than 1 indicates a greater risk of having undetectable sex steroid levels. The approach allows the analyst to describe the percentage of the population below any sex steroid hormone level of interest, as shown in Figure 1.

Figure 1.

Proportion of adolescent boys with testosterone levels below the level of detection at their follow-up visit, by percentiles of the VAT growth rate

We fit nine separate reverse-scale Cox proportional hazard models. Each model had a different hormone outcome (DHEA, testosterone or oestradiol) and a different measure of adiposity as a predictor (VAT, SAT, or VAT/SAT). For the best fitting models, we assessed the validity of the proportional hazard assumption, and reported model generated hazard ratios, parameter estimates, 95% confidence intervals and graphics.

The base model included the specific fat measure at visit 1, growth rate in adiposity between the two visits, and the following covariates: maternal education and income, birthweight and pubertal stage at visit 1. Growth rate in adiposity was computed as the difference between the adiposity measure at visit 2 and the measure at visit 1, divided by the difference in age between visit 2 and age at visit 1. Subsequent models further adjusted for visit 2 LH, fasting insulin, leptin and hepatic fat to determine whether addition of these potential mediators changed the association. We also evaluated interactions between race/ethnicity with the initial fat measure and growth rate in that specific fat measure; if no interactions were detected, models were adjusted only for the main effect of race/ethnicity. Analyses were performed using the Statistical Analysis Software (SAS) version 9.4 (SAS Institute), and all tests were two-sided with statistical significance set at P < .05.