Data were obtained from NHANES (1999–2010), a complex, multistage probability sample of the US population. These annual cross-sectional surveys are conducted by the National Center for Health Statistics (NCHS) of the Centers for Disease Control (CDC), with randomly-selected subjects undergoing anthropometric and blood pressure measurements, answering questionnaires and undergoing phlebotomy. The NCHS ethics review board reviewed and approved the survey and participants gave informed consent prior to participation. Body mass index (BMI), SBP, and laboratory measures of triglycerides, HDL-C, and fasting glucose were obtained using standardized protocols and calibrated equipment. For SBP, the mean of up to four readings taken on each individual was used. All blood samples used for analyses were obtained following a fast ≥8 hours prior to the blood draw.
Data from non-Hispanic-white, non-Hispanic-black, or Hispanic (Mexican-American/other Hispanic) adolescents 12–19 years old were analyzed. Children <12 years old were excluded since fasting values for triglycerides and glucose were only obtained in participants ≥12 years old. Subjects were excluded if they had known diabetes or unknown diabetes (fasting plasma glucose >125 mg/dL), as each of these limit insulin release. Pregnant women were also excluded, as well as individuals taking antihyperlipidemic or anti-diabetic medications as these are all likely to alter lipid and insulin levels.
We combined all data sets from the 6 two-year cycles (1999–2010) for statistical analyses to increase our total sample size. Prevalence rates of MetS were calculated by sex and race/ethnicity, according to Ford’s pediatric adaptation of the ATP III adult criteria, and mean levels (95% CI’s) of the MetS components of interest (BMI z-score, SBP, HDL, triglycerides, and glucose) as well as the surrogate outcomes (hsCRP, uric acid, fasting insulin) were calculated by sex and race/ethnicity.
A series of confirmatory factor analyses (CFA) were then performed on the five identified MetS components: BMI z-score, SBP, HDL, triglycerides, and glucose. Numerous definitions of MetS have been proposed with various components that comprise it[3
]. We selected these five components exclusively because of their inclusion in the most prominently used MetS criteria (Additional file1
: Table S1) and due to their clinical accessibility. We chose to use BMI-z-score instead of WC because BMI-z-scores can be standardized by age and sex using established CDC programs[40
]. Unfortunately, while cut-off values exist to identify elevations in WC by age in adolescents[7
], there is a lack of precise percentiles to permit such standardization of WC values in adolescents. In addition, WC as a clinical measure is prone to substantial measurement error[41
] and is not consistently used in clinical care, while BMI z-scores are readily available and are the predominant tool recommended for adiposity assessment in clinical practice[42
]. With such a factor analysis, both SBP and DBP should not be included together[4
], and we chose SBP given it is more strongly associated with insulin resistance[5
] and other outcomes[43
]. Triglycerides were log-transformed, and all variables were standardized (mean=0, SD=1) over the entire sample. The inverse of HDL was used when standardizing, so a higher factor loading score would be similar in interpretation to the other measures in the model. The variables were not standardized within groups to allow for potential overall higher standardized scores within sex- and race/ethnic-specific groups. A one-factor model formed the basis of all of the CFA’s performed; measurement errors between the five components were assumed not to be correlated. The factor loadings were of interest, indicating the magnitude of association between each component and the underlying “MetS” factor. Factor loadings >0.3 were considered clinically meaningful. Using PROC CALIS in SAS, the parameters of the CFA’s were estimated via maximum likelihood. Two multi-group one-factor CFA’s were fit:
Model 1: constrained the factor loadings to be equal across the six combinations of sex and race/ethnicity;
Model 2: allowed the factor loadings to vary across the six groups.
Chi-square tests of the equality of the factor loadings across the six groups in Model 2 were performed. The two models were compared using various fit statistics, both overall and by group. Chi-square and Akaike’s Information Criteria (AIC) were used for model comparisons; smaller chi-square and AIC values indicated a better fit. A chi-square difference test was calculated; a significant difference between two nested models implies that the model with more paths explains the data better. Other goodness of fit indices included the Root Mean Square Error of Approximation (RMSEA; >0.06 poor fit), the Standardized Root Mean Square Residual (SRMR; >0.08 poor fit), the Goodness of Fit Index (GFI; <0.90 poor fit), and the Bentler-Bonett Normed Fit Index (NFI; <0.90 poor fit).
The standardized factor coefficients from the better-fitting model were used to calculate the MetS factor score on each individual. This score can be interpreted as a Z-score (mean 0, SD=1), with higher scores representing an increased risk of MetS. Receiver operating characteristic (ROC) analysis was used to assess the ability of this new score to discriminate against the traditional MetS criteria as well as elevated levels of the three identified CVD/T2DM surrogates. Specifically, we were interested in the ability to predict elevated fasting insulin (>16 IU/mL, the 95th percentile among normal-weight adolescents), elevated hsCRP (>4.5 mg/L), and elevated uric acid (approximately the 95th percentile among lean individuals: 7.0 mg/dL for males, 5.5 mg/dL for females). In this analysis, individuals with CRP >10.0 mg/L were excluded. In addition, the ability to predict 1 and ≥2 elevations amongst these surrogates was of interest, as this indicates the more at-risk adolescents. Overall predictive performance was measured by the area under the curve (AUC) of the ROC curve, with 0.5 and 1.0 indicating no and perfect predictive ability, respectively. We considered AUC values >0.70 to be reasonably accurate and AUC >0.90 to be very accurate. Sensitivities and specificities to predict ≥2 elevations among the three surrogates were compared between the traditional MetS classification and a definition of MetS using a cutoff identified by the ROC analysis; these statistics were done on a sex and race/ethnicity-specific basis.
Statistical significance was defined as a p-value<0.05. Statistical analysis was performed using SAS (version 9.3, Cary, NC). Descriptive statistics as well as sensitivity estimates used SAS survey procedures (SURVEYMEANS, SURVEYFREQ), which accounts for the survey design when estimating standard errors to obtain population-based estimates. The CFA itself did not account for the survey design due to the inability of standard software to perform multigroup CFA’s within subpopulations while accounting for the survey design.