As part of the 2001–2006 National Health and Nutrition Examination Surveys (NHANES), the National Center for Health Statistics, Centers for Disease Control and Prevention, collected data representative of the U.S. civilian noninstitutionalized population (10). Survey participants were interviewed at home and invited to a mobile examination center (MEC) to undergo various examinations and laboratory measurements. The examination component consists of medical, dental, and physiological measurements, as well as laboratory tests administered by highly trained medical personnel. Those participants selected to undergo tests that require fasting between 8 and 24 h have appointments in the morning group. The response rates for those who participated in the examinations during 2001–2006 ranged from 76–80% (10). Among the 7,287 participants aged ≥50 years, 6,668 completed the household interview and an examination at the MEC; of those, 3224 were in the morning fasting group, with 2,925 fasting from 8 to <24 h prior to the examination.

During the home interview, participants were asked if they had ever been told by a doctor or other health professional that they had diabetes (other than during pregnancy). Based on their answer, 458 participants aged ≥50 years were classified as having diagnosed diabetes and were excluded from analyses. We excluded 193 participants who reported they did not have diabetes but had fasting glucose ≥126 mg/dL or HbA_1c ≥6.5% (48 mmol/mol) and 44 participants who had inconclusive results for fasting glucose or HbA_1c or both. The final sample size was 2,230 participants. The 1,221 persons who had a fasting glucose of 100–125 mg/dL or HbA_1c of 5.7–6.4% (39–47 mmol/mol) were classified as having prediabetes.

For our sample of older adults, variables assessed included age, sex, race/ethnicity (non-Hispanic [NH] white, NH black, Mexican-American, and other; due to the small sample size, other Hispanics are included in the “other” category) (10); household income (eight categories starting from <$20,000, then in $5,000 increments to ≥$75,000); educational attainment (less than high school, high school or GED, and more than high school); marital status (married or living with partner, divorced/widowed/separated, and never married); number of household members (actual number 1–6 and ≥7); self-reported family history of diabetes (i.e., including living and deceased, were any biological blood relatives including grandparents, parents, brothers, sisters, ever told by a health professional that they had diabetes?); measured BMI; large waist (measured circumference of ≥35 inches for women and ≥40 inches for men); high blood pressure (average of up to three readings ≥135/80 mmHg as recommended by the U.S. Preventive Service Task Force for screening of type 2 diabetes in adults or reported to be on antihypertensive medication) (11); HDL; triglycerides; total cholesterol; and a physical activity categorical variable based on metabolic equivalent hours per week in the past 30 days (0 for none, 1 for greater than 0 up to and including the median, 2 for greater than the median) for vigorous-intensity activity, moderate-intensity activity, muscle-strengthening activities, house or yard work, walking, or biking. We defined usual intake of selected dietary components (saturated fat, solid fat, high-fat dairy, meat, and refined grains) using the 24-h recall dietary assessment, averaging with a second day of recall data when available for the 2003 to 2004 and 2005 to 2006 cycles (12). All dietary data were adjusted using the MyPyramid equivalents to obtain the Healthy Eating Index (HEI) scores for each dietary component of interest (13). Although red meat has been considered to be episodic dietary intake, a recent study examining NHANES 24-h recall found it to be usual intake (14). The proportion of missing data among respondents was 7.3% for household income and <3.5% for all other variables in our analyses.

We used structural equation modeling with factor analysis, which groups intercorrelated variables into a single factor or latent construct, and path analysis, which includes the direct and indirect effects of factors previously reported associated with prediabetes as hypothesized (Fig. 1). Direct effects are depicted as an arrow emanating from an independent variable (exposure) leading and pointing to a dependent variable (outcome). For example, see the arrow between the latent constructs of socioeconomic position (SEP) and physical activity. An indirect effect is depicted as a mediating variable having an arrow pointing to it from an independent variable but also pointing to yet another dependent variable. In addition, physical activity as a mediating variable links SEP to the high blood pressure variable. A confounder, according to the use of these directed acyclic graphs, is depicted as a variable with direct effects on both the exposure and the dependent variable (15). Correlations between the measurement errors of two variables are represented by two-headed curving arrows, in which case only the measurement error terms are correlated.

In general, latent variable models reduce measurement error by having multiple indicators per latent construct, the ability to test models with multiple dependent variables, and the benefit of testing multiple integrated models simultaneously rather than factors individually. In addition, structural-equation modeling examines the direct and indirect effects of mediators on dependent variables while allowing the examination of complex associations among multiple mediators (16). Conversely, in a traditional regression model, mediators would not be included because they would block the pathway between the independent variable of interest and the dependent variable. Thus, in the structural-equation model, the independent factors and combined mediated relationships can be examined simultaneously, determining the impact of each of the dependent variables in the appropriate order. Thus, the SEM includes mediating effects without sacrificing indirect effects of interest. For each relationship in the SEM model, only data missing for either the independent or dependent variable would be missing from that equation.

Analyses proceeded in two stages. First, congruent with our hypotheses, we created and confirmed the a priori factor structure. We confirmed these latent constructs for SEP, physical activity, and dietary patterns. An assumption of this analysis is that an underlying unmeasured variable is identified by the shared variance of the observed variables. The constellations of factors that comprise poor diet and physical activity patterns may best be modeled as measures including those factors in terms of their shared variance rather than to account individually for the underlying immeasurable source. For the physical-activity construct, the observed decrease in physical activity participation among older adults may be partially due to ill-defined measurement (17). We therefore used a breadth of physical-activity domains and modeled them as a latent construct because the shared variance represents the entire pattern of physical activity that may influence the other variables and prediabetes. For the SEP construct, we used household income, level of education, marital status, and number of individuals in the household. Second, after confirmatory factor analysis, we added the other observed factors reported to be associated with prediabetes. The theoretical structural model tested is displayed in Fig. 1.

Our focus was to examine the effects of modifiable factors such as physical activity, diet, lipids, obesity, high blood pressure, and, to a lesser degree, SEP (which can only be modified with great difficulty) on prediabetes. Because age, sex, and race/ethnicity are strong, nonmodifiable confounders related to most of the other factors in the model, their direct effects, while included, are not shown in the graphic of the final model. Although family history of diabetes is nonmodifiable, it is specific to diabetes risk and therefore is examined as factor of interest.

We used SAS v9.2 (SAS Institute Inc.) for data management and descriptive statistics, along with Mplus v6.12 software for confirmatory factor analysis and testing the structural model, while accounting for the complex survey design of NHANES. P values <0.05 were considered statistically significant.

The conceptual model that specifies the relations (numbered) among concepts operationalized in this study appears in Fig. 1. A priori, we predicted 27 paths to directly and/or indirectly affect prediabetes, emanating from the 10 variables labeled. We relied on the data’s inherent temporality. For example, race, sex, and family history of diabetes are determined at birth, and age is a function of when one is born. Physical activity was reported during the past 30 days of MEC examination, and dietary intake was reported for the 24-h period prior to or up to 10 days after the MEC examination (18). We assumed these activities were “usual” patterns. Evidence from previous studies was also used to assess the determinants of the causal pathway (e.g.,family history of diabetes may lead to dyslipidemia and diabetes) (7) because prospective measures were not available in NHANES cross-sectional data.

As indices of the models’ statistical fit to the data, we used standard criteria, including comparative fit index (CFI) >0.90, root mean square error of approximation (RMSEA) < 0.08, and the standardized root mean square residual <0.06. Although not widely used, we also report the weighted root mean square residual (WRMR), which is a weighted average of the residuals—a value of <1 is recommended (19). Modification indices were used to assess specific paths for the best-fitting model, using a χ² value indicating the probability was <0.05 significance. Because the model had discrete dependent variables, the best method of estimation for the model is a robust weighted least squares, also known as the weighted least squares with mean and variance adjustment (20). With this type of model and estimation method, SEs for the standardized path coefficients are not computed. Therefore, we report the standardized estimates and the fit statistics of the models only.

Among those ≥50 years without diabetes, 51.7% had prediabetes. Compared with those with normal glucose, at the P < 0.01 level, those with prediabetes were more likely to be male (51.5 vs. 36.7%), overweight (39.4 vs. 34.6%), or obese (39.1 vs. 21.5%); to have high triglycerides >200 mg/dL (22.0 vs. 15.2%); and to have HDL <40 mg/dL (14.0 vs. 9.9%) (Table 1). Those with prediabetes had significantly higher average intakes of solid fats (5.92 vs. 4.78; P < 0.01, where higher score indicates greater percentage of total energy intake) and red meat (3.50 vs. 3.11; P = 0.03, where higher score indicates greater number of equivalents [e.g., ounces or cups] per 1,000 kcal) than those with normal glucose (Table 2).

Factor analysis confirmed the measurement portion of the model (CFI, 0.99; RMSEA, 0.02; standardized root mean square residual, 0.02) for the three latent constructs: 1) SEP (except that the number of family members did not load or contribute to the pattern of SEP identified in the factor analysis); 2) physical activity; and 3) poor diet (except that that saturated fats and processed meats did not load or contribute to the pattern of poor diet identified in the factor analysis). In the factor analysis assessment of the model, the latent constructs were not correlated with one another; however, in the structural-equation model, the latent constructs were correlated.

The best-fit structural-equation model (CFI, 0.89; RMSEA, 0.02; WRMR, 1.19) was somewhat different from our adjusted hypothesized model (CFI, 0.77; RMSEA, 0.02; WRMR, 1.56) (Fig. 2). The following changes were made for our hypothesized model to converge. First, we dropped total cholesterol and BMI from the final model based on the fit of parameters and the modification indices. Second, we dropped the direct effect of dietary patterns on HDL based on the fit of the model. Almost all factors had direct effects from age, sex, and race/ethnicity. For simplicity, these paths are not presented in Fig. 2. All paths shown are statistically significant at the 0.05 level (standardized path coefficients are given in parentheses), except for the one path of physical activity regressed on large waist circumference (P = 0.108). Higher SEP had a positive relationship with physical activity (0.317) and inverse relationships with poor diet (−0.457) and prediabetes (−0.157). The effect of SEP was greatest on poor dietary intake, indicated by the highest absolute value of a path coefficient. Family history of diabetes had a slightly greater effect on poor diet (0.111) compared with its relationships to HDL (−0.063), large waist circumference (0.066), and prediabetes (0.070). Poor diet had similar effects on triglycerides (0.148) and large waist circumference (0.146). The effect of physical activity on large waist circumference (−0.067) was not as strong as its effects on HDL levels (0.137), triglycerides (−0.136), and high blood pressure (−0.132). HDL had strong inverse correlations with triglycerides (−0.510) and large waist circumference (−0.291), being more highly associated in the former instance. Large waist circumference had a strong relationship with prediabetes (0.279) and also with high blood pressure (0.253); the relationship of high blood pressure on prediabetes was positive, but not as strong (0.122).