Participants and Procedure
Data for this study are from the Midlife in the United States (MIDUS) study, a national study of health and well-being involving a national probability sample of middle and older adults from the 48 continental states [29]. MIDUS was started in 1995–1996 (MIDUS 1), included 7,108 adults, ages 25–74, recruited through random digit dialing (RDD) and completed baseline telephone interview. Majority of the participants (89% of the total sample) in MIDUS 1 were also completed self-administered questionnaires (SAQ). The longitudinal follow-up of MIDUS was conducted in 2004–2006 (MIDUS 2), included 4,963 longitudinal participants. Similar to MIDUS 1, all the participants in MIDUS–2 completed the baseline telephone interview, in which 81% of them also completed the SAQ. To increase the racial diversity of MIDUS 2 sample, a supplemental sample consisted majority Black adults was recruited from Milwaukee County, WI (n = 592). Similar to the national sample in MIDUS 2, all the Milwaukee supplemental also completed baseline interview and majority of them (89% of the total sample) completed the SAQ. A new protocol of biomarker assessment was introduced during MIDUS 2. Participants who completed both baseline telephone interview and SAQ were eligible to participate in the biomarker assessment. In MIDUS 2, 1,255 randomly selected participants, both from the national sample and the Milwaukee supplemental sample, completed the biomarker assessment.
In 2012–2016, a new national probability sample (n = 3,577) that matched the original MIDUS sample (MIDUS 1) in terms of their sociodemographic characteristics was recruited to participate in the MIDUS Refresher study (MIDUS R). This sample was recruited to replenished the number of middle-aged adults given that the initial cohort was now older [14, 30]. Similar to MIDUS 1 and 2, participants in MIDUS R were recruited through RDD and all completed the baseline telephone interview. Majority of the participants in MIDUS R (73%) also completed the SAQ. Similar to MIDUS 2, a supplemental sample was also recruited from Milwaukee County in order to increase the racial diversity of the sample in MIDUS R (n = 508). Among the supplemental sample in MIDUS R, 299 participants (59% of the in-person interview participants) completed the SAQ. MIDUS R also include biomarker assessment protocol, with the same eligibility requirement as in MIDUS 2 (completed the baseline survey and SAQ). In MIDUS R, 863 participants (randomly selected from the national sample and the Milwaukee supplemental sample) completed the biomarker assessment.
For the current analysis, data were from 2,118 (ages 25–84; 54.9% female; 73.7% non-Hispanic White) participants who completed the biomarker assessment in MIDUS 2 and MIDUS R. The biomarker assessment protocol in MIDUS 2 and MIDUS R was identical. Participants were invited to stay overnight at one of the three regional clinical research units (CRUs; West Coast, Midwest, and East Coast). The selection of the CRU for each participant was based on the one that imposed the least travel burden. Blood and urine samples were collected during the stay. Participants provided informed consent to participate in both the baseline survey and the biomarker assessment. Additional information regarding biomarker assessment in MIDUS study can be found elsewhere [31].
Measures
Risk Factors
Seven known risk factors for CKD were included in this analysis: (1) elevated blood pressure/ BP (mean of second and third blood pressure test: systolic and diastolic blood pressure ≥ 140/90 mmHg or self-reported diagnosis of hypertension by physician; (2) elevated glycosylated hemoglobin (HbA1c ≥ 6.5%) or high fasting blood glucose (≥ 126 mg/dL) or self-reported diagnosis of type 2 diabetes by physician; (3) obese (BMI ≥ 30 kg/m2); (4) abdominal obesity (waist circumference ≥ 88 cm for women and ≥ 102 cm for men); (5) hypercholesterolemia (total serum cholesterol ≥ 200 mg/dL); (6) elevated c-reactive protein (CRP ≥ the third quartile); and (7) elevated interleukin 6 (IL6 ≥ the third quartile).
Kidney Function
Estimated glomerular filtration rate (eGFR) was estimated from serum creatinine using the CKD-EPI formula [32]. Serum creatinine was assayed from overnight fasted blood collected at the three CRUs using Roche Cobas Analyzer (Meriter Clinical Lab, Madison, WI; inter-assay coefficient of variability = 2.08%). The overall mean of eGFR was 91.2 mL/min/1.73 m2 (SD = 19.2 mL/min/1.73 m2). For further analysis, eGFR was transformed into a binary variable based on the clinical indicator of Stage 3 CKD (1 = eGFR lower than 60 ml/min/1.73 m2, n = 107 [5.1%]; 0 = the rest of participants).
Childhood SES
Childhood SES was the total score from three indicators, including (1) father (or mother in case of missing data) highest level of education (0 = < high school, 1 = graduated from high school/GED, 2 = some college or higher); (2) whether family of origin received welfare (0 = all the time/most of the time, 1 = some of the time/a little of the, 2 = never in welfare); and (3) financial level growing up (0 = a lot/somewhat/a little worse off than average family, 1 = same as average family, 2 = a lot/somewhat, a little better off than average family). The mean childhood SES score was 3.91 (SD = 1.45; range = 0 - 6). These measures of childhood SES has been shown to be a significant predictor of health outcomes in adulthood, such as allostatic load, chronic disease, and diabetes [14, 33, 34].
Covariates
Covariates in the analysis include participant’s highest formal education level (0 = no high school diploma/ GED; 1 = graduated from high school and higher) and current/ adult SES. Adult SES was the total score based on five indicators [14, 33, 34], including: (1) household-size adjusted income to poverty ratio (0 = < 150%, 1 = ≥ 150% - < 300%, 2 = ≥ 300%); (2) current financial situation (0 = worse, 1 = average, 2 = best); (3) availability of money to meet basic needs (0 = not enough money, 1 = just enough money, 2 = more money than need); and (4) difficulty level paying bills (0 = very/somewhat difficult, 1 = not very difficult, 2 = not at all difficult). Sociodemographic variables were also incorporated as covariates, including age (years), gender (female = 0, male = 1), and race/ ethnicity (minority = 0, non-Hispanic White = 1).
Statistical Analysis
Using a national probability sample of the U.S. adults, the following analysis had three primary aims (Figure 1): (1) to model the heterogeneity of comorbidity among CKD risk factors by examining the clustering of risk factors associated with age-related declines in kidney function among middle-aged and older adults; (2) to empirically test whether the clustering of comorbidity among CKD risk factors link to kidney function as a proof of concept that different characteristics of comorbidity are associated with different state of kidney functioning; and (3) to contextualize the different clustering of CKD risk factors by testing whether childhood SES, controlling for education, adult SES, age, gender, and race, was associated with the heterogeneity of comorbidity of CKD risk factors. Latent class analysis (LCA) was employed to address these research questions. A person-centered analysis such as LCA provides objective and parsimonious solutions regarding the variation in the clustering of risk factors, its impact on kidney function, and prediction by childhood SES. The analysis was divided into three steps. First, we identified the heterogeneity of the comorbidity among risk factors. Second, we examined the association between latent classes of risk factors and kidney function. Third, we tested the evidence whether childhood SES was associated with the heterogeneity of comorbidity among CKD risk factors by utilizing model-based approach LCA.
Step 1: Examination of the Heterogeneity of Comorbidity Among Risk Factors
Selection of the optimally fitting model was based on model fit statistics and selection criteria, parsimony principle, as well as theoretical interpretability. Model fit statistics and selection criteria included the Akaike information criterion (AIC), Bayesian information criterion (BIC), sample-size adjusted BIC (a-BIC), entropy, Bozdogan’s consistent AIC (CAIC), and bootstrapped likelihood ratio test (BLRT). A better fit model is indicated by lower values for the AIC, BIC, and a-BIC. In addition, higher values for entropy indicated higher classification utility. Finally, significant p-values of bootstrapped likelihood ratio test indicated improved model fit compared to models with one fewer class. Two, three, four, five, and six latent classes LCAs were compared to select the best fitting model. Model identification was conducted by using 1,000 sets of random starting values; all models were estimated using PROC LCA on SAS [35]. Two sets of parameters are of most interest from the best fitting model. The first set is the latent class membership probabilities, which indicate the distribution of the classes in the population. The second set is the item-response probabilities, which indicate the probabilities of providing certain responses to observed variables conditional on class membership [35]. These two sets of parameters are used to label and interpret the classes. The analysis was conducted using PROC LCA on SAS [35].
Step 2: Testing the Association Between the Heterogeneity of Comorbidity Among Risk Factors and Kidney Function
In the second step, we examined whether latent classes of CKD risk factors were predictive of eGFR. We used eGFR as both continuous and binary variable (1 = < 60 ml/min/1.73 m2, 0 = ≥ 60 ml/min/1.73 m2). For the analysis with eGFR as a continuous variable, the outcome from the analysis was the expected mean of eGFR for each latent class of CKD risk factors. When predicting eGFR as a binary variable, the outcome of the analysis indicated the probability of having eGFR lower than 60 ml/min/1.73 m2 for each latent class. We utilized LCA with a distal outcome to test this hypothesis using the BCH approach [36]. The LCA with a distal outcome was executed using LCA_Distal_BCH SAS macro [37].
Step 3: Examining the Association Between Childhood SES and the Heterogeneity of Risk Factors
In the final step of the analysis, we examined whether childhood SES was associated with class membership, after controlling for education level, current SES, age, gender, and race. We tested the hypothesis by utilizing model-based approach LCA with covariates [35], in which childhood SES (score) was utilized to predict the probability of belonging to certain latent class of risk factors comorbidity (relative to the reference class), controlling for the covariates. The results were presented as the odds ratios of belonging to a certain class compared to the reference class. The model-based approach LCA was conducted using PROC LCA on SAS [35].
Missing Data
Parameters in PROC LCA are estimated by maximum likelihood using an EM (expectation-maximization) procedure [35]. This procedure handled missing data when identifying the latent class indicators, assuming that data missing at random (MAR) [35].