We used latent class analysis (LCA) (7,8), the categorical analog to factor analysis, to identify subgroups by comorbid condition. We also compared their clinical characteristics and 5-year death rates. Institutional review boards at the Division of the Social Sciences at the University of Chicago and National Opinion Research Center approved data collection procedures. All respondents provided informed consent before participation in the study.

We used data from Wave 1 of the National Social Life, Health, and Aging Project (NSHAP), a longitudinal, population-based study of health and social factors of older, community-dwelling Americans (9). Wave 1, conducted between July 2005 and March 2006 in English and Spanish, consisted of 3,005 interviews of adults aged 57 to 85; it oversampled men, black and Hispanic men and women, and men and women aged 75 to 84 at the time of screening. The sampling design also took into account urbanicity, defined as the probability-proportionate-to-size (PPS) selection of US Metropolitan Statistical Areas (MSAs) and non-MSA counties. NSHAP included an in-home interview, a biomeasure collection, and a self-administered postinterview questionnaire. Details on the NSHAP survey are available elsewhere (10,11). The Wave 1 survey had a weighted response rate of 75.5%, and the postinterview questionnaire had a weighted response rate of 84% (12).

We analyzed a subsample of 750 survey participants who self-reported diagnosed diabetes or who had undiagnosed diabetes. The following question was used to determine whether participants had diagnosed diabetes: “Has a medical doctor ever told you that you have any of the following conditions: [a list of conditions includes ‘diabetes or high blood sugar’]?” Participants who had an HbA1c of 6.5% or more and who did not self-report diabetes were considered to have undiagnosed diabetes (13). HbA1c was measured via dried blood spots during the biologic sample collection; details on the blood spot collection are available elsewhere (14). We chose not to determine diabetes diagnosis by examining data on prescribed medications because diabetes medications are also used to treat glucose intolerance.

The survey also collected self-reported data on age, sex, race/ethnicity, marital status, education, and number of visits with health care professionals in the previous year. We dichotomized number of health care visits to fewer than 4 visits in the previous year and 4 or more visits. Prescribed medications, including insulin, were cataloged during the in-home interviews (15). We measured self-rated physical health by using the single-item question, “Would you say your health is excellent, very good, good, fair, or poor?” Self-rated mental health was measured by a similar question. Responses to both questions were dichotomized (excellent/very good/good vs fair/poor). During the in-home interviews, participants also self-reported whether they had difficulty with each of the following activities of daily living for more than 3 months: walking 1 block, walking across a room, dressing, bathing/showering, eating, transferring (getting in/out of bed), and toileting. We collected data on 5-year death rates for the Wave 1 sample during Wave 2 of NSHAP during 2010 and 2011.

For LCA, we used data from all 750 survey participants. We included 14 chronic conditions that are prevalent in the older population (2,16): arthritis, cancer, congestive heart failure, dementia, depression, emphysema, falls (in the previous 12 months), hypertension, incontinence (urinary or fecal), kidney disease, myocardial infarction, obesity, stroke, and thyroid disease. All conditions were self-reported except depression and obesity, and all were assessed during the in-home interview except fall history. Depression was measured through the 11-item Iowa form of the Center for Epidemiological Studies Depression scale (CES-D). A score of 9 points or more on the CES-D was identified as depression (17). Obesity (body mass index [BMI] ≥30 kg/m²] was calculated from interviewer-measured height and weight (15). Fall history was assessed in the postinterview questionnaire.

We created separate variables for each unique combination of comorbid conditions among participants who had reported data for all conditions; 508 participants, 84% of the participants who returned the postinterview questionnaire, had complete data on all comorbid conditions. At most, 508 unique combinations of comorbid conditions were possible, one for each survey participant.

We fit latent class models successively, starting with a 1-class model and then adding another class for each successive model. Using the Akaike Information Criterion (AIC), the Bayesian Information Criterion (BIC), and the sample-size adjusted BIC (ABIC), we determined the optimal number of latent classes; lower values indicated better fit. Although the BIC is the most widely used criterion for assessing LCA model fit, multiple information criteria are often used in combination to select LCA class number (18). We also determined model fit by a χ² goodness-of-fit P value greater than .05. We estimated models using full information maximum likelihood (FIML), which computes a case-wise likelihood function over the observed data by using all available information to estimate model parameters (19). FIML provides less biased and more efficient estimates than pair-wise deletion, case-wise deletion, or similar response-pattern imputation (20). We compared model fit between 1-, 2-, 3-, and 4-class models. We used Mplus version 6 (Muthén & Muthén, Los Angeles, California) to conduct the analyses.

We used χ² analysis and 1-way analysis of variance to describe class differences in sociodemographic characteristics, self-rated health, health care use, clinical characteristics, including the prevalence of multiple comorbid conditions, and 5-year death rates. We also compared classes by frequency and combinations of comorbid conditions. We used STATA version SE10.1 (StataCorp LP, College Station, Texas) to conduct these analyses. All analyses were weighted by using population weights that adjusted for the intentional oversampling of black and Hispanic participants and incorporated a nonresponse adjustment based on age and urbanicity. We adjusted standard errors for sample stratification (sampling strata independently) and clustering (sampling individuals within each of 100 primary sampling units).

The AIC, BIC, and ABIC decreased as the number of classes increased from 1- to 4-class models (Table 1). The AIC, BIC, and ABIC were all lowest with the 4-class model, but the AIC and ABIC only marginally decreased between the 3- and 4-class models, and the 4th class had only 7 members. The P value for the χ² goodness-of-fit statistic was greater than .99 for all 4 models. Thus, we selected a 3-class model to distinguish between population subgroups.

The estimated probability of having obesity, hypertension, or arthritis was at least 45% in all 3 classes (Table 2). Most participants (n = 470, 63%) were in class 1, which was characterized by the lowest probability for nearly all conditions. The estimated probability of having dementia, congestive heart failure, emphysema, kidney disease, or stroke was less than 5% in class 1. Approximately 29% of participants (n = 215) were in class 2. Class 2 had the highest estimated probabilities of cancer, hypertension, incontinence, kidney disease, and obesity. Class 3 (n = 65, 9%) had the highest estimated probabilities of arthritis, congestive heart failure, depression, emphysema, falls, myocardial infarction, stroke, and thyroid disease. The estimated probability of congestive heart failure or myocardial infarction was more than 90% for class 3.

Class 2 had the lowest average HbA1c (Class 1, 7.24%; class 2, 6.86%; class 3, 7.28%) (Table 3). A higher percentage in class 2 and class 3 reported 4 or more visits with health care professionals in the previous year than in class 1. More than half of class 2 (59%) and class 3 (78%) reported fair or poor physical health, compared with 26% of class 1. In addition, class 2 and class 3 had more difficulty than class 1 with all activities of daily living. For example, 54% of class 2 and 66% of class 3 reported difficulty walking 1 block, compared with 24% of class 1. The 5-year death rate of class 3 (33%) was almost twice the rate of class 2 (17%) and 3 times the rate of class 1 (9%).

Class 2 and class 3 had more comorbid conditions than class 1 (mean number of conditions [range]: class 1, 2.6 [0–5]; class 2, 5.5 [3–9]; class 3, 6.7 [3–9]) (Figure). Only class 1 had 0, 1, or 2 comorbid conditions, and only class 2 and class 3 had 6 or more comorbid conditions.

We found 266 combinations of comorbid conditions. Class 1 had 116 combinations, class 2 had 118 combinations, and class 3 had 32 combinations; we found no overlap of combinations between classes. The most frequent combinations of conditions were arthritis, hypertension, incontinence, and obesity (n = 18) and arthritis, hypertension, and obesity (n = 15).