Risk Factors for Mortality Among Patients With Diabetes: The Translating Research Into Action for Diabetes (TRIAD) Study

TRIAD has been described in detail elsewhere (20). In brief, six centers collaborate with 10 managed health care plans and 68 provider groups that serve ∼180,000 individuals with diabetes. Patients ≥18 years of age were sampled. Institutional review boards at each participating site approved the study. All participants provided informed consent.

A baseline survey was administered to all TRIAD participants either by computer-assisted telephone interview or in writing by mail. Medical records and health plan administrative data for each participant were also reviewed. TRIAD decedents were identified from electronic searches of the National Death Index (NDI) Plus (21). Deaths were verified by matching name, sex, date of birth, and social security number (available for 52% of participants) of the decedent with data supplied by the NDI. The sensitivity of NDI ranges from 87 to 98% (22). Different combinations of identifiers excluding social security numbers correctly identify 83 to 92% of dead persons (i.e., they classify individuals known to have died as deceased) and 92 to 99% of living persons, underscoring the accuracy of NDI Plus (23).

Vital status was determined for all TRIAD participants (n = 11,927) as of 1 January 2004. All information regarding ICD-10 codes for the underlying and contributing causes of death was derived from the NDI file. For our analyses, we included all TRIAD participants who had complete survey and medical record review data for all possible exposure variables (n = 8,733), 791 (9%) of whom were identified as having died before 1 January 2004. This excluded 27% of the original sample, 98% of whom did not consent to review of their medical records. The demographic characteristics (age, sex, race, income, education, smoking, and duration of diabetes) of the included population were similar to those of the entire cohort. Missing values for age, sex, race and ethnicity, education, income, duration of diabetes, BMI, and smoking status from the patient survey were imputed using single imputation with the transcan function in S-PLUS (edition 6.1; Insightful, Seattle, WA). Variables that were imputed had <15% missing data.

We used National Center for Health Statistics definitions to classify cardiovascular versus noncardiovascular causes of death and the American Heart Association definitions to subclassify cardiovascular causes of death according to the underlying cause (1,24). The categories were as follows: cardiac (ICD-10 code I00–09, I11, I13, I20–25, I27, and I30–52), cancer (C00–99), diabetes (E10–14), cerebrovascular disease (I60–69), respiratory (J00–09 and J19–99), all other cardiovascular disease (I10, I12, I14–19, I28–29, I53–59, and I70–99), infections (A00–99 and B00–99), accidents/suicides/assaults (V00–99, W00–99, X00–99, and Y00–99), renal failure (N17–19), influenza/pneumonia (J10–18), and others (all other ICD-10 codes).

We investigated all-cause mortality, cardiovascular mortality (ICD-10 codes I00-I99 for underlying cause of death), and noncardiovascular mortality (all other ICD-10 codes for underlying cause of death). Covariates included age; sex; race; education; income; duration and treatment of diabetes; BMI; current smoking; history of hypertension, dyslipidemia, macrovascular disease, retinopathy, nephropathy, and peripheral neuropathy, measured at baseline; and Charlson index. Macrovascular disease included (from medical record review) history of transient ischemic attack, cerebrovascular accident, angina, myocardial infarction, congestive heart failure, other coronary heart disease or coronary artery disease, or peripheral vascular disease. The Charlson index, an extensively studied, accurate, and valid measure that weights various comorbid conditions by the strength of their associations with mortality, was used to quantify the comorbidity burden (25,26).

Bivariate analyses were performed for all-cause mortality and separately for cardiovascular and noncardiovascular mortality. We summarized continuous variables using means and tested differences between means using t tests. We summarized categorical variables using frequencies and tested their differences by χ². Unadjusted hazard rate ratios (HR) were constructed using each variable singularly in a Cox proportional hazards model predicting all-cause mortality, cardiovascular mortality, or noncardiovascular mortality.

Before constructing the models, we assessed Spearman's correlations between all possible exposure variables. Although education and income were correlated (r = 0.52), we included both in the multivariable models because they may have different predictive properties when used to model health-related outcomes (27). The Charlson index was correlated with history of macrovascular disease (r = 0.49), retinopathy (r = 0.35), nephropathy (r = 0.40), and peripheral neuropathy (r = 0.37), but because the Charlson index is constructed from a weighted examination of comorbid conditions, this was not unexpected (26). We decided to keep these four variables as well at the Charlson index in the model because the index represents more than just those diseases and provides a valid measure of comorbidity (26).

To simultaneously adjust for all possible covariates predicting mortality, we constructed multivariable Cox proportional hazards models. For these analyses, all-cause mortality, cardiovascular mortality defined by underlying cause of death, and noncardiovascular mortality were modeled separately, with each compared with the total alive population. We included the variables that were significant in bivariate tests of association between each predictor and the outcomes of interest. We also included smoking, which was found in previous studies to predict mortality in persons with diabetes (2). We used Cox proportional hazards models with a dummy variable for each health plan and provider group combination to account for the clustered study design (health plan, provider group, and participant levels) and the correlation among participant characteristics within health plans and provider groups. All analyses were performed using SAS (version 9.1; Research Triangle Institute, Research Triangle Park, NC).

Of the 8,733 individuals included in our analyses, 791 (9%) died before 1 January 2004. The average length of follow-up was 3.7 years. A total of 54% of decedents were men, and the distribution of age (in years) at death was 25–44 (1%), 45–64 (27%), 65–84 (63%), and ≥85 (8%). Of decedents, 12% were Hispanic, 19% black, 51% non-Hispanic white, 10% Asian/Pacific Islander, and 8% of other race/ethnicity (Table 1). The underlying causes of death were listed as cardiac (34%), cancer (20%), diabetes (12%), other (11%), cerebrovascular disease (6%), respiratory (5%), infections (4%), all other cardiovascular disease (3%), accidents/suicides/assaults (2%), renal failure (2%), and influenza/pneumonia (1%). Of the 791 deaths, 336 (42%) had a cardiovascular cause listed as the underlying cause of death, and 455 (58%) had a noncardiovascular cause listed as the underlying cause of death; 534 deaths (68% of the total) had any cardiovascular cause listed in either part I or part II of the death certificate (ICD-10 codes I00–99).

The unadjusted characteristics of the entire TRIAD cohort with diabetes and those who died are shown in Table 2. In analyses of associations with all-cause, cardiovascular, and noncardiovascular mortality, we found that 13 variables predicted all three outcomes: older age, male sex, non-Hispanic white race (vs. Hispanic or Asian/Pacific Islander), lower education, lower income, greater duration of diabetes, lower BMI, hypertension, macrovascular disease, retinopathy, nephropathy, peripheral neuropathy, and a higher Charlson score (Table 2). For all-cause mortality, we also found associations with other race and treatment with insulin (Table 2). For cardiovascular mortality, we also found associations with treatment with insulin. Finally, for noncardiovascular mortality, we also found an association with other race (Table 2).

In fully adjusted multivariable analyses that examined risk for all-cause, cardiovascular, and noncardiovascular mortality, we found that older age, male sex, lower income, lower BMI, smoking, and nephropathy predicted all three outcomes (Table 2). For all-cause mortality, non-Hispanic white race (vs. Hispanic), longer duration of diabetes, macrovascular disease, and higher Charlson score also predicted mortality (Table 2). For cardiovascular mortality, longer duration of diabetes and macrovascular disease also predicted mortality. Noncardiovascular mortality was also associated with non-Hispanic white race (vs. black race or other) and higher Charles score (Table 2).

Our findings that older age, male sex, smoking, nephropathy, and cardiovascular disease predicted all-cause mortality are consistent with previous studies. While obesity is highly predictive of diabetes, we, like others, found that being underweight is a greater predictor of mortality among patients with diabetes (28,29). Despite an earlier finding of worse risk factor control (30), we did find that the Hispanic race was associated with lower all-cause mortality. This “Hispanic paradox” has been shown in other mortality studies (31). While both lower income and poorer education were associated with all-cause, cardiovascular, and noncardiovascular mortality in unadjusted models, only lower income remained a significant predictor in models that included both. We assume a causal pathway where income mediates the relationship between education and mortality, and thus it was expected that adjustment for income attenuated the education-mortality relationship. Income has not previously been shown to predict mortality in people with diabetes after adjustment for biological risk factors. Our results suggest that even in an insured population with good access to health care, socioeconomic circumstances remain an important predictor of health.

Because cardiovascular disease is a major contributor to all-cause mortality in diabetes, one would expect the biological predictors of cardiovascular disease, such as hypertension, dyslipidemia, and current smoking, to be associated with both all-cause and cardiovascular mortality. Indeed, previous studies have shown age, sex, duration of diabetes, treatment with insulin, high blood pressure, and smoking to be associated with cardiovascular mortality (7,8,14,32). In contrast, we did not find treatment with insulin to be predictive of cardiovascular mortality.

Unlike earlier studies, we did not find significant associations between all-cause mortality and education, treatment with insulin, or history of hypertension (7,8,14,32). A likely explanation for the absence of some previously reported predictors of cardiovascular mortality is our inclusion of additional variables in the models. For example, we included history of macrovascular disease as well as hypertension, dyslipidemia, and smoking. This has the effect of attenuating to nonsignificance the more remote factors. In addition, controlling for the Charlson index, which includes multiple comorbidities, may have attenuated the associations with some individual risk factors. The absence of some variables in these predictive models should not be interpreted to mean they are not causally important, given that our goal was to identify key predictors of mortality rather than to model causality. When we left macrovascular disease out of the model, hypertension became a significant predictor. We also observed joint confounding of smoking by all other variables in the multivariate model, which may have attenuated to nonsignificance some other causally important variables.

In our population, using only the underlying cause of death, more than one-half (58%) of the deaths were not related to cardiovascular disease. Noncardiovascular mortality has not often been studied in individuals with diabetes. Studies have reported that a longer duration of diabetes and treatment with oral hypoglycemic medications, as opposed to insulin, were associated with lower noncardiovascular mortality (7,32). We, too, found that treatment with oral antidiabetic agents with or without insulin was associated with lower noncardiovascular mortality. We hypothesized that risk factors for cardiovascular disease would not be associated with noncardiovascular mortality, and this was generally the case. As expected with competing risks for mortality, history of macrovascular disease was not significantly associated with noncardiovascular mortality.

The major limiatation of this study was its examination of a select population. Because our population consisted of those with physician-diagnosed diabetes for at least 18 months before the survey, our sample did not include individuals with new-onset diabetes. The study sites were not representative of all managed care populations and may not be generalizable to settings outside of managed care. Furthermore, the TRIAD population had good access to and quality of care, which may have attenuated the impact of biological risk factors. This high quality of care and the focus on prevention of cardiovascular disease may have accounted for the lower rate of cardiovascular mortality in the TRIAD population compared with other U.S. population samples (2).

In conclusion, we have shown the importance of demographic risk factors such as age, sex, and income on short-term, all-cause, cardiovascular, and noncardiovascular disease mortality. We have also shown that even in an insured population where diabetes and cardiovascular risk factors are well controlled (30), socioeconomic circumstance is an important independent predictor of mortality.

This study was jointly funded by Program Announcement no. 04005 from the Centers for Disease Control and Prevention (CDC) (Division of Diabetes Translation) and the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the CDC and the NIDDK. Significant contributions to this study were made by members of the TRIAD Study Group.

We acknowledge the participation of our health plan partners.