The TRIAD study has been described in detail elsewhere [28
]. In brief, the initial TRIAD cohort consisted of 11927 community-dwelling adults with diabetes ages 18
years and older, and continuously enrolled between July 2000 and August 2001 in one of 10 managed care plans in 7 states. TRIAD data included patient surveys, medical record reviews, health plan administrative claims (inpatient, outpatient and emergency room claims), and National Death Index (NDI) data. In our analyses, we limited our sample to the 8820 TRIAD participants for whom we had medical record data 18
months prior to the baseline survey. In addition, we excluded those whose age at diabetes diagnosis was less than 30
years and were treated with insulin only at the time of the baseline survey since it is likely they had type 1 diabetes. Institutional review boards at each participating site approved the study and all participants provided informed consent.
Risk predictors such as patient age, diabetes duration (years since diagnosis), and smoking status were obtained from survey responses. Hemoglobin A1C, systolic blood pressure, diastolic blood pressure, total cholesterol, HDL, and LDL were obtained from medical records, and only the most recent values within the 18
months prior to the baseline survey were used. Patients were defined as having a CHD history if at least one of the following conditions was documented in their medical records within 3
years prior to the baseline survey: angina, MI, coronary heart disease, coronary artery disease, coronary angioplasty or coronary bypass. We also obtained additional baseline information on diabetes treatment from patient surveys, and determined the use of hypertension medication, statins and co-morbid conditions (measured by the Charlson’s score [29
]) from medical records from 5 out of the original 6 research centers where the data were available to us.
For evaluation of each risk equation, we used the CHD event definition used in the study that provided the equations. To evaluate the UKPDS risk equations, we defined a CHD event as: a fatal or nonfatal MI (ICD-9-CM code of 410.xx administrative data; ICD-10 of I21-I22 NDI data); to evaluate the Framingham risk equations, we defined a CHD event as: angina pectoris, MI, coronary insufficiency, sudden death, or CHD death (ICD-9-CM code of 410.xx, 413, 411.89, 414.8 administrative data; ICD-10 of I20-I22 and I46.1 NDI data). For each CHD event, we calculated the “CHD event time” as the time from the TRIAD baseline survey to the first CHD event. Observations were censored at the end of 2003, date of death from another cause, or the date of the first health plan enrollment gap of more than two months, whichever came first.
UKPDS and Framingham CHD risk equations
We evaluated various versions of UKPDS, Framingham initial and Framingham secondary CHD risk equations (Table ). Specifically, two UKPDS risk equations were evaluated: The first equation predicts the risk of an initial CHD event for a patient with newly diagnosed diabetes and we referred to it as the incident UKPDS risk equation. The second equation takes into account diabetes duration when predicting the risk of an initial CHD event and we thus referred to it as the duration UKPDS risk equation. Framingham risk equations are gender-specific and were thus evaluated separately for men and women. We evaluated the risk equations for predicting an initial CHD event using total cholesterol categories and referred to it as Framingham-initial. We also evaluated the performance of the equations for predicting a subsequent CHD event and referred to it as the Framingham-secondary.
UKPDS and framingham risk equations
Risk score calculation and statistical methods
For each eligible participant, we calculated the absolute risk of a CHD event using each equation. Because the racial/ethnic composition of the TRIAD cohort differed from that in the UKPDS cohort, we used the “Afro-Caribbean” risk adjustment for African American patients and the “Caucasian or Asian-Indian” calculation adjustment for the remaining participants. Framingham risk equations were not adjusted for race/ethnicity. Because the Framingham-initial equations were published with the 10-year baseline survival rates, we obtained the 1–5
year baseline survival rates directly from the Framingham investigators.
We evaluated the risk equations for 1) how well they separate individuals who develop a CHD event from those who do not (discrimination) and 2) how close predicted risks are to observed risks [6
] (calibration, or goodness-of-fit (GOF)). When we examined the performance of the UKPDS and Framingham-initial CHD equations, we only included patients without a CHD history; when we examined the performance of the Framingham-secondary equations, we only included patients with a CHD history.
Discrimination was evaluated using the Harrell’s c-index for censored data (R package HMISC
available on CRAN at http://cran.r-project.org
), a statistic similar to the area under a receiver operating characteristic curve [32
]. In general, a c-index greater than 0.7 indicates good discrimination while a value of 0.5 indicates discrimination equivalent to chance. Intermediate values indicate limited discriminating utility. Calibration plots were generated and Hosmer-Lemeshow-type chi-square statistics [6
] were calculated to compare differences between predicted and observed risks based on deciles of risk scores. We conservatively defined lack of calibration as chi-square values greater than 23.2 (the 99th
percentile of chi-square distribution with 10 degrees of freedom). We also recalibrated the UKPDS and Framingham risk equations by replacing the average values of predictors and event rates in the original populations by those in the TRIAD population. Specifically, we used the method of D’Agostino et al.[6
] to recalibrate the Framingham-initial equations and the method of van Houwelingen [34
] to recalibrate the UKPDS and Framingham-secondary equations because the latter were parametric models.
To investigate the difference between study populations with regard to the effect of risk predictors, we fitted each of these equations on TRIAD data and compared the estimates of relative risk (hazard ratio) using the method described in D’Agostino et al. [6
]. Specifically, we fitted the Cox regression models and used the same CHD event definition as well as the risk predictors in the original equations. For simplicity, the models using the TRIAD data were all referred to as the TRIAD models. Regression coefficients, hazard ratio (HR) estimates, Harrell’s c-index and GOF statistics [35
] were calculated.
Missing data ranged from 1.3% (smoking) to 20.7% (HDL), and was handled in the data analysis using multiple imputation. Imputations were generated using a sequential regression imputation method via the software package IVEware, and results were combined using Rubin’s rule implemented in SAS v9.2 MIANALYZE procedure [36