The TRIAD methodology has been described in detail elsewhere.6
Six research centers collaborated with ten managed care health plans (HPs) and 68 provider groups that served approximately 180,000 geographically and ethnically diverse patients with diabetes. Institutional review boards at each participating center approved the study. All participants provided informed consent.
TRIAD enrolled 11,927 patients between July 2000 and August 2001. All were at least 18 years old, not pregnant, community-dwelling, English or Spanish speaking, and continuously enrolled in the HP for at least 18 months prior to the baseline patient survey. Medical record reviews were performed at baseline, HP administrative data were collected for 1999 through 2003, and the National Death Index (NDI) was searched for deaths occurring through 2003. We analyzed data for patients who had type 2 diabetes (excluding those with age at diagnosis under 30 years and treatment with insulin only) and with complete data for the variables investigated (N=7,439).
Across TRIAD’s ten HPs, 1,815 patients (24%) filled at least one prescription for a TZD, 773 (10%) for only rosiglitazone, 711 (10%) for only pioglitazone, and 331 (4%) for more than one TZD. In the seven HPs using both TZDs, 1,159 patients (33%) filled at least one prescription for a TZD, 564 (16%) for only rosiglitazone, 334 (10%) for only pioglitazone, and 261 (7%) for more than one TZD (). Patients filling prescriptions for more than one TZD were excluded.
We used HP administrative data to determine TZD exposure and verify drug benefits. Patient surveys and medical record reviews were used to determine if patients were treated with insulin at baseline. Subsequent initiation of insulin was determined from analysis of administrative data. We ascertained nonfatal acute myocardial infarction, stroke, or percutaneous or surgical CV intervention from HP administrative data. We used the following International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis codes (with any 4th
digit): 410 (acute myocardial infarction), 431 (intracerebral hemorrhage), 433 (occlusion and stenosis of precerebral arteries), and 434 (occlusion of cerebral arteries). Coronary revascularization procedures included operations on vessels of the heart (ICD-9-CM procedure code 36 with any 3rd
digit), coronary artery repair procedures (CPT codes 33500-33572), intracoronary stents, coronary balloon angioplasty or atherectomy (CPT codes 92980-92984, 92995, 92996). Deaths and cause of death were ascertained from NDI.7
We performed time-to-event analyses. Our study window started with the first TZD prescription and ended with an event or censoring. Event dates were the first occurrence of a CV event or procedure after the first TZD prescription was filled. Patients without CV events or procedures were censored at whichever occurred first: the date the last TZD prescription was filled plus the days supply dispensed plus 90 days (to account for any persistent biological effects of the TZD); the date the person disenrolled from the HP; the last date of service recorded in the administrative data; or TRIAD’s administrative data cut-off date. The end of the study window was re-evaluated for each outcome. Because we were concerned that the use of a random effect in the model might not adequately adjust for differences in patients across HPs, we conducted analyses after excluding the three HPs that appeared to have only one TZD on formulary (as evidenced by prescriptions filled for only one TZD). For patients in these three plans, exposure to a specific TZD could not be distinguished from other unmeasured characteristics associated with membership in that HP.
We compared groups using 2-tailed t-tests for continuous variables and χ2 tests for categorical variables. We tested the assumption of proportional hazards with graphical display and examination of the correlations between the ranked failure time variable and the Schoenfeld residuals of the independent variables. We used Cox proportional hazard multivariate models adjusted for baseline age, sex, race/ethnicity, income, history of diabetic nephropathy, history of CVD (transient ischemic attack, cerebrovascular accident, angina pectoris, myocardial infarction, coronary heart disease, congestive heart failure, or peripheral vascular disease), insulin use, and HP. Missing values for age, sex, race/ethnicity, income, and smoking were relatively infrequent (< 15% in all cases) and were imputed using single imputation with the transcan function in S-PLUS (edition 6.1; Insightful, Seattle, WA). Patients’ missing values for other variables used in this study were excluded. Statistical analyses were performed using SAS (version 9.1.3, SAS Institute, Cary, NC).