Multivariate Survival Analysis of Macrovascular Events in Veterans with Type 2 Diabetes

Many challenging diseases in medicine today are multifaceted. Type 2 diabetes affects multiple components within the body. Within each system there are multiple points of failure, and any given treatment may not affect all system components equally. Researchers are often not sure which components are sensitive to change and group these components into composite outcomes, because there is a lack of statistical methodology that can address multiple outcomes that produces inferences that are easy to interpret. In this dissertation, two modeling approaches—a parametric conditional model and a random effects proportional hazards model—are utilized for the analysis of multiple types of events, possibly recurrent, from the Veterans Affairs Diabetes Trial, a trial of type 2 diabetic patients randomized to intensive or standard glycemic control. The primary outcome of the study was time-to-first of any one of nine macrovascular events. Both methods strive to minimize information loss due to discarding data and to control the computational complexity involved in handling multiple repeated events. In the first modeling approach, a parametric conditional model is used to model three distinct types of events: myocardial infarction (MI), congestive heart failure (CHF), and cardiovascular death (CV-death). Only the time-to-first-event of each of the three event types was used in the conditional model. The results did not show a significant treatment effect but did demonstrate a positive dependence of time-to-CV-death on time-to-CHF-event. In the second modeling approach, a three-level hierarchical proportional-hazards model was developed and applied to all eight recorded primary events. Analyzed in the presence of covariates, no treatment effect was found in any of the types of events. One covariate, prior CV-event was found significant in all types of events, baseline insulin was significant in predicting CHF and invasive revascularization (IR). In conclusion, different modeling approaches are used to analyze the VADT data to examine if there might be a treatment effect that went undetected in the initial analysis due to discarding information about recurrent events. However, even when analyzing all available data, no treatment effect was found.