Cox proportional hazards regression calculated the rate of primary composite CVD events for categories of health behavior change and the health behav ior change score. Age was used as the underlying time scale in all models (29), with person-time for each participant calculated from age at study entry (baseline) to age at death or the censor date (31 December 2009), whichever came first. Clustering of individuals within GP was accounted for in all analyses by using cluster-correlated robust estimates of variance to obtain variance corrected incidence rates and rate ratios (RRs). Sex, age at study entry, baseline level of relevant health behavior, and study group were considered a priori confounders, and were included in all models. Model 1 examined whether any of the six health behavior change variables were independently associated with 5-year CVD events. In Model 2, stepwise forward regression was used to identify the health behavior changes that were most strongly associated with the composite primary CVD outcome, additionally adjusted for social class and occupation. Only health behaviors that improved model fit, determined via like lihood ratio testing, were included. Model 3 further adjusted for self-reported antihypertensive, glucose- and lipid-lowering (cardioprotective) medications at 1 year. A similar analytic approach was used to investigate the association between the health behavior change score and CVD risk.