Both BMI and waist circumference were omitted from multivariable analyses as they are likely to lie on the causal pathway between behavior change and CVD risk. To ascertain whether any behavior change variables mediate their effects through BMI or waist circumference, models were also run with and without these covariates, and the percentage change in RR associated with CVD risk for each health behavior change was assessed. Competing-risks regression estimated the risk of a composite cardiovascular end point in the presence of the competing risk of non- CVD death, while adjusting for potential confounders. The population-attributable fraction (30) estimated the proportion of CVD events that could be prevented if everyone adopted three or four health behaviors in the year after diagnosis, adjusting for all known confounders (Model 3). Because the results for all analyses were similar by trial arm, data were pooled and results presented for the whole cohort, adjusting for trial arm (study group). The relation between missing data and other variables was investigated using t tests or x tests, where appropriate. Sensitivity analyses were carried out to test the robustness of estimates: 1) multiple imputation of missing health behavior (ordered categorical variable) and self-reported drug prescription (binary variable) were carried out by conditioning via multinomial logistic regression or via logistic regression respectively, on the observed predictor variables to generate five imputed data sets (31); sensitivity analyses 2) omitting revascularization from the composite CVD end point, 3) omitting abstainers, and 4) including the ratio of polyunsaturated to saturated fats rather than the percentage of energy from total fat intake were also run. Data were analyzed using STATA version 13.1 (Stata, College Station, TX).