For continuous and categorical variables, we used t tests and chi-square tests, respectively, to determine the statistical significance of any difference in the distribution of baseline variables between treatment groups. Cumulative incidence curves of type 2 diabetes by treatment group were constructed by comparing Nelson–Aalen cumulative hazard function estimates that were calculated at different time points of the trial and by using the 2-sided log-rank test (20). In unadjusted analyses, incidence data were statistically analyzed by calculating relative risks as the ratios of the incidence density for the treatment groups, with corresponding 95% CIs. P values were derived from log-rank tests. In adjusted analyses, hazard ratios and 95% CIs were calculated by using the Cox proportional hazard model, which allowed adjustment for age, BMI (continuous variable), sex, and smoking status at baseline as covariates. We decided a priori to adjust for these diabetes risk factors regardless of whether they differed between treatment groups. Tests of proportional hazards assumptions were based on Schoenfeld residuals (21). The tests showed that the proportional hazard assumption was not violated for any variable used in the model. Modification of association by median age (65 years), sex, smoking status, and BMI tertiles at randomization was tested by using the Mantel–Haenszel test for heterogeneity in the unadjusted models. The statistical significance of the interaction between each baseline characteristic and treatment group, adjusted for other important baseline variables, was tested in Cox proportional hazards models that included this interaction and the corresponding main effect terms. None of these interactions achieved a conventional level of statistical significance (P < 0.05). Similarly, a test of interaction between selenium treatment and vitamin supplements was not significant and was therefore not included in the multivariate analysis.