All multivariate models were adjusted for potential risk factors for mortality, including age (continuous), total energy (continuous), smoking (never smokers, former smokers of ≤1 pack per day, former smokers of >1 pack per day, current smokers of ≤1 pack per day, and current smokers of >1 pack per day), education (less than high school, high school, some college, and completion of college and postgraduate college), BMI (calculated as weight in kilograms divided by height in meters squared) (<18.5, 18.5 to <25.0, 25.0 to <30.0, 30.0 to <35.0, 35.0 to <40.0, and ≥40.0), physical activity (never, rarely, 1-3 times per month, 1-2 times per week, 3-4 times per week, and ≥5 times per week), race (white, black, and other), marital status (married and not currently married), and in women, menopausal hormone therapy (never, past, and current). For finer control of smoking status, we used a multilevel variable that integrated combinations of smoking status (never, former, and current), time since quitting for former smokers (≥10 years, 5-9 years, 1-4 years, and within the last year), and smoking dose for both former and current smokers (1-10, 11-20, 21-30, 31-40, 41-60, and ≥61 cigarettes per day). Additional adjustment for other dietary variables (quintiles of polyunsaturated fatty acids, eggs, and potatoes) did not appreciably change the risk estimates, and these nutritional factors were not considered further. In a subanalysis we excluded smokers to evaluate the potential for residual confounding by smoking. Missing values for adjusting covariates were included as dummy variables in the models. Using the alternative Horvitz-Thompson method19 for missing data did not change the results appreciably.