Person-years were calculated for each participant, ranging from the date of randomization to the date of confirmed cancer diagnosis, death, or March 31, 2004, whichever occurred first. Cox proportional hazards regression models were used to calculate the relative risks (RRs) and 95% confidence intervals (CIs).45 We estimated the RRs according to quintiles of caffeine consumption and categories of caffeinated beverages and foods with adjustment for age and randomized treatment assignment (aspirin vs placebo, and vitamin E vs placebo). We also conducted a multivariate analysis that additionally adjusted for known or potential risk factors for breast cancer at baseline, including alcohol consumption (0, >0 to <10, ≥10 to <15, ≥15 to <30, or ≥30 g/d), body mass index (calculated as weight in kilograms divided by height in meters squared) (<23, ≥23 to <25, ≥25 to <27, ≥27 to <30, or ≥30), family history of breast cancer in a first-degree relative (yes or no), history of hysterectomy (yes or no), bilateral oophorectomy (yes or no), smoking status (never, past, or current), history of benign breast disease (yes or no), age at menarche (≤11, 12, 13, 14, or ≥15 years), parity (0, 1-2, 3-4, 5, or ≥6), age at first birth (≤24, 25-29, or ≥30 years), physical activity (kilocalories per week, in quartiles), total energy intake (kilocalories per day, in quintiles), multivitamin use (never, past, or current), age at menopause (<45, 45-49, 50-54, 55-59, or ≥60 years), menopausal status (premenopausal, postmenopausal, or uncertain/unknown), and postmenopausal hormone use (never, past, current <5 years, or current ≥5 years). We also conducted analyses that excluded incident cases of breast cancer diagnosed in the first 2 years of follow-up, with additional adjustment for mammography screening that was determined on the 12-month questionnaire, or that was stratified by menopausal status (premenopausal or postmenopausal), history of benign breast disease (yes or no), body mass index (<25 or ≥25), and postmenopausal hormone use (never or current). Tests for multiplicative interaction were performed by log likelihood ratio tests comparing the models with or without interaction terms.