The multivariate models included terms for age (5-year categories), cigarette smoking (never, past, and 1-14, 15-24, and ≥25 cigarettes per day), level of physical activity (metabolic equivalents per week in 5 groups), BMI (<20.0, 20.0-21.9, 22.0-23.9, 24.0-25.9, 26.0-27.9, 28.0-29.9, 30.0-31.9, 32.0-33.9, 34.0-35.9, and ≥36.0), family history of diabetes mellitus (yes or no), current use of an oral contraceptive (yes or no), history of hypertension (yes or no), and history of high cholesterol (yes or no). We retained BMI as a covariate, although it may be a confounder and a causal intermediate, to provide maximal control of possible confounding. In additional analyses adjusting for dietary factors, trans-fatty acid, glycemic load, polyunsaturated fat, and total fiber were included as continuous variables. We conducted tests of linear trend across increasing categories of alcohol consumption by assigning the median for the alcohol categories and treating the categories as a continuous variable. To assess departure from linearity, we included linear and quadratic terms (the median and the value squared) in the model. To examine whether the relation of alcohol intake to the risk of diabetes mellitus was modified by important covariates, we conducted analyses stratified by age at enrollment (<35 and ≥35 years), categories of BMI (<30.0 or ≥30.0), cigarette smoking (never, ex-smoker, and current), family history of diabetes mellitus (no or yes), and physical activity (metabolic equivalents per week, <10.5 or ≥10.5). The values for smoking and BMI were updated biennially. Because we had less power in our subgroup analyses, women with an alcohol intake of 30.0 g/d or more were excluded and teetotalers and ex-drinkers were combined. We excluded the few women with an intake of 30.0 g/d or more rather than combining them with the 15.0- to 29.9-g/d group because there was some evidence of an upturn in risk in this small group of women. Tests for interaction were conducted to evaluate whether the trends differed significantly by levels of risk factors.