Cox proportional hazards models were used to estimate risk, with censoring at first event, death, or the latest date of adjudicated follow-up through June 30, 2001. Tuna and other fish consumption were correlated (r = 0.37: P<.001), and associations of each with risk were similar to associations of tuna/other fish intake combined. Tuna and other fish consumption were therefore evaluated together (combined correlation with EPA + DHA = 0.55; P<.001).8 Fried fish/fish sandwich consumption was modestly correlated with tuna/other fish intake (r = 0.14; P < .001) and was evaluated separately. Fish intakes were evaluated as categorical (indicator) variables; given that few persons consumed fried fish/fish sandwiches 5 or more times per week (n = 63), these participants were combined with those consuming fried fish/fish sandwiches 1 to 4 times per week. To minimize potential confounding, covariates were selected based on clinical interest, previously published associations with stroke,18 or associations with exposures or outcomes in the current data set. Based on these considerations and the goal of parsimony in covariate selection, 3 final multivariate models are presented: (1) adjusted for cardiovascular risk factors (age, sex, education, diabetes, prevalent coronary heart disease, smoking status, pack-years of smoking, and aspirin use); (2) further adjusted for other risks and lifestyle factors (body mass index, leisure-time physical activity, alcohol use, and total caloric intake); and (3) further adjusted for potential confounders or mediators (systolic blood pressure and low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride, and C-reactive protein levels). Other covariates that did not appreciably alter the relations between fish intake and stroke and were therefore excluded from the final models were race, income, enrollment site, hypertension, frequent falls, exercise intensity, diastolic blood pressure, carotid intimal medial thickness, and atrial fibrillation; use of β-blockers, lipid-lowering medication, fish oil, and estrogen; fasting glucose, insulin, fibrinogen, factor VII, and factor VIII; and estimated intake of total fat, saturated fat, linolenic acid, carbohydrates, protein, fiber, wine, thiamine, vitamin A, and vitamin C. Kaplan-Meier survival methods were also used to evaluate stroke-free survival according to tuna/other fish and fried fish/fish sandwich consumption (each adjusted for the other). Intake categories were entered as ordinal variables in tests for trend as well as for evaluation of differences in baseline characteristics using linear (continuous variables) or logistic (dichotomous variables) regression. Likelihood ratio testing using multiplicative interaction terms was used to assess effect modification by age, sex, education, diabetes, coronary heart disease, hypertension, systolic blood pressure, smoking, and aspirin use. All P values are 2-tailed (α = .05). Analyses were performed using Stata 8.0 (Stata Corp, College Station, Tex).