Kaplan-Meier methods were used to evaluate CHF-free survival. Risk was estimated using Cox proportional hazards (time at risk until first event, death, or latest follow-up through June 30, 2001), with covariates based on clinical interest, associations with incident CHF (4), or associations with exposures/outcomes in the present analysis. The final model included age; gender; race; enrollment site; education; diabetes; body mass index; prevalent coronary heart disease; prevalent stroke/transient ischemic attack; and intakes of tuna/other fish, fried fish, and total calories. A second model was further adjusted for other behavioral and lifestyle factors including smoking; leisure-time physical activity; and intakes of saturated fat, fruits, vegetables, and alcohol. We also evaluated for potential mediation or confounding by levels of systolic blood pressure, diastolic blood pressure, baseline LV systolic function, low-density lipoprotein, high-density lipoprotein, triglycerides, and C-reactive protein. For parsimony in model construction, we excluded from the final models other covariates that did not materially alter relations between fish consumption and CHF risk, including annual income; treated hypertension; exercise intensity; forced expiratory volume in 1 s; carotid intimal medial thickness; atrial fibrillation; M-mode echocardiography-estimated LV mass; use of aspirin, beta-blockers, lipid-lowering medication, fish oil, and estrogen; serum fasting glucose, insulin, creatinine, and fibrinogen; and estimated intakes of total fat, carbohydrates, protein, linolenic acid, fiber, beef/pork, wine, thiamine, and vitamin C. Missing covariate values (typically <1%) were imputed using age, race, gender, diabetes, and prevalent cardiovascular disease; analyses using population medians or excluding missing data were not appreciably different. Tests for trend were evaluated with intake categories entered as ordinal variables. Potential effect modification was assessed using stratified analyses and likelihood-ratio testing (exposure times covariate term). Analyses were performed using Stata 8.0 (D.M.) (Stata Corp., College Station, Texas), with two-tailed alpha = 0.05.