We used Cox proportional hazards regression models to account for potentialeffects of other risk factors for ARM.37 Tocontrol as finely as possible for confounding by age, calendar time, and anypossible 2-way interactions between these 2 time scales, we stratified theanalysis jointly by age in months at the start of follow-up and calendar yearof the current questionnaire cycle. Multivariate models also adjusted forsmoking, body mass index, energy intake, alcohol intake, fish intake, physicalactivity (metabolic equivalents per week in quintiles in men, hours of vigorousactivity in quintiles in women), history of hypertension and high blood cholesterollevels, postmenopausal hormone use (women), and occupation (men). To adjustfor smoking, pack-years of smoking (the number of years smoked multipliedby the average number of packs of cigarettes per day) was used, since thisbest reflects the cumulative effect of smoking and is more strongly associatedwith ARM than current smoking status.4 Amongthese covariates, pack-years of smoking, body mass index, and postmenopausalhormone use were updated in every 2-year period. Dietary covariates were updatedusing cumulative averaged intake. We used SAS PROC PHREG38 (SASInstitute Inc, Cary, NC) for all analysis, and the Anderson-Gill data structure39 was used to handle time-varying covariates efficiently,with a new data record created for every questionnaire cycle at which a participantwas at risk and covariates set to their values at the time the questionnairewas returned. For all relative risks (RRs), 95% confidence intervals (CIs)were calculated. Tests for trend across categories of intake were conductedby using the median within each category as a continuous variable.40 All P values were 2-sided.