Proportional hazards regression was used to examine associations between patient characteristics and the occurrence of a stroke after MI and to examine the association between stroke and death after MI while adjusting for age, sex, Killip class, aspirin use, and warfarin use; stroke was analyzed as a time-dependent covariate. The correlation of the scaled Schoenfeld residuals with time was used to test the proportional hazards assumption. Except for family history of coronary artery disease (7.2%), anterior MI (15.8%), ST-segment elevation MI (12.8%), and presence of Q waves (21%), missing values did not exceed 5%. Because ST-segment elevation and presence of Q waves are both surrogates for MI severity, we elected to use peak creatine kinase ratio, defined as the ratio of the maximum creatine kinase value to the upper limit of normal, as the measure of MI severity. The peak creatine kinase ratio was missing in less than 5% of the observations. Fewer than 2% of persons had 1 or more missing variables for the predictors in the final multivariable models. Because the proportional hazards assumption was not met, strokes were divided into early strokes (occurring ≤ 30 days after MI) and late strokes (occurring > 30 days after MI). For late strokes, the proportional hazards assumption was met for all predictor variables (P > 0.05 for all variables). Results are reported as hazard ratios with 95% CIs. Analyses were done using the statistical software package SAS, version 8.2 (SAS Institute Inc., Cary, North Carolina).