Several limitations must be considered in interpreting our results. First, although studies have hypothesized a protective effect of COX-2 inhibitors against MI (5 - 6), including potentially beneficial effects of celecoxib (6,8), the association between celecoxib use and lower odds of nonfatal MI may have occurred by chance. Second, recall bias could have created our results if controls had better recall of their COX-2 inhibitor use than case-patients. However, analysis of other prescription analgesics did not demonstrate evidence of recall bias, and the comparisons of COX-2 inhibitors with these prescription analgesics are not likely to be biased by differential recall. Third, nonparticipation bias could occur if prevalence of analgesic use differed between participants and nonparticipants. However, if the likelihood of nonparticipation was the same for COX-2 users and for other prescription analgesic users, the comparison between COX-2 use and prescription analgesic use (which yielded similar results) should be unbiased by nonparticipation. This is consistent with our previous substudies within this study that demonstrated the same relative differences in analgesic use between participants and nonparticipants for case-patients as for controls (21). Because case-patients and controls did not differ with respect to the effect of nonparticipation, the estimate of nonparticipation on the nonselective NSAID–MI association had no effect (3). Fourth, despite adjustment for many potential confounders, above and beyond that available in administrative database studies, uncontrolled confounding is always possible in observational studies. Fifth, because we included only nonfatal MIs, a harmful effect of COX-2 inhibitors could be masked if the drugs increased the risk for fatal MI. However, the difference between rofecoxib and naproxen in the VIGOR study (10) was due to a difference in nonfatal MIs, with no difference in fatal events (11). Finally, because of limited power, we could not exclude an up to 1.9-fold increase in risk for nonfatal MI with the use of rofecoxib, nor could we directly measure the effect of 50 mg of rofecoxib or short- or long-term rofecoxib use. However, our results suggest that the difference between celecoxib and rofecoxib is not due to an isolated increase in risk from 50 mg of rofecoxib or different durations of use.