Strengths of our study include its population-based, prospective design with long-term follow-up and careful ascertainment of aspirin use, AMD subtypes, and major confounders. Major limitations include, first, the possibility of residual confounding, such as by indication, whereby CVD rather than aspirin use contributes to the increased AMD risk. We note that, after adjustment for additional CVD risk factors (BMI, blood pressure, blood total cholesterol level, diabetes mellitus, fish consumption, inflammatory markers), our findings became marginally nonsignificant (OR, 2.05; 95% CI, 0.96-4.40; P = .06), likely because of reduced power from additional covariates in the model. To address this, we performed propensity score analysis. This incorporates the same variables as an indicator variable represented by the propensity score, and adjusting for this score in the model to replace several covariables resulted in slightly stronger findings, supporting a persistent association. Furthermore, other medications associated with painful conditions and CVD (acetaminophen and β-blockers) were not associated with incident wet AMD, and adjusting for them in the multivariate model did not change our findings. Second, aspirin use was determined only at baseline and was not updated. If participants started taking aspirin after the baseline assessment, they would have been classified as nonregular users. This would bias our findings toward the null, that is, reduce the magnitude of any associations of aspirin use with incident AMD, if any association existed. Including aspirin users with less than 10 years' duration in our analyses could dilute our findings to a large extent, since our a priori hypothesis was that long-term (>10 years) aspirin use is associated with incident AMD. Our findings appear to bear this out, as shown in Figure 3, where the association became evident only after 10 to 15 years of follow-up. Cessation of aspirin use during follow-up would tend to bias toward the null as well, but since most persons with regular aspirin use continue it for life, we believe that this is unlikely to cause a large bias. Third, we have no data on the indications for aspirin use in our cohort and are unable to adjust for all possible indications (eg, rheumatologic conditions). Nonetheless, acetaminophen was not significantly associated with neovascular AMD, with a multivariate-adjusted OR of 1.04 (95% CI, 0.61-1.79). Fourth, our ascertainment of aspirin use was through questionnaire (with confirmation using medication bottles brought to examinations by participants), the assessment of geographic atrophy did not use fundus autofluorescence, the study design was observational, and the participants were mostly a homogeneous white population. Our results should thus be viewed as potentially influenced by ascertainment bias (ie, underreporting of regular aspirin use and underdetection of geographic atrophy) and nonrandom allocation of aspirin use and may not be applicable to nonwhite populations with different risk profiles for CVD and AMD. However, we anticipate the ascertainment bias from aspirin use to be nondifferential and occur equally in the groups who developed and did not develop incident AMD and thus to dilute the observed association toward the null. Our results may be influenced by the low proportion of participants who attended 15-year follow-up examinations. Participants with CVD were less likely to be followed up because of higher mortality and were more likely be regular aspirin users and develop incident AMD.7 We could have missed persons who were using aspirin and would have gone on to develop incident AMD had they lived longer. This would, however, tend to attenuate the association and bias our findings toward the null.