All statistical analyses were performed by using SAS software, version 9.1 (SAS Institute, Inc., Cary, North Carolina). Descriptive statistics (percentage, mean, and SD) were calculated for each variable. Chi-square tests were used to evaluate the association between sociodemographic characteristics and patient understanding of primary label instructions of 5 prescription medications and attendance to the auxiliary labels. In multivariate analysis, the 5 binary repeated responses of understanding per patient were modeled by using a generalized linear model with a complementary log-log link function. A generalized estimating equation approach was used to adjust model coefficients and standard errors for within-patient correlation by using PROC GENMOD (SAS Institute). Wald 95% CIs were calculated for adjusted relative risk ratios by using the robust estimate of the standard error as detailed by Liang and Zeger (25). The final multivariate model included the potential confounding variables: age, sex, race (white vs. African American), education, and number of medications currently taken daily. Although education is associated with literacy, it was examined separately but included in the final model to present conservative estimates of the effect of literacy on rates of understanding. This issue has previously been reviewed by Wolf and colleagues (26) and the same method was used in our study. Site was also entered into the model to adjust for any potential differences across study locations. In multivariate analyses, patient literacy was classified as low (sixth grade and below), marginal (seventh to eight grade), or adequate (ninth grade and higher). For the substudy analyses, chi-square tests were used to evaluate the association between sociodemographic characteristics and correct demonstration of the specified medication instructions. A multiple logistic regression model was used to examine the relationship between literacy and comprehension of the medication labels while controlling for the previously mentioned confounding variables and study site. Model fit was assessed by using the c-statistic from the receiver-operating characteristic curves and the Hosmer–Lemeshow goodness-of-fit chi-square test.