To control for differences in covariate distributions over time we produced adjusted predictive margins for prevalence and means using multinomial logistic regression (Hosmer and Lemeshow, 2000) to predict drinking status and binge drinking categories, and multiple linear regression to predict average volume including its quantity and frequency components (Graubard and Korn, 1999). Covariates were race/ethnicity (non-Hispanic white, non-Hispanic black, Hispanic, other), census region (Northeast, Midwest, South, West), education (≤high school, >high school), marital status (married or cohabiting, not married or cohabiting), and smoking status (never, former, current). Unknown or multiple race (0.2%) was grouped with the “other” category; unknown education (1.2%) was grouped with the “≤high school” category; unknown marital status (0.4%) was grouped with the “not married or cohabiting” category; and unknown smoking status (0.9%) was grouped with the “never” category. We decided, a priori, to present results of unadjusted analyses as our primary findings. Estimates of the prevalence of alcohol consumption in the US population are typically reported as unadjusted estimates. Analyses adjusted for demographics and lifestyle covariates were also performed. Adjusted results are noted as appropriate and presented as supplemental materials.
