We investigated the effects of MLDA exposure on self-reported age of onset of drinking in two types of model. First, we used ordinary logistic models (as described above) to test the association between MLDA exposure and age of drinking onset before age 16. We then extended this analysis using multinomial logistic regression to test the association between MLDA exposure and onset of drinking in the following unordered categories: under age 16, age 16–17, age 18, age 19 to 20, age 21, over age 21, and lifetime abstention, with onset under age 16 as the reference group. All analyses were conducted in STATA v.10, using ’robust’ standard errors clustered by state; these adjust for the correlation of observations within state, and capture most of the clustering of observations generated by the complex survey design. (Arellano 1987, Liang and Zeger, 1986). For all models, we report 95% confidence intervals with two-sided p-values.
