Univariate twin models were fit to raw categorical data on initiation of alcohol, cigarette, and cannabis use. Categories were created for each substance to represent never used (‘0’) and late- and early-age at first use (‘1’ and ‘2’, respectively) in order to account for the potential skewness in the continuous forms of these variables. Inclusion of non-users in the models ensured that we did not lose information from twin pairs discordant for ever using a given substance. (Among MZ pairs, discordance rates were 11.0%, 24.0%, and 27.0% for alcohol, cigarettes, and cannabis, respectively. Among DZ pairs they were 20.6%, 38.9%, and 39.7%.) In preliminary analyses, univariate models using dichotomous indicators of ever use and 2-level variables of early vs. late initiation (users only) were compared with those using the three-level variables that include non-users. Similar underlying variance structures were observed in all three modeling approaches, but models using the never/late/early categorical scheme were the most stable, providing further support for using the three-level variables. Tests of multivariate normality were conducted and assumptions were met for all three phenotypes.
