Univariate twin models were fit to raw categorical data, with timing of first alcohol use represented by ‘0’, ‘1’, and ‘2’ values (late, average, and early, respectively) and AD by a dichotomous variable. A test of multivariate normality was conducted for the variable representing timing of first use and the assumption was met. Genetic models were fitted by the method of maximum likelihood, using the structural equation modeling program Mx30. The thresholds (assessed as z-scores on the underlying standard normal distribution) were adjusted for age at the time of interview to minimize the potential influence of forward telescoping (i.e., reporting events closer to the time of interview than they actually were). The age distribution was divided into thirds (24-28, 29-31, and 32 years and older) and represented in the models by two dummy variables, with the 29-31 year-olds as the reference group. A series of sub-models examining the statistical significance of A, C and E were compared to the full model to derive the best-fitting univariate models. Sub-models were tested by calculating the difference between the -2 log likelihood fit of the full model and nested sub-model, which is distributed as chi-square for the given degrees of freedom.
