The distribution of the alcohol use behavior data was highly skewed. Preliminary analyses indicated that a log transformation was optimal for all variables, with the exception of the binary age 12 drinking outcome. The appropriate distribution for each outcome was determined using traditional fit statistics (Akaike’s Information Criterion, difference in X2) for the data modeled using normal, Poisson, negative binomial, and zero-inflated distributions. As the majority of participants had not yet initiated drinking at age 12, a zero-inflated negative binomial distribution fit the data best and was used to assess the association between the GR scores (i.e., AUD-GR and EXT-GR, separately) and age 12 initiation of drinking. A Poisson distribution best fit all other variables (age 14 and 17 frequency of intoxication and age 22 AD criteria) and was used to assess the association between the GR scores (i.e. AUD-GR and EXT-GR, separately) and age 14 frequency of intoxication, age 17 frequency of intoxication, and age 22 AD. The residual correlation within twin pairs was substantial and stronger in MZ twin pairs compared to DZ pairs; accordingly, regression models were
