The sample consisted of twin pairs, who do not constitute independent observations. We therefore fit regression models producing robust standard errors based on the clustered nature of the sample via generalized estimating equations (GEE) (Liang & Zeger, 1986) in SAS/STAT software, version 9.1.3 for Windows. We used logistic regression for dichotomous dependent measures (measures of diagnostic status, yes/no measures of early substance use and misuse), a Poisson regression for breadth of early substance experimentation, and a linear regression for the twins' own maximum drinks consumed. The latter was log transformed for analysis, although untransformed values are used for descriptive purposes.
