All analyses were done in Mplus 7.2 (Muthen and Muthen 1998–2012). Latent growth curve models were fit using maximum likelihood estimator with robust standard errors and a logit link function. Missing data were handled with full-information maximum likelihood (FIML). For latent growth curve models with binary outcomes, a continuous latent response variable y* is assumed that underlies the observed categorical outcomes y (i.e., alcohol abuse vs. no abuse) through a time-invariant threshold (see Supplementary Figure 3). The mean of the intercept (η0) of the latent response variable y* is fixed at zero for model identification and scaling (Masyn et al. 2013). Therefore, model estimated linear slope (η1) represents rates of linear growth through time over the latent response variable y*, and its exponential (eη1) represents the change of the odds of adolescents having alcohol abuse relative to no abuse for one year increase. An unconditional latent growth curve model without any covariate or predictor was first fitted using all participants to establish the most parsimonious growth model. Preliminary analyses revealed a small and significant quadratic slope, as well as a
