The primary analyses involved latent growth curve (LGC) models, structural equation models that use latent growth factors (intercept, slope, quadratic) to represent the mathematical function underlying the covariance between repeated measures. Here, we fit LGCs for each AUB in each cohort using the OpenMx package (45). First, a baseline model containing either intercept+slope (IS) or intercept+slope+quadratic (ISQ) parameters was fit to determine the general shape of the longitudinal trajectories. After comparing the difference in −2*loglikelihood between models with a chi-square test, the best fitting model was selected. Finally, this model was fit sequentially with the addition of each PGS residual as an individual-level predictor of the latent growth factors to determine how genetic influences captured by the PGSs shape AUB trajectories (Figure 2). Full information maximum likelihood approaches for model estimation were used to account for missingness and nonnormality. A multi-level application of the model (adapted from https://github.com/OpenMx/OpenMx/blob/master/inst/models/nightly/mplus-ex9.12.R) was implemented to account for the non-independence of observations in samples that included relatives by design (COGA, FinnTwin12, FTC). As >90% of participants in each cohort had initiated alcohol use before or
