The general framework of the structural equation model that was first used to assess main effects is exemplified in Fig. 1. Each model included substance use (specifically tobacco use in Fig. 1), friends’ support of substance use, and effortful control as three latent factors hypothesized to have main effects on problematic substance use in early adulthood. However, for alcohol and marijuana, multicollinearity issues emerged because substance use in adolescence was correlated with friends’ substance use at r = 0.84 or greater. The very large amount of shared variance between these two latent constructs did not allow us to make reliable predictions about the relative contribution of each factor to problematic substance use in early adulthood. A reduced model was thus developed in which a single “substance use lifestyle” factor was formed from the eight items previously used for participants’ substance use and friends’ substance use. Within these alternative models, residual errors between adjacent substance use assessment points were allowed to covary in order to improve model fit. This alternative model approach was used for the alcohol and marijuana models (see Figs. 2 and 3).
