The alcohol onset model predictors included baseline period-specific rate parameters, rate by grade, rate by gender, rate by adult supervision, and rate by exposure to peers who have already started to use alcohol. Substantively important but nonsignificant effects were retained in all models for which convergence was acceptable (Ripley et al., 2012), but in a few cases some of these effects had to be dropped (i.e., set to zero) to achieve convergence. This approach is typical for many classes of statistical models (Gelman & Hill, 2007). Wave dummies (Lospinoso, Schweinberger, Snijders, & Ripley, 2011) were examined in early model development work. Their inclusion did not materially affect substantive results and conclusions, but sometimes resulted in model convergence issues. Thus, it was decided not to attempt a systematic exploration of variation between waves for the present study.
