Due to the limits imposed by the sample size and the complexity of the possible models, it was not possible to test all potential mediators of treatment at all time periods. To determine which mediators to keep, and from which time points, each set of variables was tested as part of the LGM. Among the IPA variables, the three variables that were sensitive to Time or to Time X treatment effects were each tested (i.e., Behavioral Support for Abstinence, Attitudinal Support for Abstinence, and Number of Abstinent Friends in the social network). Of these, the best fit to the LGM included the IPA variable Number of Abstinent Friends in the social network, at the posttreatment time point (model χ2= 69.32; df= 63; p > .20; CFI=.99; RMSEA=.03). In this model the Treatment contrast significantly predicted Number of Abstinent Friends at posttreatment, which in turn predicted both slope and intercept of PDA, while the Treatment-to-PDA paths were reduced to nonsignificance.
