of age at interview to remove, in part, possible cohort effects. Finally, our model fitting strategy was not exhaustive. The CCC model is an improvement over previous separate modelling of drug use and abuse (1) because it models for the contingent nature between initiation and abuse. However, the single regression parameter (β2) specifies that the causal contributions of A, C and E on initiation are proportional to their effects on abuse. It is possible that a feedback mechanism exists between the third and first stages, such that manifesting symptoms of abuse increases the likelihood of reporting increased drug availability. In other words, diagnosed individuals may be more likely to actively seek out cannabis. To test for this possibility, we began with Model 2 and added a direct transmission path from the symptoms of cannabis abuse back to cannabis availability. This feedback model deteriorated significantly when compared to the saturated Model 1 as judged by the change in log likelihood (Δ-2LL=7.96, Δdf=2, p=0.02). This model did not perform as well as Model 2 in terms of the Bayesian Information Criterion (−3921.985). Moreover, the standardized path coefficient from symptoms to availability was very small (β=0.03).
