We pooled these epidemiological estimates in a disease model that was judged to have face validity by the GBD 2010 Expert Group and additional experts in cannabis use epidemiology, in terms of age and sex patterns for the disease, differences in incidence or prevalence between regions, and changes in these parameters over time. For these estimates we used DisMod-MR[28,45], the latest application of an incidence-prevalence-mortality (IPM) mathematical model[46], re-designed as a Bayesian meta-regression tool for GBD 2010. The IPM model was implemented as a negative-binomial rate model to ensure internal consistency between separate estimates of prevalence, incidence, remission and excess-mortality. DisMod-MR was also used to predict epidemiological estimates for regions with no available data using country random intercepts, and prevalence estimates from elsewhere in the region and respective super-regional groupings (for details of GBD regions see Table S2 ). We preferred high quality, direct epidemiological estimates. In their absence, we used predicted estimates rather than exclude regions with no epidemiological data from GBD 2010 estimates[28,45].
