As mentioned above, data from individuals who were on average 16.7 years old at Wave 1 collection were analyzed. Using data from this age group increased the reliability of the self-reported age of first use, but it also introduces the possibility that individuals might not yet have initiated use. For these potential future users, the age of first use is missing. This type of missing data is known as right censoring, and is explicitly accounted for in the likelihood function (Vermunt 1996). SMA models add the additional benefit of discriminating between long-term survivors (non-users) and individuals who have reported first use or might use in the future. This is similar to two-part random effect models where one class is used to group subjects with zero scores. Two-part random effects models are designed for symptom endorsement or severity scores, which can be modeled as a function of time. Our analyses focus on age of onset, which does not vary over time.
