Although HLM analyses were first conducted on all measures, in cases involving pretreatment, posttreatment, and follow-up data, outcomes were generally not linear in the TAU condition. Thus, MMRM, which treats time as categorical, was the primary analytic method. The simplest covariance structure not significantly different than an unstructured one was used; in all cases but one (noted below), the unstructured covariance structure was superior and used. Each analysis modeled a random effect for cohort, which allows for heterogeneity between cohorts; if it did not converge (as indicated by the Hessian matrix not positive definite), analyses without nesting by cohort were reported if there was no significant cohort or Time × Cohort effect within each arm of the study. For measures requiring full nesting due to a significant cohort or Time × Cohort effect but in which fully nested MMRM analyses did not converge, a repeated measures analysis of covariance was conducted that treated cohorts themselves as individual subjects, using cohort means as scores (termed an aggregate analysis; Hedeker, Gibbons, & Flay, 1994, p. 758). Effects found using an aggregate analysis
