The proposed Markov formulation has more behavioral content than statistical models that produce only point estimates; it explicitly accounts for flows between different categories of use. On the other hand, it has less behavioral content than do system dynamics models. Although not unimportant, this study does not include the feedback effect of prevalence on incidence. The study shows that cocaine consumption has declined little in proportion to the declines in prevalence. Further, if initiation stabilizes at its 1991 levels of about 1 million new users a year, then further declines in either prevalence or consumption are unlikely. To counteract cocaine use, initiation must be reduced, at least to, of not below, the levels of the later 1980's. Hence, continuing and possibly expanding prevention programs may be necessary. As a consequence of the persistence of heavy use, however, such prevention, while necessary, is by no means sufficient. In order to substantially alleviate cocaine- related problems, prevention programs must be supplemented by programs that encourage heavy users to quit, or at least to reduce their consumption. The model also illustrates that there is enormous inertia associated with heavy cocaine use. Given the tendency of drug epidemics to recur after a quiescent state, programs that could abort the next epidemic before it generates a large, persistent core of heavy users may be well worth their expense. In particular, maintaining the capacity to detect quickly the next epidemic and promptly implement large-scale prevention programs could be cost-effective in the long run. 1 table, 8 figures, and 22 references