The study analyzed the composition of each predictive model, compared their predictive capacity on relapse and abstinence and integrated them into an overall model. The study was based on two groups of heroin misusers (one relapsed, n = 130; one abstinent, n = 130) treated in 15 Spanish therapeutic communities. Data were collected via structured interviews and urine analyses. A logistic regression model was used to build up and cross-validate six different predictive models. Several partial models had predictive capabilities between 47 and 78 percent of abstinence and relapses, while the overall model reached 85 and 89 percent, respectively. The study suggests the advisability of integrating several models in order to more effectively prevent relapse: motivational-conditioned, situational, cognitive-affective and coping skills models. The predictive overall model is useful because it: (1) is balanced for abstinence and relapse and its predictive capability is nearly maximum; (2) is made up entirely of variables which can be modified by direct interventions; (3) reduces significantly the number of predictive variables; and (4) comes close to being an empirical demonstration of the necessity and convenience of integrating various psychological models of relapse into a single theoretical model, covering stress-coping, attitudes for abstinence and motivational-conditioned models on heroin use relapse. Tables, references