This paper presents a possible solution to the problem of identification of factors influencing long-term survival patients, using regression trees. The separation of the two classes of long-term survivors (cured patients) and of failed-to-cure patients is generalized to l* classes of survivors and is carried out via a latent variable, whose determinations are provided by the regression-tree classification. Two sets of factors are thus identified within the set of covariates: the factors influencing the prognosis and those influencing the survival classification (diagnostic factors). The relationship between the two sets is then explored, both theoretically and using an application to a data set of multiple myeloma patients. (C) 1998 John Wiley & Sons, Ltd.