Several applications of forensic genetics require probabilistic modelling of an offender database. It is difficult to accurately represent offender databases using a model, since these typically consist of groups of persons with unknown ethnic origins and proportions. We introduce a latent variable approach that overcomes this difficulty by estimating the allele frequencies of the populations comprising the offender database as well as their relative sizes from the database itself. Our approach has several potential applications, one of which is to reliably predict the number of false leads obtained in familial searching.