The use of the family history method to examine the pattern of recurrence risks for complex disorders such as autism is not straightforward. Problems such as uncertain phenotypic definition, unreliable measurement with increased error rates for more distant relatives, and selection due to reduced fertility all complicate the estimation of risk ratios. Using data from a recent family history study of autism, and a similar study of twins, this paper shows how a latent-class approach can be used to tackle these problems. New findings are presented supporting a multiple-locus model of inheritance, with three loci giving the best fit.