Standard regression models for disease incidence data can be used to test for associations between a disease and measured genetic and environmental factors and their interactions. Complications arise when the gene is not observed, requiring segregation and linkage analysis approaches, or when the candidate gene(s) are found to be highly polymorphic, as in the HLA region. We propose a Bayesian approach to the latter problem, in which the log relative risks for all alleles at a given locus are taken to be independent and exchangeable, assuming there is no preferential zygotic assortment and negligible recombination. Multi-locus problems can be addressed either by adding exchangeable interaction terms or by adopting a multivariate prior for haplotype effects. Some simulations based on our current work on family studies of IDDM are discussed.