We fitted models for the main effects of alleles at the HLA-A, B, and DR loci and their haplotypes on the risk of insulin-dependent diabetes mellitus (IDDM). Empirical Bayes methods were used, assuming independent exchangeable normal priors for effects at each locus separately and for haplotype effects. A pure main effects model, pure haplotype effects model, and a combined model were fitted using Gibbs sampling. The main effects model showed that the DR locus had the largest variation in risk between alleles, followed by the B locus; significance tests for each allele in this model were in general agreement with those in the companion paper [Langholz et al. (1995) Genet Epidemiol 12:441-453], although all relative risks were shrunk toward 1.0 in the empirical Bayes analysis. The variance estimate for pure haplotype effects was substantially larger than for any of the three main effects considered in this analysis, but in the combined model, the DR locus showed larger variability than the haplotype deviations. We confirmed that haplotype A2/B56/DR4 previously reported to be common in Finnish diabetics does indeed confer unusually high risk (relative risk = 7.6, P < 0.001), but found this to be only 1.9 times higher than predicted by its component main effects (P = 0.046). All of the other haplotypes could be adequately explained by their main effects. Empirical Bayes methods provide a natural means of dealing with the problems of multiple comparisons, multicolinearity, and sparse data that complicate the analysis of HLA data.