Several recent papers have suggested that two-locus tests of association that incorporate gene x gene interaction can be more powerful than marginal, single-locus tests across a broad range of multilocus interaction models, even after conservative correction for multiple testing. However, because these two-locus tests are sensitive to marginal associations with either marker, they can be difficult to interpret, and it is not immediately clear how to use them to select a list of the most promising markers worthy of further study. Here we apply single- and two-locus tests to 29 single-nucleotide polymorphisms (SNPs) selected from the dense marker map in the simulated Genetic Analysis Workshop 15 data spanning several candidate regions (the HLA region, the four SNPs flanking "Locus D," and two regions on the q-arm of chromosome 6). We compare the proposed two-locus likelihood ratio tests (LRT) to Akaike and Bayesian Information Criteria (AIC and BIC) for model selection, as well as AIC- and BIC-weighted measures of "SNP importance." The latter provide summary measures of evidence for association between each SNP and disease - including potential interactions with one or more other SNPs - by summing over all one- and two-SNP models. Our results suggest that the LRT using conservative p-value criteria were sensitive (but not specific) in identifying associated markers. Standard AIC and BIC criteria were similarly sensitive but not specific. On the other hand, the AIC- and BIC-weighted importance measures yielded a specific but not very sensitive rule for SNP selection. Algorithms incorporating gene x gene interaction to prioritize markers for follow-up will require further development.