of the generalized linear model, which permits adjustment for covariates and handling of both dichotomous and quantitative phenotypes (Lou and others 2008). A key advantage of PGMDR is that the method can handle different pedigree structures and sizes simultaneously in the presence of various patterns of missing data. In our study, by using PGMDR, three-locus, four-locus and six-locus interaction models were detected in the AA sample, and two-locus and three-locus interaction models were detected in the pooled sample (Table 6). It is intriguing that rs1317286, an SNP in the CHRNA3, which was associated with both SQ and HSI in the pooled sample, was included in four of these five best interaction models. Therefore, the interactive effects detected by PGMDR are potentially valuable, such that they not only lend support to the single-locus results, but also shed significant light on the joint effects among SNPs in the CHRNA5/A3/B4 gene cluster, which could be missed in single-locus analysis. Nevertheless, larger and more rigorous replication studies are necessary to establish such interactions. As pointed by Milne et al. (2008), replication in additional independent samples of observed gene–gene interaction is crucial. In particular, caution should be exercised during the replication because differences of LD
