Association testing was performed for three traits: DSM-IV CD and the two heavy cocaine-use subtypes to identify SNPs that exerted a significant main effect or that, in pair-wise combinations, exerted a joint (epistatic) effect on the traits. GEE logistic regression was used in main effects tests to correct for correlations among related individuals in families. All SNP pairs were tested for association in epistatic analyses. We compared the findings of pair-wise joint effects detected using three methods: GEE logistic regression [SAS Institute, Inc., 2008], regular logistic regression [Purcell et al., 2007] and a bioinformatics method called BOOST [Wan et al., 2010]. For each SNP pair, BOOST estimates two logistic models: one uses the two SNPs as covariates; the other uses the two SNPs and their cross-product as covariates. An upper bound is calculated to approximate the ratio of the two logistic models, which aims to assess the additional effect that the cross-product contributes. The bound has been shown to be tight, and most nonsignificant interactions can be pruned. In all three models, age, sex and the first PCA dimension calculated
