To evaluate how well we succeeded in increasing the explained variance in phenotype, we evaluated two logistic regression models for nicotine dependence in the COGEND dataset: one with interactions and one without. Predictors included the 11 SNPs from the 3 two-locus models that were validated in the ACS data and the 4 two-locus models that had logistic regression interaction term p values <0.01: rs16969968, rs2133965, rs3787138, rs13277524, rs3743075 from Table 2 and rs2292977, rs17483548, rs667282, rs680244, rs1500948, rs2611603 from Table 4. The logistic regression model without interactions accounted for 5.15% of the trait variance and resulted in a c statistic of 0.622. In contrast, the model including the same predictors, but adding the 7 pair-wise interaction terms listed in the tables, accounted for 9.49% of the trait variance and resulted in a c statistic of 0.662.
