One simple filter is using only SNPs that have shown a main effect (Sullivan, et al., 2009; Thomas, 2010a, 2010b). SNPs with main effects are more likely than other SNPs to be relevant to disease in the first place and thus also more likely to be differentially relevant conditional on environmental experience. Other potentially useful filtering mechanisms include testing for variance differences conditional on genotype (Pare, Cook, Ridker, & Chasman, 2010) as well as selecting SNPs associated with differences between monozygotic twins (Elashoff, Cantor, & Shain, 1991; Magnus, Berg, Borresen, & Nance, 1981). After filtering, one is left with a subset of SNPs probabilistically enriched for GxE interactions. Even better, filtering could be conducted in one sample (or a meta-sample) and the GxE interaction tested in a separate sample, to avoid capitalization on chance such as when the filtering mechanism and the GxE testing mechanism are non-independent but performed within the same sample.
