Most pathway analysis tools utilize one association signal per gene. While expression arrays yield a single p-value for each gene, SNP arrays include multiple signals per gene, some of which are correlated. As such, some studies use the minimum SNP-level p-value within a gene as the operative signal [4, 25, 33, 34]; however, this approach will not detect additive effects among SNPs with moderate individual association. For methods that combine SNP-level signals, including those based on the truncated product method [14], LD must be accounted for to prevent highly-correlated SNPs from biasing gene-level significance. Strategies to accomplish this include discarding SNPs that depart from LD at a preset threshold [25, 26, 39] and adapting principal component analysis to extract the most independent signals within a gene [10, 11, 26]; unfortunately, these methods can eliminate substantial information. Alternatively, the SNP ratio test [40] and the “set-based analysis” in PLINK [41] use phenotype permutation to naturally correct for biases introduced by LD and gene size; however, these tools require raw genotype data and are computationally demanding, making them better suited for studies
