Discussion
GSA is useful method in interpreting the result from a GWAS. A systematic evaluation of its performance is of paramount interest to the GWAS community, as the method is getting popular. Here, we compared the performance of two such methods using the common datasets and gene set databases. While GSA-SNP behaved predictably, i-GSEA4GWAS produced too many hits for most of the test settings. For example, i-GSEA4GWAS reported 3.8- and 6.5-fold more hits, respectively, for GO and KEGG, with the imputed dataset than with the unimputed one. Imputation is such a useful practice that augments the power of a genotype dataset, and ideally, gene set analyses can benefit from it. Our study warns that one must be cautious in applying i-GSEA4GWAS to an imputed dataset. As we pinpointed above concerning the 'top 5%' threshold as the potential cause of the high hit rates of i-GSEA4GWAS, it would be interesting to re-evaluate its performance with lower thresholds. Currently, the threshold is not available for the users to change it. It would have been better if the user could choose it at will. For GSA-SNP, we recommend using an imputed dataset if at all possible. GSA-SNP allows the user to choose k in assigning the k-th best p-value to a gene. We recommend using k = 2 instead of k = 1, as the latter inflates the scores for some genes, diminishing the power of GSA.