We compared the results derived from PrediXcan with those from two widely-used gene-based tests, namely VEGAS49 and SKAT50,51. VEGAS aggregates information from the full set of SNPs within a gene and accounts for LD using simulations from the multivariate normal distribution. SKAT is a kernel-based association test that evaluates the regression coefficients of the SNPs within a gene by a variance component score test in a mixed model framework. We generated BED-formatted files for SNPs and genes (as defined by GENCODE v12) and mapped SNPs that met post-imputation QC parameters to gene regions using bedtools. The use of an offline Perl implementation for VEGAS allowed us to examine the dependence of the results from this approach on LD information through the use of the actual genotype data (versus the default HapMap CEU reference panel data). We developed an R-based pipeline that invokes the SKAT package (version 1.0.1) that is publicly available from CRAN. We generated a Q-Q plot showing the distribution of gene-level p-values for association with RA (for genes outside the HLA region) derived from each gene-based test to test for systematic departure from the null expectation (of uniform p-values).
