Operating characteristics of the rank-based inverse normal transformation for quantitative trait analysis in genome-wide association studies