Abstract
Sequence kernel association test (SKAT) has become one of the most commonly used nonburden tests for analyzing rare variants. Performance of burden tests depends on the weighting of rare and common variants when collapsing them in a genomic region. Using the systolic and diastolic blood pressure phenotypes of 142 unrelated individuals in the Genetic Analysis Workshop 18 data, we investigated whether performance of SKAT also depends on the weighting scheme. We analyzed the entire sequencing data for all 200 replications using 3 weighting schemes: equal weighting, Madsen-Browning weighting, and SKAT default linear weighting. We considered two options: all single-nucleotide polymorphisms (SNPs) and only low-frequency SNPs. A SKAT default weighting scheme (which heavily downweights common variants) performed better for the genes in which causal SNPs are mostly rare. This SKAT default weighting scheme behaved similarly to other weighting schemes after eliminating all common SNPs. In contrast, the equal weighting scheme performed the best for MAP4 and FLT3, both of which included a common variant with a large effect. However, SKAT with all 3 weighting schemes performed poorly. Overall power across all causal genes was about 0.05, which was almost identical to the type I error rate. This poor performance is partly due to a small sample size because of the need to analyze only unrelated individuals. Because a half of causal SNPs were not found in the annotation file based on the 1000 Genomes Project, we suspect that performance was also affected by our use of incomplete annotation information.