Sequence kernel association test for quantitative traits in family samples

Abstract

A large number of rare genetic variants have been discovered with the development in sequencing technology and the lowering of sequencing costs. Rare variant analysis may help identify novel genes associated with diseases and quantitative traits, adding to our knowledge of explaining heritability of these phenotypes. Many statistical methods for rare variant analysis have been developed in recent years, but some of them require the strong assumption that all rare variants in the analysis share the same direction of effect, and others requiring permutation to calculate the P-values are computer intensive. Among these methods, the sequence kernel association test (SKAT) is a powerful method under many different scenarios. It does not require any assumption on the directionality of effects, and statistical significance is computed analytically. In this paper, we extend SKAT to be applicable to family data. The family-based SKAT (famSKAT) has a different test statistic and null distribution compared to SKAT, but is equivalent to SKAT when there is no familial correlation. Our simulation studies show that SKAT has inflated type I error if familial correlation is inappropriately ignored, but has appropriate type I error if applied to a single individual per family to obtain an unrelated subset. In contrast, famSKAT has the correct type I error when analyzing correlated observations, and it has higher power than competing methods in many different scenarios. We illustrate our approach to analyze the association of rare genetic variants using glycemic traits from the Framingham Heart Study.

Figure 1

Distribution of the p-values for famSKAT, famBT, unrSKAT and SKAT from the null simulation with LD between adjacent SNPs 0.5 and proportion of unrelated individuals 0%.

Figure 2

Power comparisons of famSKAT, famBT and unrSKAT. Empirical power calculated at level of 0.001. The sample consists of sib pairs and unrelated individuals. The total sample size in each scenario is 1000, and the total number of SNPs analyzed is 20. In each panel, +/−/0 indicates the proportion of SNPs with positive effects, negative effects and no effects.

Figure 3

Q-Q plots for famSKAT in the genome-wide sliding window analysis on four glycemic traits. The p-values were plotted as minus log base 10 p-values. The genomic control factor λ_GC was computed as the ratio of median chi-square statistics with 1 df corresponding to observed and expected p-values.

Figure 4

Run time of famSKAT, famBT and SKAT in analyzing 20 SNPs.