Rare-variant association analysis: study designs and statistical tests

Abstract

Despite the extensive discovery of trait- and disease-associated common variants, much of the genetic contribution to complex traits remains unexplained. Rare variants can explain additional disease risk or trait variability. An increasing number of studies are underway to identify trait- and disease-associated rare variants. In this review, we provide an overview of statistical issues in rare-variant association studies with a focus on study designs and statistical tests. We present the design and analysis pipeline of rare-variant studies and review cost-effective sequencing designs and genotyping platforms. We compare various gene- or region-based association tests, including burden tests, variance-component tests, and combined omnibus tests, in terms of their assumptions and performance. Also discussed are the related topics of meta-analysis, population-stratification adjustment, genotype imputation, follow-up studies, and heritability due to rare variants. We provide guidelines for analysis and discuss some of the challenges inherent in these studies and future research directions.

Figure 1

Data-Processing and Analysis Flow Chart for Sequencing-Based Association Studies Explanations of these steps are given in Box 1. The following abbreviation is used: QC, quality control.

Figure 2

Using Single-Variant Tests to Estimate the Proportion of Heritability Explained by Significantly Associated Low-Frequency and Rare Variants Dashed lines represent the true proportion of heritability explained by low-frequency and rare variants, and solid lines represent the estimated (by single-variant tests) observed proportion of heritability due to significantly associated low-frequency and rare variants at level α = 5 × 10⁻⁸. From top to bottom, the three curves correspond to the situation when a low-frequency (0.5% ≤ MAF < 5%) or rare (MAF < 0.5%) variant is ten times more (r = 10), four times more (r = 4), or equally (r = 1) likely to be causal than a common variant. (A) Effect sizes of causal variants are assumed to be constant regardless of MAF: $\beta =\theta$. (B) Effect sizes of causal variants of rare or low frequency are assumed to be a decreasing function of MAF: $\beta =\theta \left|{\log }_{10}\text{MAF}\right|\mathrm{∕}\left|{\log }_{10}0.05\right|$. The parameter θ is set at θ = 0.183, which provides power = 0.8 at level α = 5 × 10⁻⁸ when the sample size is 50,000 and the MAF is 0.05.