Why SKAT Does Not Examine Rare Variants Individually
Why Rare Variants Are Analyzed in Sets
Common variants have sufficient sample sizes within each genotype group to allow association testing at individual loci. In contrast, rare variants are carried by very few individuals, most of whom possess the reference allele; consequently, statistical power for single-variant tests is insufficient even when a true effect exists. Therefore, biologically related variants are aggregated into sets at the gene, pathway, or regulatory region level to assess their collective association with the phenotype.
Aggregation does not automatically enhance information quality. The choice of which variants to include and the application of functional annotation and minor allele frequency criteria determine the hypothesis being tested. Mixing unrelated variants dilutes the signal, while selecting favorable sets post hoc inflates the type I error rate.
Distinct Questions Addressed by Burden Tests and SKAT
Burden tests aggregate rare alleles within a set into a single score. They are powerful when the effects of included variants share similar directions and magnitudes. Conversely, when protective and risk effects are mixed or non-causal variants are prevalent, mutual cancellation or increased noise may occur.
SKAT combines variant-specific scores within a kernel to capture signals even when effect directions and magnitudes differ. This flexibility does not imply superiority in all scenarios. When true effects cluster in the same direction, burden tests may be more powerful.
SKAT-O explores the continuum between burden tests and SKAT to identify an optimal balance for the data. Although the selection process is corrected, it is not an automatic solution for correcting misdefined variant sets or phenotype models.
Core Components of the Analytical Design
Variant weights typically assign greater values to rarer alleles, but frequency alone cannot guarantee functional importance. It is necessary to predefine whether loss-of-function, missense, and regulatory variants are included in the same set, and to verify annotation errors and transcript selection. The minor allele frequency (MAF) threshold must be defined consistently across internal and external reference cohorts.
For binary phenotypes, caseโcontrol imbalance affects results; for quantitative traits, distribution and outliers exert influence. Failure to adjust for ancestry principal components, relatedness, and sequencing batch may cause population structure or technical artifacts to mimic genetic signals. Additionally, the presence of individuals duplicated across multiple cohorts violates the assumption of independence.
Small Example
Suppose that rare coding variants in an enzyme gene influence disease risk. If most reduce function and act in the same direction, a burden test may effectively aggregate effects for detection. However, if some increase function while others decrease it, and unrelated missense variants are mixed, SKAT may be more appropriate.
Even if the set-level p-value is significant, it does not identify which variant is causal or how many phenotypes it explains. Follow-up with leave-one-out analysis, variant-level data, replication in independent cohorts, and functional assays is required.
Multiple Testing and Reproducibility
Examining multiple genes, annotation masks, MAF cutoffs, and phenotypes renders each combination a distinct analytical opportunity. A primary analysis must be defined a priori, and multiple testing correction must be applied relative to the total number of exploratory tests. Significance observed in the discovery cohort should be replicated in independent datasets with both matching and distinct ancestries, and effect heterogeneity should be assessed.
Rare variants are sensitive to sequencing platforms and variant calling quality. If cases and controls are generated using different technologies, batch artifacts may accumulate within the set; therefore, joint quality control and callability comparisons are required.
Common Misconceptions
- A significant gene set does not imply the identification of diagnostic genes or clinical actionability.
- Although SKAT-O combines multiple models, it does not encompass all genetic architectures.
- Large sample sizes cannot eliminate bias if phenotype definition and ancestry representation are inadequate.
- Annotation-based masks represent biological hypotheses rather than objective preprocessing criteria.
Interpretive Caveats
SKAT-family methods are population-level association tools. The results do not determine the pathogenicity of individual variants or diagnose specific patients. Reproducible evidence is provided when the set definition, weights, phenotype model, ancestry/relatedness, and overall analysis plan are disclosed together.
Reading in Context
Variant-level functional evidence is linked to functional assays and variant interpretation, while large-scale population study design is connected to the Broad Medical and Population Genetics concept.