Interpreting SKAT Results and Avoiding Common Errors

What a Significant SKAT Result Means

A significant SKAT result is evidence against the null hypothesis that the tested variant set has no association with the phenotype, conditional on the fitted null model and the analysis choices. It is not, by itself, evidence that every variant is associated, that the variants act in one direction, or that any particular variant causes the phenotype. Those distinctions matter because SKAT aggregates evidence across variants without requiring all their effects to have the same sign.

A useful interpretation keeps three things separate: the association question being tested, the weighting and variant set used to define the test, and the statistical calibration used to obtain its p-value. A small p-value addresses the first question under the latter two choices. It does not validate those choices automatically.

For a quantitative trait, a null model commonly specifies the phenotype mean using covariates. For a binary trait, it commonly specifies the probability of the outcome using a logistic model. In either case, the null model is fitted without the tested variant effects. The score vector summarizes how the observed phenotype residuals align with the genotypes under that null model. SKAT combines the score components into a nonnegative quadratic statistic.

For clarity, use the convention that \(\omega_j\) is the nonnegative kernel weight applied to the squared score for variant \(j\). If \(U_j\) is that variant's score component, then the statistic is:

The set of variants, genotype coding, weights, covariates, and phenotype model all help define what this statistic tests. A reported SKAT p-value is therefore not a property of a gene name alone; it pertains to a particular analysis specification.

Why the Null Distribution Is a Mixture

Under \(H_0\), and under suitable regularity conditions, the score vector is approximately multivariate normal in large samples. The quadratic statistic then has an asymptotic weighted chi-square distribution. If \(\lambda_1,\ldots,\lambda_r\) are the nonzero eigenvalues of the matrix defining the weighted quadratic form under the null, its limiting distribution is:

The approximation symbol matters. In general, this is not an exact finite-sample distribution, especially for logistic models or analyses with limited information. The eigenvalues incorporate the null score covariance and the chosen kernel weights. They need not be equal, so an ordinary chi-square distribution with degrees of freedom equal to the number of variants is generally not the correct calibration.

The basic justification follows from diagonalizing the quadratic form. Write the null score vector as approximately normal with mean zero and covariance matrix \(\Sigma\), and write the statistic as \(Q=U^\mathsf{T}WU\), where \(W\) is the nonnegative weight matrix. The transformed vector \(W^{1/2}U\) has covariance \(W^{1/2}\Sigma W^{1/2}\). Diagonalize that covariance matrix with an orthogonal matrix. In the resulting coordinates, the quadratic form is a sum of eigenvalues times squared, approximately independent standard normal variables. Squaring each standard normal gives a chi-square variable with one degree of freedom, yielding the mixture above. This derivation explains the asymptotic form; it does not make that form an exact finite-sample result.

Check How Weights Enter the Statistic

Weights encode how strongly different variants contribute to the kernel. They may be chosen using prespecified information such as allele frequency or functional annotation, but they are not estimated effect directions. A variant with a larger \(\omega_j\) contributes more to \(Q\) for the same magnitude of \(U_j\). Changing the weights changes the test statistic and its null eigenvalues, and therefore changes the test being performed.

Software and descriptions may express weights differently. In the convention used here, \(\omega_j\) multiplies \(U_j^2\). If an implementation instead multiplies the genotype column by a factor \(a_j\) before forming the kernel, the resulting squared contribution is proportional to \(a_j^2 U_j^2\). These conventions agree only when \(\omega_j=a_j^2\). Check the implementation's definition rather than assuming that a quantity called a “weight” enters linearly.

Assess the Null Model Before Trusting the P-Value

The mixture calibration is conditional on a suitable null model. The model should represent the phenotype under \(H_0\), including relevant covariates and the appropriate trait distribution. For example, a binary outcome generally requires a model for its probability rather than treating the outcome as an unconstrained quantitative measurement. Covariates should be chosen for the study question and design, and genotype and phenotype data should be checked for alignment, coding, and missingness problems.

Misspecification can distort the score covariance and thus the reference distribution used for the p-value. In genetic data, population structure is one potential concern: if ancestry-related differences influence both allele frequencies and the phenotype, omitting relevant ancestry adjustment can create apparent score alignment. A significant result under an inadequate null model is not repaired simply by applying a sophisticated mixture calculation. Conversely, adding covariates mechanically without considering the study design is not a guarantee of validity.

Read the P-Value From the Correct Reference Distribution

A SKAT p-value is the upper-tail probability, under the null reference distribution, of a statistic at least as large as the observed \(Q\). In practice, mixture-distribution methods use the relevant eigenvalues to evaluate or approximate this tail probability. The calculation must use eigenvalues corresponding to the fitted null model and the actual test weights. Substituting a standard chi-square distribution merely because \(Q\) is nonnegative is not justified.

Even when a mixture is appropriate asymptotically, finite-sample accuracy depends on the model, sample size, phenotype, and calibration method. Thus, a reported mixture p-value should be described as a result under the stated calibration, not as an exact probability guaranteed by the formula in every dataset. This is particularly important when information is limited. Use a calibration method appropriate to the design, and report it clearly.

Common Errors and Responsible Reporting

A p-value is not the probability that \(H_0\) is true, and \(p=0.0037\) does not mean there is a 0.37% probability that the observed association is a false positive. It is a tail probability calculated assuming the null and the specified calibration. Statistical significance also does not establish causation, clinical importance, or replication.

A region-level result should be reported with enough information to define the test: the variant set and inclusion rules, the phenotype model and covariates, the weight scheme, the statistic and p-value, and the calibration method. If many regions, kernels, or weight schemes were tested, the interpretation must also account for the multiplicity relevant to the reported claim. A nominal p-value for one test does not by itself control the error rate across a search involving many tests.

Common errors include interpreting SKAT as a one-direction test, naming a specific causal variant from a set-level p-value, applying weights twice, omitting important null-model structure, and treating a mixture approximation as exact for any sample size. Each error changes either the scientific claim or the validity of its uncertainty assessment. A careful report states what was tested, what the result supports, and what remains unresolved.