Testing Multiple Kernels and Variant Weighting Schemes

Why Test More Than One Kernel?

A gene can contain variant classes that differ in frequency, functional annotation, or expected effect pattern. A kernel emphasizes particular similarities among individuals and therefore gives more weight to some patterns of genetic variation than others. An annotation-informed kernel or a different prespecified weighting scheme can be useful when that choice reflects a scientific hypothesis. But trying several kernels and reporting whichever gives the smallest p-value is a different procedure: the choice itself depends on the observed data.

The distinction matters because the usual p-value for one SKAT kernel is calibrated for that one test. It does not automatically account for a search over several kernels. The tests are often correlated because they use the same individuals, variants, and phenotype, but correlation does not make the selection cost disappear. The analysis must either define a single kernel in advance or calibrate the comparison or selection procedure.

Score Statistics and Their Null Distributions

Let \(U\) be the vector of variant score statistics after accounting for the covariate model, and let \(V\) be its covariance matrix under the null hypothesis of no association for the tested gene. For a positive semidefinite kernel \(K\), the kernel score statistic is

The kernel determines how the score evidence is aggregated. For a diagonal variant-weighting scheme, larger diagonal entries give the corresponding variant scores more influence. More general kernels can encode annotation-based similarities or combine several variant classes. The matrix must be positive semidefinite for the standard nonnegative weighted-mixture interpretation.

Under the null, the usual score-test calibration relies on the null mean and covariance model being appropriate and on the score vector having its stated null distribution. In the standard asymptotic formulation, \(U\) is approximately multivariate normal with mean zero and covariance \(V\). The eigenvalues of \(V^{1/2}KV^{1/2}\), denoted \(\lambda_1,\ldots,\lambda_r\), then determine the reference distribution:

Thus each kernel has its own null distribution and p-value. Comparing raw \(Q_K\) values across kernels is generally not a fair comparison: their eigenvalues, and therefore their null scales, can differ. A p-value accounts for the null distribution of its own statistic, but a collection of p-values still requires care if one is chosen after seeing them.

Combining Prespecified Kernels

Suppose \(K_1,\ldots,K_m\) represent scientifically motivated variant classes or weighting schemes. One option is to specify nonnegative combination weights \(a_1,\ldots,a_m\) before examining the association evidence and define

If each \(K_k\) is positive semidefinite and each \(a_k\geq 0\), then \(K_{\mathrm{comb}}\) is positive semidefinite. The combined statistic is one kernel test, calibrated using the eigenvalues of \(V^{1/2}K_{\mathrm{comb}}V^{1/2}\). It is not calibrated by averaging the individual p-values. A prespecified combination can distribute sensitivity across patterns, but it may be less sensitive than a well-matched individual kernel.

Comparing Kernels Without Selecting on the Outcome

If the scientific goal is to compare prespecified kernel tests, report the set of tests transparently, including their definitions, weighting schemes, statistics, and calibrated p-values. When inferential claims are made across the set, account for multiplicity. Bonferroni is valid under arbitrary dependence: for \(m\) tests, compare each p-value with \(\alpha/m\), or report \(p_{\mathrm{adj}}=\min(1,mp)\) for an individual result. Holm’s step-down method is also valid under arbitrary dependence and can be less conservative.

These adjustments control the family-wise error rate for the stated family of tests when the individual p-values are valid. In this setting, that family might be all the kernel tests examined for one gene. The family should reflect the actual search: trying additional weights, annotation subsets, or alternative kernels and omitting them from the reported count understates the multiplicity.

Calibrating a Data-Driven Kernel Choice

A different strategy is to define the selection rule in advance and calibrate the whole rule under the null. For example, suppose the rule is to select the kernel with the smallest p-value. Each null replicate must reproduce the dependence among the candidate tests: generate or resample outcomes under the null model, retain the observed genotype and covariate structure as appropriate, recompute every kernel statistic and p-value, and record the smallest p-value (or an equivalent maximum statistic). The adjusted p-value is the null probability of obtaining a result at least as extreme as the observed selected result.

A parametric bootstrap can implement this procedure when outcomes can be simulated from the fitted null model. The null model and covariance structure must be appropriate, and nuisance parameters are typically estimated under the null. A permutation approach is appropriate only when the permutation scheme is justified under the null—for example, it must respect the study design and any relevant covariates. Neither method is automatically valid merely because it uses resampling.

Practical Choices and Common Pitfalls

There is no universally best weighting scheme. Weights encode assumptions about which variants should contribute most strongly, and different kernels can emphasize different alternatives. A careful analysis distinguishes a kernel chosen from external biological or statistical reasoning from one selected after inspecting association results. If the latter choice is made, the selection procedure belongs in the calibration.

A common mistake is: “I tested six kernels, the smallest p-value is \(0.02\), and I will report it as a \(0.02\) result because the kernels are correlated.” Correlation can reduce the multiplicity penalty relative to treating tests as independent, but it does not establish that there is no penalty. The correction is to use a valid multiplicity procedure, such as Holm or Bonferroni, or a joint calibration that preserves the dependence among tests.

• Do not compare raw statistics alone. Each kernel can have a different null distribution; use calibrated p-values or a defined joint statistic.
• Do not average p-values to create a combined-kernel p-value. Combine the kernels or statistics according to a specified rule, then calibrate that rule’s null distribution.
• Do not report only the winning kernel. State how many kernels and weighting schemes were considered and how the choice was handled.
• Do not assume a prespecified weighted sum is automatically optimal. Its validity and power depend on the kernel, the null calibration, and the effect pattern.

The central decision is whether the analysis tests one kernel, uses a fixed combination, or searches among candidates. These are distinct procedures. Clear specification and matching calibration make the resulting evidence interpretable, while leaving biological interpretation of an association to the broader study context.