Why Kernel Tests Detect Heterogeneous Rare-Variant Effects

Why Directional Heterogeneity Matters

A gene can contain rare variants whose effects on a trait point in different directions. One variant may be associated with a higher trait value, while another is associated with a lower value. A signed burden combines variant information into one direction-sensitive quantity. When variant effects oppose one another, their contributions can cancel before the test measures the evidence. A kernel test instead uses a quadratic combination of variant-level scores: squaring makes each component contribute without retaining its sign.

This difference is the central reason a variance-component or kernel test can detect a gene-level signal that a signed burden may miss. It is not a claim that kernel tests are always more powerful. When effects are consistently aligned, a burden can combine their evidence efficiently. The methods respond differently because they summarize a different feature of the variant-level evidence.

Variant-Level Scores and Two Ways to Combine Them

To make the comparison precise, consider a quantitative trait \(y\), an \(n\)-row vector of trait values, and a matrix \(C\) containing an intercept and any covariates in the null model. Let \(g_j\) be the \(n\)-row genotype vector for variant \(j\), coded, for example, as the number of alternate alleles. Under a linear null model, assume \(y=C\gamma+\varepsilon\), with correctly specified mean, independent errors of mean zero and variance \(\sigma^2\), and a full-column-rank \(C\). The null residual-maker is \(M_C=I-C(C^\mathsf{T}C)^{-1}C^\mathsf{T}\); it removes the fitted covariate component.

A covariate-adjusted score component for variant \(j\) is \(U_j=g_j^\mathsf{T}M_Cy\). It measures the alignment between the genotype pattern, after accounting for covariates, and the trait residuals. A positive or negative value records the direction of that alignment. In practice, score components may be scaled by their null standard deviations, and their joint covariance must be handled in the test calibration.

These expressions describe the basic contrast, not a complete recipe for inference. A test also needs the null variance or null distribution of its statistic. Variant scores can be correlated, for example because genotypes at nearby variants are in linkage disequilibrium. A valid calibration accounts for this covariance; it must not assume independent scores merely because the statistic is written as a sum.

For transparent numerical comparisons below, use standardized scores \(Z_j\) with null mean zero and null variance one, equal weights, and independent null scores. In this simplified setting, the signed burden statistic is \(Z_B=(\sum_j Z_j)/\sqrt{m}\), and the quadratic statistic is \(Q=\sum_j Z_j^2\). Under the stated independent standard-normal null, \(Z_B\) has a standard normal distribution and \(Q\) has a chi-square distribution with \(m\) degrees of freedom. These assumptions make the examples calculable; real analyses require an appropriate covariance-aware calibration.

Why Opposing Effects Can Cancel in a Burden

A signed burden adds contributions before evaluating how far the result is from its null expectation. If one contribution is positive and another negative, the sum can be small even when both individual scores are large. A quadratic statistic first removes the sign of each component by squaring and then adds the resulting nonnegative values. Consequently, opposite directions do not cancel in this statistic.

The contrast is visible algebraically. For two equally weighted components of equal magnitude \(a\) and opposite signs, the burden sum is zero, but the quadratic sum is positive. With the same two magnitudes and the same sign, both methods register evidence; the burden can use the shared direction directly.

This is a statement about how the statistics combine scores, not a guarantee about p-values. The null variability of each statistic differs, and the quadratic statistic generally spends degrees of freedom to remain sensitive to several possible directions. Its observed value must be judged against its own null distribution rather than compared numerically with a burden statistic.

A Short Justification: Squaring Preserves Heterogeneous Signal

The same point can be expressed in terms of expected score patterns. Suppose standardized scores have an alternative mean vector \(\delta=(\delta_1,\ldots,\delta_m)\), with components representing the directions and magnitudes of the variant-level signals. Ignoring covariance for this illustration, the signed sum has mean \(\sum_j\delta_j\), whereas the quadratic statistic responds to the squared components. Opposing values can make the first quantity vanish while leaving the second large.

This calculation explains the cancellation mechanism, but it is not a full power derivation: power also depends on the score covariance, effect sizes, sample size, weights, and the test’s null calibration. In particular, a quadratic statistic responds to random variation as well as true signal. It therefore pays a calibration and variability cost for being sensitive to multiple directions.

When the Advantage Can Disappear

A kernel test is not uniformly superior to a burden test. If many variants have effects in a common direction, the signed burden can pool them coherently. In contrast, a quadratic test does not use the shared sign to make a single directional contrast; it evaluates the combined squared evidence against a null distribution that reflects the many component contributions. Under aligned effects, this can make a burden more efficient.

The weighting scheme also matters. If a truly associated variant receives a very small weight, its contribution to a weighted quadratic statistic is reduced. Conversely, including many variants with little or no association can add noise and weaken evidence. Weight choices should be determined in a principled way, rather than selected after inspecting which choice gives the smallest p-value without accounting for that selection.

Correlation among genotypes is another important qualification. When variants are correlated, their score components are correlated too. The raw sum of squared scores can still be formed, but its null distribution is not generally the independent chi-square distribution used in the examples. In a real analysis, the covariance structure and other design features must be reflected in calibration. Failure to do so can produce an incorrect p-value even if the observed statistic was calculated correctly.

Finally, a significant quadratic test is a gene-level association result, not a signed estimate of one gene-wide effect. Because the statistic discards component signs, it does not by itself say which variants are associated, whether their effects increase or decrease the trait, or whether any variant has a causal effect. Those questions require additional analyses and careful interpretation.

Common Mistake

Wrong reasoning: “A kernel statistic squares every score, so it must be more powerful than a burden whenever rare variants are tested.” This confuses protection against cancellation with universal power. A quadratic test retains evidence from opposite signs, but it also has its own null variability and can be less efficient when effects are aligned. Adding unassociated variants or using unsuitable weights can further reduce sensitivity.

Correction: Match the interpretation to the effect pattern the statistic can detect. A signed burden is most naturally suited to a coherent direction under its weights; a quadratic test can capture departures spread across multiple directions. Neither is guaranteed to win for every gene. Their p-values require valid, test-specific null calibration, and a significant kernel result does not reveal a common direction or prove causation.