Directional Effect Assumptions in Burden Tests

Why Direction Matters for a Burden Score

A burden test compresses information across variants in a set into one score. This can be efficient: if included alleles tend to shift the phenotype in the same direction, their contributions can reinforce one another in that aggregate. But the benefit is conditional, not automatic. It depends on how alleles and phenotypes are coded, which variants are included, how they are weighted, and how their score contributions covary.

The relevant claim is about statistical alignment. After allele harmonization, effects in a common direction have coefficients with the same sign on the chosen phenotype scale. A positive coefficient means an increase in the coded phenotype per additional copy of the coded allele; reversing the allele coding reverses that coefficient’s sign. A common direction is therefore meaningful only after both allele and phenotype orientations have been fixed.

The derivation below uses the score-vector form of a burden test. Let \(U=(U_1,\ldots,U_m)^\mathsf{T}\) be the vector of variant score components for a set of \(m\) variants, and let \(V\) be its covariance matrix under the null model. Let \(\beta=(\beta_1,\ldots,\beta_m)^\mathsf{T}\) denote the corresponding effect parameters, and let \(w=(w_1,\ldots,w_m)^\mathsf{T}\) be the fixed weight vector. The scalar burden score is \(U_B=w^\mathsf{T}U\). This notation describes the score test’s local behavior; the precise phenotype model determines the score components and covariance.

Deriving the Alignment Condition

The standardized burden score is \(Z_B=U_B/\sqrt{w^\mathsf{T}Vw}\). Under the null approximation, its mean is zero and its variance is one. Under the local alternative, the score vector has mean \(V\beta\). Thus the burden score has mean \(w^\mathsf{T}V\beta\), while its variance remains approximately \(w^\mathsf{T}Vw\). Squaring the standardized score gives a one-degree-of-freedom test statistic whose noncentrality parameter is

This quantity separates the signal collected in the numerator from the null variability in the denominator. A large absolute value of \(w^\mathsf{T}V\beta\), relative to the burden’s standard deviation, produces a larger noncentrality parameter and, under the approximation, greater power at a fixed significance level.

A particularly clear sufficient condition arises when variant scores are uncorrelated, so \(V\) is diagonal with positive entries \(v_j\), and all weights are nonnegative. Then

If all nonzero included effects have the same sign, the terms in this sum have the same sign wherever \(w_j>0\). They add rather than cancel. If some weights are zero, the corresponding variants contribute no burden signal. This condition explains the directional advantage in a simple setting, but it should not be overgeneralized: with correlated scores, off-diagonal entries of \(V\) also affect the aggregate, and common signs alone do not guarantee that every contribution to \(w^\mathsf{T}V\beta\) is positive.

One further way to understand the role of weights follows from the Cauchy–Schwarz inequality. Write \(a=V^{1/2}w\) and \(b=V^{1/2}\beta\). Then \(w^\mathsf{T}V\beta=a^\mathsf{T}b\), so

Equality holds when \(a\) is proportional to \(b\), or equivalently when the burden direction is proportional to the effect direction in this covariance-scaled space. This does not mean that a researcher can know the optimal weights in advance: the effects are unknown. It does show why a burden test can lose efficiency even when effects share a sign, if its weights poorly match their relative magnitudes or the score covariance.

Worked Examples

When the Directional Argument Is Credible

The calculation is only as meaningful as its assumptions. Several checks are especially important in genomic applications:

• Allele orientation: All effects must refer to consistently coded alleles. If one variant is coded for the alternate allele and another for the opposite allele, a shared biological direction can appear to have mixed signs. Allele harmonization is essential before describing a direction.
• Phenotype scale: Effect signs must refer to the same phenotype definition and orientation. For example, a larger value must consistently mean the same thing throughout the analysis.
• Variant inclusion: The claim concerns the variants actually included, not every variant in a gene or region. A set may contain neutral variants, variants with effects in another direction, or variants whose association is due to linkage rather than a direct effect.
• Weight choice: Weights should be justified in advance or estimated without using the same outcome information in a way that invalidates the test calibration. Incorrectly large weights can increase score variance without a commensurate increase in mean signal.
• Score covariance: Correlation, population structure, and nuisance estimation can affect \(V\). The test’s covariance must reflect the fitted null model and the data structure well enough for its reference distribution to be credible.

Diagnostics can assess coding and model fit, but they cannot prove a shared direction when individual rare variants have little information. Review allele harmonization and phenotype definitions; inspect variant-level estimates and their uncertainty where feasible; and check whether the fitted null model and score covariance are appropriate. A few noisy signs are not decisive evidence that effects truly point in different directions. Conversely, a significant burden result alone does not establish that every included variant has the same effect direction.

How the Burden Target Differs From Alternatives

A burden test asks whether a prespecified weighted aggregate has a nonzero effect. Its one-dimensional direction is its strength when the underlying effects align with that aggregate. A collection of separate single-variant tests does not impose this shared-direction structure, but it also does not pool evidence into one score and can require a multiple-testing adjustment. Other set-level approaches can target patterns of variant effects that are not well represented by one shared burden direction. Their detailed statistics and calibration are distinct; the key choice is the scientific alternative the analysis is meant to detect, not a universal ranking of procedures.

The directional assumption is thus a modeling choice about the alternative, not a requirement that nature must obey. It is sensible when biology, functional evidence, or prior evidence supports effects in one coded direction and the set and weights reflect that target. It is less secure when the set is broad or the direction is uncertain. In that case, a burden result may be less sensitive than expected, and a different set-level target may be more appropriate.

Common Mistake and Limitations

Incorrect reasoning: “All variants have positive estimated effects, so the burden test must have high power.” This conclusion is too strong. Estimated signs can be noisy, the weights may be poorly matched to the true effects, and covariance affects both the burden’s mean and its variance. Even under a shared direction, the noncentrality can be small if effects are weak or the burden is inefficient.

Correction: State the coding and model assumptions, then assess alignment through \(w^\mathsf{T}V\beta\) relative to \(w^\mathsf{T}Vw\). In an actual analysis, \(\beta\) is unknown, so the calculation explains the test’s behavior under alternatives rather than certifying that those alternatives hold. The local normal approximation can also be inaccurate in very sparse data or under a poorly specified null model. A valid test calibration and a plausible effect-direction assumption are separate requirements.

The central limitation is that one burden score reduces a multi-variant pattern to a single dimension. If the true effect pattern is far from the chosen burden direction, the projection can be weak even when some variants are associated. This tutorial’s result explains the favorable case; it does not imply that all rare-variant associations are directional or that a burden test is always preferable.