Annotation-Rich Omnibus Tests With STAAR

Why an Annotation-Rich Omnibus Test?

A rare-variant set can contain variants with different predicted functional importance and different effect patterns. Some may act in the same direction, making a burden test effective; effects may instead vary in direction, favoring a variance-component test such as SKAT. Evidence may also be concentrated in only a few variants, which variant-level aggregation can help detect. Functional annotations provide information for weighting variants, but no single annotation or test is guaranteed to be optimal.

STAAR addresses these uncertainties by constructing annotation-informed variant weights and combining evidence across a collection of weighting schemes and test types. Its omnibus result is a single calibrated p-value for the set-level null hypothesis, not an estimate of the effect of every variant and not proof that a particular annotation is biologically correct. The introductory material on rare-variant set tests provides context for burden and variance-component approaches; here the focus is how annotation-informed choices are assembled into omnibus evidence.

From Annotations to Variant Weights

Suppose a set has \(m\) variants. For variant \(j\), let \(G_j\) denote its genotype information and let \(U_j\) denote its score contribution under a fitted null model that includes the outcome and relevant covariates. A test may use a weight vector \(w=(w_1,\ldots,w_m)\). In a SKAT-type component, for example, a common quadratic statistic has the form \(Q_w=\sum_{j=1}^{m}w_j^2U_j^2\). The weights affect how much each variant contributes to the statistic; its null distribution also depends on the covariance of the score vector.

In annotation-informed weighting, a functional score can modify a baseline weight, such as one that gives greater weight to rarer variants. A simplified illustration is \(w_j=b_j f_j\), where \(b_j\) is a baseline weight and \(f_j\) is a nonnegative, suitably scaled annotation value. This illustrates the role of annotation; it is not a universal STAAR transformation. Implementations use specified annotations, transformations, and weight choices. A weight is not an estimated effect: it says how the test is constructed, not what the data have established about the variant.

Different annotations can yield different weight vectors. A conservation score might emphasize variants conserved across species, while a predicted-impact score might emphasize variants expected to alter protein function. Since the scores may have different scales and meanings, they must be processed according to a defined method before being used. A zero, a missing value, or an unexpectedly reversed scale can change the role of a variant substantially; these are analysis choices to check, not details that an omnibus p-value can repair.

What the Omnibus Combines

For each selected annotation-informed weighting scheme, STAAR can obtain p-values from different kinds of set tests. A burden component is sensitive to effects that aggregate coherently; a SKAT component can detect heterogeneous variant effects; and an ACAT-V component combines evidence across variant-level tests. The selected collection and its implementation must be specified. The key idea is to avoid committing all inference to one annotation or one assumed effect pattern.

The p-values in the omnibus combination answer tests of the same overall null: no association between the variant set and the outcome, conditional on the null model's covariates. They are not independent merely because they come from different annotations or test types. The annotation-specific p-values will often be correlated because they reuse the same individuals, variants, and outcome. STAAR uses a Cauchy combination approach to combine these component p-values while accommodating dependence under broad conditions.

Why the Cauchy Combination Can Be Useful

The transformation \(C(p)=\tan\{\pi(0.5-p)\}\) maps small p-values to large positive values, p-values near 0.5 to values near zero, and p-values near 1 to large negative values. Thus, a particularly small valid component p-value can contribute strongly to \(T\). The inverse transformation maps \(T\) back to the unit interval as the omnibus p-value.

A short justification explains the combination. If a component p-value is exactly uniform under its null, then for \(0<u<1\), \(P(C(p)\leq t)=P(p\geq 0.5-\arctan(t)/\pi)=0.5+\arctan(t)/\pi\), the cumulative distribution function of a standard Cauchy random variable. A weighted sum of independent standard Cauchy variables, with nonnegative weights summing to 1, is also standard Cauchy. Under dependence, that exact independent-sum argument does not apply. The Cauchy combination procedure is useful because its tail behavior often remains a good approximation under broad dependence, but calibration should still be checked for the analysis at hand.

The STAAR omnibus p-value is evidence against the set-level null across the included tests and annotations. It does not identify which component drove the result, establish a causal mechanism, or prove that every included variant is associated. Component results can help describe where evidence arose, but choosing a reported component after inspecting the omnibus result requires care: that selection does not replace the omnibus calibration.

Assumptions and a Practical Sequence

A small omnibus p-value is meaningful only if the underlying component p-values are valid for the intended null. In a typical score-test setting, the null model must be appropriate for the outcome, covariates, and sampling structure, and the genotype and annotation data must be aligned to the same variants and individuals. The outcome model, relatedness adjustment where required, variant filters, set definition, and annotation transformations are part of the analysis specification.

The method also relies on using valid p-values for each chosen test and applying the intended omnibus combination. If an annotation is selected because it gave the smallest p-value and then treated as though it had been fixed in advance, the selection is not accounted for by a combination over a different, incomplete set of tests. Likewise, missing or incorrectly aligned annotation values can silently create unintended weights. Omnibus testing addresses uncertainty among the included choices, not choices that were omitted or data-processing mistakes.

Diagnostics, Alternatives, and Limitations

Before interpreting the omnibus result, check that variant identifiers and annotation rows match exactly; that no annotation has been accidentally reversed or left on an incompatible scale; and that the component p-values are finite and valid. Review the specified test collection and combination weights, and confirm that the same set-level null and outcome are being tested across components. Where possible, assess null calibration using suitable simulations or resampling that preserve relevant features of the data. A suspicious excess of small p-values under null simulations can indicate problems with the null model, genotype handling, or calibration rather than a successful annotation strategy.

STAAR differs from committing to a single annotation-weighted SKAT or burden test: the single test may be more focused when its assumptions are well justified, whereas an omnibus can be more robust to uncertainty about annotations and effect architecture. It also differs from SKAT-O, which adapts across a burden-to-variance-component family, and from ACAT-O, which combines a specified family of p-values. STAAR's distinguishing emphasis here is extending such omnibus evidence across annotation-informed weights as well as test types. These methods are not interchangeable; their component families and null calibrations should be understood before comparing results.

Omnibus testing has costs. Combining many choices can reduce power relative to a correctly specified, prespecified single test; changing the annotation collection or combination weights can change the result. A significant p-value does not correct annotation bias, establish causal effects, localize a causal variant, or account automatically for testing many genes. It supports an association conclusion only for the set, phenotype, and analysis that were tested. Annotation-informed tests are most useful when the data and biological rationale support the annotations and when the full test construction is reported transparently.

• Confusing the two kinds of weights. Variant weights \(w_j\) determine variant contributions within component tests; combination weights \(a_k\) determine how component p-values enter the omnibus statistic.
• Comparing raw statistics across annotations. A larger weighted quadratic statistic need not mean a smaller p-value because its null distribution changes with the weights.
• Treating the component p-values as independent. They commonly reuse data and are dependent. The omnibus combination is designed to accommodate broad dependence, but valid component calibration remains essential.
• Over-interpreting a significant result. An omnibus p-value supports evidence against the set-level null for the included tests; it does not name a causal variant or prove that a chosen annotation is correct.