Region-Based Burden Tests For Rare Structural Variants

Why Test a Regional Burden?

A rare structural variant may be too uncommon to assess reliably on its own. A regional burden test asks a different question: is carrying at least one qualifying structural variant in a defined gene or genomic region associated with the outcome? This can combine evidence across multiple rare events, but the result depends on exactly which events count and how carriers are defined.

A burden test is therefore not simply a count of every call near a gene. The region, event classes, quality requirements, size limits, and overlap rule must be specified before examining the association results. The analysis then reduces each participant’s qualifying calls to a regional carrier indicator. This tutorial focuses on that unweighted, binary carrier burden; it does not assign different weights to different variants.

Specify the Region and Inclusion Rules

The genomic region and event rules determine the meaning of \(B_i\). A reproducible analysis states the genome build and interval coordinates, including whether the endpoints are included; the gene or region definition; the eligible structural-variant classes; any size thresholds; the variant-level quality criteria; and what it means for an event to overlap the region. For example, an analysis might use a fixed gene interval, retain only quality-filtered deletions of at least 50 base pairs, and count a deletion if its harmonized interval overlaps the gene by at least one base pair.

The exact choices depend on the scientific question and the reliability of the calls. What matters for interpretation is that they are explicit and applied consistently to all participants. Calls that fail the rules do not contribute to the burden. If a participant has several qualifying calls, the binary burden remains \(B_i=1\): the usual carrier-burden coding does not count that participant several times. A participant without a usable genotype for the region should not automatically be treated as a noncarrier; the analysis needs a prespecified missingness rule.

Model and Hypothesis for a Binary Outcome

For a binary outcome \(Y_i\), such as case status, a common model is logistic regression. Let \(C_i\) denote the prespecified covariates for participant \(i\), such as age, sex, or ancestry-related variables appropriate to the study design. The model is

Here \(\operatorname{logit}(p)=\log\{p/(1-p)\}\). The coefficient \(\beta_B\) is the regional carrier effect conditional on the included covariates. The null hypothesis of no regional association is \(H_0:\beta_B=0\); a two-sided alternative is \(H_a:\beta_B\ne 0\). A test may use a likelihood-ratio, score, or Wald statistic, with calibration chosen to suit the sample size and carrier count. In a sufficiently large sample under the model, the usual one-degree-of-freedom reference distribution is appropriate; sparse-carrier settings require particular care with that approximation.

The effect measure is the odds ratio \(\exp(\beta_B)\). It compares the odds of the outcome for carriers with the odds for noncarriers who have the same covariate values under the model. It is not a risk ratio, and an association does not establish that carrying a variant caused the outcome.

Worked Examples of Regional Association

Why the Burden Definition Matters

A regional result is a result for the chosen set of qualifying events and the carrier indicator derived from them. If the region boundary, quality thresholds, event classes, or overlap rule changes, the set of carriers can change, and so can the estimate and test. Those choices should therefore be reported alongside the association result, rather than treating a gene name alone as a complete definition of the analysis.

The burden also makes a substantive assumption: it combines the included events into one carrier-versus-noncarrier contrast. This is most interpretable when a single regional effect is a useful summary of the qualifying events. A positive estimate does not identify which event drives the association, and a null estimate does not show that every included event has no effect. The test concerns the aggregate indicator, not each individual structural variant.

A Short Derivation of the Table-Based Estimate

For a \(2\times2\) table with \(a\) carrier cases, \(b\) carrier controls, \(c\) noncarrier cases, and \(d\) noncarrier controls, the estimated carrier odds among cases are \(a/c\), and the estimated carrier odds among controls are \(b/d\). Their ratio is

For independent cell counts in a sufficiently large sample, the approximate variance of the log odds ratio is the sum of the reciprocals of the four cell counts. One way to see the form is to write \(\log(\widehat{\mathrm{OR}})=\log(a)+\log(d)-\log(b)-\log(c)\). The large-sample variance of the log of a count is approximately the reciprocal of its expected count; adding the four contributions gives

Taking the square root yields the standard error used for the approximate Wald interval. The approximation depends on adequate cell information. When a cell is very small, the interval and test can be poorly calibrated, so sparse-carrier methods or an appropriately chosen likelihood-based analysis may be preferable.

Common Mistakes

• Changing inclusion rules after seeing the association. Choosing a boundary or quality threshold because it gives a smaller p-value undermines the stated test. Define and document the rules before interpreting the association.
• Counting calls instead of participants. Treating someone with three qualifying calls as three independent carriers changes the unit of analysis. For a binary regional burden, that participant has \(B_i=1\), once.
• Calling an odds ratio a risk ratio. An odds ratio of 2.10 means the modeled odds are 2.10 times as high, not that the probability or risk is necessarily 2.10 times as high.
• Interpreting a non-significant result as proof of no effect. A large p-value does not establish equivalence or rule out meaningful effects; consider the interval and the information available from the carrier count.
• Treating every call without a positive burden as a noncarrier. A failed or missing genotype is not automatically evidence of absence. Apply a clear genotype-usability rule before assigning \(B_i=0\).