Ancestry Matching When Rare Allele Frequencies Differ

Why Ancestry Matching Matters for Rare-Variant Studies

A gene-level test combines information across variants, often to detect a difference in rare-variant burden between cases and controls. But the variants in a gene need not have the same frequencies in every ancestry group. If cases and controls have different ancestry compositions, their pooled burdens can differ even when cases and controls have the same burden within each ancestry group. A gene-level test may then reflect the recruitment imbalance rather than an association with the phenotype.

This is a design problem as well as an analysis concern. Planning recruitment so cases and controls have comparable ancestry composition can reduce the opportunity for this particular spurious signal. The plan must also identify the population the study is intended to represent: matching changes which comparisons are well supported, and can change the population to which a result applies.

How Unequal Ancestry Composition Creates a Pooled Difference

Let \(k\) index ancestry strata. Write \(\pi_{Dk}\) for the proportion of cases in stratum \(k\), and \(\pi_{Ck}\) for the proportion of controls there. For a gene containing a specified set of variants, let \(Q_k\) be the sum of their allele frequencies in stratum \(k\). If the burden is the total number of alternate alleles across those variants, the expected burden per diploid person in stratum \(k\) is \(2Q_k\). This expectation follows by adding the expected allele count, \(2q\), for each variant with allele frequency \(q\); it does not require the variants to be independent.

Suppose the gene has no case-control difference within any stratum, so that cases and controls in stratum \(k\) have the same expected burden \(2Q_k\). The pooled case-control difference in expected burden is nevertheless

The equation shows the design mechanism directly: a pooled difference can arise when the stratum proportions differ and the gene’s aggregate allele frequency also differs across strata. This is not evidence that ancestry itself causes the phenotype, nor does it establish that a particular observed result is confounded. It shows why matching or stratifying recruitment is important before interpreting a pooled gene-level signal.

Designing Recruitment Strata and Matching

An ancestry-matching plan starts by choosing strata relevant to the study population and the available data. Researchers may use genetic ancestry information, carefully selected recruitment populations, or both. Broad self-identified categories alone may not capture genetic differences relevant to allele frequencies, while very fine categories can leave too few cases or controls per stratum. The categories and their purpose should be stated in advance, and recruitment should use consistent definitions for cases and controls.

When the goal is to compare cases with controls at the case ancestry distribution, a direct allocation rule is to make the control proportions match the case proportions. If there are \(K\) strata, \(n_{Dk}\) cases in stratum \(k\), \(n_D=\sum_k n_{Dk}\) total cases, and a planned total of \(n_C\) controls, the target control allocation is

In practice, these values may not be whole numbers, so allocations are rounded while keeping the overall total and the intended proportions as close as feasible. If exact matching is not possible, the discrepancy should be recorded and the analysis plan should compare cases and controls within strata or standardize to a clearly stated common ancestry distribution. Matching can be done by stratum or, where feasible, by selecting controls with similar ancestry profiles to cases. It does not require pairing each case with one control.

Matching, Stratified Recruitment, and the Target Population

Matching and stratified recruitment are related but not identical ideas. Matching sets a goal for comparability between cases and controls. Stratified recruitment ensures that enough participants are recruited within each relevant group to make within-stratum comparisons possible. A study may recruit fixed numbers from each stratum, deliberately oversample a smaller group, or follow the case proportions. The right choice depends on the scientific target and recruitment feasibility.

If sampling deliberately changes the ancestry composition, a crude pooled contrast gives more influence to strata that are more common in the sample. To describe a target population with known stratum weights \(w_k\), one can instead form a standardized contrast by combining within-stratum contrasts using those weights. The weights are nonnegative and sum to one. For a burden outcome, let \(\mu_{Dk}\) and \(\mu_{Ck}\) be the expected case and control burdens in stratum \(k\). The standardized contrast is

This formula describes a comparison standardized to the chosen stratum distribution. It does not make the data representative of a target population automatically: each stratum needs adequate observations, the within-stratum comparisons must be meaningful, and the target weights must be justified. Matching to the cases’ distribution and standardizing to a broader population are different design and interpretation choices.

Justification and Practical Planning

The mixture equation also gives a useful design justification. If cases and controls have the same ancestry composition, then \(\pi_{Dk}=\pi_{Ck}\) for every stratum. Substituting this equality into the pooled-difference expression gives

Thus, under the stated within-stratum null and comparable measurement, exact balance removes this particular composition-driven difference in expected burden. It does not rule out other sources of a gene-level signal, such as technical differences or genuine within-stratum association. Nor does it guarantee that every individual variant has been balanced if the chosen strata are too broad to capture relevant population structure.

Before recruitment, investigators should check likely case and control availability by stratum, define how ancestry will be assessed, and decide whether the target is the case distribution or another population distribution. They should plan for sufficient numbers in each stratum rather than relying only on a large overall sample. If rare strata are undersampled, estimates can be unstable; if strata are excessively fine, recruitment may be impractical. Matching is a balance between capturing frequency-relevant structure and retaining enough participants to estimate within-stratum contrasts.

Common Pitfalls

• Assuming the pooled difference must be biological. A larger pooled burden in cases is not by itself evidence of disease association. Check whether case and control ancestry compositions differ and whether allele frequencies could vary across those groups.
• Treating broad labels as precise genetic strata. A convenient category may not capture the ancestry variation relevant to rare-allele frequencies. State how strata are defined and what information they are intended to represent.
• Matching overall totals but ignoring stratum counts. Equal total numbers of cases and controls do not imply ancestry balance. Compare the distribution across strata, not just the total sample sizes.
• Changing the target population without saying so. Equal recruitment from strata can be efficient for within-stratum comparisons, but an equally weighted pooled result may not describe a population in which the strata are unequally represented. State the target weights and interpret the result accordingly.
• Assuming matching solves every problem. Ancestry balance addresses one design pathway for spurious association. It does not ensure comparable sequencing, callability, phenotype measurement, or other aspects of the study.