Why Examine Ancestry-Stratified Results?
A gene-level test can show a case-control difference because carrier frequencies differ across ancestry groups and the case and control samples contain different proportions of those groups. In that situation, a pooled comparison can suggest a gene association even when the carrier proportion is similar among cases and controls within each ancestry group. Conversely, a pooled result can conceal associations that differ across groups.
Ancestry-stratified carrier summaries are a diagnostic tool: they show whether a pooled signal is also visible within the groups being compared. They do not, by themselves, establish the cause of an association or guarantee that ancestry has been adequately controlled. Interpretation also depends on how ancestry groups were defined, the precision of the estimates, and the gene-level test used.
For a simple carrier-versus-noncarrier comparison, the stratum-specific odds ratio compares the odds of carrier status in cases with the odds in controls. An odds ratio of 1 indicates equal carrier odds in the two groups in that stratum. The pooled odds ratio instead compares the corresponding odds after combining ancestry strata; because the group composition may differ, it need not represent the within-stratum association.
A Pooled Signal Created by Different Ancestry Composition
Let \(k\) index ancestry strata, \(D\) denote case status, and \(G=1\) denote meeting the carrier rule for a gene. The pooled carrier proportion among cases is a weighted average of the case carrier proportions within strata. Its weights are the proportions of cases in each stratum. The control carrier proportion has its own weights, determined by the ancestry composition of controls.
Even if the within-stratum carrier proportions are equal for cases and controls, the two weighted averages can differ when case and control ancestry compositions differ and carrier proportions vary across strata. The pooled contrast then reflects both the ancestry-specific carrier frequencies and the different weights. This is a form of confounding by ancestry composition, not evidence on its own of a within-ancestry gene association.
Worked Example: A Pooled Difference With No Within-Stratum Difference
Suppose a hypothetical study has two ancestry strata. In stratum A, 8 of 80 cases and 2 of 20 controls meet the carrier rule. In stratum B, 1 of 20 cases and 4 of 80 controls meet it. The carrier proportions match between cases and controls within each stratum: 10% in A and 5% in B.
After pooling, the case carrier proportion is \(9/100=0.09\), and the control carrier proportion is \(6/100=0.06\). The pooled odds ratio is
The pooled comparison suggests higher carrier odds among cases. But within A, the odds ratio is \((8\times18)/(72\times2)=1\), and within B it is \((1\times76)/(19\times4)=1\). Here cases are more often from A, where carriers are more common, while controls are more often from B. The pooled odds ratio is therefore misleading as evidence for a within-stratum association. It does not show that ancestry caused the observed pattern; it shows why the stratum-specific summaries are needed to interpret it.
Comparing Pooled and Adjusted Evidence
A practical diagnosis compares the pooled gene-level result with results that account for ancestry. For a binary phenotype, one possible model for a carrier-based gene test is a logistic regression containing a gene-level carrier term and ancestry covariates. A more complex gene-level test, such as a variant-set test that allows effects to differ across variants, requires its corresponding ancestry-adjusted procedure. A carrier summary is not a replacement for the test's specified model.
Here \(A\) represents the chosen ancestry covariates or stratum indicators, and \(\mathbf{Z}\) represents other prespecified covariates. In this model, \(\beta_G\) is the conditional log odds ratio for carrier status under the model. The model assumes an appropriate link and covariate specification, and that the analyzed observations and genotype calls are suitable for the intended association analysis. Adjusting for ancestry can reduce confounding only to the extent that the ancestry variables capture the relevant structure and are modeled adequately. It does not automatically address other sources of bias or poor data quality.
For stratified \(2\times2\) tables, a Mantel–Haenszel summary odds ratio is one way to combine stratum-specific comparisons. Let \(a_k\) and \(b_k\) be the carrier and noncarrier counts among cases in stratum \(k\), and \(c_k\) and \(d_k\) the corresponding counts among controls. With \(n_k=a_k+b_k+c_k+d_k\), its estimate is
This is a summary conditional on the chosen strata, not a guarantee that the strata adequately control ancestry structure. It is also a single summary: substantial differences between stratum-specific estimates may make that summary hard to interpret.
Worked Example: A Similar Association Within Both Strata
In a hypothetical study, stratum A contains 40 carriers and 160 noncarriers among 200 cases, and 20 carriers and 180 noncarriers among 200 controls. Stratum B contains 10 carriers and 90 noncarriers among 100 cases, and 5 carriers and 95 noncarriers among 100 controls.
The stratum-specific odds ratios are
Both estimates point in the same direction and are of similar magnitude. The pooled carrier counts are 50 of 300 cases and 25 of 300 controls, giving a pooled odds ratio of \((50\times275)/(250\times25)=2.200\). The Mantel–Haenszel estimate is
The similar pooled and stratified estimates are not a diagnosis of confounding; they indicate that this particular comparison does not show the pooled signal disappearing after stratification. These estimates alone do not establish a causal effect or statistical significance. Uncertainty, the gene-level test, and the adequacy of ancestry control still matter.
Residual Structure Within Broad Ancestry Groups
An apparent association within a broad ancestry stratum does not prove that ancestry has been fully controlled. The stratum may contain finer population structure related both to case status and to carrier frequency. For example, a broad group may combine subgroups with different rare-variant frequencies, while cases and controls have different proportions of those subgroups. Principal components or other ancestry measures may help describe continuous structure, but adjustment must still be appropriate to the data and study design.
Worked Example: A Broad Stratum Hides Finer Imbalance
Consider one hypothetical broad ancestry stratum with 500 cases and 500 controls. Within it are high- and low-carrier-frequency subgroups. Among cases, 450 are from the high-frequency subgroup and 50 from the low-frequency subgroup; among controls, the corresponding counts are 50 and 450. Suppose the carrier proportion is 10% in the high-frequency subgroup and 2% in the low-frequency subgroup for both cases and controls.
The resulting counts are 45 carriers among 450 high-frequency cases and 1 among 50 low-frequency cases. Among controls they are 5 among 50 high-frequency participants and 9 among 450 low-frequency participants. Thus, within each finer subgroup, the carrier proportions match: 10% and 2%, respectively. But the broad-stratum totals are 46 carriers among cases and 14 among controls. The broad-stratum odds ratio is
This broad-stratum result is compatible with residual confounding by the finer structure: cases are disproportionately drawn from the subgroup with a higher carrier proportion. The example does not establish that real ancestry measures were inadequate in any particular study. It demonstrates why a broad ancestry label alone may not resolve all population structure.
Sparse Strata and Unstable Gene-Level Results
Rare-variant carrier counts may be very small in one or more ancestry groups. A zero cell can make the ordinary sample odds ratio equal to zero or infinite, and estimates based on one or two carriers can change sharply when a single observation changes. Small counts also limit the information available for detecting differences and make asymptotic approximations less dependable. A large-looking estimate in a sparse stratum should therefore be read alongside its counts and uncertainty, not as strong evidence by itself.
Worked Example: A Zero Cell Does Not Give a Precise Effect
Suppose a hypothetical small ancestry stratum has 2 carriers and 6 noncarriers among 8 cases, and 0 carriers and 40 noncarriers among 40 controls. The ordinary odds ratio is
The denominator is zero, so the ordinary estimate is infinite; it is not a finite, precise estimate of the association's magnitude. If a single control is instead classified as a carrier, the odds ratio becomes \((2\times39)/(6\times1)=13\). That remains large, but it is very different from an infinite estimate and still rests on only three carriers in total. The appropriate interpretation is that the data are sparse and the effect estimate is unstable. An exact or otherwise suitable small-sample analysis may be needed for inference, but no method can make the small stratum contain more information than it does.
A Practical Diagnostic Sequence
When reviewing a gene-level result, inspect counts as well as p-values. A small p-value does not reveal whether the signal is distributed across ancestry strata, concentrated in one group, or driven by a few observations. Use the test that matches the study's gene-level procedure; a simple carrier-based odds ratio is illustrative, but it may not represent a multi-variant burden or kernel test.
Record case and control totals for each ancestry group. Note groups with very few participants or carriers before interpreting estimates.
For each group, report carrier and noncarrier counts and the corresponding proportions for cases and controls. Do not report proportions without their denominators.
Check whether the pooled association weakens, persists, or varies across strata after appropriate adjustment. Examine uncertainty as well as point estimates.
Consider whether the chosen categories or continuous ancestry covariates may leave finer structure uncontrolled. Interpret the result with the limitations of those measures in view.
Describe the pattern as consistent with, or not suggestive of, ancestry confounding in the examined summaries. Do not treat the diagnosis as proof of a causal mechanism.
Common Pitfall
Wrong reasoning: “The pooled gene-level p-value is small, so the gene is associated with case status regardless of ancestry.” This skips the possibility that ancestry composition and ancestry-specific carrier frequencies generated the pooled contrast.
Correction: Examine ancestry-stratified carrier counts and compare the pooled result with an appropriately ancestry-adjusted analysis. If the evidence is concentrated in a sparse group, or changes substantially after adjustment, report that limitation. If the signal persists within well-supported strata, that is more reassuring about this particular confounding concern, but it is not proof against all confounding or proof of causation.
Check Your Understanding
Use the definitions and diagnostic logic in this tutorial to answer the following questions.
- Why can the pooled case carrier proportion differ from the pooled control carrier proportion when the within-stratum carrier proportions match?
- What does it suggest if a pooled odds ratio is large but stratum-specific estimates are close to 1?
- Why might an association within a broad ancestry group still reflect residual population structure?
- In a sparse stratum with a zero cell, why should an infinite ordinary odds-ratio estimate not be interpreted as a precise effect?
- If an ancestry-adjusted gene-level result remains associated, what does that establish—and what does it not establish?