Admixture Mapping Versus SNP-Based Association

Two Association Tests, Two Different Questions

In an admixed cohort, people inherit chromosome segments from different ancestral populations. At a particular genomic region, the number of chromosome copies assigned to one ancestral source can vary between people. That variation gives rise to admixture mapping: testing whether local ancestry at a region is associated with a phenotype. A different question is whether a particular allele is associated with the phenotype. That is the focus of SNP-based association.

These tests can produce different results at the same region. Local ancestry can be associated with a phenotype because a causal allele differs in frequency between ancestral source populations, even when that causal allele was not directly measured. A SNP test instead asks whether variation in the tested SNP’s allele count is associated with the phenotype. It can detect an association when that SNP is causal or tags a causal variant, whether or not it is informative about ancestry.

The distinction matters for interpreting results. A local ancestry association is not a claim that ancestry itself causes a phenotype, nor does it identify a causal variant. A SNP association is not proof that the tested SNP is causal. Both are statistical associations, conditional on the covariates and model used.

Definition: At a genomic region, let \(L_i\) be the number of the two chromosome copies in individual \(i\) assigned to a specified ancestral source, so \(L_i\) is a local ancestry dosage of 0, 1, or 2. Let \(G_i\) be the number of copies of a specified effect allele at a tested SNP, also coded 0, 1, or 2. An admixture-mapping test evaluates association with \(L_i\); a SNP-based association test evaluates association with \(G_i\).

Null Hypotheses and Model Setup

For a continuous phenotype, a simple adjusted linear model makes the contrast explicit. Let \(Y_i\) be the phenotype for person \(i\), and let \(C_i\) denote covariates other than the intercept, such as global ancestry and relevant technical or clinical variables. The covariates should be chosen for the study design and the intended association estimand.

Y_i = \alpha + \beta_L L_i + \gamma^\mathsf{T} C_i + \varepsilon_i

In this model, the admixture-mapping null hypothesis is \(H_0:\beta_L=0\); the alternative is \(H_a:\beta_L\ne 0\) for a two-sided test. The coefficient \(\beta_L\) describes the modeled mean phenotype difference per additional chromosome copy assigned to the specified source, holding the included covariates fixed.

For a SNP-based test, replace \(L_i\) by \(G_i\):

Y_i = \alpha + \beta_G G_i + \gamma^\mathsf{T} C_i + \varepsilon_i

Here the SNP null is \(H_0:\beta_G=0\), and \(\beta_G\) is the modeled mean phenotype difference per additional copy of the effect allele, conditional on the covariates. These are separate one-predictor tests. If local ancestry and genotype are both included, the coefficient for either predictor has a conditional interpretation given the other; that joint model asks a different question and is not required to define the basic comparison.

For the usual ordinary least-squares t-test to have its exact small-sample reference distribution, the model must have a correctly specified conditional mean, independent errors with mean zero and constant variance, and normally distributed errors. The design matrix must have full column rank, so the tested predictor is not an exact linear combination of the other predictors. In genomic studies, relatedness, residual heteroskedasticity, or other dependence may call for a mixed model or another variance procedure rather than the simple independent-error model. For a binary phenotype, a logistic model is typically used; the tested coefficient is then on the log-odds scale, and its null is still that the relevant predictor coefficient is zero.

What Each Test Can Signal

Feature	Admixture mapping	SNP-based association
Tested predictor	Local ancestry dosage at a region	Effect-allele count at a particular SNP
Null hypothesis	No conditional association with local ancestry dosage	No conditional association with SNP allele dosage
Typical signal source	A region where ancestry-linked genetic differences track the phenotype	A tested or tagged allele whose dosage tracks the phenotype
Key marker information	Markers informative about ancestral origin to infer local ancestry across the genome	Genotype data for the tested SNP, measured or imputed with adequate quality
What a significant result does not establish	The causal variant, causal mechanism, or a causal effect of ancestry	That the tested SNP itself is causal

An ancestry-informative marker (AIM) has allele frequencies that differ between the ancestral source populations relevant to the analysis, making it useful for distinguishing their contributions. An admixture-mapping scan generally needs enough well-distributed, informative marker data to estimate local ancestry along the genome. It does not require that every causal variant be genotyped. The signal may therefore localize a region without naming the responsible allele.

In contrast, the SNP used in a SNP-based test need not be an AIM. The important requirement is accurate genotype information at that SNP, or a sufficiently accurate imputation, and enough variation in its allele dosage to test association. A SNP with little frequency difference between source populations can still be associated with the phenotype. Conversely, an AIM can help infer ancestry but need not itself be associated with the phenotype.

Worked Example: Testing Local Ancestry Dosage

Suppose a hypothetical study analyzes a continuous biomarker in \(n=51\) unrelated participants. At one region, \(L_i\) counts chromosome copies assigned to source A. The fitted model includes an intercept, \(L_i\), and one global ancestry covariate. Assume the linear model’s conditions for the ordinary t-test are reasonable. The estimated local ancestry coefficient is \(0.40\) biomarker units per source-A chromosome copy, with \(SE=0.16\).

There are three fitted coefficients, so the residual degrees of freedom are \(df=51-3=48\). To test \(H_0:\beta_L=0\), the test statistic is:

t = \frac{\hat{\beta}_L-0}{SE(\hat{\beta}_L)} = \frac{0.40}{0.16} = 2.50

For a two-sided test with \(df=48\), the p-value is \(2P(T_{48}\ge 2.50)\), approximately \(0.01589\). At the nominal level \(\alpha=0.05\), this is evidence against the no-local-ancestry-association null for this region. It is not, by itself, evidence that the result meets a genome-wide significance threshold after scanning many regions.

Using the \(t\) critical value of approximately \(2.011\), a 95% confidence interval is:

0.40 \pm 2.011(0.16) = 0.40 \pm 0.3218 = [0.0782,\ 0.7218]

Under the model, the interval estimates a positive conditional mean difference of about \(0.0782\) to \(0.7218\) biomarker units per additional source-A chromosome copy. The interval does not identify a causal variant or prove that the ancestry-associated region is the biological mechanism.

Worked Example: A Different Result From a SNP Test

In the same hypothetical cohort and with the same covariates, suppose a tested SNP has an estimated effect-allele coefficient of \(0.18\) biomarker units per allele copy, with \(SE=0.12\). Assume the SNP genotypes vary in the cohort, the design matrix has full rank, and the model conditions stated above are adequate.

The SNP test statistic is \(t=0.18/0.12=1.50\), with \(df=48\). The 95% interval is:

0.18 \pm 2.011(0.12) = 0.18 \pm 0.2413 = [-0.0613,\ 0.4213]

This interval includes zero, so the two-sided test does not reject \(H_0:\beta_G=0\) at the nominal 0.05 level. The estimate is a conditional difference per effect-allele copy, not per ancestry copy. It is therefore not contradictory for the local ancestry test to be significant while this SNP test is not: the tests use different predictors and target different associations. The SNP may not tag the relevant causal variation well, or the local ancestry signal may reflect a combination of variants rather than this SNP.

Why Ancestry-Informative Markers Matter

Local ancestry is inferred from genetic marker patterns, not observed directly as a phenotype measurement. Marker informativeness affects how well those patterns distinguish ancestral sources. If marker allele frequencies are very similar across sources, the marker contributes little information about which source transmitted a chromosome segment. Reliable admixture mapping therefore depends on an appropriate set of reference data and informative markers distributed sufficiently across the genome.

This requirement differs from the SNP test’s. A poorly informative marker set can make local ancestry estimates imprecise even when genotypes at a particular SNP are excellent. Conversely, local ancestry may be estimated adequately in a region while the SNP selected for association is missing, poorly imputed, or too rare to estimate its effect precisely. Marker quality and sample size affect power, but they do not change what the two null hypotheses mean.

Worked Example: Choosing Marker Data for the Question

Consider a hypothetical two-source setting. At marker M1, the allele frequency is \(0.70\) in source A and \(0.20\) in source B. The absolute frequency difference is \( |0.70-0.20|=0.50\), so M1 is potentially informative for distinguishing the sources. At marker M2, the corresponding frequencies are \(0.52\) and \(0.48\), giving \( |0.52-0.48|=0.04\); M2 contributes much less ancestry discrimination on its own.

If the research question is whether a region’s local ancestry dosage tracks a phenotype, the analysis needs an adequate set of informative markers such as M1, distributed so that ancestry can be inferred across the region. A single informative marker is not automatically sufficient to establish ancestry along an entire segment. If the question instead concerns association with the allele at a particular SNP, the study needs accurate genotypes at that SNP and enough allele variation. M2 could still be a valid SNP association target if it is directly tested and varies in the cohort; its small source-frequency difference does not make the SNP test invalid.

Derivation and Interpretation of the Regression Test

The t-test used above follows from the ordinary least-squares model. Let \(X\) be the design matrix containing the intercept, the tested predictor, and the covariates, and let \(q\) be its rank. Under the stated linear-model assumptions, the least-squares coefficient estimate has sampling variance proportional to the error variance. Estimating that variance from residuals gives the coefficient standard error. Under \(H_0\), the estimate divided by its standard error has a t distribution with \(n-q\) degrees of freedom:

t = \frac{\hat{\beta}-0}{SE(\hat{\beta})} \sim t_{n-q} \quad\text{under }H_0

Here \(\hat{\beta}\) is \(\hat{\beta}_L\) for an admixture-mapping test or \(\hat{\beta}_G\) for a SNP test. The numerator is the estimated conditional effect under the fitted model, and the denominator expresses its uncertainty in the same units. This provides a test of each predictor’s coefficient, not a direct comparison of whether one biological mechanism is more important than another.

Diagnostics, Failure Modes, and Alternatives

Before interpreting either test, check that the predictor is estimable and measured well enough for the question. For a local ancestry test, inspect whether the ancestry-informative marker coverage and reference data are appropriate for the source populations represented in the cohort. Poor coverage or mismatched references can impair ancestry inference. For a SNP test, examine genotype or imputation quality, allele frequency, and the number of informative genotype observations. A coefficient based on very little variation can be unstable.

For both tests, check model fit and calibration: review residual patterns for a continuous-trait linear model, assess whether covariates are appropriate, and account for relatedness or other dependence when needed. Compare ancestry distributions and predictor variation across the analyzed participants, and investigate whether a tested predictor is strongly collinear with included covariates. A nominal p-value from a genome-wide scan must also be interpreted in light of the many regions or SNPs tested and the dependence among tests.

The main alternative is not simply “use the other test.” SNP-based association can offer variant-level resolution when the relevant allele is measured or well tagged. Admixture mapping can prioritize a broader region when ancestry-linked allele-frequency differences create a detectable local ancestry signal, including when the causal variant is not directly observed. The methods can be complementary, but their results answer different questions. A joint analysis may be useful when the scientific aim is to assess local ancestry and genotype together; interpreting such conditional effects requires attention to their correlation and to whether the data can distinguish them.

Common Mistake

Wrong reasoning: “The local ancestry test is significant, so ancestry causes the phenotype and the causal SNP must be ancestry-informative.”
Correction: The test shows an association between local ancestry dosage and the phenotype conditional on the model’s covariates. It can reflect unmeasured variants whose frequencies differ by ancestry, residual confounding, or other model limitations. It neither establishes a causal effect of ancestry nor requires the causal SNP itself to be an AIM.
Wrong reasoning: “A SNP is not an AIM, so it cannot be used in association testing.”
Correction: SNP association tests the SNP’s genotype dosage. Ancestral frequency contrast is relevant to ancestry inference, not a prerequisite for a valid test of genotype association.

Key takeaway: Admixture mapping tests local ancestry dosage; SNP-based association tests a particular allele dosage. Their null hypotheses, signals, marker requirements, and effect interpretations differ. Neither test alone proves causation or identifies every causal variant.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions.

In a local ancestry model, what does the null hypothesis \(H_0:\beta_L=0\) state?
Why can a SNP be useful for SNP-based association even if its allele frequencies are similar in the ancestral source populations?
A study has well-genotyped SNPs but sparse, weakly informative ancestry markers. Which test’s key data requirement is more directly at risk, and why?
Why does a significant local ancestry association not establish that ancestry causes the phenotype?

Solutions. (1) Conditional on the modeled covariates, the phenotype’s mean does not change with local ancestry dosage under the specified model. (2) The SNP test uses the tested allele dosage; a source-population frequency difference is not required, although enough genotype variation and accurate measurement are needed. (3) Admixture mapping is more directly at risk because it requires marker information that distinguishes ancestral origin across the genome. (4) Local ancestry can track unmeasured variants or residual confounding; the association is not by itself a causal effect of ancestry.