Why allele frequency affects association power
A genetic association can be estimated less precisely in one ancestry group than another even when the underlying per-allele effect is the same. One reason is that allele frequencies differ: if nearly everyone has the same genotype, there is little genotype variation with which to estimate an association. Less variation generally means a larger standard error and lower power at a given sample size.
This is a statement about precision, not automatically about biology. A smaller test statistic or a non-significant result in one group does not, by itself, show that the genetic effect differs between groups. To make a fair comparison, keep the effect scale and allele coding consistent, and account for the amount of genotype information available in each group.
From genotype variance to standard error
For a quantitative outcome, consider the within-group linear regression model \(Y_i=\mu+\beta G_i+\varepsilon_i\), where \(Y_i\) is the outcome for participant \(i\), \(\mu\) is an intercept, and \(\beta\) is the mean outcome difference per additional effect-allele copy. Assume the regression model is appropriate, the errors have mean zero conditional on genotype, have common variance \(\sigma^2\), and are independent across participants. For the simple model with an intercept, the conditional variance of the least-squares estimate is \(\sigma^2/S_{GG}\), where \(S_{GG}=\sum_i(G_i-\bar{G})^2\). Thus its standard error is \(\sigma/\sqrt{S_{GG}}\).
If genotypes are sampled from a population in Hardy–Weinberg equilibrium, their variance is \(2p(1-p)\), and \(S_{GG}\) is approximately \(n\,2p(1-p)\) in a sufficiently large sample. This gives the useful planning approximation \(SE(\hat{\beta})\approx\sigma/\sqrt{n\,2p(1-p)}\). The approximation uses the population genotype variance in place of the observed sample variance; it is not a substitute for calculating the actual regression standard error when the data are available.
The quantity \(n\,2p(1-p)\) summarizes how much genotype variation is expected in the sample. It is sometimes useful to compare groups through this quantity, but it is not a literal count of independent carriers. In a model that includes covariates, information about \(G\) depends on the variation left after accounting for those covariates. For a fixed design matrix \(C\), define \(M_C=I-C(C^\mathsf{T}C)^{-1}C^\mathsf{T}\), assuming the columns of \(C\) are linearly independent. In the linear model, the corresponding variance is \(\sigma^2/(G^\mathsf{T}M_CG)\). If genotype is strongly predictable from the covariates, this residualized genotype variation can be much smaller than its unadjusted variance.
Worked examples
Worked Example: Comparing precision at two allele frequencies
Consider hypothetical independent ancestry groups, each with \(n=10{,}000\) participants. Suppose the same quantitative outcome has residual standard deviation \(\sigma=1\) outcome unit in each group, the per-allele effect is \(\beta=0.10\) outcome units per allele, and genotypes follow Hardy–Weinberg equilibrium. Ignore covariates for this planning calculation.
In the first group, let \(p=0.30\). Its genotype variance is \(2(0.30)(0.70)=0.42\), so the expected total genotype variation is \(10{,}000(0.42)=4{,}200\). The approximate standard error is \(1/\sqrt{4{,}200}=0.01543\) outcome units per allele.
In the second group, let \(p=0.05\). Its genotype variance is \(2(0.05)(0.95)=0.095\), giving expected total genotype variation \(10{,}000(0.095)=950\). The approximate standard error is \(1/\sqrt{950}=0.0324443\), or \(0.03244\) outcome units per allele.
The expected standardized effects, \(\beta/SE\), are approximately \(0.10/0.01543=6.48\) and \(0.10/0.0324443=3.08\), respectively. At a two-sided significance level of \(\alpha=0.05\), a normal-approximation test has substantially greater power in the first group. The effect was set to be identical in both groups; the difference in expected evidence arises from genotype variation, not from a different \(\beta\).
Worked Example: A noisy estimate is not proof of a different effect
Suppose two hypothetical, independent groups each have \(n=5{,}000\), and each has residual standard deviation \(\sigma=1\) outcome unit. Assume Hardy–Weinberg equilibrium and no covariate adjustment. In the first group, \(p=0.25\); in the second, \(p=0.05\). For illustration, suppose the observed per-allele estimates are \(0.15\) and \(0.05\) outcome units per allele.
For \(p=0.25\), the genotype variance is \(2(0.25)(0.75)=0.375\), and the approximate standard error is \(1/\sqrt{5{,}000(0.375)}=1/\sqrt{1{,}875}=0.02309\). For \(p=0.05\), the variance is \(2(0.05)(0.95)=0.095\), and the standard error is \(1/\sqrt{5{,}000(0.095)}=1/\sqrt{475}=0.04588\) outcome units per allele.
The estimates differ by \(0.15-0.05=0.10\) outcome units per allele. Under the stated independence assumption, the standard error of their difference is approximately \(\sqrt{0.02309^2+0.04588^2}=\sqrt{1/1{,}875+1/475}=0.05137\) outcome units per allele. The observed difference is therefore about \(0.10/0.05137=1.95\) such standard errors.
This calculation illustrates how sampling variation can produce different point estimates, especially when one estimate is less precise. It does not establish that the true effects are equal or different. In particular, a small or non-significant association estimate in the lower-frequency group would not demonstrate that the effect is absent there. The calculation also relies on independent estimates; shared participants or correlated estimates require a covariance term.
Worked Example: Planning sample size for equal power
Suppose a hypothetical study seeks 80% power for a two-sided test at \(\alpha=0.05\) to detect a per-allele effect of \(0.10\) outcome units, with residual standard deviation \(\sigma=1\) outcome unit. Assume the simple linear model, Hardy–Weinberg equilibrium, and no covariate adjustment. For a normal approximation, an 80% target requires a noncentrality parameter of about \(z_{0.975}+z_{0.80}=1.96+0.842=2.802\).
Setting \(\beta\sqrt{n\,2p(1-p)}/\sigma\) equal to 2.802 and solving for \(n\) gives \(n\approx 2.802^2\sigma^2/\{\beta^2\,2p(1-p)\}\). At \(p=0.25\), this is \(2.802^2/(0.10^2\times0.375)=7.851204/0.00375\), or about 2,094 participants after rounding up. At \(p=0.05\), it is \(2.802^2/(0.10^2\times0.095)=7.851204/0.00095\), or about 8,265 participants after rounding up.
Under these assumptions, the lower-frequency group needs nearly four times as many participants to reach the same approximate power for the same effect. This is a planning approximation: exact power depends on the test, the realized genotype distribution, and whether the model and variance assumptions hold.
Justification and interpretation
The standard-error relationship follows directly from least squares. With an intercept and a fixed genotype vector, the slope estimate is \(\hat{\beta}=\sum_i(G_i-\bar{G})(Y_i-\bar{Y})/S_{GG}\). Substituting \(Y_i=\mu+\beta G_i+\varepsilon_i\) cancels the intercept and gives \(\hat{\beta}-\beta=\sum_i(G_i-\bar{G})\varepsilon_i/S_{GG}\). Conditional on the genotypes, independent errors with variance \(\sigma^2\) therefore imply \(\operatorname{Var}(\hat{\beta}\mid G)=\sigma^2/S_{GG}\). If genotype variance is \(2p(1-p)\), then the expected scale of \(S_{GG}\) grows with \(n\,2p(1-p)\), explaining the standard-error approximation.
For a two-sided Wald test under the same approximate normal model, the test statistic under a true effect \(\beta\) is approximately \(N(\delta,1)\), where \(\delta=\beta/SE(\hat{\beta})\). With critical value \(z_{1-\alpha/2}\), the approximate power is \(P(Z>z_{1-\alpha/2})+P(Z<-z_{1-\alpha/2})\) for \(Z\sim N(\delta,1)\). Equivalently, it is \(1-\Phi(z_{1-\alpha/2}-\delta)+\Phi(-z_{1-\alpha/2}-\delta)\), where \(\Phi\) is the standard normal cumulative distribution function. For a fixed effect and residual variance, lower genotype variance reduces \(|\delta|\) and usually reduces power.
Assumptions, diagnostics, and alternatives
The formula \(2p(1-p)\) describes genotype variance under Hardy–Weinberg equilibrium: the two allele copies are independently sampled with allele frequency \(p\). If that assumption is not appropriate, use the observed genotype variance rather than treating the formula as guaranteed. Hardy–Weinberg deviations may arise for several reasons, including technical problems, relatedness, or population structure; a deviation is a prompt for investigation, not a diagnosis on its own.
In real association models, covariates and other design features matter. Ancestry adjustment can reduce the genotype variation available for estimating a within-group effect if genotype is strongly associated with the adjustment variables. The relevant quantity in the linear model is then residualized genotype variation, not simply \(n\,2p(1-p)\). For a binary outcome or a different regression model, information depends on that model's likelihood and outcome distribution; the linear-model standard-error formula should not be applied unchanged.
Useful checks include comparing allele frequencies and genotype counts by group, inspecting the observed genotype variance, checking missingness and dosage quality, and calculating standard errors from the fitted model. At low allele frequency, very few individuals may carry two copies of the allele. A normal approximation may then be unreliable, and dosage uncertainty or genotype errors can further reduce information. Strong correlation between genotype and covariates is another warning that the simple planning calculation may overstate precision.
| Approach | What it uses | When it is useful or limited |
|---|---|---|
| Hardy–Weinberg planning approximation | \(n\), \(p\), and the assumed residual variance | Useful for comparing expected precision or planning sample sizes; depends on the model and equilibrium assumptions. |
| Observed genotype variance | The variation in recorded genotype dosages | Reflects the sample more directly; does not alone account for covariate adjustment or dosage uncertainty. |
| Fitted-model standard error | The model's information after covariate adjustment | Usually the appropriate precision summary for the specified association model; can still be unreliable if the model or data quality is poor. |
A key distinction is between a frequency difference and an effect difference. If the same effect allele, per-allele scale, outcome definition, and regression model are used, changing \(p\) changes the expected genotype variance and hence precision; it does not mechanically change the true regression coefficient. Effects can genuinely differ across groups, but frequency differences alone cannot establish that. Nor can a significant result in one group and a non-significant result in another establish a difference: those outcomes can occur simply because the groups have different standard errors.
Common mistake
Wrong reasoning: “The association is significant in the group with the common allele and not significant in the group with the rarer allele, so the genetic effect must be weaker in the latter.” This compares significance labels rather than effect estimates and their uncertainty.
Correction: First compare the estimated effects on a consistent per-allele scale and examine their standard errors. A lower allele frequency can mean less genotype variation and lower power even if the true effects are equal. Conversely, frequency alone does not rule out real effect differences. The evidence must be interpreted using the model, its assumptions, and the uncertainty of the estimates.
Check Your Understanding
Try each question, then check the solutions below.
- Under Hardy–Weinberg equilibrium, what is the genotype variance when \(p=0.10\), and how does it compare with the variance at \(p=0.50\)?
- For \(n=2{,}000\), \(p=0.20\), and \(\sigma=2\) outcome units, calculate the approximate standard error of the per-allele slope.
- Two groups have the same sample size and residual variance, but one has a lower allele frequency. What is the expected direction of the difference in standard errors, under the stated model assumptions?
- Why does significance in one ancestry group but not another fail to show that the true effects differ?
Solutions: (1) The variance is \(2(0.10)(0.90)=0.18\); at \(p=0.50\), it is \(2(0.50)(0.50)=0.50\), the maximum. (2) The approximation gives \(2/\sqrt{2{,}000(0.18)}=2/\sqrt{360}=0.1054\) outcome units per allele. (3) If the lower frequency is farther from 0.50, its genotype variance is smaller, so its standard error is expected to be larger, all else equal. (4) Significance depends on both effect size and precision. Different standard errors can produce different significance outcomes even when the true effects are equal.