Why Single-Variant Tests Have Low Power for Rare Alleles

Why Rarity Limits Single-Variant Tests

A single-variant association test asks whether the genotype at one variant is associated with a trait. When the allele is rare, only a small fraction of participants are expected to carry it. As a result, the data contain relatively little information for estimating its association, even if the per-allele effect is substantial. Increasing sample size can help, but the gain depends on how rare the allele is and how large its effect is.

The key distinction is between an effect and the evidence available to estimate it. A large effect does not ensure high power if very few participants carry the allele. Conversely, a less rare allele can yield more informative genotype variation in the same sample, making an effect easier to detect.

Allele Frequency and the Number of Carriers

For an autosomal variant under Hardy–Weinberg equilibrium, the genotype probabilities are \((1-p)^2\) for zero copies, \(2p(1-p)\) for one copy, and \(p^2\) for two copies. The probability of carrying at least one copy is therefore \(2p(1-p)+p^2=2p-p^2\). In a sample of \(N\) independent participants, the expected number of carriers is \(N(2p-p^2)\). For a rare allele, \(p\) is small, so this is approximately \(2Np\).

An expected count is not a guarantee about the observed sample. A sample can contain fewer carriers than expected, or none. Under the same assumptions, the probability that one participant is not a carrier is \((1-p)^2\); the probability that no one in the sample carries the allele is consequently \((1-p)^{2N}\). This possibility is important when interpreting power calculations based on a smooth, large-sample approximation.

How Genotype Variation Determines Information

To make the link between rarity and power precise, consider a quantitative trait and an additive linear model. For each participant, write \(Y=\mu+\beta G+\varepsilon\), where \(\mu\) is the intercept, \(\beta\) is the mean trait difference per additional allele copy, and \(\varepsilon\) is an independent error with mean zero and variance \(\sigma^2\). Assume participants are independent, genotypes are measured without error, and the model's mean relationship is linear in dosage. For the usual standard-error calculation, also assume constant residual variance; normal errors support the normal-theory test used below.

Under Hardy–Weinberg equilibrium, the two allele copies can be represented as independent Bernoulli variables with probability \(p\) of carrying the allele. Their sum is \(G\), so its variance is the sum of their variances:

For a rare allele, this variance is approximately \(2p\). A variant with lower frequency therefore produces less spread in \(G\): most participants have dosage zero, and relatively few have dosage one or two. In a linear regression, the estimated slope is less precise when the predictor has little variation.

Here is a short justification of the standard-error relationship. Conditional on the observed genotypes, the least-squares slope has variance \(\sigma^2/\sum_{i=1}^{N}(G_i-\bar{G})^2\). For a sufficiently large sample drawn from the stated genotype distribution, the sum of squared deviations is approximately \(N\operatorname{Var}(G)\). Substituting the genotype variance gives the approximate standard error and the standardized effect:

The quantity \(\delta\) is the approximate mean of the test statistic under the specified alternative, when the null standard error is used. The formula shows the central information pattern: holding the effect and residual variation fixed, the standardized effect grows like the square root of sample size and roughly like the square root of allele frequency when the allele is rare. Equivalently, the sample size needed for a fixed power grows approximately in inverse proportion to \(p\).

From the Standardized Effect to Power

For a two-sided test at significance level \(\alpha\), the large-sample normal approximation rejects \(H_0\) when \(|Z|>z_{1-\alpha/2}\), where \(Z\) is the standardized test statistic and \(z_{1-\alpha/2}\) is the corresponding standard-normal quantile. Under the alternative, approximate \(Z\sim N(\delta,1)\). The rejection probability is the sum of the lower-tail and upper-tail probabilities:

where \(\Phi\) is the standard normal cumulative distribution function. The formula includes both tails because the test is two-sided. When the effect is negative, the same power results after replacing \(\delta\) by its absolute value.

Sample Size, Frequency, and Effect Size Together

The standardized-effect formula can compare study designs directly. For example, doubling \(N\) multiplies \(\delta\) by \(\sqrt{2}\), not by two. Doubling a small allele frequency has approximately the same effect on \(\delta\) as doubling \(N\), provided the other quantities remain fixed. A larger effect increases \(\delta\) in direct proportion to \(|\beta|/\sigma\). These relationships are useful for understanding why modest changes in any one design feature may not overcome an extremely low allele frequency.

What the Approximation Does—and Does Not—Say

The formulas identify a basic limitation of single-variant tests: low allele frequency means low genotype variance and, usually, few carriers. Both reduce the information available to estimate an association. A larger sample, a larger effect, or lower residual variation can improve power, but the gain should be assessed against the actual allele frequency and the intended test threshold. For example, more stringent significance thresholds require a larger critical value and therefore more information for the same power.

These calculations presume that the allele frequency is relevant to the study participants and that genotype calls are accurate. A frequency estimate from a different population may not predict the study's carrier count. Genotype uncertainty or missingness can further reduce usable information. Power also does not establish that an association is causal; it concerns the probability of rejecting a statistical null under a specified alternative and model.

Common Mistake: Confusing a Large Effect With High Power

A common error is to reason, “The allele changes the trait a lot, so the test must have high power.” The effect enters the standardized effect, but so does the square root of the available genotype information. If only a few participants carry the allele, the estimated effect can remain highly uncertain. The correction is to assess the effect relative to residual variation and multiply by \(\sqrt{N\,2p(1-p)}\), then evaluate power at the test's significance threshold. Also check both tails in a two-sided calculation: for instance, \(1-\Phi(-0.275)=\Phi(0.275)\), not \(1-\Phi(0.275)\).

A second caution is that a normal approximation can be misleading when counts are very small. The approximation is useful for explaining how power changes with \(N\), \(p\), and effect size, but sparse genotype categories can make test-statistic distributions discrete or poorly approximated by a standard normal distribution. In that setting, the specific test and its calibration matter.