What an Allele-Frequency Difference Describes
An allele-frequency comparison is a comparison of population summaries, not a comparison of individual people. In pharmacogenetics, the distinction matters because a frequency may help describe the genetic composition of a defined cohort, but it does not tell us which allele a particular patient carries or what that patient’s drug exposure will be.
A frequency is meaningful only after the population and the allele have been specified. A study’s recruited participants, a clinical-trial cohort, and a target population for a PBPK application are not automatically the same population. Estimates from one group can inform another only when the relationship between the groups is justified.
The population definition should state how participants were recruited or selected and what makes them members of the group being described. For example, a frequency estimated among patients at one hospital is directly a description of those sampled patients. Treating it as an estimate for all patients in a country requires an additional sampling or transportability argument. A large sample does not correct a systematic mismatch between the sample and the intended target.
Comparing Estimates and Their Uncertainty
For two independently sampled populations, a simple comparison is the difference \(\hat f_{A,1}-\hat f_{A,2}\). The sign indicates which sample has the larger estimated frequency, and the magnitude gives the difference in proportion of allele copies. A difference of \(0.05\), for example, is five percentage points, not a five percent relative increase.
Uncertainty depends on how the data were sampled. Under a simple model in which allele copies are independent draws from a population with frequency \(f\), the allele count is binomial with \(2n\) trials. The estimated frequency has approximate standard error \(\sqrt{\hat f(1-\hat f)/(2n)}\). For two independent samples, the estimated difference has approximate standard error equal to the square root of the sum of the two estimated variances. This calculation is not automatically appropriate when individuals are related, the sampling is clustered, or allele copies are not independent under the sampling design.
The formula follows by treating the two sample estimates as independent. Each estimate is a count divided by \(2n_g\), so its variance under the binomial model is \(f_{A,g}(1-f_{A,g})/(2n_g)\). The variance of a difference of independent estimates is the sum of their variances. Replacing the unknown population frequencies with the sample estimates gives the displayed estimated standard error. A normal-approximation interval for the difference is the estimated difference plus or minus 1.96 estimated standard errors; it can be unreliable when counts are small or frequencies are near zero or one.
Worked Example: Comparing Two Sample Frequencies
Suppose two hypothetical, independently sampled cohorts each contain unrelated diploid participants typed at the same autosomal locus. In cohort 1, 80 copies of allele \(A\) are observed among 400 allele copies; in cohort 2, 45 copies are observed among 300. Assume the cohorts are representative samples of the two populations of interest and that independent allele-copy sampling is a reasonable approximation.
The estimates are \(\hat f_{A,1}=80/400=0.200\) and \(\hat f_{A,2}=45/300=0.150\). Thus, the estimated difference is \(0.200-0.150=0.0500\), or 5.00 percentage points, with the larger estimate in cohort 1.
The estimated standard error is \( \sqrt{[0.200(0.800)/400]+[0.150(0.850)/300]} \). The two terms are \(0.000400\) and \(0.000425\), so the result is \(\sqrt{0.000825}=0.02872\).
A 95% normal-approximation interval is \(0.0500 \pm 1.96(0.02872)\), or \(0.0500 \pm 0.05630\). Its endpoints are approximately \(-0.00630\) and \(0.1063\). The interval includes zero, so these data do not estimate the difference precisely enough to distinguish it from zero using this approximation at the 95% interval level. The interval describes uncertainty under the sampling model; it does not establish that the population frequencies are equal, nor does it account for selection bias or an incorrect population definition.
Composition Can Explain an Overall Difference
A population-wide estimate can change because the allele frequency differs within constituent groups, because the groups occur in different proportions, or both. This is why a comparison of broad labels alone can be misleading. If the available data allow it, compare frequencies within measured, relevant strata as well as reporting the overall estimates. The strata must be defined using information that is actually measured, not inferred from a broad label.
Worked Example: Different Mixtures, Same Within-Group Frequencies
Consider two hypothetical cohorts, each with 1,000 unrelated diploid participants. In cohort X, 800 participants are in measured subgroup 1 and 200 in subgroup 2. The observed allele frequencies are 160/1,600 = 0.100 in subgroup 1 and 120/400 = 0.300 in subgroup 2. In cohort Y, the subgroup sizes are reversed: 200 participants in subgroup 1 and 800 in subgroup 2. The corresponding counts are 40/400 = 0.100 and 480/1,600 = 0.300.
For cohort X, the overall estimate is \((160+120)/(1,600+400)=280/2,000=0.140\). For cohort Y, it is \((40+480)/(400+1,600)=520/2,000=0.260\). The overall estimates differ by 0.120, or 12 percentage points, even though the observed frequencies are identical within each measured subgroup.
Here, the overall difference arises from the subgroups’ different proportions: cohort X is 80% subgroup 1, while cohort Y is 20% subgroup 1. This example does not show that every population difference is compositional. It shows why an overall comparison alone cannot identify its source. A broad ancestry label would not reveal these measured subgroup proportions, and using that label as if it did would conceal the actual composition of the cohorts.
For PBPK applications, the relevant question is not whether a cohort carries a particular broad ancestry label. It is whether the cohort’s measured genetic and clinical composition is appropriate for the intended application. Broad labels can encompass substantial variation, overlap across people, and fail to describe a particular recruited sample. They are therefore inadequate substitutes for measured cohort composition. This does not mean ancestry-related information can never be relevant; it means that a label should not be treated as a precise allele-frequency measurement or as an individual’s genotype.
A Frequency Is Not an Individual Prediction
An allele frequency gives the fraction of allele copies in a population. It is not the fraction of people who carry the allele, and it does not identify the genotype of any particular person. Even translating an allele frequency into expected genotype frequencies requires additional assumptions, such as Hardy-Weinberg equilibrium in the population being modeled. Genotype or diplotype, functional interpretation, and clinical context are separate information needed for individual-level conclusions.
Worked Example: From an Allele Frequency to a Limited Population Prediction
Suppose a hypothetical autosomal allele \(A\) has estimated frequency \(\hat f_A=0.200\) in a defined cohort, with the other allele denoted \(a\). If Hardy-Weinberg equilibrium is a reasonable model for that population, the expected genotype proportions are \(A/A: 0.200^2=0.0400\), \(A/a: 2(0.200)(0.800)=0.320\), and \(a/a: 0.800^2=0.640\).
Under those assumptions, the expected proportion carrying at least one \(A\) copy is \(0.0400+0.320=0.360\), or 36.0%. It is not 20.0%, because 20.0% is the fraction of allele copies that are \(A\), not the fraction of individuals carrying \(A\). This is a model-based population expectation, not a determination of any person’s genotype. A person’s genotype requires individual genotyping or other valid individual-level evidence.
Even the genotype distribution is not a drug-response prediction by itself. A pharmacogenetic interpretation also depends on which variant was measured, its functional evidence, and how the relevant biological effect is represented. An allele-frequency estimate does not establish an individual’s enzyme activity or drug exposure.
Diagnostics, Alternatives, and Limitations
Before interpreting a difference, check that the cohorts were genotyped at the same locus with comparable variant definitions and quality control, and that the allele counted as \(A\) is coded consistently. Confirm the denominators, missing-data handling, and whether the samples include related individuals. Review recruitment and ascertainment: a clinic-based cohort may differ from a population sample because of disease status, access to care, or study eligibility. Such differences can affect how well the estimate represents its intended target.
A useful comparison depends on the question. A raw overall comparison describes the two sampled mixtures. A stratified comparison shows estimates within measured subgroups, while a standardized comparison can estimate what the overall frequencies would be under a shared set of subgroup weights. A pooled estimate may be suitable for a clearly defined combined cohort, but it can hide differences between its constituent groups. No statistical adjustment repairs a poorly defined target population or creates information about an unmeasured subgroup.
| Approach | What it describes | Main caution |
|---|---|---|
| Separate overall estimates | Observed allele frequencies in each sampled cohort | Differences may reflect cohort composition or selection |
| Within-stratum estimates | Frequencies within measured subgroups | Small strata can produce imprecise estimates |
| Standardized estimates | Frequencies after applying a common set of subgroup weights | The chosen weights define the comparison and must be reported |
| Pooled estimate | Frequency in the combined set of observed allele copies | Can conceal differences and may not represent either target population |
A frequency comparison also has limits that remain after careful estimation. Sampling uncertainty is only one source of uncertainty; measurement error, variant misclassification, nonrepresentative recruitment, and population structure can matter as much or more. When estimates differ, report their definitions, denominators, and uncertainty before proposing an explanation. Do not infer that a frequency difference causes a difference in drug response without evidence connecting the variant to function and the function to the clinical outcome.
Common Pitfall
Incorrect reasoning: “Allele \(A\) has frequency 0.20 in this broad ancestry group, so an individual identified with that ancestry has a 20% chance of carrying \(A\).” This confuses allele copies with people, treats a broad label as a precise population definition, and turns a group estimate into an individual prediction.
Correction: State the frequency as an estimate for a specified cohort and report how it was obtained. If a genotype proportion is needed, specify and assess the additional model assumptions. For an individual, use individual-level genetic information and an appropriate functional interpretation; do not substitute a population frequency or broad ancestry label.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A hypothetical diploid sample of 250 unrelated people has 75 copies of allele \(B\) among 500 observed copies. What is the estimated allele frequency?
- Two independently sampled cohorts have estimated frequencies 0.30 and 0.20. Explain what their difference of 0.10 means and name one conclusion it does not support by itself.
- Two cohorts have different overall frequencies but identical frequencies within each of two measured subgroups. Give one reason their overall estimates could differ.
- Why is it incorrect to interpret an allele frequency of 0.20 as saying that 20% of individuals carry at least one copy of the allele?
Solutions. (1) \(\hat f_B=75/500=0.150\), or 15.0% of observed allele copies. (2) The first sample’s estimated frequency is 10 percentage points higher than the second’s. This comparison alone does not establish why they differ, whether the difference is statistically distinguishable from zero, or whether it predicts an individual’s response. (3) Their subgroup proportions may differ, so the overall weighted averages can differ despite equal within-subgroup estimates. (4) The frequency counts allele copies, not people. The proportion of individuals carrying at least one copy depends on the genotype distribution and requires additional assumptions or individual genotype data.