Why Ancestry Differences Need Careful Interpretation
A pharmacogenomic association found in one study population may not transfer unchanged to another. The relevant allele may be more or less frequent, a measured marker may tag it differently, drug exposure may differ, or the genotype-specific treatment effect may genuinely vary. These possibilities have different statistical meanings and require different evidence.
Ancestry labels need particular care. Genetic ancestry describes patterns of genetic similarity and can be represented continuously or through study-specific groupings. It is not interchangeable with race, ethnicity, nationality, or an individual’s social experience. Group comparisons can reveal where an estimate may be less applicable, but a difference between groups does not by itself identify a biological cause. Social, clinical, environmental, and healthcare factors may also differ.
Allele Frequencies and Linkage Patterns
For a biallelic variant, the allele frequency is the proportion of observed allele copies that carry the allele of interest. With \(n\) diploid individuals and genotype counts \(N_{CC}\), \(N_{CT}\), and \(N_{TT}\), the frequency of \(C\) is \((2N_{CC}+N_{CT})/(2n)\). A frequency difference can affect how many people carry a pharmacogenomic allele, even if the allele’s effect on response is identical across populations.
A nearby marker is often used to identify or predict a functional variant. How well it does so depends on linkage disequilibrium (LD), the non-random association of alleles at two loci. LD can differ across populations, so a marker that is a useful proxy in one population may be a weak proxy in another. For two biallelic loci with alleles \(C\) and \(T\), let \(p_C\) and \(p_T\) be their frequencies and \(p_{CT}\) the frequency of chromosomes carrying both. Define \(D=p_{CT}-p_Cp_T\). The squared allele correlation is \(r^2=D^2/[p_C(1-p_C)p_T(1-p_T)]\), when both loci are polymorphic.
Worked Example: Comparing Allele Frequencies
Suppose two hypothetical study populations each include 100 unrelated individuals. Genotypes at a pharmacogenomic variant are counted as follows.
| Population | CC | CT | TT |
|---|---|---|---|
| A | 36 | 48 | 16 |
| B | 16 | 48 | 36 |
The frequency of \(C\) in A is \((2(36)+48)/(2(100))=120/200=0.60\). In B it is \((2(16)+48)/200=80/200=0.40\). The observed difference, A minus B, is \(0.20\), or 20 percentage points.
For an approximate interval, treat the allele copies as independent Bernoulli observations, as is justified approximately under random mating and the usual large-sample conditions. The estimated standard error of the difference is \(SE=\sqrt{0.60(0.40)/200+0.40(0.60)/200}=0.0490\). A 95% Wald interval is \(0.20\pm1.96(0.0490)=(0.104,0.296)\). Under these assumptions, the data are consistent with a higher \(C\) frequency in A by about 10.4 to 29.6 percentage points. This comparison describes these population samples; it does not explain why the frequencies differ.
Worked Example: A Marker Tags Differently Across Populations
Consider phased chromosome data for a functional allele \(C\) and marker allele \(T\), with 100 chromosomes observed in each hypothetical population. The four haplotype counts are:
| Population | CT | Ct | cT | ct |
|---|---|---|---|---|
| A | 30 | 10 | 10 | 50 |
| B | 20 | 10 | 20 | 50 |
In A, \(p_C=(30+10)/100=0.40\), \(p_T=(30+10)/100=0.40\), and \(p_{CT}=0.30\). Thus \(D=0.30-(0.40)(0.40)=0.14\), and \(r^2=0.14^2/[(0.40)(0.60)(0.40)(0.60)]=0.0196/0.0576=0.3403\).
In B, \(p_C=(20+10)/100=0.30\), \(p_T=(20+20)/100=0.40\), and \(p_{CT}=0.20\). Therefore \(D=0.20-(0.30)(0.40)=0.08\), and \(r^2=0.08^2/[(0.30)(0.70)(0.40)(0.60)]=0.0064/0.0504=0.1270\).
The marker and functional allele are less strongly correlated in B by this measure. A marker-based association may consequently be less informative about the functional allele in B, even if the functional allele has the same pharmacologic effect in both populations. These calculations assume phased haplotypes are correctly measured; uncertainty in the haplotype estimates should also be considered in real data.
Separate Exposure Differences From Effect Differences
Drug exposure is not the same quantity as treatment effect. Exposure measures how much drug reaches or remains in the body, often summarized by a pharmacokinetic measure such as area under the concentration-time curve. Response is an outcome under a treatment strategy. Differences in dose, adherence, co-medications, body size, organ function, or care can produce exposure differences. Genetic variants affecting metabolism may contribute, but a group-level exposure contrast alone does not establish a genetic mechanism.
When comparing exposure, use the same dose or clearly specify the dosing strategy, measurement schedule, and exposure summary. If those conditions differ, a group contrast may mix pharmacokinetics with treatment and measurement differences. Likewise, compare pharmacogenomic effects using the same outcome definition, follow-up time, treatment contrast, genotype coding, and effect scale. A difference in a risk difference need not match a difference in an odds ratio or hazard ratio.
Worked Example: Comparing Drug Exposure
Suppose a hypothetical pharmacokinetic study measures area under the concentration-time curve (AUC), in mg·h/L, after the same prescribed dose. There are 64 participants per ancestry group. The sample mean AUC is 12 mg·h/L in A and 10 mg·h/L in B; both sample standard deviations are 4 mg·h/L.
Assuming independent participants and using a large-sample normal approximation, the estimated difference is \(12-10=2\) mg·h/L. Its standard error is \(SE=\sqrt{4^2/64+4^2/64}=\sqrt{0.50}=0.707\) mg·h/L. A 95% interval is \(2\pm1.96(0.707)=(0.614,3.386)\) mg·h/L.
The sample suggests higher mean AUC in A under the specified dosing and measurement conditions. This is evidence about exposure, not proof that ancestry causes the difference or that the drug has a different genotype-specific effect. Investigators would need to consider factors such as adherence, renal function, co-medications, and relevant genetic variation before attributing a mechanism.
Assessing Whether Pharmacogenomic Effects Differ
A direct comparison of pharmacogenomic effects asks whether the genotype-related treatment contrast differs across the specified populations. For a binary genotype \(G\), treatment \(Z\), and outcome \(Y\), define \(\delta_a\) in ancestry group \(a\) as the difference between treatment effects for \(G=1\) and \(G=0\). For example, on the mean-difference scale:
The cross-population contrast is \(\delta_A-\delta_B\). It is a difference between two genotype-by-treatment interactions, not simply a comparison of group averages. In a model with binary \(Z\) and \(G\), fitting their interaction within each group estimates \(\delta_a\) on the additive mean scale when the model is correctly specified. A further comparison tests whether those interaction estimates differ. For other outcomes or scales, the model and interpretation must be adapted.
Worked Example: Comparing Treatment-By-Genotype Effects
In a hypothetical randomized study, lower symptom-change scores indicate greater improvement. Assume treatment and comparator are assigned within each ancestry group, genotype is measured consistently, and the outcome is recorded at the same follow-up time. Estimated treatment-by-genotype contrasts on the mean-difference scale are \(\hat{\delta}_A=-4.0\) points with \(SE=1.2\), and \(\hat{\delta}_B=-1.0\) point with \(SE=1.0\).
The difference is \(\hat{\delta}_A-\hat{\delta}_B=-4.0-(-1.0)=-3.0\) points. If the two estimates are independent and approximately normal, its standard error is \(SE=\sqrt{1.2^2+1.0^2}=\sqrt{2.44}=1.562\) points. The test statistic for no difference is \(z=-3.0/1.562=-1.92\), giving a two-sided \(p\)-value of about 0.055. A 95% interval is \(-3.0\pm1.96(1.562)=(-6.06,0.062)\) points.
The estimated genotype-related treatment contrast is 3 points more negative in A, but the interval includes zero and the two-sided test does not cross the 0.05 threshold. The data do not establish a cross-population difference; nor does this result establish that the effects are identical. Randomization supports causal treatment contrasts within the studied groups, but it does not make ancestry a randomized exposure or remove limitations from unmeasured differences between those groups.
Why Marginal Results Can Differ Without Effect Heterogeneity
A group-level response contrast can vary solely because genotype frequencies vary. To see this, suppose the outcome follows \(E(Y\mid G)=\mu+\beta G\), where \(G\) is allele dosage coded 0, 1, or 2, and the same \(\mu\) and \(\beta\) apply in both populations. Under random mating with allele frequency \(p_a\), \(E(G\mid a)=2p_a\). Therefore:
The population means can differ even though the per-allele coefficient \(\beta\) is identical. Conversely, similar population means do not show that the genotype-specific effects are equal. This is why marginal comparisons and interaction comparisons answer different questions.
Transporting Estimates and Common Pitfalls
Transporting a pharmacogenomic estimate means applying an estimate from a source population to a defined target population. A transport calculation requires more than knowing that ancestry labels differ. It requires relevant effect modifiers to be measured, sufficient overlap in those modifiers between source and target, comparable treatment and outcome definitions, and assumptions that the conditional effects used in the calculation apply to the target. Broad ancestry categories may not capture the important variation within either population.
Worked Example: Standardizing a Treatment Effect to a Target Mix
Suppose hypothetical, comparable studies estimate treatment effects on the same mean outcome scale: \(-5\) points in group A and \(-2\) points in group B, with negative values indicating improvement. A target population is estimated to comprise 25% group A and 75% group B. If these group-specific effects apply to the target members of each group, a composition-standardized effect is \((0.25)(-5)+(0.75)(-2)=-1.25-1.50=-2.75\) points.
This weighted mean describes the target mix under the stated assumptions. It is not automatically a valid transport estimate merely because the group proportions are known. If treatment effects also vary by age, dose, clinical severity, or other factors whose distributions differ between source and target, those factors may need to be standardized as well. The calculation also assumes comparable treatment strategies and outcome measurement.
A frequent mistake is to reason, “The allele frequency differs, so the drug effect must differ.” Frequency changes can alter the number of people carrying an allele and hence a population-average result; they do not by themselves show that an individual-level genotype effect differs. The correction is to estimate the relevant genotype-by-treatment contrast within each population and compare those contrasts on the same scale. Another mistake is to interpret a non-significant interaction test as proof of equal effects. A wide interval may include both no difference and clinically important differences; examine the estimate and its uncertainty.
Before interpreting a cross-population result, check that genotype and phenotype definitions are consistent, that ancestry characterization is appropriate to the question, and that treatment, dose, follow-up, and covariate adjustment are comparable. Consider population structure and social or clinical differences that could distort observational comparisons. When many drugs, outcomes, variants, or subgroup contrasts are examined, multiplicity also matters; selective attention to the smallest \(p\)-value can exaggerate evidence.
Specify whether the comparison concerns allele frequency, LD, exposure, a marginal outcome, or a genotype-specific treatment effect.
Use consistent variant coding, treatment contrast, outcome, follow-up, and effect scale across groups.
Report estimates and intervals, evaluate design and measurement differences, and avoid treating ancestry group as a causal mechanism.
Explain which source-to-target assumptions are needed and which effect modifiers or population differences remain unaddressed.
Check Your Understanding
Answer each question using the distinctions and assumptions developed above.
- Two populations have different frequencies of a pharmacogenomic allele. What can be concluded about the number of carriers, and what cannot be concluded about the allele’s treatment effect?
- Why could a marker-based association be weaker in one population even if the functional variant has the same effect in both?
- In a randomized study, what contrast should be compared across populations to assess whether genotype modifies the treatment effect differently?
- Why does a non-significant test of a cross-population interaction not establish that the effects are equal?
- Give two assumptions needed before a weighted average of group-specific effects can be interpreted as an estimate for a target population.