Why Population Structure Matters in Genetic PK Analyses
A genetic variant can appear to predict clearance even when it has no direct effect on the pharmacokinetic (PK) parameter of interest. One route to this misleading association is population stratification: ancestry-related population structure is associated both with the frequency of the variant and with factors that affect PK. A pooled comparison of genotype groups can then mix a genuine within-ancestry genotype contrast with differences in the ancestry composition of those groups.
The concern is not that ancestry itself must be a biological cause of a PK difference. Ancestry-related structure can track other genetic variants, clinical characteristics, environmental exposures, or care patterns. Any of these may be related to the PK outcome. The key statistical issue is that genotype groups are not comparable with respect to relevant background factors unless the analysis accounts for them adequately.
Race and ethnicity categories, as discussed in the earlier material on race and ethnicity categories as measured covariates, are not interchangeable with genetic ancestry. They may be useful measured covariates for particular questions, but they do not directly identify genetic ancestry or remove all ancestry-related structure. Conversely, genetic ancestry measures do not capture every social or clinical factor that may be relevant to PK.
A PK Model and the Meaning of Adjustment
For illustration, let \(G_i\) denote a genotype coding for individual \(i\), such as carrier status, and let \(A_i\) denote an ancestry-related covariate. In a population PK model, a covariate effect on clearance may be represented on the log scale as
Here \(CL_i\) is individual clearance, \(\theta_0\) is the log typical clearance at the reference covariate values, \(\beta_G\) is the genotype-associated contrast conditional on \(A_i\), and \(\gamma_A\) is the ancestry-covariate contrast conditional on genotype. The individual random effect \(\eta_{CL,i}\) represents unexplained between-individual variation. The concentration data are ordinarily modeled through the PK structural model; this equation describes how clearance is related to covariates and individual variation.
This specification is useful only under assumptions appropriate to the analysis. The covariate relationships must be represented adequately (including nonlinearities or interactions when justified), the genotype and ancestry covariates must be measured well enough for the intended contrast, and the random effect must not retain a systematic mean association with omitted determinants of genotype and clearance. The model must also have enough genotype overlap across ancestry values to distinguish the two effects. Including a random effect does not, by itself, control population stratification: a mean-zero random-effect distribution cannot substitute for a systematically omitted covariate.
If \(G_i\) is binary and \(\beta_G\) is estimated on the log-clearance scale, the conditional clearance ratio for carriers versus noncarriers at the same \(A_i\) is \(\exp(\beta_G)\). This is an adjusted association under the specified model, not automatically a causal effect of changing genotype. Causal interpretation requires additional assumptions about confounding, selection, measurement, and the biological question.
How Omitted Structure Produces Bias
A simple linear model makes the source of bias explicit. Suppose the conditional mean of a log-scale PK outcome \(Y\) is
where \(A_i\) is a binary ancestry indicator, \(\delta\) is its outcome contrast, and the error has conditional mean zero given \(G_i\) and \(A_i\). If \(A_i\) is omitted and \(Y_i\) is regressed on \(G_i\) alone, the population slope is
The first equality is the population least-squares slope. Substituting the full model into its numerator gives \(\operatorname{Cov}(G,Y)=\beta_G\operatorname{Var}(G)+\delta\operatorname{Cov}(G,A)\), because the error is uncorrelated with \(G\) under the stated conditional-mean assumption. Dividing by \(\operatorname{Var}(G)\) gives the result. For binary \(G\), the covariance ratio equals \(P(A=1\mid G=1)-P(A=1\mid G=0)\). Thus the bias is the ancestry outcome contrast multiplied by the ancestry-composition difference between genotype groups.
This decomposition also clarifies important exceptions. If ancestry is unrelated to the genotype in the analyzed sample, the covariance term is zero. If ancestry is associated with genotype but has no outcome association after other modeled covariates are considered, then \(\delta=0\) and this particular source of bias is absent. In practice, both relationships may be present, and the analysis must address them rather than assume either is negligible.
Worked Example: A Pooled Association Reverses the Within-Group Contrast
Consider hypothetical data on log clearance, \(Y=\log(CL/(1\ \mathrm{L/h}))\), in two ancestry-related groups. Assume the cell means below are exact for this illustration. Genotype \(G=1\) is the carrier group, and \(G=0\) is the noncarrier group.
| Ancestry group | Genotype | Number | Mean log clearance |
|---|---|---|---|
| Group 0 | 0 | 80 | 1.00 |
| Group 0 | 1 | 20 | 0.90 |
| Group 1 | 0 | 20 | 1.40 |
| Group 1 | 1 | 80 | 1.30 |
Within group 0, the carrier-minus-noncarrier contrast is \(0.90-1.00=-0.10\). Within group 1, it is \(1.30-1.40=-0.10\). Both groups therefore show the same negative genotype contrast in this hypothetical data.
Pooling over ancestry gives a different result. The noncarrier mean is \((80(1.00)+20(1.40))/100=1.08\). The carrier mean is \((20(0.90)+80(1.30))/100=1.22\). The pooled contrast is \(1.22-1.08=0.14\), which has the opposite sign from either within-group contrast.
On the clearance scale, the pooled contrast corresponds to a ratio of \(\exp(0.14)=1.150\), while either within-group contrast corresponds to \(\exp(-0.10)=0.9048\). The reversal occurs because carriers are much more concentrated in group 1, which has higher mean log clearance. The pooled comparison therefore does not isolate a genotype contrast.
Estimating a Conditional Genotype Contrast
When the model is appropriate and there is adequate overlap, including ancestry-related structure can separate the genotype contrast from differences in group composition. In a simple additive model, this means estimating genotype and ancestry coefficients together. In a population PK analysis, the same principle is applied within the structural and random-effects model rather than by treating fitted individual clearances as error-free observations.
Worked Example: Interpreting the Adjusted Clearance Effect
Use the four cell means in the preceding hypothetical example and fit the additive model \(Y=\beta_0+\beta_GG+\gamma_AA+\varepsilon\), with \(A=0\) for group 0 and \(A=1\) for group 1. The cell means have the form \(\beta_0=1.00\), \(\beta_G=-0.10\), and \(\gamma_A=0.40\): for group 0, the predicted means are \(1.00\) and \(1.00-0.10=0.90\); for group 1, they are \(1.00+0.40=1.40\) and \(1.00-0.10+0.40=1.30\). These match all four supplied cell means.
The adjusted genotype coefficient is therefore \(\hat{\beta}_G=-0.10\) log units. Exponentiating gives \(\exp(-0.10)=0.9048\). Under this model, the carrier group has an estimated typical clearance about \(0.905\) times that of the noncarrier group, or about \(9.52\%\) lower, conditional on ancestry group. The result describes a conditional association in these hypothetical data; it does not establish a causal effect.
The fitted additive model assumes the genotype contrast is the same in both ancestry groups. That assumption is satisfied by these cell means. If the within-group contrasts differed, a genotype-by-ancestry interaction could be needed, and a single common \(\beta_G\) would not summarize both contrasts adequately.
Ancestry adjustment can use an appropriate measured ancestry variable, a set of genetic ancestry components, or analyses that compare genotype groups within ancestry strata. These approaches answer related but not identical questions. A single group indicator can be too coarse if structure varies within groups; genetic components can represent continuous axes of variation but may not capture all outcome-related differences. A stratified analysis makes the within-group comparisons explicit, but can become imprecise when strata are small. Adjustment is useful only if the covariates represent relevant structure and the data support the comparisons being made.
Overlap, Diagnostics, and Failure Modes
A conditional genotype effect is learned from comparisons among people who differ in genotype but are comparable on the ancestry covariates in the model. If a genotype is almost exclusive to one ancestry group, the model may have little information to distinguish a genotype contrast from an ancestry contrast. If a genotype category is entirely absent in a stratum, the within-stratum contrast is not estimable there without additional assumptions or extrapolation.
Worked Example: When Adjustment Cannot Create a Missing Comparison
Suppose a hypothetical study has 100 participants in each of two ancestry groups. In group 0, all 100 are noncarriers; in group 1, 40 are noncarriers and 60 are carriers. Assume the mean log clearances are 1.00 for group 0 noncarriers, 1.40 for group 1 noncarriers, and 1.30 for group 1 carriers.
Within group 1, the observed carrier contrast is \(1.30-1.40=-0.10\) log units. In group 0, there is no carrier group, so a within-group genotype contrast cannot be calculated. A model with genotype and ancestry indicators can use the group 1 comparison, but it cannot learn a genotype contrast for group 0 directly. Any claim that the contrast is identical in group 0 relies on the model's common-effect assumption, not on observed carriers in that group.
The practical conclusion is not that adjustment is impossible in every such data set. It is that the source and limits of information must be stated: overlap is incomplete, the estimated common contrast is supported by the group with both genotypes, and conclusions about the other group require extrapolation. If there is no genotype variation at all within the full sample, genotype and ancestry effects cannot be separated from these data.
Useful checks include tabulating genotype frequencies across ancestry covariates, examining whether the genotype groups overlap across the covariate range, and comparing estimated contrasts with and without ancestry adjustment. A large change after adjustment signals that composition matters, although it does not by itself prove that the adjusted estimate is unbiased. For PK models, inspect whether the covariate specification is plausible and whether the fitted model captures the concentration-time data across relevant genotype and ancestry groups. Apparent adjustment is not reassuring if model fit is poor or the covariate effects are inadequately represented.
A common mistake is to reason, “The genotype coefficient remains statistically significant after adding race, so population stratification is ruled out.” This is incorrect. A measured race or ethnicity category may be an imperfect proxy for ancestry-related structure, and residual structure or other confounding may remain. The correction is to define what the covariate measures, assess overlap and model adequacy, and describe the estimated effect as conditional on the covariates actually included. Statistical significance alone is not a test that all relevant structure has been controlled.
Choosing and Interpreting an Analysis
The principal alternatives are an unadjusted pooled comparison, covariate adjustment, and within-stratum estimation. The pooled analysis is simple but vulnerable when genotype and ancestry are associated and ancestry predicts PK. Covariate adjustment can use information across strata and estimate a common conditional effect, but depends on covariate and model specification. Stratified estimates make heterogeneity visible and reduce reliance on a common-effect assumption, but can be unstable or unavailable when some genotype-by-stratum cells are sparse. None of these approaches automatically resolves unmeasured structure or poor overlap.
Interpretation should identify the scale and comparison. A coefficient on log clearance becomes a clearance ratio after exponentiation; an ancestry-adjusted coefficient compares genotype groups conditional on the modeled ancestry covariates. It is not a universal biological effect across all populations, and it does not imply that ancestry is itself the mechanism. If genotype effects plausibly vary across ancestry backgrounds, report that a common-effect model is an assumption and evaluate whether the data support separate contrasts.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If a variant is more common in one ancestry group and that group has higher typical clearance for unrelated reasons, what direction of bias might arise in a pooled genotype comparison?
- In the omitted-variable expression, suppose \(\delta=0.20\) log units, \(P(A=1\mid G=1)=0.70\), and \(P(A=1\mid G=0)=0.30\). Calculate the bias in the unadjusted genotype slope.
- A fitted genotype coefficient for log clearance is \(-0.15\). What clearance ratio does it imply, and how should the comparison be described?
- Why does adding a random clearance effect not automatically control ancestry-related confounding?
- If one genotype is absent from an ancestry stratum, what limitation should be stated about an adjusted common genotype effect?
Solutions. (1) The pooled estimate may be biased upward because the genotype group enriched in the higher-clearance ancestry group inherits part of that group's higher outcome mean. The exact direction depends on which genotype is enriched and the direction of the ancestry-outcome contrast. (2) The ancestry-composition difference is \(0.70-0.30=0.40\), so the bias is \(0.20(0.40)=0.08\) log units. (3) The ratio is \(\exp(-0.15)=0.8607\). Conditional on the model's covariates, the genotype group has estimated typical clearance about \(0.861\) times that of the reference genotype group, or about \(13.9\%\) lower; this is an adjusted association, not necessarily a causal effect. (4) A random effect represents unexplained individual variation under its distributional assumptions. It does not represent a systematic omitted mean difference associated with both genotype and ancestry. (5) The within-stratum genotype contrast is not observed there. The common adjusted effect may be informed by other strata, but applying it to the stratum without that genotype relies on the common-effect assumption and constitutes extrapolation.