Identifiability of Subgroup Effects in Sparse PK Data

Why Subgroup Effects May Be Hard to Separate

A fitted population pharmacokinetic model can return a numerical estimate for a sex, race, genotype, or co-medication coefficient even when the data do not meaningfully identify that effect. For example, if every participant with a particular genotype also takes the same inhibitor, the observed concentration differences cannot by themselves reveal which covariate accounts for the difference. Sparse concentration sampling adds a separate difficulty: even when the covariates vary independently, sampling times may provide little information about the parameter they are proposed to affect.

The key question is not simply whether a subgroup is represented or whether an optimizer converged. It is whether the observed design contains information that can distinguish the parameter of interest from other model parameters. Identifiability should therefore be considered before interpreting subgroup estimates.

Definition: A parameter or parameter contrast is structurally identifiable when distinct values of it imply distinct probability distributions for the observations under the specified model and design. It is practically identifiable when the available data contain enough information to estimate it with useful precision. Structural non-identifiability cannot be repaired by collecting more observations under the same design; practical identifiability may improve with more informative data.

These definitions apply to the model and design together. A coefficient may be identifiable under a design with covariate overlap and informative sampling, but not under a design in which covariates always occur together or observations are made only at times insensitive to that coefficient. Estimability is not a permanent property of a covariate name.

Representing Subgroup Effects in a PK Model

Suppose a model describes clearance, \(CL_i\), for participant \(i\). A simple covariate model for two binary covariates might be

\log(CL_i)=\log(CL_{\mathrm{ref}})+\beta_S S_i+\beta_M M_i+\eta_i,

where \(CL_{\mathrm{ref}}\) is the typical clearance for the reference categories, \(S_i\) indicates one sex category, \(M_i\) indicates use of a co-medication, \(\beta_S\) and \(\beta_M\) are covariate effects on log clearance, and \(\eta_i\) is an individual deviation. In this parameterization, \(e^{\beta_S}\) is the clearance ratio for the sex contrast when medication status is held fixed; \(e^{\beta_M}\) is the medication-associated ratio when sex is held fixed. These interpretations require that the corresponding contrasts can be separated by the data.

Sparse PK data enter through the concentration model. Concentrations depend on clearance together with other parameters, such as volume, absorption rate, dose history, and sampling time. Thus, a subgroup coefficient may be hard to estimate for either of two broad reasons: its covariate pattern is confounded with another model term, or the observed concentrations are weakly sensitive to the parameter it changes. Both can occur at once.

Covariate Overlap and Structural Identifiability

A useful first check is to tabulate participant counts for combinations of the proposed covariates. Counts should be at the participant level, not the concentration-record level: ten samples from one participant do not provide the same covariate replication as ten participants. For binary covariates, compare the number of participants in each joint category. For multi-level race or genotype variables, examine the corresponding cross-classification and note empty or very small cells.

The algebraic issue can be seen by temporarily considering a model for a participant-level response \(Y_i\) with mean \(a+\beta_S S_i+\beta_M M_i\). Its design matrix has a column of ones, a column for \(S_i\), and a column for \(M_i\). If \(S_i=M_i\) for every participant, the last two columns are identical. The data can then inform \(\beta_S+\beta_M\), but not the two coefficients separately.

Key criterion: In a linear model, separate coefficients require the relevant design-matrix columns to be linearly independent after accounting for other model terms. In a nonlinear PK model, the analogous local check uses the sensitivity of predicted observations to the parameters: if two parameter changes produce indistinguishable changes in the observations, their effects cannot be reliably separated by that design.

Worked Example: Sex and Co-medication Always Occur Together

Consider a hypothetical study with 30 participants. Fifteen are in the \(S=0\) category and do not take the medication; the other 15 are in the \(S=1\) category and all take it. Suppose the intended log-clearance model includes both \(\beta_S S_i\) and \(\beta_M M_i\).

For every participant, \(S_i=M_i\). The covariate part of the model is therefore

\beta_S S_i+\beta_M M_i=(\beta_S+\beta_M)S_i.

Only the sum \(\beta_S+\beta_M\) changes the model predictions. For instance, \(\beta_S=0.10\) and \(\beta_M=0.20\) give the same covariate contribution as \(\beta_S=0.25\) and \(\beta_M=0.05\), because both sums equal \(0.30\). The corresponding estimated clearance ratio between the two observed groups could be \(e^{0.30}=1.350\), but the data do not establish how much of that contrast belongs to sex and how much to medication.

The 30 participants and any number of concentration records from them cannot resolve this exact aliasing while the covariates remain perfectly aligned. Separate interpretation requires additional information or a design containing participants who break the alignment, such as participants in both sex categories with and without the medication.

With imperfect rather than exact alignment, the coefficients may be structurally identifiable but still highly correlated and imprecise. A cross-tabulation reveals whether comparisons exist; the size and balance of the cells indicate how much support those comparisons have. A large total sample does not guarantee useful overlap if nearly all participants occupy only a few combinations.

Subgroup Size, Sampling Time, and Practical Identifiability

Even with covariate overlap, small subgroups usually provide less information about a subgroup contrast. The number of participants matters because participants, rather than repeated records from one individual, supply independent covariate assignments. Repeated concentrations can improve the information contributed by a participant, but the gain depends on their timing, residual error, and the model.

One way to make this dependence explicit is to use a local approximation. Suppose a transformed observation has mean \(a+q\beta\), where \(\beta\) is a subgroup effect and \(q\) is the sensitivity of the observation to that effect at its sampling time. If observations have independent error with standard deviation \(\sigma\), a small \(|q|\) means that changing \(\beta\) barely changes the expected observation. In a PK model, sensitivities vary with time and with the current parameter values; a concentration taken at a time that is informative for one parameter may be weakly informative for another.

\text{Approximate } SE(\hat{\beta}) \ \propto\ \frac{\sigma}{|q|\sqrt{n}},

This proportionality is a simplified illustration, not a universal standard-error formula for nonlinear mixed-effects models. It captures two general features: more participants tend to increase information, while weak sensitivity or high observation variability reduces it. In a subgroup contrast, the smaller subgroup often limits precision, and uncertainty also depends on how well the reference group is estimated.

Worked Example: A Small Genotype Subgroup and Weak Sensitivity

Suppose a hypothetical study has 4 genotype carriers and 20 non-carriers. To illustrate practical identifiability, use a simplified local model for a transformed concentration, with a known sensitivity \(q\) to the genotype effect and independent residual error of \(\sigma=0.20\) log-concentration units. Assume \(q\) is the same for each participant and other model parameters are held fixed. For a two-group contrast under these assumptions, the approximate standard error is

SE(\hat{\beta})\approx \frac{\sigma}{|q|}\sqrt{\frac{1}{n_1}+\frac{1}{n_0}}.

If the sampling time gives \(q=0.50\) and \(n_1=4,\ n_0=20\), then

SE(\hat{\beta})\approx \frac{0.20}{0.50}\sqrt{\frac{1}{4}+\frac{1}{20}} =0.40\sqrt{0.30}=0.219.

Now suppose the available sampling time has much weaker sensitivity, \(q=0.05\), with the same participants and residual variability. Then

SE(\hat{\beta})\approx \frac{0.20}{0.05}\sqrt{0.30} =4.00(0.5477)=2.191.

The subgroup counts have not changed, but the approximate uncertainty is ten times larger because the observation is ten times less sensitive to the effect. This calculation assumes a simplified model with known sensitivity and no uncertainty in other parameters; an actual PK analysis must account for the full concentration model and its parameter correlations. It nevertheless shows why a subgroup count alone cannot establish practical estimability.

A related concern is competition with other PK parameters. If, over the observed sampling window, a change in clearance produces nearly the same concentration pattern as a change in volume, absorption, or another covariate effect, the model may estimate a broad range of parameter combinations similarly well. Sparse sampling can leave the concentration-time profile insufficiently described to distinguish those patterns. A subgroup effect that is well represented in the covariate table can still be weakly identified by the concentration data.

Checking Overlap Across Several Subgroups

For sex, race, genotype, and co-medication effects considered together, inspecting one covariate at a time can be misleading. A genotype may be represented overall but occur almost exclusively in one race category; a medication may be used mostly in one sex category. The relevant question is whether the data include the combinations needed for the contrasts in the proposed model.

Worked Example: Genotype Effects With and Without Race Overlap

Consider hypothetical participant counts for two race categories and a binary genotype:

Race category	Non-carriers	Carriers	Total
A	18	2	20
B	2	18	20

Both genotype groups and both race categories occur in the study, but the covariates are strongly aligned. A model with separate race and genotype effects can in principle distinguish them because all four cells are nonempty, but the within-race genotype comparisons rely on only 2 participants in each rare cell. The nominal total of 40 participants therefore overstates the support for separate contrasts.

For illustration, suppose the hypothetical cell means for log clearance are 2.00 and 2.10 for non-carriers and carriers in category A, and 2.30 and 2.40 in category B. The within-category genotype contrasts are

2.10-2.00=0.10,\qquad 2.40-2.30=0.10.

Their agreement is compatible with a common genotype contrast of \(0.10\) log-clearance units in this simplified table. Its clearance ratio is \(e^{0.10}=1.105\). However, the small carrier cell in category A and small non-carrier cell in category B make those comparisons weakly supported; this arithmetic does not establish that a full PK model will estimate the effect precisely.

If the two rare cells were empty, carriers would occur only in category B and non-carriers only in category A. The race and genotype columns would then be perfectly aligned, as in the previous example, so the separate effects would not be identifiable from these data. In practice, inspect both empty cells and sparse cells, and consider how the available concentration times inform the parameter contrast within each populated cell.

A Practical Identifiability Workflow

State the contrast.
Specify the PK parameter and the comparison, such as a genotype-associated clearance ratio at fixed medication status. A vague request to estimate a “subgroup effect” does not define which model terms need to be separated.

Count participants in joint covariate categories.
Tabulate combinations of the relevant sex, race, genotype, and medication categories. Flag empty and very small cells; do not substitute concentration-record counts for participant counts.

Check for exact or near aliasing.
Ask whether two covariate columns, or a covariate and another model term, are identical or nearly redundant. If they are, determine whether only a combined contrast is supported.

Review sampling information.
Check whether the sampling times and dose histories make predictions sensitive to the parameter of interest, and whether they can distinguish it from other PK parameters.

Interpret the fitted model in light of the design.
Use uncertainty, parameter correlations, and stability under reasonable model or design checks as evidence about practical identifiability. A successful numerical fit alone is not evidence that separate subgroup effects are well determined.

What to Do When Effects Are Not Well Identified

The appropriate response depends on the source of the problem. Exact aliasing calls for more information that breaks the covariate alignment; more participants with the same alignment do not separate the effects. Sparse but nonempty cells may benefit from additional participants in those combinations. Weak sensitivity may call for sampling times chosen to be informative for the parameter contrast, or for a more modest model that does not try to estimate effects the data cannot support.

A simpler model can estimate a combined contrast when that is scientifically meaningful, but its interpretation must match the model. Removing one covariate does not prove that the remaining coefficient represents that covariate alone; it may absorb differences associated with the omitted covariate. Similarly, combining categories can improve numerical stability while changing the scientific question. External information or informative prior distributions can stabilize estimates, but then the result depends partly on that added information rather than on the study data alone.

These approaches are not interchangeable. Adding participants addresses limited subgroup size only when the added participants contribute useful overlap and concentration information. Changing sampling times addresses weak PK sensitivity. Simplifying the covariate model can reduce the number of contrasts but cannot recover a distinction that the design never observed. Each remedy should be tied to the particular identifiability limitation.

Diagnostics and Common Pitfalls

Warning signs include an empty covariate combination, extreme imbalance across combinations, very large uncertainty for a subgroup coefficient, strong parameter correlations, estimates that change substantially under small model changes, and difficulty obtaining stable estimates. These signs are not individually proof of non-identifiability. They prompt investigation of whether the data support the requested contrast and whether the model is asking the data to distinguish effects with similar consequences for concentrations.

Counting records instead of participants. Repeated samples from a few carriers do not create many independent carrier assignments. Report subgroup counts by participants as well as the sampling information.
Equating nonempty cells with adequate support. A single participant in a cell technically breaks exact aliasing, but generally provides little basis for a precise subgroup comparison.
Equating convergence with identifiability. An optimizer can return one point on a flat or poorly determined region. Numerical completion is not evidence that the data distinguish the parameter.
Interpreting a combined contrast as two effects. If the covariates are aligned, the data may estimate their sum or another combined contrast. They do not thereby estimate each component separately.
Assuming a large overall sample resolves weak sensitivity. If observations barely change when the parameter changes, a large sample may still provide limited information, especially when the relevant subgroup is small.

The main alternatives are to add independent covariate overlap, improve the concentration-time design, reduce the number of unsupported contrasts, or use external information with its role made explicit. No one option is automatically best: the choice depends on whether the limitation is exact covariate aliasing, sparse subgroup representation, weak concentration sensitivity, or some combination.

Key takeaway: Before interpreting a subgroup coefficient, establish that the relevant participant combinations are represented and that the observed concentrations are sensitive to the parameter contrast. Subgroup counts, covariate overlap, and sampling times answer different parts of that question.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

Every participant with genotype \(G=1\) takes medication \(M=1\), and every participant with \(G=0\) has \(M=0\). Can a model with separate genotype and medication main effects identify both coefficients from this design? What combination can it identify?
A study has 3 carriers, 27 non-carriers, and many concentration records per carrier. Why do the record count and participant count provide different information about a genotype effect?
In the simplified approximation \(SE(\hat{\beta})\propto \sigma/(|q|\sqrt{n})\), what happens to the standard error if \(|q|\) is halved while \(\sigma\) and \(n\) stay fixed?
All four cells in a race-by-genotype table are nonempty, but two cells contain only one participant each. Is exact aliasing ruled out? Does that guarantee a precise effect estimate?

Solutions. (1) No. Since \(G_i=M_i\) for every participant, the model contribution is \((\beta_G+\beta_M)G_i\); only the sum is identified. (2) The 3 participants provide only 3 independent genotype assignments. Their repeated records may improve information about each carrier’s concentration profile, but do not provide the same covariate replication as many independent carriers. (3) The approximate standard error doubles because it is inversely proportional to \(|q|\). (4) Nonempty cells rule out that particular exact alignment, but single-participant cells give little support for within-category contrasts, so precise estimation is not guaranteed.