Subpopulation Mixtures Hidden by a Single-Population Model

Why Subject-Level Clusters Matter

A population model can fit the overall concentration distribution reasonably well and still systematically miss different sets of individuals in opposite directions. For example, concentrations may be underpredicted for one group of people and overpredicted for another. When those directions recur across an individual's observations, they can indicate that the population contains distinct kinetic subgroups that a single-population model is averaging together.

The key evidence is not simply that individual parameter estimates vary. Interindividual variability is expected in population pharmacokinetics. The diagnostic question is whether the variation looks like one continuous distribution, or whether subjects repeatedly fall into separated patterns that a latent-class model could represent more plausibly. A cluster is a modeling hypothesis, not proof that a biological subgroup exists.

For example, suppose \(C_{ij}\) is the observed concentration for individual \(i\) at observation \(j\), and \(\hat C_{ij}\) is the corresponding model prediction. The standardized prediction residual is \(r_{ij}=(C_{ij}-\hat C_{ij})/SE_{ij}\), where \(SE_{ij}\) is the standard deviation of the observation error under the model. A positive value means the observed concentration exceeds the prediction; a negative value means it is below. Residual patterns across time and individuals can help localize model failure, as in the earlier material on residual patterns, but the subgroup question requires summarizing whether particular individuals show a consistent direction or parameter pattern.

What a Mixture Adds to the Model

Let \(Z_i\) be an unobserved component indicator with possible values \(1,\ldots,K\), and let \(\pi_k=P(Z_i=k)\), where the component probabilities satisfy \(\pi_k\geq0\) and \(\sum_{k=1}^K\pi_k=1\). For a clearance model, one possible specification is \(CL_i=CL_k\exp(\eta_i)\) conditional on \(Z_i=k\), with \(\eta_i\mid Z_i=k\sim N(0,\omega_k^2)\). Here \(CL_k\) is the typical clearance in component \(k\), while \(\omega_k^2\) describes continuous variation within that component. The likelihood contribution for an individual is a weighted sum of component likelihoods:

The individual component is unknown, so the model does not assign each person with certainty. Instead, the component probabilities and component-specific parameters are estimated from the population data, and an individual's observations provide evidence about their component membership. Component labels are interchangeable: relabeling components does not change the model. A mixture also needs enough information to distinguish components; simply adding components can describe noise or sampling irregularities rather than meaningful subgroups.

The main alternative is a single continuous distribution for the random effect, such as \(\eta_i\sim N(0,\omega^2)\). Other plausible explanations include an omitted covariate that shifts a parameter, a continuous but skewed or heavy-tailed random-effect distribution, or structural and residual-model errors. These explanations can produce superficially similar outliers, so residual separation alone is not sufficient to choose a mixture.

Why a Mixture Can Look Different From One Normal Distribution

A mixture can have greater overall variability than either of its component distributions, and can even have two modes. A short derivation shows why separation between component centers matters. Consider two equally weighted normal distributions for a log-parameter, with means \(-a\) and \(a\), and common standard deviation \(\sigma\). Their mixture density is \(f(x)=\{g(x-a)+g(x+a)\}/2\), where \(g\) is the density of \(N(0,\sigma^2)\). By symmetry, \(f'(0)=0\). Differentiating the normal density twice gives \(g''(u)=\{u^2/\sigma^4-1/\sigma^2\}g(u)\). Therefore:

Because \(g(a)>0\), the center \(x=0\) is a local minimum when \(a>\sigma\), and the symmetric mixture then has two modes. When \(a<\sigma\), it has a local maximum at the center. This result is an illustration for this specific equal-weight, equal-variance case, not a general rule for deciding whether a fitted mixture is justified. Unequal weights or variances change the shape, and real parameter estimates have uncertainty.

A single normal distribution can still accommodate a wide range of values by increasing its variance. But it may then assign too much probability to intermediate parameter values and too little to separated clusters. A mixture can represent both the within-component spread and the between-component separation. Its greater flexibility is useful only when the data contain enough information to support the additional parameters.

Connecting Clusters to Kinetic Parameters

Residual clusters become more interpretable when they correspond to a coherent parameter pattern. For example, at comparable doses and sampling times, individuals with persistently higher-than-predicted concentrations may have lower clearance than the single-population model predicts. Conversely, those with persistently lower-than-predicted concentrations may have higher clearance. The kinetic interpretation depends on the structural model and the observations available; residual direction by itself does not identify which parameter is wrong.

An individual's fitted parameter estimate is a diagnostic summary, not a direct observation of their true parameter. Estimates may be imprecise, especially when the data provide little information about that parameter. Look for separation that is consistent across relevant parts of the profile, not just a pair of extreme estimates. Also ask whether a measured covariate or a known design feature could explain the pattern more simply than a latent class.

Comparing Explanations and Checking for Failure

A mixture should be compared with reasonable alternatives, not only with a restrictive single normal model. The alternatives answer different questions and imply different population structures.

Evidence against a proposed mixture includes components that disappear under small changes to initial values, a component represented by very few individuals, highly uncertain component parameters, or component assignments that depend strongly on a few observations. A model may also appear to separate subjects because the sampling design is different between groups. Check whether dose, observation timing, assay conditions, and relevant covariates are balanced or adequately represented. In addition, inspect predictions and residuals by individual and over time: a subgroup in clearance should not be used to explain a repeating time-pattern that instead points to a structural problem.

Formal comparison needs care. Mixture models add parameters, and under the null hypothesis of a single component some mixture parameters, including a component's weight or its parameter separation, lie on boundaries or are not identified. Consequently, a routine likelihood-ratio test using a standard chi-squared reference distribution is not automatically valid. Information criteria, simulation-based assessment, stability checks, and predictive performance can contribute evidence, but none turns a statistical component into a proven biological class. Report the uncertainty and the assumptions behind the comparison.

Common Mistake and Practical Interpretation

The practical value of detecting a plausible mixture is that it can prevent a single typical parameter and a large variability estimate from concealing different prediction patterns. Yet extra complexity is not automatically beneficial. A subgroup model is most useful when its components are estimable, its predictions are better supported, and its interpretation remains cautious. If these conditions are not met, the single-population model may be the more defensible summary, with the unresolved pattern clearly reported.