Diagnostics for Latent Metabolizer Mixtures

Why Mixture Diagnostics Need More Than One View

A fitted fast–slow clearance mixture is useful only if it captures meaningful pharmacokinetic (PK) differences without leaving systematic patterns unexplained. A pooled plot can conceal a poor fit: fast-class observations may tend to lie above their predictions while slow-class observations lie below, even if the overall average residual is near zero. Conversely, a visually separated set of observations does not by itself establish that the fitted classes are adequate.

Diagnostics should therefore examine class-specific patterns, individual predictions, and the distribution of possible observations implied by the model. The classes remain latent: each individual has probabilities of belonging to the fast or slow class, rather than a directly observed label. This matters because assigning each individual to the class with the largest probability discards uncertainty and can make a diagnostic picture look more definite than the evidence warrants.

Model Quantities Used in the Diagnostics

Let \(i\) index individuals, \(j\) index observations within an individual, and \(k\in\{F,S\}\) denote the fast and slow classes. Let \(z_i\) be the latent class for individual \(i\), and let \(\pi_{ik}=P(z_i=k)\) be the class probability implied by the membership model and that individual's covariates. The probabilities satisfy \(\pi_{iF}+\pi_{iS}=1\). The class is constant across that individual's observations.

Write \(f_{ikj}\) for the model-predicted concentration at the \(j\)th observation time for individual \(i\), conditional on class \(k\) and the individual's random effects. The observed concentration is \(y_{ij}\). Conditional on class and random effects, the observation model specifies a density \(g(y_{ij}\mid f_{ikj},\ldots)\). The dots allow for the specified residual-error model and relevant design variables.

For an individual's complete record \(y_i\), the fitted likelihood under class \(k\) is obtained by integrating over that individual's random effects. The posterior class probability is

where \(L_{ik}\) is the likelihood of the individual's record under class \(k\), after integrating over random effects. Thus, \(\tau_{iF}+\tau_{iS}=1\). These probabilities depend on both the membership model and the observed record; they are not observed phenotypes.

The precise individual prediction depends on how random effects are handled. For example, it may use their conditional distribution given the individual's record rather than a single point estimate. Whatever convention is used should be stated and applied consistently. Individual predictions are useful for asking whether the model can describe an individual's observed trajectory; they are not the same target as predictions for a new individual, whose random effects and class are unknown.

Probability-Weighted Class-Stratified Checks

Class-stratified checks ask whether the observations assigned probability to a class behave as that class's fitted PK model implies. Instead of putting each individual wholly into the class with the larger \(\tau_{ik}\), use \(\tau_{ik}\) as a weight. At a given time or within a suitable time bin, a weighted observed mean for class \(k\) is \(\sum_i\tau_{ik}y_{ij}/\sum_i\tau_{ik}\), and the corresponding weighted prediction mean uses \(\hat f_{ikj}\) in place of \(y_{ij}\). The comparison is descriptive; it does not turn latent class membership into an observed variable.

Compare these summaries across time, dose, or other design features relevant to the PK model. A persistent difference between class-weighted observations and class-specific predictions can indicate that the fitted class curves do not capture the systematic pattern. Also inspect within-class spread: agreement in weighted means can coexist with large or structured individual discrepancies.

Individual Residual Patterns and Their Limits

Plotting \(e_{ij}=y_{ij}-\hat f_{ij}\) against time, predicted concentration, or dose can reveal structure that class averages hide. In a mixture-specific review, also examine residuals against posterior class probabilities and compare them with the class-specific predictions. Look for persistent runs of positive or negative residuals over an individual's profile, or for patterns that recur among individuals with substantial probability in the same class. A residual trend may indicate a mismatch in the structural PK model, the error model, or another aspect of the fitted model; it does not identify the cause on its own.

A key qualification is that \(\tau_{ik}\) and the individual predictions are informed by \(y_i\). Consequently, a good-looking individual fit is not independent evidence that the model predicts new individuals well. Posterior-probability weighting avoids the extra distortion caused by forced hard labels, but it does not remove this dependence. Treat individual-level displays as conditional descriptions of the records used to fit or classify individuals. Assessing predictions for new individuals requires a predictive distribution that accounts for their unknown class and random effects.

Predictive Distributions Preserve Class Uncertainty

A mean prediction alone does not describe the range or shape of responses the model considers plausible. For an observation from a new individual, the predictive distribution averages over class membership, random effects, and residual error. For a future observation from an individual with an existing record, the class weights are instead updated to the posterior probabilities given that record, and uncertainty in the individual's random effects must also be considered. In either case, the distribution is a mixture of class-specific predictive distributions, not generally a single distribution centered at the mixture mean.

A simple derivation shows why class separation contributes to predictive variance. Suppose, for illustration, the class-specific predictive distributions at a fixed design point are Normal with means \(\mu_F,\mu_S\) and the same variance \(\sigma^2\). Let the relevant class probabilities be \(\tau_F,\tau_S\), summing to one. The mixture mean is \(\bar\mu=\tau_F\mu_F+\tau_S\mu_S\). By conditioning on class,

The first term is the within-class predictive variance. The second is the between-class contribution, which is large when both classes remain plausible and their predictions differ. Replacing the mixture with a Normal distribution having the same mean and variance preserves these two moments but can still misrepresent tail probabilities.

Assumptions, Warning Signs, and Alternatives

These checks have meaning only relative to a clearly specified model. In particular, the mixture assumes that each individual belongs to one class for the relevant record; class membership probabilities are specified by the membership model; and the class-specific PK and residual models describe observations conditional on class and random effects. The likelihood calculation also relies on the stated dependence structure, commonly conditional independence of observations given the individual's class and random effects. If these assumptions are inappropriate, apparent class-specific residual patterns may reflect the model setup rather than a simple need to alter class predictions.

Warning signs include persistent time-ordered residual patterns within a class, class-weighted observations that depart consistently from class-specific predictions, or predictive distributions that assign very little probability to recurring types of observed profiles. Also inspect whether conclusions change markedly when individuals with uncertain posterior membership are included with their probability weights. If apparent class differences are driven only by a small number of highly weighted individuals, report that limitation rather than presenting the pattern as broadly representative.

A pooled population prediction is simpler and can summarize average behavior, but may hide opposing class-specific discrepancies. A hard-assignment display is easy to read, but treats uncertain classifications as certain and can move individuals between groups under small changes in their records. Probability-weighted summaries preserve membership uncertainty, while predictive distributions additionally represent within-class response variability. None of these displays proves that the mixture is the unique or correct explanation of PK differences; they diagnose implications of the fitted model. They also do not replace careful assessment of parameter uncertainty, which a plug-in predictive calculation may omit.