Reporting and Predicting With Metabolizer Mixtures

From a Fitted Mixture to an Interpretable Report

A fitted fast–slow metabolizer mixture is useful only if its results are reported in a way that preserves what the model says. That means distinguishing the prevalence of each class from an individual’s probability of belonging to a class, presenting the pharmacokinetic parameters for each class, and carrying the mixture through to population predictions. In particular, a population prediction should average predictions across classes; it should not generally replace their clearances with a single average clearance.

The formulas below use a two-class model. Let \(C_i\in\{F,S\}\) denote the latent class of individual \(i\), where \(F\) and \(S\) refer to the fast- and slow-clearance classes. Let \(x_i\) denote covariates used to model class membership, and let \(\pi_F(x_i)\) and \(\pi_S(x_i)\) be the corresponding class probabilities, with \(\pi_F(x_i)+\pi_S(x_i)=1\). Let \(\theta_F\) and \(\theta_S\) collect class-specific PK parameters. The class-specific PK model may also include individual random effects.

This distinction matters in reporting. A statement such as “30% are fast metabolizers” is a population summary for a specified covariate distribution. For an individual with given membership covariates, the corresponding prior probability is \(\pi_F(x_i)\). Once that individual’s PK measurements are used, their posterior probability of fast-class membership may be larger or smaller. A posterior probability is not proof of a person’s biological class; it quantifies uncertainty under the fitted model.

What to Report and How to Express Uncertainty

A clear report states the class definitions and ordering, how membership probabilities depend on covariates, the estimated class prevalence for a specified target population, and the class-specific PK parameters with units. It should also identify which parameters are shared across classes. If prevalence varies with covariates, report the membership model and give probabilities for clinically meaningful covariate profiles or a defined population, rather than presenting one number as universal.

For each estimate, report an uncertainty measure, such as a standard error and confidence interval. A conventional approximate 95% Wald interval for an estimate \(\hat\psi\) with standard error \(SE(\hat\psi)\) is \(\hat\psi\pm1.96SE(\hat\psi)\). This approximation relies on the estimator being reasonably close to normally distributed and the standard error being reliable. For a prevalence constrained to \([0,1]\), a direct Wald interval can extend outside the allowable range; an interval calculated on a transformed scale, such as the logit scale, or a suitable bootstrap interval may be preferable. Confidence intervals describe uncertainty in the estimated parameter under repeated sampling and the model; they are not 95% probabilities that a fixed parameter lies in the particular reported interval.

A class-specific interval is not an interval for the difference between classes. To quantify that difference, calculate it directly and account for covariance between estimates. For example, if both clearances are estimated from the same mixture fit, the uncertainty in \(\widehat{CL}_S-\widehat{CL}_F\) depends on their covariance as well as their individual standard errors. Similarly, a population prediction’s uncertainty depends on uncertainty in the class parameters and mixing proportions, and may also include random-effect and residual variability depending on the prediction target.

Individual Class Probabilities

To update class membership using an individual’s PK data \(y_i\), combine the prior class probability with the likelihood of those data under each class. Here \(f_k(y_i\mid\theta_k)\) denotes the class-specific likelihood, including integration over any individual random effects in the model. The posterior probability of fast-class membership is

The slow-class probability is one minus this quantity. This expression shows why a high prior probability does not guarantee a high posterior probability: sufficiently stronger support from the PK data for the other class can change the assignment. Conversely, the posterior depends on both the data and the membership model, not just on how well the observations match a class-specific PK curve.

Population Predictions Must Preserve the Mixture

Let \(g(t;\theta_k,b)\) be the model-predicted concentration at time \(t\) for class \(k\), with individual random effect \(b\). For a population prediction conditional on covariates \(x\), first average the prediction within each class over that class’s random-effect distribution, then weight by the class probabilities:

If the prediction uses typical individuals with random effects fixed at their reference values, the integral is replaced by the corresponding class-specific typical prediction. If the target is a future measured concentration rather than the expected concentration, residual variability must also be included in the predictive distribution. These are different targets: the expected population concentration does not itself include a random residual deviation.

The mixture average generally cannot be replaced by a prediction evaluated at the prevalence-weighted average clearance. The reason is that the PK prediction is nonlinear in clearance. In general, for a nonlinear function \(g\), a weighted average of \(g\) evaluated at class-specific parameters is not equal to \(g\) evaluated at the weighted-average parameter.

Uncertainty, Diagnostics, and Alternatives

Prediction uncertainty should match the question being asked. A point prediction can be accompanied by a confidence interval for the estimated population mean, or by a prediction interval for future observations; those intervals are not interchangeable. A practical way to propagate parameter uncertainty through a nonlinear mixture prediction is to use a parametric bootstrap: simulate parameter estimates under the fitted model, recompute the class probabilities and predictions for each replicate, and summarize the resulting prediction distribution. If predicting future individuals, the simulation should also draw class membership, within-class random effects, and—when predicting observed measurements—residual error.

Check whether the reporting summary remains credible in light of model uncertainty. Weakly separated classes can make class-specific estimates imprecise and individual posterior probabilities diffuse. A very small estimated class can make its parameter estimates and prevalence unstable. If estimates change substantially under reasonable alternative initializations, or the class labels are not consistently ordered, reporting a single confident “fast” and “slow” interpretation is misleading. Label the classes by a prespecified parameter ordering, report uncertainty, and avoid treating a thresholded posterior assignment as known truth. These checks complement, rather than replace, the mixture diagnostics and identifiability assessment.

The main alternative to a latent mixture prediction is a single-population model that represents clearance with one distribution or one typical value. Such a model may be suitable when the mixture is unsupported or when the prediction target is explicitly a single-population approximation. But it does not preserve distinct class-specific response patterns. Likewise, averaging clearances and then predicting is a convenient shortcut, not generally an equivalent mixture calculation. When metabolizer classes are part of the fitted model and relevant to the target population, report the class-specific predictions and their prevalence-weighted combination.

A useful report makes the target population and prediction target explicit, gives class definitions and prevalence, presents class-specific PK estimates with uncertainty and units, and explains how individual posterior probabilities were obtained. It should distinguish a population mean prediction from an individual prediction and from a future-observation prediction. Those distinctions prevent a concise summary from concealing either class uncertainty or the variability implied by the model.