Interpreting Uncertain Metabolizer-Class Assignments

From Class Probabilities to Individual Assignments

In a two-class NLME model, each individual belongs to a latent metabolizer class, such as fast or slow clearance. The class is not directly observed: it is inferred from the individual’s concentration-time data and any covariates in the membership model. Consequently, an individual assignment is generally uncertain. A fitted probability of 0.70 for fast-class membership does not mean the individual is known to be fast; it means the fitted model gives the fast class 70% of the conditional membership probability, given the available data and model.

The distinction matters in reporting and decision-making. A hard label such as “fast” is convenient for a table, but it hides how strongly the data support that label. A probability preserves that information and allows a threshold to be chosen for a specific purpose.

The word “posterior” here describes updating class membership probabilities using individual data. In a standard fitted NLME analysis, the calculation usually plugs in estimated parameters; it does not by itself include uncertainty in those parameter estimates. If a fully Bayesian analysis instead averages over a parameter distribution, its membership probabilities can also reflect parameter uncertainty.

How the Posterior Probability Is Calculated

Let \(\pi_{iF}=P(Z_i=F\mid x_i,\hat{\theta})\) be the model-based class probability before using individual \(i\)’s observed response data, and define \(\pi_{iS}=1-\pi_{iF}\). Let \(L_{iF}\) and \(L_{iS}\) be the likelihoods of the observed data under the fast and slow classes, respectively. In an NLME model, each class likelihood is based on the class-specific model and accounts for the random effects according to the model’s likelihood calculation. Bayes’ rule gives

Thus, an individual’s class probability combines two sources of information: the prior class probability supplied by the membership model and how well each class explains that individual’s data. A class with a larger prior probability need not have the larger posterior probability if the observed data fit the other class sufficiently better.

This calculation assumes that the membership probabilities and class-specific likelihoods are defined by the fitted model, and that the individual’s data contribute through the stated likelihood. It also relies on the model’s assumptions about conditional independence across individuals and the within-class response and random-effects distributions. A probability is therefore a model-based summary, not a direct measurement of a person’s biological class.

Thresholds, Hard Labels, and Abstention

A hard assignment requires a rule that converts probabilities into labels. The maximum-posterior-probability (MAP) rule assigns fast when \(q_i\geq 0.5\) and slow otherwise. This rule selects the class with the larger fitted posterior probability. It is a natural default when the two kinds of classification error have equal cost, but it does not imply that probabilities just above 0.5 represent strong evidence.

A higher confidence threshold can be used when a label should be issued only if support is stronger. For example, with threshold \(t=0.80\), assign fast if \(q_i\geq 0.80\), assign slow if \(q_i\leq 0.20\), and leave the individual unclassified otherwise. The middle range is an abstention region: it is an explicit acknowledgement that neither class has enough probability for the chosen confidence rule. Reporting the number left unclassified is essential; otherwise, a high apparent accuracy among assigned individuals can conceal very low coverage.

For a set of \(N\) individuals, useful summaries at a stated threshold include the number assigned to each class, the number abstaining, and the coverage (the proportion receiving a label). If labels are assigned to a subset \(A\), the sum of the assigned class probabilities gives the expected number correctly classified under the fitted probabilities. Dividing that sum by \(|A|\) gives the expected proportion correct among assigned individuals. These are model-based expectations, not observed accuracy against known latent labels.

Why the MAP Rule Uses a 0.5 Threshold

The MAP rule follows from a simple loss calculation. Suppose a false fast assignment (assigning fast when the individual is slow) has cost \(C_{FP}\), and a false slow assignment has cost \(C_{FN}\). Conditional on the individual’s data, the expected loss from assigning fast is \(C_{FP}(1-q_i)\), because the probability the individual is actually slow is \(1-q_i\). The expected loss from assigning slow is \(C_{FN}q_i\). Choose fast when its expected loss is no greater:

With equal costs, \(C_{FP}=C_{FN}\), the threshold is 0.5, giving the MAP rule. If falsely calling a slow individual fast is more costly, then \(C_{FP}>C_{FN}\) and the threshold rises. This derivation also shows why no threshold is universally optimal: the appropriate rule depends on the consequences of each error and on whether abstention is allowed.

Expected Class Counts Preserve Uncertainty

Hard labels can distort a summary of a group when many individuals have probabilities near 0.5. If \(q_i\) is the fitted fast-class probability for each individual, the expected number in the fast class is \(\sum_{i=1}^{N}q_i\), and the expected number in the slow class is \(\sum_{i=1}^{N}(1-q_i)\). These are fractional expected counts, not claims that a fractional person exists. They summarize the model’s membership probabilities across individuals.

Interpretation, Comparisons, and Common Pitfalls

Three reporting approaches answer different questions. The probability \(q_i\) communicates the fitted uncertainty for each individual. A MAP label provides a complete two-class partition under equal error costs, but conceals how close probabilities are to the threshold. A high-confidence rule communicates only stronger assignments and should be accompanied by its threshold and coverage. Expected class counts summarize group membership without forcing each individual into a single class.

These approaches are not substitutes for a model with observed class labels. Nor do posterior class probabilities establish that a latent-class model is scientifically appropriate; that question concerns model specification and evidence for the mixture. A class probability is conditional on the chosen model, including its class definitions, class-specific response distributions, covariate effects, and fitted parameters.

A frequent mistake is to say, “This individual is fast because their probability is 0.55.” The correction is: “Under the fitted model, this individual has a 0.55 probability of fast-class membership; the MAP rule would label them fast, but the assignment is uncertain.” Another mistake is to call the fraction labelled fast the estimated population prevalence. Thresholded counts depend on the threshold and individual probabilities; prevalence is represented by the membership model, while the sum of posterior probabilities is a data-updated expected count. They need not be equal.

Probabilities may also be sensitive to model assumptions or parameter estimates, especially when the classes overlap or the individual data are sparse. A reported value with many decimal places does not remove that uncertainty. Report enough precision for the decision, state the classification rule, and avoid presenting a hard label as ground truth. The detailed diagnostics used to assess the adequacy of a latent mixture are a separate question from interpreting its individual class probabilities.