Assessing Class Separation and Rare Classes

Why Class Support Needs More Than a Convergence Check

A fitted latent class joint model can converge to a numerical solution and still provide weak evidence for some of its classes. Two classes may have highly overlapping posterior membership probabilities; one class may account for very few subjects; or a class may contain too little event information to estimate its hazard parameters precisely. These are distinct problems. A model can have well-separated classes but a very small class, or a reasonably prevalent class whose members are difficult to distinguish.

The assessments here build on posterior class probabilities and mixture estimation. They do not establish that the chosen number of classes is correct, and they do not replace checks of the longitudinal and event-model assumptions. Their purpose is narrower: to describe how strongly the fitted data support the estimated class assignments, class prevalences, and class-specific parameters.

The word “effective” matters: \(N_g^{\mathrm{eff}}\) is a sum of fractional memberships, not a count of people known to belong to class \(g\). The fitted prevalence is model-based and depends on the observed data, the model specification, and the fitted solution. If each subject is instead assigned to the class with the largest posterior probability, the resulting hard-assigned counts discard uncertainty and can tell a different story.

Measuring Posterior Separation

A simple subject-level measure of assignment uncertainty is posterior entropy. For subject \(i\), define \(H_i=-\sum_{g=1}^{G}\tau_{ig}\log(\tau_{ig})\), with a term \(0\log(0)\) defined as 0 by continuity. Entropy is smallest, at 0, when one class has posterior probability 1. It is largest, at \(\log(G)\), when the probabilities are equal across the \(G\) classes. Averaging over subjects and dividing by \(\log(G)\) gives normalized entropy between 0 and 1; lower values indicate more concentrated posterior probabilities. Some software reports a separation index that reverses this direction, so always check the definition used.

Entropy summarizes posterior concentration, not whether the model is true. It can also conceal detail: the same overall average may arise from a mixture of nearly certain and highly ambiguous subjects, or from moderate uncertainty for nearly everyone. Inspecting the distribution of the largest posterior probability, or the posterior probabilities themselves, complements an average summary.

Prevalence and the Meaning of a Rare Class

The prevalence estimate \(\hat{\pi}_g\) is obtained by averaging posterior membership probabilities over subjects. This follows directly from the fitted model's expected class indicators. If \(I(C_i=g)\) equals 1 when subject \(i\) belongs to class \(g\) and 0 otherwise, its conditional expectation given the observed data and fitted parameters is \(E[I(C_i=g)\mid\mathcal{D}_i,\hat{\theta}]=\tau_{ig}\). Summing these expectations gives the posterior expected number of members, \(\sum_i\tau_{ig}\); dividing by \(n\) gives the estimated proportion.

A low fitted prevalence is a warning to investigate, not an automatic reason to remove a class. Small classes may represent real heterogeneity, but estimates for them can be sensitive to a few subjects, to model assumptions, or to the fitted solution. Also inspect uncertainty in the prevalence estimate where available, and whether it changes substantially across plausible model fits. The posterior effective size itself is not a standard error or a formal measure of certainty.

Event Support and Parameter Precision

A class's effective size does not show how much information it supplies for its event model. For example, a class may include many subjects but few observed events, or many events may occur among a small effective class. A useful descriptive quantity is the posterior-weighted event count \(\sum_i \delta_i\tau_{ig}\), where \(\delta_i=1\) denotes an observed event and \(\delta_i=0\) denotes censoring. This is a fractional support summary, not a replacement for fitting the survival likelihood: event times, censoring, covariates, and risk-set contributions also affect the information available for hazard parameters.

Parameter precision should be assessed directly. Examine standard errors, confidence intervals, correlations among estimates, and signs of numerical instability, such as a nearly singular information matrix or estimates that change markedly across reasonable starting values or resamples. A large standard error relative to the estimate often signals weak information, but no universal cutoff separates “supported” from “unsupported.” For survival parameters, a small event count is an important warning, not a complete calculation of precision.

A Joint Assessment

The following summaries answer related but non-interchangeable questions. Consider them together rather than collapsing them into one pass/fail score.

These measures should be compared across classes and across plausible fitted solutions. A class with high entropy may still have a substantial prevalence; a class with low entropy may still be rare. Likewise, a class may have a stable estimated prevalence but imprecise event parameters. Numerical thresholds for entropy, effective size, or event counts are context-dependent and should not be treated as universal rules.

Diagnostics, Alternatives, and Limits

A practical review begins with the posterior probability distribution: inspect the largest posterior probability per subject and whether uncertainty is concentrated in particular classes. Then report effective class sizes alongside hard-assigned counts, making clear which definition is being used. For each class, examine observed and posterior-weighted event counts, parameter standard errors and intervals, and whether estimates are sensitive to reasonable fitting choices. A warning sign is a combination such as a very small effective class, very few weighted events, and a large standard error for its event parameters. That combination points to limited support more convincingly than any one summary alone.

These checks address a different issue from local-maxima diagnosis. Multiple starting values and comparison of fitted likelihoods help identify optimization problems; they do not by themselves show that the resulting classes are well separated or precisely estimated. Conversely, diffuse posterior probabilities do not prove optimization failed. Distinguishing these explanations requires checking both the behavior of the fitting procedure and the class-support summaries.

Hard classification is a tempting alternative: assign each subject to the class with the largest \(\tau_{ig}\), then count subjects or events. It is easy to describe, but suppresses uncertainty and can make a borderline subject count as fully belonging to one class. Effective sizes preserve fractional membership, while posterior entropy summarizes the concentration of those probabilities. Neither provides a direct test of whether the number of classes is correct; model-choice criteria and sensitivity to class number address a different question. Nor does good separation establish that the class-specific longitudinal or event models are adequate.

A common mistake is to see a rare class with a few events and conclude either that it must be spurious or that its estimated risk is a reliable description of a distinct subgroup. Neither conclusion follows from the count alone. The correction is to report prevalence and posterior uncertainty, examine the event information and parameter precision, and qualify class-specific interpretations when evidence is sparse. Classes are model components, not automatically identifiable natural groups.