What a Latent Class Represents
In a latent class joint model, a person’s longitudinal measurements and event-time information contribute to estimating an unobserved class. The model can use classes to represent distinct patterns in the data, but this does not by itself establish that people belong to objectively distinct biological types. A class is first a component of a statistical model: its meaning depends on the model, the variables included, and the data available.
Let \(G_i\) denote the unobserved class for person \(i\), with \(G_i\in\{1,\ldots,K\}\). Write \(Y_i\) for the person’s longitudinal data, \(T_i\) for observed follow-up time, and \(\delta_i\) for the event indicator, where \(\delta_i=1\) denotes an observed event and \(\delta_i=0\) denotes censoring. Let \(X_i\) denote included covariates. A mixture model represents the observed-data contribution as a sum over possible classes:
Here, \(\pi_{ic}\) is the model probability of class \(c\) given the covariates, and \(f_c\) is the class-specific contribution for the observed longitudinal and event-time data. When class probabilities do not depend on covariates, \(\pi_{ic}\) is simply the class proportion \(\pi_c\). The likelihood sums over classes because \(G_i\) is not observed.
This distinction matters because several different data-generating situations can yield similar fitted classes. A mixture may approximate a continuous range of trajectories, represent unmodeled covariate differences, or absorb a feature omitted from the class-specific model. Even when the classes are useful summaries, their boundaries need not correspond to natural boundaries in the population. The earlier material on label switching also matters: class labels such as “1” and “2” are names for components, not intrinsic identities.
Posterior Probabilities Are Not Certain Labels
After fitting the model, the posterior probability that person \(i\) belongs to class \(c\), conditional on the observed data and fitted parameters, is
The probabilities \(\tau_{ic}\) sum to one across classes. A person with a largest posterior probability of \(0.62\) is often assigned to the corresponding class for a summary, but that hard assignment is a decision rule, not a new observation of \(G_i\). The probability is conditional on the model and its fitted parameters; it does not account automatically for uncertainty in those parameters, uncertainty about the number of classes, or model misspecification.
Worked Example: A Posterior Probability Is Not Certainty
Suppose a hypothetical two-class model gives a person prior class probabilities \(\pi_{i1}=0.40\) and \(\pi_{i2}=0.60\). The observed data have class-specific likelihood contributions \(f_1=0.30\) and \(f_2=0.10\). Calculate the posterior probability for class 1.
The unnormalized contributions are \(0.40(0.30)=0.12\) for class 1 and \(0.60(0.10)=0.06\) for class 2. Therefore,
The observed data shift the model-based probability toward class 1, despite its smaller prior probability. The result supports saying that class 1 is more probable than class 2 under this fitted model. It does not justify saying that the person is known with certainty to be a member of class 1, nor that there is a 66.7% chance of belonging to a biological type defined independently of the model.
Hard classification also changes summaries. If a report counts people assigned to a class, a person with posterior probabilities \(0.51\) and \(0.49\) contributes one whole person to the first class, while a person with probabilities \(0.99\) and \(0.01\) contributes the same amount. The assignments hide how differently certain those cases are.
Worked Example: Hard Counts Versus Posterior-Weighted Counts
For four hypothetical people, posterior probabilities for class A are \(0.91, 0.58, 0.52,\) and \(0.14\). Compare the hard-assignment count with the posterior-weighted count for class A.
Under the maximum-posterior rule, the first three people are assigned to A and the fourth to the other class. The hard count is therefore 3 people. The posterior-weighted count is the sum of the probabilities:
The weighted value \(2.15\) is the model-based expected number in A for these four people, conditional on the fitted model and data. It is not a literal fractional number of people, and it is not an observed census of biological types. The difference between 3 and 2.15 also shows why a hard count can give a misleading impression of how clearly the sample separates into classes.
Class Membership Does Not Explain Every Dependence
Some latent class joint models assume conditional independence between the longitudinal process and event time given class and covariates. Other specifications also include shared or class-specific random effects; in those models, independence may be assumed only after conditioning on both class and random effects. These are model assumptions, not consequences of introducing classes. In particular, a class model does not guarantee that all residual association between longitudinal measurements and event risk disappears within each class.
A useful way to see the distinction is the covariance decomposition. For a longitudinal summary \(Y\) and an event-related variable \(Z\), such as an indicator of being event-free at a fixed time, the law of total covariance gives
The first term is average within-class dependence. The second is dependence arising from differences in class-specific means. Class membership accounts for the second term, but it does not make the first term vanish. The decomposition holds when the relevant expectations and covariances exist; using a bounded event indicator \(Z\) avoids requiring a finite mean for an event time.
Worked Example: Dependence Can Remain Within Classes
Consider a hypothetical population split equally between two classes. Let \(Y\) be a biomarker summary and let \(Z=1\) indicate remaining event-free through a specified follow-up time. In class 1, suppose \(E(Y\mid G=1)=0\), \(P(Z=1\mid G=1)=0.80\), and \(\operatorname{Cov}(Y,Z\mid G=1)=0.20\). In class 2, suppose the corresponding values are \(10\), \(0.20\), and \(0.20\). Calculate the overall covariance and separate its two components.
The overall means are \(E(Y)=0.5(0)+0.5(10)=5\) and \(E(Z)=0.5(0.80)+0.5(0.20)=0.50\). Within-class dependence contributes \(0.5(0.20)+0.5(0.20)=0.20\). The class-specific mean differences contribute
Thus the overall covariance is \(0.20-1.50=-1.30\). Directly, \(E(YZ)=0.5\{0(0.80)+0.20\}+0.5\{10(0.20)+0.20\}=1.20\), so \(\operatorname{Cov}(Y,Z)=1.20-(5)(0.50)=-1.30\), as required. Class differences account for a substantial component, but positive dependence remains within each class. Treating class membership as the entire explanation would miss that feature.
Assumptions, Diagnostics, and Alternatives
Interpretations of class-specific patterns rely on the joint model being an adequate description of the observed data. Important assumptions include a suitable number and form of classes, correctly specified class-specific longitudinal and event-time models, and appropriate handling of censoring and missing longitudinal measurements. The required censoring or missingness assumptions depend on the likelihood being used; they should not be presumed merely because a mixture model has been fitted. If a model asserts conditional independence within classes, residual dependence checks are especially relevant.
Several warning signs call for restraint in interpretation:
- Many people have posterior probabilities close to one another across classes, so hard assignments are unstable.
- Class-specific residual checks reveal systematic patterns or remaining association between longitudinal outcomes and event indicators.
- Class profiles or membership proportions change substantially under reasonable choices of class number, trajectory shape, or model specification.
- A small or weakly separated class is described with more biological certainty than its estimates support.
- Estimated class patterns are being treated as causal effects. A class-specific association does not by itself establish that class membership causes an outcome.
The main alternative is not always another finite mixture. A model with continuous heterogeneity, such as random effects, may be more appropriate when variation appears gradual rather than clustered. A mixture with within-class random effects can represent both between-class pattern differences and variation among people in the same class. Neither option is universally superior: the choice should reflect the scientific question and be checked against the data, rather than inferred from an appealing class interpretation.
A common mistake is to reason, “The model found two classes, so the sample contains two known biological types, and class membership explains the dependence.” The correction is to make three separate claims only when supported: the fitted model represents the data with two components; an individual’s posterior probabilities quantify model-based classification uncertainty; and within-class residual dependence must be assessed rather than assumed away. These distinctions make class results useful without making them more definitive than the model warrants.
State the fitted class structure and what observed processes contribute to class estimation.
Use posterior probabilities when describing ambiguous membership; identify hard assignment as a summary rule.
Assess residual patterns and dependence, especially when conditional independence is part of the model.
Present classes as model-based descriptions unless independent scientific evidence supports a stronger interpretation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A person has posterior probabilities \(0.54\) and \(0.46\) for two classes. What does assigning the person to the first class say, and what does it not say?
- For five people, the posterior probabilities for class A are \(0.80, 0.70, 0.50, 0.30,\) and \(0.10\). Find the posterior-weighted count for A. How many people would be assigned to A by the maximum-posterior rule?
- In the covariance decomposition, what does each term represent? Does a nonzero within-class term contradict the use of latent classes?
- Give one reason a fitted class should not automatically be called a biological subtype and one diagnostic that could make its interpretation less credible.
Solutions: (1) The first class has the larger posterior probability under the fitted model, so the assignment follows a maximum-posterior rule. It does not establish certain membership or a verified biological identity. (2) The weighted count is \(0.80+0.70+0.50+0.30+0.10=2.40\). Three people have probabilities greater than \(0.50\), so three are assigned to A; the person at \(0.50\) is tied and requires a stated tie rule. (3) The first term is average within-class dependence; the second is dependence due to differences in class-specific means. A nonzero first term is possible and indicates that classes do not account for all dependence. (4) Classes may approximate continuous heterogeneity or absorb omitted structure, so a fitted component is not proof of a natural type. For example, unstable class profiles across reasonable model specifications or systematic within-class residual patterns weaken the interpretation.