Latent Classes as Unobserved Trajectory Subpopulations

Why Represent Trajectory Heterogeneity With Latent Classes?

A single longitudinal trajectory can summarize an average pattern while concealing important differences between people. For example, one subpopulation might show a gradual increase in a biomarker, while another shows a decline. If those patterns are also associated with different event-time distributions, modeling the longitudinal outcome and event outcome together may require representing more than one kind of subject.

A latent class joint model represents this heterogeneity with a finite set of unobserved subpopulations. Each subject belongs to one class, and that class determines a joint distribution for the subject’s longitudinal measurements and event outcome. The class is not directly recorded; it is a model-based description of recurring patterns in the data.

The word joint matters: the model represents the longitudinal process and event outcome together within each class. This definition does not, by itself, require that the two outcomes be independent given class membership. Their relationship within a class is part of the class-specific joint distribution. A latent class model therefore makes a different claim from a model that simply fits separate trajectory and survival summaries and treats them as unrelated.

What a Class Means—and What It Does Not Mean

A class is a discrete, unobserved category in the statistical model. Its membership is not the same thing as a recorded grouping variable such as treatment assignment, clinic, or disease subtype. Those groups have observed labels. A latent class does not: the model describes how likely different trajectory-and-event patterns are in the population without asserting that each subject’s class is directly known.

Nor does a class mean that every subject in it has identical measurements or event times. Class-specific trajectories and event distributions describe patterns, not exact individual outcomes. Measurement variation and other subject-to-subject variation can remain within a class. In a model with only discrete classes, the classes are the model’s representation of heterogeneity; a model may also include continuous subject-level random effects, but those are a distinct modeling component.

A continuous random effect, such as a subject-specific intercept or slope, represents a graded deviation drawn from a distribution. A latent class instead assigns subjects to a finite number of categories with different distributions or patterns. For example, “higher and increasing” versus “lower and decreasing” can be represented as two discrete classes; a random slope represents continuous variation in the rate of change. Neither representation automatically implies the other, and a class should not be interpreted as a measured biological type merely because its fitted trajectory has a convenient label.

Mixtures for Event-Time Outcomes

The same mixture idea applies to event-time distributions. Write \(S_g(t)=P(T_i>t\mid C_i=g)\) for the survival probability in class \(g\), and \(S(t)=P(T_i>t)\) for the population survival probability. If class membership is fixed and the class proportions describe the target population, then survival at time \(t\) is the weighted average of class-specific survival probabilities.

This is a population-level summary, not a class assignment for any particular subject. It also shows why class-specific event patterns matter: a mixture population can contain subjects with substantially different risks even when a single population survival curve can be calculated.

Why the Mixture Hazard Changes Over Time

A further consequence is that the population hazard is generally not a fixed average of the class-specific hazards. Let \(\lambda_g(t)\) be the hazard in class \(g\), and suppose the class-specific survival functions are differentiable. The hazard is the instantaneous event rate among subjects still event-free at time \(t\). By the definition of a hazard, the population hazard is the negative derivative of population survival divided by population survival.

The second equality follows because \(S_g'(t)=-S_g(t)\lambda_g(t)\), so differentiating \(S(t)=\sum_g\pi_gS_g(t)\) gives \(S'(t)=-\sum_g\pi_gS_g(t)\lambda_g(t)\). Dividing its negative by \(S(t)\) yields the displayed expression. The effective weights at time \(t\) are proportional to \(\pi_gS_g(t)\), not just to the baseline proportions \(\pi_g\). Classes with greater survival contribute more subjects to the event-free population at later times.

Observed Groups and Latent Classes Are Different

An observed group can be used to describe or stratify a population, but it is not itself a latent class. The distinction remains important even if the observed groups have different mixtures of latent trajectory patterns. A difference in group averages might reflect different class composition, different class-specific trajectories, or both. Those possibilities are not interchangeable, and a mixture description alone does not establish a causal effect of group membership.

Assumptions, Diagnostics, and Limitations

A finite latent class representation is useful only when its assumptions are defensible for the scientific question and data. In its basic form, it assumes that each subject belongs to one of \(G\) mutually exclusive and exhaustive classes, with class membership stable over the observation period. It also assumes that the chosen class-specific joint distributions provide a reasonable description of longitudinal and event outcomes within each class. The class proportions refer to the population being modeled; they need not describe a different population with a different composition.

Observed data may be censored or may have incomplete longitudinal measurements. The model’s assumptions about how those observations arise must be appropriate for the study design; the mixture definition does not make censoring or missingness ignorable. Likewise, representing outcomes jointly does not automatically guarantee that all sources of dependence have been captured.

Useful checks include comparing observed longitudinal patterns with the fitted class-specific trajectories, examining whether class-specific event patterns are compatible with the observed event experience, and checking whether important systematic patterns remain unexplained. Nearly identical class trajectories or event distributions may suggest that the proposed categories are not meaningfully distinct. A class supported by very little data can yield unstable summaries. These observations are warnings to investigate the specification and data support, not automatic tests that prove a class model is right or wrong.

A finite mixture is also an approximation: the population may have continuously varying trajectories rather than a small number of distinct subpopulations. Classes can sometimes make a continuous distribution look artificially discrete, and their interpretation can depend on how the model is specified. A single-population joint model with continuous random effects is an alternative way to represent heterogeneity. Recorded-group analyses answer a different question when group labels are observed. The choice among these approaches should reflect the scientific target and the patterns the model is meant to describe; the existence of fitted classes alone does not establish real, sharply separated subpopulations.

Common Mistakes

• Treating a latent class as an observed group. A class is unobserved and inferred through a model; a treatment arm or clinic label is directly recorded. They may be related, but they are not the same variable.
• Calling a class a continuous random effect. A finite class label describes discrete categories. A continuous random effect describes graded subject-level deviation. One cannot be substituted for the other without changing the model.
• Assuming class membership makes outcomes independent. The joint distribution within a class need not factor into separate longitudinal and event distributions. Independence would be an additional assumption.
• Averaging class hazards with baseline proportions at every time. The population hazard weights classes by their contributions to the event-free population at that time, as well as by their baseline proportions.
• Interpreting a mixture as proof of biological types or causation. Classes are model-based summaries of heterogeneity. Their labels and fitted patterns do not by themselves establish biological categories or causal effects.