Reporting Latent Class Joint Model Results

What a Results Report Must Make Clear

A latent class joint model results section has to communicate two things at once: what longitudinal and event patterns the fitted model describes, and how uncertain the model’s class structure is. Reporting only a table of coefficients can obscure whether classes are prevalent, distinguishable, and adequately estimated. Reporting only a narrative such as “one group progressed faster” can conceal the scale, uncertainty, and assumptions behind that claim.

A useful report presents a connected set of results: the analysis population and event follow-up; the selected model and its estimation checks; estimated class proportions; class-specific longitudinal trajectories and hazard summaries; posterior classification uncertainty; and model-selection evidence. The conclusions should describe patterns conditional on the fitted model, not assert that the classes are known biological types or that class membership causes an outcome.

Reporting principle: A latent class is a component of a probability model. Class proportions, trajectories, and hazards are estimated features of that model; an individual’s class membership is uncertain unless the posterior probability is essentially one. State both the estimated pattern and the uncertainty relevant to it.

State the Model and Its Interpretive Conditions

Use enough detail for a reader to understand what the estimates refer to. Let \(C_i\) denote the latent class for participant \(i\), with \(G\) possible classes, and let \(\pi_g=P(C_i=g)\) be the class proportion when membership probabilities do not depend on covariates. If membership depends on covariates, describe the membership model and clarify whether reported proportions are conditional or standardized. A class-specific latent mean trajectory may be written \(m_g(t)\). For example, a joint model may link the event hazard to the underlying marker trajectory through

h_i(t\mid C_i=g)=h_{0g}(t)\exp\{\boldsymbol{\gamma}_g^\mathsf{T}\boldsymbol{Z}_i+\alpha_g m_i(t)\}.

Here \(h_{0g}(t)\) is the class-specific baseline hazard, \(\boldsymbol{Z}_i\) contains stated covariates, and \(\alpha_g\) is the class-specific association between the latent marker level and the hazard on the log-hazard scale. This is an illustration of a common specification, not a requirement that every fitted model use this exact form. Report the actual longitudinal and event submodels, the link between them, the time scale, and any covariate adjustment. State whether the association is with current latent marker level, a slope, or another feature.

Interpretation relies on the fitted model being a reasonable representation of the data: the chosen number and form of classes must be defensible; the longitudinal and event submodels must be adequately specified; estimation must have reached a satisfactory solution; and the censoring and missing-data assumptions must be appropriate for the analysis. In the usual joint likelihood, each person’s contribution sums over classes and integrates over any random effects. Conditional factorization of the longitudinal and event contributions is justified only under the dependence structure specified by that model. A report should name those assumptions rather than imply that the fitting procedure eliminates them.

Class labels are arbitrary. In a two-class model, “class 1” and “class 2” can exchange labels without changing the fitted mixture. Use descriptive labels only after defining them by a reproducible feature, such as the lower versus higher estimated slope, and keep that ordering consistent across tables and figures. This avoids making a label appear to be an intrinsically identified population identity.

Report Class Proportions and Trajectories

Report estimated class proportions with uncertainty when available, and say how they were obtained. These are mixing probabilities in the model, not necessarily the fractions produced by assigning each participant to their most probable class. If the model includes covariates in class membership, a single proportion may not exist without specifying a covariate distribution or standardization target.

Worked Example: Describing the Estimated Class Structure

Suppose a hypothetical analysis includes 240 participants and selects a two-class joint model. The estimated mixing proportions are \(\hat{\pi}_1=0.65\) and \(\hat{\pi}_2=0.35\), with bootstrap 95% confidence intervals of 0.59 to 0.71 and 0.29 to 0.41, respectively. The intervals complement one another because \(\pi_2=1-\pi_1\) in a two-class model.

The fitted model therefore assigns an estimated 65% of the population mixture to class 1 and 35% to class 2. Multiplying the estimates by the sample size gives expected class counts of \(240(0.65)=156\) and \(240(0.35)=84\). These are not necessarily the observed numbers assigned by maximum posterior probability. If hard assignment yields 150 and 90 participants, those counts describe a classification rule applied to the posterior probabilities, not a replacement for the estimated mixture proportions.

A careful report would identify the intervals as bootstrap intervals, state how the bootstrap was conducted, and explain how labels were aligned across resamples. Without label alignment, switching labels can make class-specific bootstrap summaries misleading.

For each class, present trajectories on the original measurement scale when possible. Give the time origin and units, and distinguish a population-level latent mean trajectory from individual trajectories or raw observed values. If a trajectory is summarized by an intercept and slope, clarify the reference time for the intercept. A plot should show uncertainty bands or intervals where useful, while avoiding visual suggestions of precision unsupported by the data.

Worked Example: Translating Trajectory Estimates Into Context

Suppose a hypothetical two-class model estimates mean biomarker trajectories in milligrams per liter, with time \(t\) measured in years since enrollment:

m_1(t)=9.8+1.1t,\qquad m_2(t)=10.2+2.8t.

At two years, the fitted class 1 mean is \(m_1(2)=9.8+1.1(2)=12.0\) mg/L. The fitted class 2 mean is \(m_2(2)=10.2+2.8(2)=15.8\) mg/L. The slopes are 1.1 and 2.8 mg/L per year, respectively. If model-based 95% confidence intervals for the two-year means are 11.5 to 12.5 mg/L and 14.9 to 16.7 mg/L, report these as uncertainty intervals for the fitted means, not as ranges containing 95% of individual measurements.

The class 2 trajectory rises more steeply in this fitted model. That statement is descriptive. Whether the slope difference is statistically distinguishable requires its own uncertainty calculation, and the estimated trajectories do not establish that membership causes biomarker progression.

Summarize Event Associations and Classification Uncertainty

For an event submodel, report the hazard scale, the contrast represented by each coefficient, the time scale, adjustment variables, and uncertainty. If the association parameter multiplies the current latent marker, its exponent is the hazard ratio for a one-unit higher underlying marker at a given time, conditional on the model’s other terms. It is not automatically a comparison of the baseline hazards between classes. When class-specific baseline hazards are allowed, comparing class-specific risks requires more than comparing association coefficients.

Worked Example: Reporting Class-Specific Hazard Associations

Suppose the fitted association estimates for a one mg/L higher current latent marker are \(\hat{\alpha}_1=0.262\) with \(SE=0.086\), and \(\hat{\alpha}_2=0.077\) with \(SE=0.061\). Under the stated proportional-hazards association, the class-specific hazard ratios are exponentiated coefficients.

For class 1, \(\exp(0.262)=1.300\). A Wald 95% interval on the log-hazard scale is \(0.262\mathbin{\pm}1.96(0.086)\), or \(0.0934\) to \(0.4306\). Exponentiating gives a hazard-ratio interval of approximately 1.098 to 1.538. The Wald statistic is \(0.262/0.086=3.047\), giving a two-sided \(p\)-value of about 0.0023. In context, conditional on the specified model, a one mg/L higher current latent marker is associated with an estimated 30% higher instantaneous event hazard in class 1; the interval is approximately 10% to 54% higher.

For class 2, \(\exp(0.077)=1.080\). The log-scale interval is \(0.077\mathbin{\pm}1.96(0.061)\), or \(-0.0426\) to \(0.1966\), which exponentiates to approximately 0.958 to 1.217. The Wald statistic is \(0.077/0.061=1.262\), with two-sided \(p\) approximately 0.207. The interval includes 1, so these data do not provide strong evidence of a nonzero class 2 association at the conventional 0.05 level. They do not prove the association is absent.

Posterior classification summaries describe how strongly the observed data support individual class assignments. Define \(\tau_{ig}=P(C_i=g\mid\text{observed data for person }i)\). The posterior probability follows from Bayes’ rule: the prior class probability is multiplied by that person’s class-specific observed-data likelihood, then normalized across classes. Report a distribution or summary of these probabilities, not only hard assignments. A mean maximum posterior probability, for example, is easy to understand but can hide a subset of highly uncertain participants.

Normalized entropy: For \(G\) classes, person \(i\)’s normalized posterior entropy is \(E_i=-\sum_{g=1}^{G}\tau_{ig}\log(\tau_{ig})/\log(G)\), with \(0\log(0)\) defined as 0. It ranges from 0 for a nearly certain assignment to 1 for equal posterior probabilities across all classes. The sample mean summarizes average uncertainty; it does not measure whether the classes are scientifically meaningful.

The range follows from the entropy function: \(-\sum_g \tau_{ig}\log(\tau_{ig})\) is zero when all probability is concentrated in one class and is maximized at \(\tau_{ig}=1/G\), where it equals \(\log(G)\). Dividing by \(\log(G)\) therefore scales the measure to 0–1. This is a measure of posterior allocation uncertainty, not uncertainty in the estimated class proportions or model parameters.

Worked Example: Quantifying One Participant’s Classification Uncertainty

For a hypothetical participant in a two-class model, suppose the posterior probabilities are 0.72 for class 1 and 0.28 for class 2. The normalized entropy is

E_i=-\frac{0.72\log(0.72)+0.28\log(0.28)}{\log(2)} \approx\frac{0.59295}{0.69315}=0.85545.

The normalized entropy is approximately 0.855, indicating substantial allocation uncertainty for this participant relative to the maximum possible uncertainty in a two-class model. The maximum-posterior rule would assign the participant to class 1, but that label should not be reported as if it were certain. A report can communicate both the assigned class and the posterior probabilities, or summarize uncertainty across all participants using the mean entropy and the distribution of maximum posterior probabilities.

Present Model-Selection Evidence and Diagnostics

State why the reported number of classes and trajectory shapes were retained. A compact comparison can include the maximized log likelihood, number of estimated parameters, AIC, BIC, and relevant substantive or diagnostic considerations. Lower AIC or BIC indicates a preferred balance under that criterion among the candidates compared; it does not prove that the selected model is true or give the probability that it is correct. In mixture models, regularity conditions behind simple likelihood-ratio reference distributions may fail, so do not report a routine chi-square likelihood-ratio test for class number without a method justified for that setting.

Worked Example: Reporting Information-Criterion Evidence

Suppose two candidate models are fitted to the same 240 participants and the same observed-data likelihood. The one-class model has \(\ell=-1872\) and \(q=7\) parameters; the two-class model has \(\ell=-1840\) and \(q=12\). For AIC, \(AIC=-2\ell+2q\). Thus the one-class AIC is \(-2(-1872)+2(7)=3744+14=3758\), and the two-class AIC is \(-2(-1840)+2(12)=3680+24=3704\).

For BIC, \(BIC=-2\ell+q\log(n)\), with \(n=240\) here. Since \(\log(240)\approx5.48064\), the one-class BIC is \(3744+7(5.48064)\approx3782.36\), and the two-class BIC is \(3680+12(5.48064)\approx3745.77\). Both criteria favor the two-class candidate among these two fitted models. A balanced report adds whether the two-class solution replicated across starting values, whether classes were adequately represented, and whether within-class residual checks and trajectory credibility were acceptable.

Diagnostics are part of the results, not optional technical detail. Report whether multiple starting values converged to a comparable optimum, whether estimates were finite and plausible, and whether any class was very small or poorly separated. Mention checks of longitudinal residual behavior and event-model fit, along with sensitivity to plausible class counts and trajectory forms. If a class disappears, parameters are unstable, or results change materially across reasonable specifications, qualify the substantive claims accordingly.

Common Reporting Errors, Alternatives, and Limits

Confusing mixing proportions with assigned counts. A proportion such as \(\hat{\pi}_g\) is a model estimate; the number assigned by a maximum-posterior rule depends on a classification decision.
Reporting only the most likely class. A hard label discards posterior uncertainty. Give posterior summaries or explain that assignments are uncertain.
Interpreting a hazard coefficient without its scale. State the marker unit and whether the association is with its latent current value, slope, or another modeled quantity.
Calling a class a discovered natural subtype. A fitted mixture can summarize heterogeneity, but the estimated classes depend on model form, covariates, measurement schedule, and selected class count.
Treating a lower information criterion as proof. It is comparative evidence within the candidate set, not a test that validates the model or establishes the truth of its classes.

Alternatives should be discussed in relation to the scientific aim. A continuous random-effects joint model can represent heterogeneity without dividing participants into discrete latent groups. A single-class model can be preferable when additional classes are unstable or add little useful description. A different class count or trajectory specification may fit better but can change proportions and interpretations. These approaches answer related but not identical modeling questions; avoid implying that one is universally superior.

Even a well-fitting latent class joint model has limitations. Estimated classes are conditional on assumptions about class membership, trajectory shape, within-class variation, event association, and censoring. Classification uncertainty can be substantial, especially for individuals with overlapping patterns. Class-specific estimates may also be imprecise when a class is rare or has few events. Avoid causal language unless the design and assumptions support causal inference, and distinguish uncertainty about a person’s class from uncertainty about the fitted parameters.

Key takeaway: A defensible report links each claim to its model-based quantity and uncertainty: proportions describe the mixture, trajectories describe class-specific longitudinal patterns, hazard summaries describe conditional event associations, and posterior probabilities describe allocation uncertainty. None alone establishes fixed, causal, or naturally occurring subgroups.

Check Your Understanding

Answer each question, then compare with the solutions below.

A two-class fit estimates \(\hat{\pi}_1=0.40\). What does this proportion represent, and what does it not necessarily equal?
A class-specific marker association is 0.20 per mg/L on the log-hazard scale. Give the corresponding hazard ratio and interpret its unit.
For a two-class participant, posterior probabilities are 0.50 and 0.50. What is the normalized entropy, and what does it indicate?
Both AIC and BIC are lower for a three-class candidate than for a two-class candidate. Name two additional considerations before interpreting the extra class.

Solutions. (1) It is the estimated mixture probability of class 1 under the fitted model, not necessarily the fraction assigned to class 1 by a maximum-posterior rule. (2) \(\exp(0.20)=1.221\); conditional on the specified model, a one mg/L higher latent marker is associated with an estimated 22.1% higher hazard in that class. (3) \(E_i=-[0.5\log(0.5)+0.5\log(0.5)]/\log(2)=1\), the maximum allocation uncertainty for two classes. (4) Check convergence across starting values, class size and separation, residual and event-model fit, and sensitivity of substantive conclusions; lower criteria alone do not establish a meaningful or stable class.