Evaluating Class-Specific Trajectory Credibility

What Makes a Trajectory Difference Credible?

A latent class joint model can estimate a distinct longitudinal mean trajectory for each class. But a smooth curve is not, by itself, evidence that the classes truly have different trajectories. A contrast may be estimated precisely near a time where both classes have abundant measurements and poorly supported at the end of follow-up, where few members of one class remain observed. It can also change substantially when a small number of subjects or an unusual measurement pattern drives the fit.

Here, trajectory credibility means that the estimated class-specific difference is interpretable in light of its uncertainty and the data supporting it. It is not a separate test with a universal cutoff. The assessment combines the model-based contrast and its uncertainty with the distribution of follow-up times, class-specific information, and sensitivity to reasonable changes in the analysis. The aim is not to require every class to have identical measurement schedules, but to determine where the estimated differences are genuinely informed by observed data.

The contrast is conditional on the model, the covariate value \(x\), and the way the longitudinal outcome is defined. It describes a model-based difference in mean outcome, not a causal effect of class membership. A class is latent rather than a directly assigned treatment group. The interpretation also depends on the fitted joint model: its class structure, longitudinal mean specification, event submodel, and assumptions about the observation and event processes.

Uncertainty in a Contrast Is Only One Part of the Assessment

For a specified time and covariate value, the estimated contrast is a function of fitted parameters. If \(\widehat{\boldsymbol\theta}\) is the parameter estimate and \(\widehat{\Sigma}\) its estimated covariance matrix, the delta method gives an approximate variance:

A pointwise Wald interval is then \(\widehat{\Delta}_{gh}(t;x)\pm 1.96\,SE\{\widehat{\Delta}_{gh}(t;x)\}\) for an approximate 95% interval. This interval concerns the contrast at that particular time; it is not a 95% confidence band covering the whole trajectory. If the scientific question concerns a difference over a time range, pointwise intervals alone do not control the chance of at least one false positive across all examined times. A simultaneous band or a prespecified summary of the contrast over time is more appropriate for that question.

The covariance calculation should reflect the fitted joint model, including uncertainty in parameters that determine the trajectories. It should not be computed as if each subject's class were observed without error. Posterior classification uncertainty and class separation, addressed in the earlier material on those subjects, matter here because uncertain class membership can reduce the information available for separating class-specific means. A subject-level bootstrap that refits the full model can provide a useful check on the covariance-based uncertainty, provided label switching is handled consistently across fits.

Check Class-Specific Support Across Follow-Up

Support is not captured by the total sample size alone. Examine how many measurements occur at different follow-up times and how those measurements are distributed across classes. Useful summaries include counts in prespecified time windows, subject-level follow-up ranges, and plots of observed measurements or measurement times by class. Because class membership is uncertain, such summaries can use posterior probabilities \(w_{ig}=P(C_i=g\mid\text{observed data})\) as weights rather than assigning every subject to a single most likely class.

For example, the sum \(\sum_i w_{ig}\) is the total posterior weight for class \(g\) in a group of subjects or a time window. A complementary descriptive measure is the Kish effective sample size:

This measure is small when the total weight is concentrated in only a few subjects. It is a diagnostic summary, not a replacement for the likelihood-based uncertainty estimate, and it does not account for all dependence or model uncertainty. Report the time window and the weighting rule when using it. Counts of measurements and counts of distinct subjects should also be distinguished: many repeated records from a few people do not provide the same support as measurements spread across many people.

Support can deteriorate near the end of follow-up for several reasons, including administrative study end, dropout, or events that prevent further longitudinal measurement. A joint model uses its specified event and longitudinal submodels to address dependence represented by those components. That does not make unsupported regions empirically rich: the estimated late trajectory may still rely strongly on model assumptions. The earlier material on residual checks within latent classes can help identify poor fit, but satisfactory residual checks do not by themselves demonstrate adequate late-time coverage.

Interpolation, Extrapolation, and Measurement Patterns

A contrast at a time where both classes have observations nearby is generally more directly informed by data than a contrast far beyond the observed follow-up. This distinction is not absolute: a simple correctly specified model can estimate beyond observed times, while a flexible model may be unstable even within the observed range. Still, a trajectory that continues to separate after one class has almost no observed measurements is especially dependent on the chosen functional form and other model assumptions. Describe the range of observed times for each class, not merely the overall study range.

Also examine whether measurement patterns differ in ways that concentrate information unevenly. For instance, if subjects with worsening outcomes are measured more often, the data may overrepresent their trajectories unless the observation process is appropriately handled. If observation timing or missingness assumptions are not represented in the model, apparent class differences may partly reflect those patterns. A useful diagnostic is to compare measurement counts and timing across classes and across relevant outcome or event histories, while avoiding the mistaken conclusion that unequal counts automatically invalidate a model.

A Practical Assessment and Its Limits

A disciplined assessment can be organized as a sequence. First define the contrast, time range, covariate values, and outcome scale that answer the scientific question. Next inspect uncertainty in that contrast using the fitted covariance matrix and, where feasible, a full subject-level bootstrap. Then assess class-specific measurement coverage and effective support over the relevant times. Finally, examine sensitivity to plausible changes in follow-up windows, sample size, and measurement pattern, keeping the estimand fixed wherever possible.

This assessment complements rather than replaces other checks. A formal test of whether classes differ in event association addresses an event-process parameter, not whether longitudinal trajectory differences are supported throughout follow-up. Residual checks assess aspects of fit; model selection compares candidate class structures; and sensitivity analysis asks whether conclusions persist under alternatives. None alone answers every credibility question. In particular, a small p-value for a trajectory contrast cannot compensate for almost no observations in one class at the relevant time.

There is no universal minimum effective sample size or observation count that makes a trajectory credible. What is adequate depends on the model complexity, measurement variability, distribution of follow-up, class separation, and the purpose of the comparison. A class-specific mean can be estimable even when some subjects have uncertain membership, but small or poorly separated classes tend to make trajectory estimates less stable. Extrapolation, misspecified mean structure, and unmodeled observation processes remain limitations even when a numerical interval is available.

Common Pitfall