Class-Specific Covariate Effects

One Covariate, Three Different Roles

In a latent-class joint model, a baseline covariate can be related to several parts of the data-generating model. It may predict which latent trajectory class a subject belongs to, be associated with that subject’s longitudinal measurements within a class, and be associated with the subject’s event hazard within a class. These are distinct effects, even when the same covariate appears in all three components.

For example, a baseline risk factor might be associated with greater odds of belonging to a class with a particular trajectory. Within that class, it might also be associated with a lower measured outcome over time, while separately being associated with a higher event hazard. The coefficients are not interchangeable: they have different scales, condition on different quantities, and answer different questions.

The model below uses a baseline-category multinomial logit for class membership, a linear mean model for the longitudinal outcome, and a proportional-hazards model for the event. These are useful standard choices, not requirements of every latent-class joint model. Their assumptions must be assessed in the application.

Specify the Three Components Separately

Take class 1 as the reference class. For \(k=2,\ldots,K\), model class membership by

Here \(a_k\) is the change in the log odds of class \(k\) versus class 1 for a one-unit increase in \(x_i\). Thus, \(\exp(a_k)\) is the multiplicative change in those relative odds per one-unit increase. It is not, by itself, a change in the probability of belonging to class \(k\). The class probabilities depend on the predictors for all classes.

For a simple longitudinal component, suppose the conditional mean is linear in time and in the covariate:

The coefficient \(\gamma_k\) is the difference in mean outcome per one-unit difference in \(x_i\), at the same time and within class \(k\). Without a time-by-covariate interaction, the model assumes the covariate difference in mean outcome is constant over time. If the scientific question calls for a changing difference, one can add \(\eta_k x_i t_{ij}\); then the covariate contrast at time \(t\) is \(\gamma_k+\eta_k t\).

For the event component, a class-specific proportional-hazards model is

Here \(\lambda_{0k}(t)\) is the baseline hazard in class \(k\), and \(\theta_k\) is the log hazard ratio per one-unit increase in \(x_i\), conditional on class and the covariates in the model. The proportional-hazards assumption means that this covariate’s hazard ratio does not vary with time within class \(k\). The hazard describes the instantaneous event rate among subjects still event-free at time \(t\); it is not a probability of an event by time \(t\).

Worked Example: Class-Membership Odds

Worked Example: Longitudinal and Event Effects

Why the Effects Are Distinct

The distinction follows directly from the model equations. For a one-unit change in \(x_i\), the membership log-odds contrast is

so exponentiating gives the relative-odds ratio \(\exp(a_k)\). In the longitudinal mean model, the corresponding contrast at fixed class and time is

In the event model, the ratio of hazards at the same time and in the same class is

These calculations also show why a class-specific predictor value cannot be mistaken for a covariate effect. The odds ratio for a one-unit covariate contrast comes from the coefficient \(a_k\), not from exponentiating the entire class-membership linear predictor for just one covariate value.

Comparison, Diagnostics, and Limitations

A model allowing separate effects can be compared conceptually with more restricted specifications. A common-slope model constrains, for example, \(\gamma_k\) to be equal across classes; a model omitting \(x_i\) from the longitudinal component sets those effects to zero; and a membership-only model allows \(x_i\) to predict class without a within-class outcome association. These restrictions encode different scientific assumptions. If uncertainty about a restriction matters, estimate and report the relevant contrast with its uncertainty; model comparison procedures must account for the latent-class estimation setting rather than being applied mechanically.

Diagnostics should be matched to each component. Check whether the class-membership relationship is plausible over the observed covariate range and whether some covariate patterns nearly determine class membership. Within classes, inspect whether the longitudinal mean form and residual behavior are adequate, including whether a covariate-by-time interaction is needed. For the event component, assess whether the proportional-hazards assumption is tenable and whether each class has enough events and covariate variation to estimate its effect. Check convergence and sensitivity to starting values as well: unstable estimates can make apparent class-specific differences unreliable.

A coefficient in this model is an adjusted association under the specified model, not automatically a causal effect. Class membership is latent, and conditioning on a latent class is not the same as randomizing subjects to classes. Causal interpretations require additional design and identification assumptions. Also, if \(x_i\) is poorly supported in some classes, the corresponding class-specific effect may be imprecise or driven by extrapolation. A numerical estimate does not ensure that each class contains adequate information.

Class-specific effects are useful when the same covariate may relate differently to latent trajectory groups, measured outcomes, and event risk. Their flexibility is also a limitation: more parameters require more information, and results depend on the chosen class structure and component models. Report which component each coefficient belongs to, its contrast and units, and the class conditioning used in its interpretation.