Interpreting Class-Specific Trajectories and Hazards

Reading a Trajectory and an Event Hazard Together

A latent class joint model gives more than a collection of regression coefficients. It describes classes with different longitudinal patterns and different event-risk patterns. Interpretation therefore requires looking at the fitted outcome curve and the event hazard together. A class whose mean outcome rises over time may also have a larger event hazard, but neither pattern alone tells the full story: the trajectories might cross, or the hazard contrast might change over follow-up.

The goal is to describe the fitted profiles on interpretable scales and at relevant times. This is not the same as claiming that a trajectory causes an event, or that every individual assigned to a class follows its mean curve. The classes represent model-based subpopulations; individual outcomes and event times can vary within each class.

Some fitted models report a mean curve averaged over the random-effects distribution within class; others report a curve at a specified random-effects value or covariate pattern. Similarly, a hazard prediction may be conditional on a specified profile. Those details matter: comparisons are most interpretable when the same convention and covariate pattern are used for every class. In the examples below, predictions are stated to be on a common, specified basis.

A useful event-risk summary is the class-specific cumulative hazard \(H_c(t)=\int_0^t h_c(u)\,du\). The corresponding class-specific survival function is \(S_c(t)=\exp\{-H_c(t)\}\), when the hazard describes the event process under consideration. These quantities describe survival conditional on class under the model; they are not the overall survival curve for a population mixture of classes.

Assumptions Behind a Class Profile

Interpretation presumes that the fitted class structure is a useful representation of the data and that the model’s longitudinal and event components are adequately specified. In the latent class formulation, individuals belong to one of the modeled classes, and the class-specific curves and hazards are conditional descriptions. If the model also includes random effects within classes, the reported mean and hazard must be interpreted according to how those random effects were handled.

The conditional-independence assumption used in this setting says that, given class and the modeled random effects or other shared latent quantities, the longitudinal measurements and event process are independent. It does not say that they are unconditionally independent. The class-specific profiles are therefore meaningful only in relation to the fitted joint model and its assumptions.

• Time and units must be clear. State the outcome scale and the time unit. A slope in outcome units per year is not interchangeable with a slope per month.
• Hazard comparisons must be aligned. Compare hazards at the same follow-up time and under the same covariate or random-effects setting.
• Censoring and observation must be handled appropriately. The event and longitudinal likelihoods rely on the model’s assumptions about censoring and the observation process; informative loss not represented by the model can distort the fitted profiles.
• Do not over-interpret beyond observed support. A curve far beyond the observed follow-up may reflect the chosen functional form more than reliable information in the data.

Compare Curves Before Comparing Coefficients

A coefficient is a compact description of one model component, but the clinical interpretation usually concerns predicted outcome levels, changes, and event rates. For example, a linear mean curve \(m_c(t)=\beta_{0c}+\beta_{1c}t\) has an intercept \(\beta_{0c}\) in outcome units and a slope \(\beta_{1c}\) in outcome units per time unit. A contrast between slopes does not by itself say which class has the larger mean at a clinically relevant time; that depends on both intercepts and slopes.

Likewise, if a class-specific hazard is expressed on a log scale, exponentiating a contrast gives a hazard ratio. A hazard ratio describes a relative instantaneous event rate, not a difference in event probabilities. Absolute hazards and, when appropriate, cumulative hazards or class-specific survival can help convey the size of the difference in context.

Quantifying Event-Risk Contrasts

Suppose class-specific hazards are reported at a particular time \(t\), with \(h_A(t)\) and \(h_B(t)\) in events per person-year. The absolute hazard difference is \(h_B(t)-h_A(t)\), also in events per person-year. The hazard ratio is \(h_B(t)/h_A(t)\), when \(h_A(t)>0\). These are distinct summaries: a large ratio can correspond to a small absolute difference when both hazards are small.

For a proportional-hazards contrast with class-specific log hazards \(\eta_A(t)\) and \(\eta_B(t)\), the ratio is obtained by exponentiating the difference. This follows because a log hazard is the natural logarithm of the hazard:

If the model allows the class contrast to vary over time, this ratio must be evaluated at specified times; a single hazard ratio would conceal that variation. Regardless of the parameterization, distinguish a pointwise hazard from an interval risk. The hazard is an instantaneous rate conditional on remaining event-free, whereas the probability of an event over an interval depends on the hazard over that whole interval.

Build a Clinically Meaningful Joint Profile

The interpretation should bring the outcome and event components together without treating one as proof of the other. A clear profile might say that one class has a persistently higher symptom mean and a higher fitted hazard at specified times. That is more informative than reporting, for example, only that one slope differs from zero or that a hazard coefficient is positive. But the profile remains conditional on the fitted model, its class definition, and the prediction settings.

Uncertainty, Diagnostics, and Alternatives

A profile should include uncertainty when it is available. Pointwise confidence intervals for mean curves can show how precisely the class means are estimated at particular times; uncertainty bands or intervals for hazards can do the same for event risk. A contrast such as a hazard ratio should ideally have its own interval, because uncertainty in two estimated hazards is not captured by simply comparing their separate intervals. An interval that includes the null ratio of 1 does not prove the classes have equal hazards; it indicates that the data and model have not ruled out that value at the corresponding confidence level.

Check fitted curves and hazards over the observed follow-up, rather than relying on a table of coefficients alone. Look for implausible curvature, abrupt changes unsupported by data, crossing curves hidden by a single slope summary, and hazards that become extreme near the edge of follow-up. Confirm that displayed comparisons use the same time scale, units, covariate pattern, and random-effects convention. If class-specific estimates are imprecise, avoid confident clinical labels even if the point estimates look widely separated.

Alternative summaries answer different questions. A trajectory-only clustering approach may describe outcome patterns but does not use the event process to define the joint likelihood. A model with one pooled event hazard can obscure class-specific risk differences. Conversely, reporting only the class-specific event coefficients can hide how longitudinal means change over time. Joint profiles are useful because they present both fitted components, but they do not replace model checking or resolve uncertainty in class assignment.

A common mistake is to read the hazard ratio as a risk ratio: “class H has four times the probability of an event.” The correction is that the ratio compares instantaneous rates among individuals still event-free at the specified time, conditional on the model’s settings. Another mistake is to infer causation from the alignment of a rising mean curve and higher hazard. The correction is to describe the fitted association and avoid claims about what would happen under an intervention. Finally, a class-specific hazard is not automatically the hazard for the full population; averaging or combining class-specific hazards requires accounting for the class mixture and who remains event-free over time.