Canonical Shared Random-Effects Joint Model

Why Specify the Two Processes Together?

A longitudinal marker is measured repeatedly, while an event time records when an outcome occurs or, for some subjects, how long they remain event-free before follow-up ends. If the marker’s underlying trajectory is related to event risk, analyzing the measurements and event times as unrelated outcomes can fail to represent that connection. The canonical shared random-effects model expresses the two processes together: a linear mixed model describes the marker, and a hazard model uses the marker’s latent trajectory. Subject-level random effects enter that trajectory and are therefore shared by both parts.

The model is a specification, not a claim that the marker causes the event. Its usefulness depends on whether the assumed trajectory, hazard relationship, and data-collection conditions are scientifically plausible. The earlier material on linear mixed models and survival models supplies the standard component ideas; here the focus is how they form one basic joint model.

The Canonical Specification

Let \(Y_{ij}\) be the observed marker for subject \(i\) at measurement time \(t_{ij}\), and let \(T_i\) denote the event time. Write \(m_i(t)\) for subject \(i\)’s latent, or underlying, marker value at time \(t\). A basic linear mixed model for the measurements is

Here \(\mathbf{x}_i(t)\) and \(\mathbf{z}_i(t)\) are specified covariate vectors, \(\boldsymbol{\beta}\) is a vector of population-level coefficients, and \(\mathbf{b}_i\) is a vector of subject-specific random effects. The residual \(\epsilon_{ij}\) represents measurement-level variation around the latent trajectory. In the usual basic specification, \(\mathbf{b}_i\sim N(\mathbf{0},\mathbf{D})\) and \(\epsilon_{ij}\sim N(0,\sigma^2)\). Random effects are independent across subjects, and the residuals are independent across measurements and independent of the random effects.

The event submodel specifies the hazard at time \(t\), conditional on the subject’s random effects and event-model covariates:

In this expression, \(h_0(t)\) is the baseline hazard, \(\mathbf{w}_i\) is a vector of event-model covariates, \(\boldsymbol{\gamma}\) their coefficient vector, and \(\alpha\) the coefficient linking the latent marker value to the hazard. Since \(m_i(t)\) contains \(\mathbf{b}_i\), those same subject-level random effects influence the marker measurements and the event hazard. The canonical specification uses the latent marker trajectory as the route through which the random effects enter the hazard; adding other direct random-effect terms would define an extension.

Assumptions and Interpretation

A model statement is meaningful only together with its assumptions. The basic version requires a suitable mean trajectory: the chosen fixed- and random-effect terms should represent how the marker changes over time. It assumes a distribution for the random effects and a residual distribution and dependence structure appropriate to the measurements. The event submodel assumes the hazard has the stated baseline form and that covariates and the latent marker affect it through the stated linear predictor. For a proportional-hazards form, the covariate and marker contributions multiply the baseline hazard through exponentiation.

The event-time and marker processes are linked through the shared random effects and the latent trajectory. The model does not treat each noisy measurement as the exact marker value at every time. It also assumes that the event process, conditional on the modeled covariates and random effects, has no additional unmodeled dependence on the marker measurements. Censoring is ordinarily assumed independent of the event time conditional on the modeled information; informative censoring requires additional modeling. These are substantive assumptions, not automatic consequences of fitting the equations.

The parameter \(\alpha\) describes the model’s association between the latent marker level and the hazard, conditional on the other terms. For two otherwise comparable subjects at time \(t\), a latent marker difference of \(\Delta m\) implies a hazard ratio \(\exp(\alpha\Delta m)\), provided their event-model covariates are equal. This is a relative instantaneous hazard, not a difference in event probabilities and not, by itself, a causal effect.

Why the Shared Structure Matters

The connection can be seen by substituting the longitudinal trajectory into the hazard. This also gives a short justification for the model’s “shared” description:

The final term contains the subject’s random effects. A subject with a different random intercept or slope can therefore have a different latent marker path and, through that path, a different hazard. The model describes a dependence structure; it does not assert that the random effects themselves are observed or that this structure is the only plausible one.

Diagnostics, Alternatives, and Limitations

Diagnostics should examine both parts of the specification and the adequacy of their connection. For the longitudinal component, inspect residual patterns over time and across fitted values, and assess whether the assumed mean trajectory and random-effect structure capture systematic variation. For the event component, use appropriate survival-model diagnostics to assess the baseline and covariate structure, including whether the proportional-hazards form is reasonable. Also check numerical convergence and whether estimates are stable to plausible changes in the trajectory or baseline specification.

Warning signs include curved residual patterns when a linear time trend was specified, residual spread that changes markedly with fitted values, influential subjects, or an estimated random-effect covariance that is unstable. Event-model warning signs include systematic departures from the assumed hazard structure or strong time patterns left unexplained by the model. A convergence warning is not evidence that the model fits adequately. Sparse measurements near the event time, very few events, or weak marker variation can also make the association difficult to estimate.

A common alternative is to fit a survival model using observed marker measurements as time-varying covariates. That approach uses the recorded values directly and does not, in this canonical form, represent a latent trajectory with shared random effects. Separate longitudinal and survival analyses are another option, but they do not encode this shared structure in a single model. These alternatives may be useful for different questions and data conditions; they are not automatically equivalent to the joint model.

The canonical model has limits. A linear marker trajectory may be too restrictive; a single current latent value may not capture an association with cumulative exposure or rate of change. The proportional-hazards form may also be unsuitable. Results can depend on the random-effect distribution, measurement schedule, and assumptions about censoring and dependence. Finally, the fitted association is observational and conditional on the specified model; it does not establish a causal effect of changing the marker.

Common Mistake

A frequent error is to substitute the most recent observed marker \(Y_{ij}\) directly for \(m_i(t)\), then interpret the resulting hazard as if the model had used the latent trajectory. The reasoning is that the observed value is “the marker,” so it must equal the value entering the hazard. The correction is that the canonical model distinguishes the latent process from measurement error: \(Y_{ij}=m_i(t_{ij})+\epsilon_{ij}\). The hazard formula uses \(m_i(t)\), and the random effects connect that trajectory to the longitudinal observations. Another error is to call \(\exp(\alpha)\) the hazard ratio for any marker difference; it is the ratio for a one-unit difference in the marker’s stated units, with other terms held fixed.