Random-Effects Covariance and Apparent Link Sensitivity

Why the Random-Effects Covariance Matters

In a joint model, the longitudinal submodel describes each subject’s underlying biomarker trajectory, while the event submodel relates the event hazard to that trajectory or to a feature of it. Subject-specific deviations from the population trajectory are represented by random effects. Their covariance structure determines not only how much trajectories vary, but also how intercept deviations and slope deviations tend to occur together. Because the hazard link uses the latent trajectory, a covariance assumption can affect the estimated association even when the link itself has not changed.

This is a distinct question from comparing a current-value link with a slope or cumulative-exposure link. Here, the target is the sensitivity of the fitted trajectory and hazard-link estimate to assumptions about the distribution of the random effects. A useful comparison changes those assumptions in paired fits while holding the link and other modeling choices fixed.

Model and Covariance Structures

Definition: For subject $i$ at visit time $t_{ij}$, a linear longitudinal submodel with a random intercept and slope is $$ Y_{ij}=m_i(t_{ij})+\epsilon_{ij}, \qquad m_i(t)=\beta_0+\beta_1t+b_{0i}+b_{1i}t. $$ Here $Y_{ij}$ is the observed biomarker, $m_i(t)$ is its latent trajectory, $\beta_0,\beta_1$ are population fixed effects, and $b_{0i},b_{1i}$ are subject-specific random effects. A common assumption is $\begin{pmatrix}b_{0i}\\b_{1i}\end{pmatrix}\sim N\left(\begin{pmatrix}0\\0\end{pmatrix},D\right)$, independently across subjects, with $$ D=\begin{pmatrix}\tau_{00}&\tau_{01}\\\tau_{01}&\tau_{11}\end{pmatrix}. $$ The residuals $\epsilon_{ij}$ are commonly assumed independent of the random effects, mean zero, and normally distributed with variance $\sigma^2$, conditional on the model covariates and random effects.

The diagonal entries of $D$ are the random-intercept and random-slope variances. The off-diagonal entry $\tau_{01}$ is their covariance. If both variances are positive, their correlation is $\rho=\tau_{01}/\sqrt{\tau_{00}\tau_{11}}$. A positive covariance means that subjects with higher-than-average intercepts tend, on average, to have higher-than-average slopes; a negative covariance means the reverse. The intercept is the deviation at $t=0$, so its interpretation depends on the time origin.

Three common specifications are a random-intercept model, with $\tau_{11}=0$; a random-intercept-and-slope model with diagonal $D$, setting $\tau_{01}=0$; and a random-intercept-and-slope model with unstructured $D$, estimating all three distinct covariance parameters. These are substantive restrictions, not merely different ways of writing the same model. In particular, a diagonal covariance does not mean that subjects have no slope variation; it means that intercept and slope deviations are assumed uncorrelated.

How Covariance Changes the Distribution of Trajectories

Let $z(t)=(1,t)^\mathsf{T}$. The random part of the trajectory is $z(t)^\mathsf{T}b_i=b_{0i}+tb_{1i}$. Its variance and its covariance between two times follow directly from the covariance matrix:

\operatorname{Var}\{m_i(t)\}=z(t)^\mathsf{T}Dz(t) =\tau_{00}+2t\tau_{01}+t^2\tau_{11}, \qquad \operatorname{Cov}\{m_i(t),m_i(s)\}=z(t)^\mathsf{T}Dz(s).

Thus, setting $\tau_{01}=0$ removes the linear-in-time term from the variance curve. The population mean $\beta_0+\beta_1t$ need not change as a mathematical consequence, but refitting the model can change the estimated fixed effects as well as the estimated random-effects distribution. More importantly for a joint model, different distributions of latent trajectories imply different distributions of the quantity entering the hazard.

For example, with a current-value association, a typical conditional hazard is

h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\mathsf{T}w_i+\alpha m_i(t)\},

where $h_0(t)$ is the baseline hazard, $w_i$ is a vector of baseline covariates, $\gamma$ is its coefficient vector, and $\alpha$ is the association parameter per one biomarker unit. Other association links use a different specified trajectory feature, but the same point applies: the random-effects distribution determines the distribution of the linked feature. In estimation, the event contribution is integrated over the random effects, weighted by the longitudinal data contribution. Changing $D$ changes that integration and can therefore change $\hat{\alpha}$, even when the link formula is held fixed.

Worked Examples

Worked Example: Comparing Trajectory Variance Under Two Covariance Assumptions

Suppose time is measured in years from baseline, and two fitted random-intercept-and-slope models have the same estimated variances: $\hat{\tau}_{00}=4$ biomarker-units squared and $\hat{\tau}_{11}=0.09$ biomarker-units squared per year squared. The unstructured fit estimates $\hat{\tau}_{01}=0.30$ biomarker-units squared per year; the diagonal fit fixes it at zero. The implied intercept-slope correlation in the unstructured fit is $0.30/(\sqrt{4}\sqrt{0.09})=0.30/(2\times0.30)=0.50$.

At $t=2$ years, the random-effects contribution to trajectory variance is $4+2(2)(0.30)+2^2(0.09)=4+1.20+0.36=5.56$ biomarker-units squared under the unstructured fit. Under the diagonal fit it is $4+2^2(0.09)=4.36$ biomarker-units squared.

At $t=5$ years, the corresponding variances are $4+2(5)(0.30)+5^2(0.09)=4+3.00+2.25=9.25$ and $4+5^2(0.09)=6.25$ biomarker-units squared. The positive covariance increases trajectory dispersion at these later times. This calculation concerns between-subject random-effects variation; it does not include measurement-error variance $\sigma^2$.

Worked Example: Paired Fits and the Hazard Association

Consider hypothetical joint-model fits to the same data, using the same current-value link, time origin, covariates, baseline-hazard specification, and estimation settings. Fit A estimates an unstructured random-intercept-and-slope covariance and gives $\hat{\alpha}=0.42$ per biomarker unit. Fit B sets the intercept-slope covariance to zero and gives $\hat{\alpha}=0.27$ per biomarker unit. Suppose the biomarker scale is unchanged between fits.

For a one-unit higher latent current biomarker value, the conditional hazard ratios implied by the two fitted coefficients are, respectively, $\exp(0.42)=1.522$ and $\exp(0.27)=1.310$. Thus the point estimates differ by $0.42-0.27=0.15$ log-hazard units per biomarker unit, and the corresponding one-unit hazard-ratio estimates differ by about $1.522-1.310=0.212$.

These are model-based associations, not causal effects. Because the link, scale, and other stated specifications were held fixed, this paired comparison indicates that the estimated association is sensitive to the covariance restriction. It does not establish which covariance structure is correct; uncertainty and adequacy of each fit also need consideration.

Worked Example: A Covariance Assumption Changes Subject-Specific Shrinkage

A simple calculation illustrates how a random-slope variance can affect an estimated latent trajectory. Suppose the intercept deviation is known to be zero for this illustration, and a subject has one biomarker residual of $r=1$ unit above the fixed mean at $t=2$ years. Write $r=2b_{1i}+\epsilon$, with $\epsilon\sim N(0,1)$. Under a normal prior $b_{1i}\sim N(0,v)$, the posterior mean of the slope deviation is $v(2)/(1+4v)\,r$. This follows from the normal conditional-mean formula: the covariance of $b_{1i}$ with $r$ is $2v$, and the variance of $r$ is $1+4v$.

If the slope variance is $v=0.09$, the posterior mean is $(2\times0.09)/(1+4\times0.09)\times1=0.18/1.36=0.1324$ units per year. If the assumed variance is $v=0.04$, it is $0.08/1.16=0.0690$ units per year. The same observed residual is shrunk more strongly toward zero under the smaller between-subject slope variance.

At year 2, these posterior slope deviations imply fitted deviations of $2(0.1324)=0.2648$ and $2(0.0690)=0.1380$ biomarker units. Under a current-value link with $\alpha=0.40$ per unit, the hazard ratio associated with the difference between these two fitted values is $\exp\{0.40(0.2648-0.1380)\}=\exp(0.05072)=1.052$. This is an illustration of the trajectory-to-link mechanism, not a full joint-model posterior calculation: in a joint fit, the event history also informs the random effects.

Designing a Paired Sensitivity Analysis

A paired comparison is informative only if it isolates the covariance choice. Before fitting, state the common specifications and the one feature allowed to differ. For example, compare an unstructured $D$ with a diagonal $D$, keeping the fixed-effects design, time origin, residual model, hazard link and its units, baseline-hazard specification, covariates, analysis sample, and numerical integration settings the same. If a random-intercept-only model is also considered, identify it as a more restrictive change because it removes slope variation as well as intercept-slope covariance.

Specify the common model.
Write down the longitudinal mean structure and the precise hazard link, including the biomarker scale and time origin.

State the covariance contrast.
For example, permit $\tau_{01}$ in one fit and constrain it to zero in the other, while retaining both random-effect variances.

Fit both models and check their numerical behavior.
Confirm convergence, inspect covariance estimates and uncertainty, and check whether estimates are on or near parameter-space boundaries.

Compare trajectory and link summaries.
Report relevant trajectory variation over follow-up and the association estimates on the same scale, with uncertainty for each fit.

A model with $\tau_{01}=0$ is nested within an unstructured covariance model when the variances are otherwise unrestricted. However, a likelihood-ratio test alone does not answer whether the hazard association is robust to the covariance choice. The sensitivity question concerns how the relevant fitted quantities change across defensible specifications, not just whether a restriction can be rejected. If comparing likelihoods, account for the fact that a zero covariance is an interior restriction when the variances are positive; restrictions that set a variance to zero instead involve a boundary and require more care.

Diagnostics, Alternatives, and Limitations

Check whether the covariance estimates are supported by the visit schedule and observed trajectories. Useful warnings include a slope variance estimated near zero, an intercept-slope correlation near $1$ or $-1$, a non-positive-definite covariance estimate, large standard errors, or convergence warnings. Such results may indicate weak information for distinguishing covariance structures, especially when there are few repeated measurements per subject. Compare model-implied trajectory variability across the observed follow-up rather than interpreting $\tau_{01}$ in isolation; its contribution depends on time origin and units.

A diagonal random-intercept-and-slope model is a direct alternative to the unstructured model when the correlation is poorly estimated or independence is scientifically defensible. A random-intercept-only model is another alternative, but it removes slope heterogeneity and should not be presented as merely setting the correlation to zero. A more flexible covariance structure may be considered when supported by the data and design, but increased flexibility can make estimation unstable. These alternatives do not replace checking the common mean trajectory and hazard-link assumptions.

The paired approach has limits. A change in $\hat{\alpha}$ is not by itself proof that either covariance structure is wrong: both estimates also reflect finite-sample uncertainty and the joint model’s assumptions. Conversely, similar point estimates do not prove that the trajectory distributions are equivalent. Report the fitted covariance estimates, the association estimates with uncertainty, and the exact restrictions. Keep the link and biomarker scale fixed so that a covariance comparison is not confounded with a change in the meaning of the association coefficient.

Common Pitfall

A common mistake is to say, “The hazard-link coefficient changed, so the effect of the biomarker changed.” That treats the association estimate as though it were independent of the latent-trajectory model. In a joint model, changing the random-effects covariance changes the distribution and subject-specific estimation of the linked trajectory feature; the coefficient can shift even though the specified link is identical. The correction is to describe the result as covariance-structure sensitivity of the fitted association estimate, report the paired specifications and common scale, and avoid a causal interpretation.

Key takeaway: The random-effects covariance controls how subject-specific trajectories vary together over time. In a joint model, that distribution feeds into the event likelihood, so paired fits that differ only in covariance assumptions are needed to assess whether a hazard-link estimate is robust to this modeling choice.

Check Your Understanding

Answer each question, then compare with the solutions below.

In a random-intercept-and-slope model, what does setting $\tau_{01}=0$ assume, and what does it leave unrestricted?
Suppose $\tau_{00}=1$, $\tau_{01}=-0.10$, and $\tau_{11}=0.04$. Calculate the random-effects variance of the trajectory at $t=2$.
Why should a paired covariance-sensitivity analysis keep the association link and biomarker scale fixed?
A fitted random-slope variance is nearly zero. What does this suggest, and why is it not equivalent to a zero intercept-slope covariance?

Solutions. (1) It assumes the random intercept and random slope are uncorrelated, while allowing both variances to be positive and estimated. (2) Use $\tau_{00}+2t\tau_{01}+t^2\tau_{11}=1+2(2)(-0.10)+2^2(0.04)=1-0.40+0.16=0.76$ trajectory-units squared. (3) Otherwise a changed coefficient could reflect a changed link definition or scale rather than the covariance assumption, making the comparison ambiguous. (4) It suggests little estimated between-subject slope variation under that fit; setting $\tau_{01}=0$ instead removes only covariance and does not force the slope variance to zero.