Bayesian Covariance Priors For Multiple Markers

Why Covariance Priors Matter

In a multimarker joint model, each marker can have its own random intercepts and slopes, and those subject-specific effects may be correlated across markers. Measurement errors from different markers may also be correlated at the same visit. These covariance structures help determine how information is shared across markers: they influence estimated subject trajectories, uncertainty in those trajectories, and consequently the survival association estimates that depend on them.

A covariance prior is not just a technical device for making estimation possible. With many markers, the number of variances and correlations can grow quickly, while the data may provide limited information about some of them. Prior choices therefore affect regularization and can matter for inference, especially when markers are highly correlated or sparsely observed.

Separate the Covariance Structures

Let \(Y_i(t)\) be the vector of \(K\) marker measurements for subject \(i\) at time \(t\). A simple longitudinal model is

Here \(\mu_i(t)\) denotes the fixed-effects mean, \(b_i\) is a vector of subject-specific random effects, \(Z_i(t)\) maps those effects to the marker measurements, and \(\varepsilon_i(t)\) is the residual measurement error. In a joint model, the event-time component may condition on latent marker trajectories derived from this longitudinal model. The covariance of \(b_i\) and that of \(\varepsilon_i(t)\) are different parameters with different interpretations.

For example, if \(b_i\) stacks a random intercept and slope for each of several markers, its covariance matrix \(D\) describes between-subject variation and association in those latent effects. A residual covariance matrix \(R\) can instead describe contemporaneous measurement-error association across markers after conditioning on the modeled trajectories. If residual errors are assumed independent across markers, \(R\) is diagonal. A separate serial residual model may be needed if errors from different visits remain associated; a same-visit cross-marker matrix alone does not specify temporal dependence.

A Separated Prior for Scales and Correlations

A common construction separates the standard deviations from the correlation matrix. For a \(q\)-dimensional random-effect vector, write

where \(\tau_j>0\) is the standard deviation of random effect \(j\), and \(\Omega\) is a \(q\times q\) correlation matrix. Thus \(D_{jj}=\tau_j^2\) and \(D_{jk}=\tau_j\tau_k\Omega_{jk}\). The same separation can be used for a positive-definite residual covariance matrix, with residual standard deviations in \(S\) and residual correlations in \(\Omega\).

One possible scale prior is \(\tau_j\sim\text{half-}t_\nu(0,A_j)\): a Student \(t\) distribution with \(\nu\) degrees of freedom and scale \(A_j\), restricted to positive values. The scale \(A_j\) must be meaningful in the units of random effect \(j\); a random slope measured per year and a random intercept measured in marker units need not use the same scale. Other choices, such as half-normal priors, are also possible. The essential practice is to state the scale and check that it allows plausible variation without placing substantial prior mass on implausibly large values.

A standard correlation prior is the LKJ prior. For a positive-definite correlation matrix \(\Omega\), its density is proportional to \(|\Omega|^{\eta-1}\), where \(\eta>0\) is the shape parameter. At \(\eta=1\), the density is constant over the space of positive-definite correlation matrices. Values \(\eta>1\) increasingly favor matrices nearer the identity; values \(0<\eta<1\) put relatively more density near the boundary. This is a joint distribution on the matrix, not a collection of independently specified pairwise correlations: the entries must form a valid correlation matrix.

What the Correlation Prior Implies

The LKJ prior’s behavior can be seen explicitly for two markers. Their correlation matrix has the form

and determinant \(1-\rho^2\). Thus the marginal density for \(\rho\) under the two-dimensional LKJ prior is proportional to \((1-\rho^2)^{\eta-1}\). Normalizing it gives

The normalizing constant follows because the integral of \((1-\rho^2)^{\eta-1}\) from \(-1\) to \(1\) is \(B(\tfrac12,\eta)=\sqrt{\pi}\Gamma(\eta)/\Gamma(\eta+\tfrac12)\). This calculation also shows why a uniform prior on each correlation is not a general solution for a matrix with three or more variables: arbitrary pairwise values may fail to define a positive-semidefinite matrix.

How Covariance Choices Affect Survival Inference

Suppose a subject-specific log hazard contains a weighted combination of two latent marker values, \(g_i(t)=\alpha_1m_{i1}(t)+\alpha_2m_{i2}(t)\), where \(\alpha_1\) and \(\alpha_2\) are marker association coefficients. If uncertainty in those marker values is represented by variances \(v_1,v_2\) and covariance \(c_{12}\), then the variance of the weighted combination is

This identity follows by expanding the variance of a linear combination and using \(\operatorname{Cov}(m_{i1},m_{i2})=c_{12}\). It makes clear that cross-marker covariance can increase or decrease uncertainty in a hazard predictor, depending on the signs of the weights and covariance. The formula concerns uncertainty in the weighted latent values; it is not by itself a formula for the marginal hazard, which also involves the nonlinear exponential link and integration over latent quantities.

Choosing and Checking Covariance Priors

The separated scale-and-correlation construction is often useful because it makes the prior’s components interpretable and permits scale choices in the measurement units. It does not eliminate the need for careful modeling. In particular, the random-effect covariance and residual covariance should each be specified according to the scientific and measurement structure, rather than assigned the same matrix automatically.

An inverse-Wishart prior is a traditional prior for a whole covariance matrix. It can be convenient, but its degrees-of-freedom and scale parameters jointly affect variances and correlations. As a result, apparently simple choices can induce unexpected marginal priors on standard deviations or correlations. The separated construction makes these components more directly controllable, though it also requires specifying scale priors and a correlation prior. Neither construction repairs a likelihood that cannot distinguish random-effect variation from residual variation.

Before fitting, check the following:

• Units and scaling. The prior scales for standard deviations should match the units of the corresponding intercepts, slopes, or residuals. Centering time and scaling markers can make these choices easier to interpret.
• Matrix validity. The implied covariance and correlation matrices must be symmetric and positive semidefinite. A positive-definite parameterization ensures nonsingularity, but singular covariance matrices remain valid mathematical covariance matrices.
• Prior predictive behavior. Simulate standard deviations, correlation matrices, longitudinal trajectories, and implied survival predictors from the prior. Check whether these quantities are scientifically plausible before conditioning on data.
• Posterior behavior. Inspect trace plots, effective sample sizes, convergence diagnostics, and posterior summaries for standard deviations and important correlations. Strong concentration near a singular boundary or poor mixing can signal weak information or difficult geometry.
• Sensitivity. Refit with scientifically defensible alternatives for the scale priors or LKJ shape parameter. Material changes in association estimates or uncertainty indicate that covariance assumptions matter to the conclusions.

When markers are strongly collinear or some subjects have few measurements, their covariance components may be weakly identified. A posterior correlation near zero does not establish conditional independence, and a broad posterior should not be summarized as a precise absence of association. Likewise, a prior favoring the identity matrix can stabilize estimation but may understate real cross-marker dependence if chosen too strongly.

Common Pitfalls and Limits

• Calling every valid covariance matrix positive definite. The general validity condition is positive semidefiniteness. Positive definiteness is an additional nonsingularity condition that many prior parameterizations impose.
• Confusing random-effect and residual correlations. Correlated subject-specific trajectories do not imply correlated measurement errors, and vice versa. They represent different sources of dependence.
• Assigning independent priors to all pairwise correlations. In dimensions greater than two, separately plausible correlation values may not form a valid correlation matrix. Use a joint matrix prior or another construction that enforces matrix validity.
• Interpreting the prior as data evidence. A prior preference for modest correlations is a modeling choice, not an empirical finding. Report the prior and examine its influence on posterior inference.

Covariance priors cannot resolve fundamental lack of information. If measurements are sparse, markers nearly redundant, or the longitudinal design provides little slope information, posterior covariance estimates can remain prior-sensitive. Priors also do not substitute for appropriate residual structure, model checks, or a clear account of which latent marker features enter the survival model.