Block Covariance Structures for Marker Random Effects

Why the Random-Effects Blocks Matter

For multiple longitudinal markers, each individual can have a marker-specific trajectory that differs from the population mean. A random intercept represents an individual’s deviation in marker level, and a random slope represents an individual’s deviation in rate of change. When several markers are measured on the same individual, these deviations may be related: an individual whose first marker is unusually high may also tend to have an unusually high second marker, or the individual’s slope for one marker may be associated with the other marker’s intercept.

A block covariance matrix expresses these relationships directly. Its diagonal blocks describe variation and intercept–slope association within each marker. Its off-diagonal blocks describe association between random effects for different markers. This is a subject-level covariance structure: it is distinct from covariance among measurement errors recorded at the same visit.

For two markers, using the ordering \((b_{i10},b_{i11},b_{i20},b_{i21})\), the matrix is

Each diagonal block contains the variance of the marker’s random intercept, the variance of its random slope, and their covariance. The off-diagonal block \(D_{12}\) contains four covariances: marker 1 intercept with marker 2 intercept, marker 1 intercept with marker 2 slope, marker 1 slope with marker 2 intercept, and marker 1 slope with marker 2 slope. These quantities need not be equal to their transposed counterparts in the same block, because \(D_{12}\) is not generally symmetric.

From Random Effects to Covariance Between Measurements

Write the random-effects part of marker \(k\)’s trajectory at time \(t\) as \(z(t)^\mathsf{T}b_{ik}\), where \(z(t)=(1,t)^\mathsf{T}\). A typical longitudinal model also includes a population mean and a measurement error:

Here \(x_{ik}(t)^\mathsf{T}\beta_k\) describes the mean trajectory. The covariance calculation below concerns the random-effects contribution only. For markers \(k\) and \(\ell\), at times \(t\) and \(s\), its covariance is

This follows because the fixed mean terms are nonrandom and the covariance matrix of \((b_{ik}^\mathsf{T},b_{i\ell}^\mathsf{T})^\mathsf{T}\) contains \(D_{k\ell}\) as its cross block. For two random intercepts and slopes, write \(D_{k\ell}=\begin{pmatrix}d_{00}&d_{01}\\d_{10}&d_{11}\end{pmatrix}\). Expanding the product gives

The time arguments matter. For example, a cross-marker covariance can be positive at baseline but smaller or negative later if the intercept–slope or slope–slope terms offset the intercept covariance. The units also matter: an intercept variance has units of marker squared, while a slope variance has units of marker squared per time squared. Centering time changes the interpretation of the intercept and changes the numerical covariance entries, although it does not change the underlying fitted trajectories when the model is reparameterized consistently.

Validity and Common Covariance Structures

A covariance matrix must be symmetric and positive semidefinite: for every vector \(a\), \(a^\mathsf{T}Da\geq 0\). This condition ensures that no linear combination of the random effects has negative variance. It also constrains the off-diagonal blocks; they cannot be chosen independently of the diagonal blocks without regard to validity. In practice, estimation procedures parameterize or constrain \(D\) so the fitted matrix satisfies this requirement.

An unstructured covariance matrix estimates all distinct entries subject to symmetry and positive semidefiniteness. A block-diagonal structure instead sets every cross-marker block to zero, asserting that the markers’ random effects are uncorrelated within individuals. A block-diagonal assumption is a modeling restriction, not a consequence of using marker-specific random effects.

Intermediate structures are possible. For example, a model may retain cross-marker intercept covariances while constraining some cross-marker slope covariances, or impose a parsimonious parameterization across markers. Such restrictions should be justified by the scientific question and the available data, rather than selected solely because they produce a convenient fit.

Assumptions, Diagnostics, and Limitations

A standard Gaussian random-effects specification assumes that \(b_i\) has a mean-zero multivariate normal distribution with covariance \(D\), and that different individuals’ random-effects vectors are independent, conditional on modeled covariates. The covariance calculations themselves require finite second moments and the stated covariance structure; normality is an additional distributional assumption often used for likelihood-based fitting. The choice of random-effects design, such as an intercept and a linear slope for each marker, is also part of the model.

A useful check is to inspect the estimated diagonal variances, correlations, and eigenvalues of \(D\). A variance estimated at or near zero, a correlation near \(1\) or \(-1\), or a nearly singular matrix can indicate that the data do not support the requested flexibility. Such results may also arise from weakly informative time patterns, too few individuals, limited overlap in measurement schedules, or a poorly scaled time variable. A convergence message alone does not establish that each covariance parameter is well identified.

When diagnosing a fit, compare estimates under scientifically defensible structures and check whether conclusions are sensitive to those restrictions. Examine standard errors and uncertainty intervals for covariance parameters, not only point estimates. If an unstructured matrix is unstable, a more parsimonious structure may be appropriate, but it changes the assumptions and may affect uncertainty in trajectory estimates. Conversely, setting cross-marker blocks to zero simply to avoid a difficult fit can conceal meaningful within-individual dependence.

The random-effects covariance is not a substitute for a suitable model of measurement error. It captures dependence induced by shared subject-level trajectories; residual covariance describes additional dependence among observations after those trajectories are accounted for. These are distinct sources of covariance. The present structure also does not, by itself, specify how marker histories relate to a time-to-event outcome; that requires a separate survival-model component.

Common Mistake

A frequent error is to read \(D_{12}\) as a correlation matrix and interpret \(0.12\) as a correlation of 0.12. It is a covariance, with units equal to the product of the associated random effects’ units. To obtain a correlation, divide that covariance by the product of the two standard deviations. Another error is to assume that a positive intercept covariance guarantees positive covariance between the markers at every time. The time-dependent expression \(d_{00}+d_{01}s+d_{10}t+d_{11}ts\) shows why that need not hold.