Covariance Structures for Repeated Pseudobulk Measures

Why the Within-Donor Covariance Matters

A donor may contribute pseudobulk samples at several conditions or collection occasions. Those measurements are not independent simply because each is a separate sample: observations from the same donor can share biological features, and the amount of dependence can differ across pairs of occasions. A covariance structure describes that dependence. It affects how repeated observations contribute to estimated condition means and, in turn, the standard errors used for inference.

The focus here is the covariance among repeated measures, not the mean trend across occasions. The earlier material on random intercepts for donors gives a familiar starting point: a donor-level random effect induces positive within-donor correlation. The key question is what happens when the equal-correlation assumption implied by that model is too restrictive.

For a Gaussian model, one can write \(\mathbf{Y}_i=\mathbf{X}_i\boldsymbol{\beta}+\boldsymbol{\varepsilon}_i\), where the fixed-effects design matrix \(\mathbf{X}_i\) specifies the mean model and \(\operatorname{Cov}(\boldsymbol{\varepsilon}_i)=\mathbf{V}_i\). Donors are typically assumed independent of one another, so the full covariance matrix is block diagonal, with one block per donor. A covariance structure specifies the form of each block.

Pseudobulk counts are often modeled with count distributions rather than a Gaussian response. Covariance statements must therefore be tied to the model scale: the following Gaussian calculations apply directly to a modeled continuous response or to an appropriate working or latent scale, not automatically to raw counts. In a count model, the likelihood and covariance assumptions must be compatible with that model.

Compound Symmetry and Its Random-Intercept Connection

Compound symmetry (CS) assumes that every observation has the same marginal variance and every distinct pair of observations from one donor has the same covariance. If the marginal variance is \(v\) and the common covariance is \(c\), the \(m\)-occasion covariance matrix is:

When \(v>0\), the common pairwise correlation is \(\rho=c/v\). For this matrix to be a valid covariance matrix, its eigenvalues must be nonnegative: \(v-c\), with multiplicity \(m-1\), and \(v+(m-1)c\), with multiplicity one. Thus the CS parameters must satisfy \(v-c\geq 0\) and \(v+(m-1)c\geq 0\). The common special case \(c\geq0\) and \(c\leq v\) satisfies these conditions.

A random-intercept model produces CS under specific assumptions. Suppose \(Y_{ij}=\mathbf{x}_{ij}^\mathsf{T}\boldsymbol{\beta}+b_i+e_{ij}\), where \(b_i\) has mean zero and variance \(\tau^2\), the errors \(e_{ij}\) have mean zero and variance \(\sigma^2\), and the donor effects and errors are mutually independent. Assume also that errors are independent across occasions and have equal variance. Then each response has variance \(\tau^2+\sigma^2\), while two different responses from the same donor share covariance \(\tau^2\). Equal residual variances and independent residuals are essential: with unequal variances or residual correlation, this simple random-intercept construction does not imply CS.

More Flexible Structures

A more flexible choice is an unstructured covariance matrix. It estimates a separate variance for each occasion and a separate covariance for each pair. For \(m\) occasions, that is \(m(m+1)/2\) distinct parameters: \(m\) variances and \(m(m-1)/2\) covariances. The estimates must still form a symmetric positive semidefinite matrix. This flexibility can represent unequal variances and pair-specific dependence, but costs many parameters and can be difficult to estimate with few independent donors.

Other structures impose different, more economical patterns. For example, an autoregressive structure of order one (AR(1)) for equally spaced occasions assigns a common marginal variance and makes correlation decrease geometrically with the separation between occasions: the correlation at lag \(k\) is \(\rho^k\), with \(|\rho|<1\). It can represent greater dependence between nearby measurements without estimating every pairwise covariance. Independent residuals are the limiting simple alternative with zero off-diagonal covariance. These are substantive assumptions; the most complex model is not automatically the best choice.

One can compare structures through their implied covariance patterns, whether the design supports estimation of their parameters, and whether fitted models show plausible residual dependence. Likelihood-based comparisons require models fitted to the same response and mean structure, and their inferential conditions must be met. A numerical fit criterion alone cannot establish that a covariance pattern is biologically appropriate.

How Covariance Assumptions Change Standard Errors

For a linear model with known covariance matrix \(\mathbf{V}\), generalized least squares estimates the fixed effects using \(\widehat{\boldsymbol{\beta}}=(\mathbf{X}^\mathsf{T}\mathbf{V}^{-1}\mathbf{X})^{-1}\mathbf{X}^\mathsf{T}\mathbf{V}^{-1}\mathbf{Y}\). Under the model, its covariance matrix is \((\mathbf{X}^\mathsf{T}\mathbf{V}^{-1}\mathbf{X})^{-1}\); a standard error is the square root of the corresponding diagonal entry. In practice, covariance parameters are estimated, and the reported model-based standard errors generally treat the fitted structure as correct, subject to the method’s estimation and inference approximations.

A useful derivation shows the effect of CS on a balanced average. For one donor, let \(\bar{Y}_i=m^{-1}\sum_{j=1}^{m}Y_{ij}\). The variance of a sum is the sum of all variances and covariances, so:

If \(n\) independent donors each contribute all \(m\) measurements under the same structure, the variance of the average of their donor averages is \(v\{1+(m-1)\rho\}/(mn)\). Treating all \(mn\) observations as independent would instead use \(v/(mn)\), which is too small when \(\rho>0\). This derivation assumes a balanced design, equal variances, common correlation, and independent donors. With unequal variances or pair-specific covariances, the relevant variance is computed from the full matrix and the particular contrast weights.

Diagnostics, Limitations, and Common Pitfalls

Covariance structure is rarely known with certainty. Useful checks include inspecting within-donor residual patterns by occasion and separation, comparing observed pairwise residual associations with those implied by the fitted model, and checking whether estimated variances or correlations are near parameter boundaries. Such checks can be noisy when donor numbers are small. A complex unstructured matrix may fail to converge or yield unstable estimates because there is insufficient independent donor information to estimate all its parameters. Conversely, a converged CS model can still be inadequate if residual variances differ by occasion or correlations clearly vary across pairs.

A covariance matrix must be symmetric and positive semidefinite, but it need not be positive definite. For instance, if two components of a random vector are exactly the same random variable, their covariance matrix is singular and positive semidefinite. A fitting method that requires an inverse may need positive definiteness; that computational requirement should not be mistaken for the definition of a covariance matrix.

Common mistake: “There are many repeated samples, so treating them as independent gives a more precise estimate; each sample adds information.” The correction is that observations from the same donor share dependence, so the information gained from another measurement depends on its covariance with existing measurements. Positive within-donor correlation often makes the independent-observation standard error too small. A more flexible covariance model can change that standard error, but does not necessarily increase it for every coefficient or contrast: the direction and size of the change depend on the covariance pattern and the design.

Covariance modeling also has practical limits. Estimating many parameters consumes information, especially when the number of independent donors is modest. If a structure is misspecified, model-based standard errors can be inaccurate even when the mean model is reasonable. A robust variance approach can be an alternative in suitable settings, but it has its own sample-size and design requirements. No covariance choice compensates for a design with too few independent donors or for a poorly specified mean model.