Cholesky Parameterization Of Random-Effect Covariance in Population PK

Why Parameterize a Random-Effect Covariance Matrix?

A population PK model may assign a vector of subject-level random effects to each individual. For example, effects on log clearance and log volume can be represented by \(\eta_i=(\eta_{CL,i},\eta_{V,i})^T\), with mean zero and covariance matrix \(\Omega\). The diagonal entries of \(\Omega\) are the random-effect variances; its off-diagonal entries are covariances. The earlier material on interpreting the clearance–volume random-effect correlation explains how covariance and correlation describe the association between two effects.

During model estimation, the elements of \(\Omega\) must vary while still defining a valid covariance matrix. A Cholesky parameterization expresses \(\Omega\) as the product of a lower-triangular matrix and its transpose. The factor's entries are not, in general, themselves the variances or covariances. Instead, multiplying the factor out shows exactly how they combine to produce those quantities.

The distinction in the definition matters. A covariance matrix must be positive semidefinite, but it need not be positive definite: a zero variance, for example, is possible. The standard parameterization with strictly positive Cholesky diagonal elements specifically represents positive-definite matrices. Many estimation implementations use that parameterization because it provides a convenient way to maintain a positive-definite \(\Omega\) while estimating its elements. A singular covariance matrix is a limiting or separately handled case, not a matrix represented by a strictly positive-diagonal factor.

How the Factor Maps to Variances and Covariances

For two random effects, write the lower-triangular factor as \(L\). Multiplication gives the covariance matrix directly:

Thus \(\Omega_{11}=l_{11}^2\) and \(\Omega_{22}=l_{21}^2+l_{22}^2\) are the two variances, while \(\Omega_{12}=\Omega_{21}=l_{11}l_{21}\) is the covariance. The second variance depends on two factor elements; it is not just \(l_{22}^2\). This is a frequent source of misreading.

For \(q\) random effects, let \(L=(l_{ij})\) be a \(q\times q\) lower-triangular matrix, so \(l_{ij}=0\) when \(j>i\). Entry-by-entry multiplication gives

For a diagonal entry, this becomes \(\Omega_{ii}=\sum_{k=1}^{i}l_{ik}^2\). For an off-diagonal entry, it is a sum of products of the factor elements in the two relevant rows. Therefore, changing one factor element can affect more than one entry of \(\Omega\).

Recovering the Factor From a Covariance Matrix

For a two-effect positive-definite covariance matrix, the same equations can be solved in the other direction. Write its entries as \(\Omega_{11}\), \(\Omega_{12}\), and \(\Omega_{22}\). From the first variance and covariance equations, \(l_{11}=\sqrt{\Omega_{11}}\) and \(l_{21}=\Omega_{12}/l_{11}\). The remaining diagonal element follows from the second variance equation:

The positive square roots are used for the positive-diagonal factor. The quantity under the final square root must be positive for this two-dimensional positive-definite case. This calculation follows by matching the entries in \(LL^T\); it is not an additional model assumption.

More Than Two Random Effects

With three or more effects, each covariance is still a row-wise sum of products. The triangular structure makes the calculation systematic: the first row contributes to the first variance, the next row adds its own diagonal-square term to its variance, and subsequent rows follow the same pattern. The following example shows both diagonal and off-diagonal entries, including a negative covariance.

Using the Factor During Estimation

A covariance matrix for \(q\) random effects has \(q(q+1)/2\) distinct entries: \(q\) variances and \(q(q-1)/2\) covariances. A lower-triangular factor has the same number of free elements. A common estimation strategy is to estimate the entries of \(L\) and form \(\Omega=LL^T\) at each parameter evaluation. If the diagonal elements are constrained to be strictly positive, each resulting \(\Omega\) is positive definite. Some implementations achieve this by estimating an unconstrained parameter \(a_i\) and setting \(l_{ii}=\exp(a_i)\); this is a computational choice, not part of the definition of the Cholesky factorization.

The factor entries are computational parameters, whereas model interpretation generally uses the variances, standard deviations, covariances, and correlations derived from \(\Omega\). For a pair of effects \(i\) and \(j\), the correlation is \(\Omega_{ij}/\sqrt{\Omega_{ii}\Omega_{jj}}\). It should be calculated from the reconstructed covariance matrix, not inferred from a single off-diagonal element of \(L\).

The Cholesky approach is one way to estimate a full covariance structure. A diagonal structure instead sets all off-diagonal covariances to zero, and a block-diagonal structure allows covariance within specified groups but fixes covariances between groups at zero. Those choices reduce the number of covariance parameters and may be preferable when the data cannot support a full structure. Directly estimating covariance-matrix entries is another possible approach, but it requires constraints to ensure that the result remains a valid covariance matrix. A Cholesky factor encodes that constraint through its construction when its diagonal entries are strictly positive.

Diagnostics, Limitations, and Common Mistakes

A valid factorization does not guarantee that the random-effect structure is well supported or that the model has estimated it reliably. Check that the factor is lower triangular, that the diagonal convention used by the estimation method is clear, and that multiplying the reported factor by its transpose reproduces the reported \(\Omega\). If reported covariance or correlation summaries disagree with that reconstruction, first check the ordering of the random effects and the software's reporting convention.

Very small diagonal factor estimates, large uncertainty, or unstable changes across model runs can indicate that the data provide weak information about some random-effect directions. A factor may still produce a mathematically valid covariance matrix in that situation, but the estimates can be imprecise and sensitive to the model or data. Examine standard errors, parameter correlations, convergence information, and whether simplifying to a diagonal or block-diagonal structure is scientifically defensible. The factorization itself cannot resolve weak identifiability, sparse observations, or misspecification of the structural PK model.

• Mistake: treating factor elements as covariance entries. In two dimensions, \(l_{21}\) is not the covariance; the covariance is \(l_{11}l_{21}\). Reconstruct \(\Omega\) before interpreting its entries.
• Mistake: reading the second variance as \(l_{22}^2\). The correct expression is \(l_{21}^2+l_{22}^2\), as follows directly from multiplying the second row of \(L\) by itself.
• Mistake: assuming every covariance matrix has a strictly positive-diagonal factor. That statement holds for positive-definite \(\Omega\). A singular positive-semidefinite covariance matrix may require a factor with a zero diagonal element.
• Mistake: believing a valid factor proves the covariance structure is appropriate. Cholesky parameterization helps maintain a valid matrix; it does not establish that a full covariance structure is estimable or scientifically warranted.

The main practical limitation is that the elements of \(L\) can be less directly interpretable than those of \(\Omega\). In addition, changing the ordering of random effects changes the displayed factor, even though the covariance structure can be expressed in the reordered coordinates. Report the ordering and interpret the reconstructed covariance matrix in the model's stated effect order.