Unstructured Working Correlation In GEE

Why Use an Unstructured Working Correlation?

Repeated measurements from the same person or experimental unit are generally associated. In generalized estimating equations (GEE), the mean model describes the expected outcome, while a working correlation describes how outcomes within a cluster are related for the estimating procedure. An unstructured working correlation allows each distinct pair of measurement occasions to have its own correlation. It makes no assumption that measurements close in time are more alike, or that all pairs are equally correlated.

That flexibility has a cost: each pair-specific correlation must be estimated from the data. With many measurement occasions, there can be many association parameters, and some may be poorly supported, especially when clusters are few or observations are missing. The key is to understand both what the structure permits and what data it needs.

Definition: For clusters measured at the same $m$ occasions, an unstructured working correlation is a matrix $R(\alpha)$ with ones on its diagonal and a distinct working correlation $\alpha_{jk}$ for each pair of different occasions $j$ and $k$. The matrix is symmetric, so $\alpha_{jk}=\alpha_{kj}$. Apart from the requirements for a valid correlation matrix, the structure imposes no equality or time-lag pattern among the off-diagonal entries.

For cluster $i$, let $Y_i$ be its vector of repeated outcomes and $\mu_i$ the corresponding vector of marginal means. Let $A_i$ be the diagonal matrix of marginal variances specified by the chosen outcome distribution and mean model. The working covariance matrix used in GEE is commonly written

V_i=A_i^{1/2}R_i(\alpha)A_i^{1/2}.

Thus, the unstructured specification concerns the correlation component $R_i(\alpha)$, not the marginal means or the diagonal variances. The GEE framework uses this working covariance to account for within-cluster association when estimating mean-model parameters. It does not require introducing a separate correlation parameter for every individual cluster: when clusters share the same measurement occasions, the pair-specific parameters are ordinarily shared across clusters.

How Many Association Parameters?

There are $m(m-1)$ off-diagonal matrix cells, but symmetry means each pair is represented twice. The diagonal entries are all fixed at one and are not association parameters. Therefore, the number of distinct working correlation parameters is the number of ways to choose two occasions from $m$:

q=\binom{m}{2}=\frac{m(m-1)}{2}.

Justification: For the first occasion there are $m-1$ possible partners, for the second there are $m-2$ partners not already counted, and so on, down to one partner for the penultimate occasion. The total is $$ (m-1)+(m-2)+\cdots+1=\sum_{r=1}^{m-1}r=\frac{m(m-1)}{2}. $$ Equivalently, choose two different occasions from $m$, which gives $\binom{m}{2}$ unordered pairs.

Worked Example: Counting Parameters for Four Visits

Suppose hypothetical subjects are measured at four scheduled visits. An unstructured working correlation has

q=\frac{m(m-1)}{2} =\frac{4(4-1)}{2} =\frac{12}{2}=6

distinct association parameters. They correspond to the six pairs: visits 1–2, 1–3, 1–4, 2–3, 2–4, and 3–4. The correlation for visits 1 and 4 need not equal the correlation for visits 1 and 2, even though the former visits are farther apart. The structure is flexible, but it does not by itself explain why any estimated pair has its particular value.

Reading and Checking an Unstructured Matrix

For three occasions, a typical unstructured working correlation matrix is

R(\alpha)= \begin{pmatrix} 1 & \alpha_{12} & \alpha_{13}\\ \alpha_{12} & 1 & \alpha_{23}\\ \alpha_{13} & \alpha_{23} & 1 \end{pmatrix}.

Each off-diagonal entry is a working correlation for a particular pair of occasions. The entries must be between $-1$ and $1$, but that condition alone is not enough. All of the entries together must form a positive-definite correlation matrix for the usual inverse-based GEE calculations. For a three-occasion matrix, positive definiteness requires positive leading principal minors; in particular, its determinant must be positive. For larger matrices, all eigenvalues must be positive (equivalently, all leading principal minors must be positive for a symmetric matrix).

Worked Example: Checking a Three-Visit Matrix

Suppose an illustrative analysis produces the following hypothetical working correlations: 0.30 for visits 1 and 2, 0.10 for visits 1 and 3, and 0.40 for visits 2 and 3. The proposed matrix is

R= \begin{pmatrix} 1 & 0.30 & 0.10\\ 0.30 & 1 & 0.40\\ 0.10 & 0.40 & 1 \end{pmatrix}.

Its leading principal minors are $1$, $1-(0.30)^2=0.91$, and the determinant

\det(R) =1+2(0.30)(0.10)(0.40)-(0.30)^2-(0.10)^2-(0.40)^2 =1+0.024-0.09-0.01-0.16 =0.764.

All three are positive, so this matrix is positive definite. The values are therefore jointly compatible with a valid correlation matrix. Checking only that each entry lies between $-1$ and $1$ would not establish this compatibility.

In practice, the working correlations are estimated from within-cluster information, often using standardized residuals from the current mean model. This makes the support for each pair important. If only a small number of clusters have observed outcomes at both visits in a particular pair, that pair-specific correlation can be unstable. If the observed pattern is too sparse, an estimate may also fail to produce a positive-definite matrix; an algorithm may then need a constrained estimate or another numerical remedy.

When Flexibility Becomes Impractical

The quadratic parameter count is the central limitation. With six occasions, for instance, an unstructured structure requires $6(5)/2=15$ correlations; with ten occasions it requires $10(9)/2=45$. The data do not contribute equally to all of these estimates. Each pair is informed by clusters in which both relevant occasions are observed. Missingness or irregular schedules can leave some pairs with much less information than others.

Worked Example: Many Visits and Sparse Pair Information

Consider a hypothetical study with 24 independent clusters and six scheduled occasions. The unstructured structure requires

q=\frac{6(6-1)}{2}=15

association parameters. If all 24 clusters have complete measurements, each of the 15 occasion pairs has observations from 24 clusters. Now suppose that missing measurements leave only three clusters with both occasions 2 and 6 observed. The correlation for that pair is then informed by just three within-cluster pairs, not by all 24 clusters. By contrast, another occasion pair could still be observed together in all 24 clusters.

This is a warning about the information available, not a formula that guarantees a particular amount of precision. The count of 15 parameters alone does not prove that the analysis will fail, but a small number of clusters combined with uneven pair availability makes unstable or infeasible correlation estimates more plausible. It is important to inspect cluster counts for occasion pairs and to check that the fitted working matrix is valid.

An unstructured working correlation is most straightforward when clusters have a common set of occasions and enough clusters contribute information to each pair. With unequal cluster sizes or missing visits, the intended occasion labels and the available pairwise observations need to be clear. A common approach is to define the structure for the scheduled occasions and use the corresponding submatrix for each cluster’s observed outcomes. This does not create information for a pair that is rarely observed.

The structure is also specific to the occasions represented in the data. An unstructured estimate for a pair of scheduled visits does not provide a correlation rule for an unobserved visit or a new time gap. Unlike a model based on temporal distance, it offers no built-in pattern for extending dependence beyond the observed occasion pairs.

How It Differs From More Restrictive Structures

Independence, exchangeable, and AR(1) working correlations use more restrictive patterns. They trade pair-specific flexibility for fewer association parameters. The table summarizes the distinction; the parameter counts refer to clusters measured at $m$ occasions.

Working structure	Pattern among distinct occasions	Association parameters	Main trade-off
Independence	All off-diagonal correlations are zero	0	Simple, but does not represent within-cluster correlation in the working matrix
Exchangeable	All off-diagonal correlations are equal	1	Parsimonious, but cannot represent pair-specific dependence
AR(1)	Correlation follows a common lag-based pattern	1	Parsimonious for ordered, regularly spaced occasions, but imposes a pattern with lag
Unstructured	Each occasion pair has its own correlation	$m(m-1)/2$	Flexible, but requires substantially more information and a valid joint matrix

These structures are not claims that the real dependence must exactly follow their patterns. They are alternative working descriptions. An unstructured specification can be useful when the number of occasions is modest and the data can support pair-specific estimates. Its flexibility is not automatically an advantage when estimates are unstable or the matrix is difficult to estimate.

Common Mistakes and Diagnostics

Counting both halves of the matrix. There are $m(m-1)$ off-diagonal cells, but symmetry makes the number of distinct parameters $m(m-1)/2$.
Assuming every pair has equal correlation. That is the exchangeable pattern, not the unstructured one. Unstructured correlations may differ across every pair.
Assuming valid individual entries guarantee a valid matrix. Correlations must also be jointly compatible, so check positive definiteness.
Treating the pair count as the whole feasibility question. Check how many clusters actually contribute to each pair, especially with missing visits or unequal cluster sizes.

A practical warning sign is a fitted matrix that is not positive definite or is numerically close to singular. Such a matrix may be difficult to invert reliably. Very different pair estimates based on very different numbers of clusters are another reason for caution. These diagnostics do not establish that one alternative is always preferable; they show when the data may not support the unstructured detail.

Key takeaway: An unstructured working correlation assigns one parameter to every distinct pair of occasions, giving $m(m-1)/2$ association parameters. Use it only when the repeated-measure design and available cluster information can support those pair-specific estimates and a valid correlation matrix.

Check Your Understanding

Answer each question, then compare with the solutions below.

How many distinct working correlation parameters are needed for five occasions?
For three occasions, explain why three off-diagonal correlations do not mean there are six distinct association parameters.
A pairwise working correlation estimate is 0.20, and another is 0.50. Does that alone establish that the full proposed correlation matrix is valid? Explain.
A study has eight scheduled visits but only a few clusters with both visits 1 and 8 observed. What is the concern for an unstructured working correlation?
What is the main difference between an unstructured and an exchangeable working correlation?

Solutions. (1) $5(5-1)/2=10$ parameters. (2) The matrix is symmetric, so the entry for visits 1 and 2 is the same parameter as the entry for visits 2 and 1; only the three unordered pairs are distinct. (3) No. The correlations must be jointly compatible, and the complete matrix must be positive definite. (4) The visit 1–8 correlation is informed by only those few clusters, so its estimate may be unstable even though the design has eight occasions; the structure requires $8(7)/2=28$ pair parameters overall. (5) Unstructured allows each pair to have its own correlation, whereas exchangeable constrains all distinct pairs to share one correlation.