Why the Working Correlation Is a Design Choice
In GEE, the marginal mean model describes how the expected outcome changes with covariates, while the working correlation represents the assumed pattern of association among outcomes from the same cluster. Choosing a working correlation is therefore not the same as choosing the scientific mean model. The choice matters because it affects how within-cluster observations are weighted when estimating the mean-model coefficients. A suitable structure can make estimation more efficient; a poor one can lose efficiency or create practical estimation problems.
The choice should begin with the study design: how many measurements are taken in each cluster, whether their timing is regular, and what dependence pattern is plausible. It should not be made by selecting whichever structure produces the most appealing coefficient or smallest standard error. The working correlation is a device for organizing dependence in the estimating equations, not a claim that the selected pattern is the true data-generating mechanism.
A standard GEE uses the estimating equation $$ \sum_{i=1}^{K}D_i^{T}V_i^{-1}(Y_i-\mu_i)=0, $$ where \(K\) is the number of independent clusters and \(D_i\) is the derivative of \(\mu_i\) with respect to the mean-model coefficients. Thus \(V_i^{-1}\) determines how residual information from each cluster contributes to the coefficient estimates. Changing \(R_i(\alpha)\) changes those weights, even though it does not change the definition of the marginal mean model.
A Practical Selection Framework
Four familiar choices cover many repeated-measures applications. Their suitability depends on the measurement schedule and the expected shape of dependence, not simply on which one is most flexible.
| Working structure | Typical pattern represented | Useful when | Main cost or caution |
|---|---|---|---|
| Independence | All within-cluster correlations are set to zero for the working calculation. | A simple baseline is needed, or a dependable nonzero pattern is difficult to specify. | It ignores dependence in the weighting, even when repeated outcomes are associated. |
| Exchangeable | Every pair within a cluster has the same working correlation. | Measurements have no meaningful ordering, or a roughly common association is plausible. | It does not represent association that declines with time or lag. |
| AR(1) | For equally spaced ordered occasions, correlation at lag \(h\) is \(\rho^h\). | Measurements are regularly spaced and association plausibly decreases with lag. | Ordinary discrete AR(1) depends on occasion lag, not actual elapsed time when visits are irregular. |
| Unstructured | Each distinct pair of occasions has its own working correlation. | There are a small, fixed number of occasions and enough information to estimate the pairwise correlations. | The number of correlation parameters grows quickly with the number of occasions. |
A useful first question is whether order and elapsed time should matter. If all pairs are expected to be similarly associated, exchangeable may be a reasonable compact representation. If association is expected to weaken as the gap between observations grows, an AR(1) form is often suitable for regular schedules. For irregular visit times, a continuous-time decay structure may be more appropriate if it is available; using discrete AR(1) would treat visits one occasion apart as equally separated even when their actual time gaps differ.
Next consider the cluster size and number of occasions. An unstructured correlation for \(m\) occasions has \(m(m-1)/2\) distinct off-diagonal correlations. That flexibility can be useful when \(m\) is modest, but it demands substantially more information than a one-parameter structure. Large or variable cluster sizes can also make complex structures difficult to estimate. Independence, exchangeable, and AR(1) impose stronger restrictions and use fewer correlation parameters. Parsimony is valuable when justified, but a simple structure should not be chosen merely because it is easy to fit if it badly contradicts the design or plausible dependence.
Why a Working Choice Can Still Be Useful
The key justification is that GEE targets the marginal mean through residuals whose expectation is zero when the mean model is correct. If clusters are independent and \(E(Y_i-\mu_i\mid X_i)=0\), then, for a fixed valid working covariance matrix, $$ E\left\{D_i^{T}V_i^{-1}(Y_i-\mu_i)\mid X_i\right\} =D_i^{T}V_i^{-1}E(Y_i-\mu_i\mid X_i)=0. $$ Thus the estimating equation has expectation zero even if \(V_i\) does not equal the actual covariance matrix, provided the mean model is correct and the matrices used are well-defined. This is why a working correlation can be misspecified without automatically changing the marginal estimand.
This argument is an expectation calculation, not a guarantee that every working structure performs equally well in a finite sample. A better approximation to the actual dependence can improve efficiency, while a poor approximation can make estimates less efficient. Estimating the correlation structure can also be unstable if it has too many parameters for the available clusters. GEE inference must use variance calculations appropriate to clustered data; the working correlation alone does not justify treating all observations as independent. The detailed variance estimators are a separate topic.
Worked Example: Regular Follow-Up With Declining Dependence
Suppose a hypothetical study measures a continuous outcome at four equally spaced monthly visits for each participant. Before analysis, the investigators consider it plausible that adjacent visits are more strongly associated than visits farther apart. They use \(\rho=0.60\) as an illustrative working value for a discrete AR(1) structure.
For AR(1), the working correlation between occasions \(j\) and \(k\) is \(\rho^{|j-k|}\). The lag-one correlation is \(0.60^1=0.60\); the lag-two correlation is \(0.60^2=0.36\); and the lag-three correlation is \(0.60^3=0.216\). The implied correlation matrix is $$ R_i= \begin{pmatrix} 1&0.60&0.36&0.216\\ 0.60&1&0.60&0.36\\ 0.36&0.60&1&0.60\\ 0.216&0.36&0.60&1 \end{pmatrix}. $$
This pattern matches the stated design: visits are equally spaced, and the anticipated association declines with the number of monthly intervals. Exchangeable correlation would instead assign \(0.60\) to every off-diagonal entry, including the first-to-fourth visit pair, which does not represent the anticipated decline. Unstructured correlation could represent different pairwise values, but four occasions require \(4(4-1)/2=6\) correlation parameters rather than the single AR(1) parameter. If the data contain many independent participants, unstructured may be considered as a flexible alternative; if information is limited, AR(1) is a more economical choice consistent with the design. These calculations motivate a candidate structure; they do not establish that the true correlations equal the illustrative values.
Worked Example: Irregular Visit Times
Consider a hypothetical cohort with repeated measurements at 0, 1, 4, and 10 months. Suppose subject-matter knowledge suggests that association decays with actual elapsed time. An illustrative continuous-time exponential working correlation is $$ \operatorname{Corr}(Y_{ij},Y_{ik})=\exp\{-\lambda|t_{ij}-t_{ik}|\}, $$ where \(\lambda>0\) is a decay rate per month. Taking \(\lambda=0.25\) per month, a one-month gap gives \(\exp(-0.25)=0.7788\), a three-month gap gives \(\exp(-0.75)=0.4724\), and a six-month gap gives \(\exp(-1.50)=0.2231\).
The first-to-second visit gap is one month, so its working correlation is approximately \(0.779\). The second-to-third gap is three months, so its correlation is approximately \(0.472\). The third-to-fourth gap is six months, so its correlation is approximately \(0.223\). These differ because the actual time gaps differ.
Discrete AR(1) would use occasion lags instead: every adjacent pair would have correlation \(\rho\), regardless of whether its separation were one, three, or six months. That is a questionable representation if elapsed time is the scientifically relevant basis for dependence. If the available GEE implementation supports a continuous-time structure and the decay assumption is plausible, it is a natural candidate. If not, investigators should state the approximation used and assess its adequacy rather than describe irregularly spaced visits as equally spaced.
Using Data and Diagnostics Without Overinterpreting Them
Design knowledge should narrow the candidates; the observed data can then reveal obvious incompatibilities. One descriptive check is to examine residual association by occasion pair or by time lag after fitting the intended marginal mean model. Pearson residuals are often used for this purpose: for outcome \(Y_{ij}\), a Pearson residual is the observed outcome minus its fitted mean, divided by the fitted marginal standard deviation. A pattern of larger residual association for nearby times and smaller association for distant times is qualitatively consistent with a decay structure. Similar associations across pairs are more consistent with exchangeability.
Such displays are noisy, especially when the number of independent clusters is small. They should be read as diagnostics, not as proof of a particular population correlation. A pattern driven by a few clusters, outliers, an inadequate mean model, or missing visits can mislead. If correlation estimates are near their allowable boundaries, the estimated matrix is not positive definite, or fitting produces warnings, a complex structure may be unsupported. Simplifying the structure can be sensible, but the resulting choice should still be scientifically defensible.
Formal comparison criteria, including QIC, can help compare candidate GEE fits when used appropriately. They are aids rather than automatic selectors: small differences need not imply meaningful superiority, and a criterion does not override the visit schedule or the need for stable estimation. Comparisons are most interpretable when the candidate fits use the same outcome data and the same marginal mean model. If the mean model differs, a change in fit criterion cannot be attributed solely to the working correlation.
Worked Example: Choosing a Parsimonious Structure for Short Clusters
Suppose a hypothetical cluster-randomized study measures an outcome at three scheduled occasions for each of many clinics. The occasions are equally spaced, but investigators expect each pair within a clinic to have a broadly similar association because the outcome reflects a stable clinic-level environment. An exchangeable structure is a reasonable primary candidate. For an illustrative common correlation of \(0.35\), its matrix is $$ R_i= \begin{pmatrix} 1&0.35&0.35\\ 0.35&1&0.35\\ 0.35&0.35&1 \end{pmatrix}. $$
With three occasions, an unstructured matrix requires \(3(3-1)/2=3\) separate pairwise correlation parameters, while exchangeable requires one. If the scientific expectation is indeed similar association for all pairs, the extra flexibility of unstructured correlation has no clear design-based motivation. The equal spacing alone does not require AR(1): AR(1) is preferable when association is expected to decline with lag, whereas the stated expectation here is approximately constant association.
The investigators could inspect residual association by pair after fitting the mean model. If the first-to-third association is consistently much weaker than adjacent-occasion associations, that would challenge exchangeability and support considering AR(1). If estimates are noisy and vary sharply across pairs, the evidence may be insufficient to justify three unstructured parameters. The decision remains a compromise among the expected pattern, empirical diagnostics, and stability, not a search for the most flexible matrix by default.
Common Mistakes and Limits
- Choosing by label alone. “Longitudinal data” do not automatically require AR(1). The actual spacing and plausible decay pattern determine whether that structure is apt.
- Treating unstructured as universally safest. Flexibility costs parameters. With many occasions or limited information, the pairwise correlations may be unstable or the working matrix may fail to be usable.
- Assuming independence means outcomes are independent. Independence is a working correlation choice for the estimating procedure. It does not prove that outcomes within a person or cluster are unrelated.
- Selecting the structure that gives the smallest standard error. This rewards a favorable result rather than a defensible dependence model. Choose based on design, plausible association, and stable estimation; use appropriate clustered-data inference.
- Using a correlation diagnostic to repair a poor mean model. Residual patterns can reflect an incorrect mean structure as well as dependence. Check that the marginal mean model is scientifically reasonable before interpreting residual correlations.
No working correlation can compensate for dependence between clusters when the analysis assumes independent clusters, or for a seriously misspecified marginal mean. The structures are also approximations: real dependence may vary with covariates, time, or cluster characteristics in ways that the chosen form cannot capture. A transparent analysis states the primary choice, why it fits the design, what plausible alternatives were considered, and whether diagnostics raised concerns.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A study has five equally spaced monthly measurements per person. Investigators expect adjacent measurements to be more associated than measurements several months apart. Name a natural working structure and explain why exchangeable is less aligned with this expectation.
- For an AR(1) structure with \(\rho=0.50\), calculate the working correlation at lag three.
- Measurements occur at days 0, 2, and 12. Why might ordinary discrete AR(1) be a poor representation if dependence decays with elapsed time?
- How many distinct off-diagonal correlations does an unstructured matrix require for six occasions? What is the main practical concern?
- True or false: choosing independence in GEE establishes that outcomes within each cluster are statistically independent. Correct the statement.
Solutions. (1) AR(1) is a natural candidate because it represents decreasing correlation with increasing occasion lag; exchangeable assigns the same correlation to every pair. (2) The lag-three correlation is \(\rho^3=0.50^3=0.125\). (3) Discrete AR(1) treats every one-occasion gap alike, although the intervals are two days and ten days; a time-decay structure may better reflect the stated dependence. (4) The count is \(6(6-1)/2=15\). Estimating 15 correlations may be unstable or impractical if the data provide limited information. (5) False. Independence is the working correlation used in the estimating procedure; it does not establish that the actual within-cluster outcomes are independent.