Why the Working Correlation Matters
Generalized estimating equations (GEE) describe a marginal mean while allowing outcomes within a cluster to be correlated. A working correlation is the correlation structure used to weight each cluster’s contribution to the estimating equation. It is called “working” because it need not equal the true within-cluster correlation for the estimating equation to target the correct marginal mean parameter. But a wrong working correlation can change how much weight different observations receive, and therefore reduce precision.
The distinction is important: robustness of the mean estimate to correlation misspecification is not the same as robustness of its standard error, nor does it mean that the choice of correlation is irrelevant. The claims here require a correctly specified marginal mean, independent clusters, and valid large-sample robust inference. Correlation misspecification alone does not necessarily bias the estimate under these conditions; an incorrect mean model can.
Here \(K\) is the number of independent clusters. Common working correlations include independence, exchangeable correlation (one common within-cluster correlation), and autoregressive correlation (correlation that declines with separation in time). A working covariance must be valid for the cluster size in question, including being positive definite so its inverse exists.
Why the Target Can Survive a Wrong Correlation
Suppose the mean model is correct at \(\beta_0\), so \(E(Y_i)=\mu_i(\beta_0)\). Treat \(D_i\) and \(V_i\) as fixed at that parameter for the expectation argument. The expected contribution from cluster \(i\) is $$ E\left[D_i^\mathsf{T}V_i^{-1}\{Y_i-\mu_i(\beta_0)\}\right] =D_i^\mathsf{T}V_i^{-1}\left[E(Y_i)-\mu_i(\beta_0)\right]=0. $$ This calculation uses the mean, not the true covariance. Thus even if \(V_i\ne\Omega_i\), the estimating equation is unbiased at \(\beta_0\). Under regularity conditions—independent clusters, finite moments, an identifiable parameter, and a nonsingular limiting sensitivity matrix—the solution can consistently estimate \(\beta_0\) as the number of clusters grows.
This is a large-sample statement, not a guarantee that every finite-sample estimate is unchanged by the working correlation. Different \(V_i\) values generally produce different weights and therefore different estimates in a particular data set. The key point is that, when the mean is correct, those estimating equations still have the same target in expectation.
Worked Example: A Zero-Mean Score With the Wrong Correlation
Suppose each hypothetical cluster has two outcomes, both with marginal mean \(\mu\), and variance 1. The true within-pair correlation is 0.6, but the analyst uses a working independence covariance \(V=I\). For an intercept-only mean, \(D=(1,1)^\mathsf{T}\), so the cluster’s contribution to the estimating equation is $$ D^\mathsf{T}V^{-1}(Y-\mu D)=(Y_1-\mu)+(Y_2-\mu). $$ At the true mean, each residual has expectation zero. Therefore the expected contribution is \(0+0=0\), despite the incorrect working correlation.
The actual variance of the pair’s residual sum is not the working independence variance: $$ \operatorname{Var}(Y_1+Y_2)=1+1+2(0.6)=3.2, \qquad \operatorname{Var}_{\text{working}}(Y_1+Y_2)=1+1=2. $$ The working correlation has not changed the expected score, but it has failed to represent the score’s variability. This is why a valid robust variance estimate is needed.
Robust Variance and Efficiency
The covariance structure affects the sampling variance of the GEE estimator. Define the sensitivity and variability contributions, evaluated at the target, by $$ H=\sum_{i=1}^{K}D_i^\mathsf{T}V_i^{-1}D_i, \qquad J=\sum_{i=1}^{K}D_i^\mathsf{T}V_i^{-1}\Omega_iV_i^{-1}D_i. $$ The large-sample covariance is approximated by the sandwich \(H^{-1}JH^{-1}\). In practice, \(\Omega_i\) is not known; the empirical sandwich replaces the unknown variability contribution with cluster score outer products. This accounts for the observed between-cluster variability without requiring the working correlation to be correct. Its justification is asymptotic, so the number of independent clusters—not merely the number of individual outcomes—needs to be adequate.
A model-based standard error instead treats the working covariance as the actual covariance, effectively using \(H^{-1}\) (with any appropriate dispersion factor). If \(V_i\ne\Omega_i\), that standard error may be too large or too small. Robust inference means using the sandwich variance and its corresponding standard errors and tests; it does not mean the estimate is immune to mean-model errors or that sandwich estimates are reliable with very few clusters.
Efficiency refers to the estimator’s sampling variance: a more efficient estimator has smaller variance under the conditions being compared. If the true covariance is known and used for weighting, the resulting GEE is optimal among a broad class of linear unbiased estimating equations under standard regularity conditions. A misspecified working covariance may give less efficient weighting. It does not follow that every plausible working correlation improves precision over independence in every design, or that the most complicated working structure is best in a finite sample.
Worked Comparison of Working Weights
The following simplified example makes the efficiency cost visible. Use a Gaussian outcome with identity link and marginal mean \(\mu\), variance 1, and independent clusters. There are 100 single-outcome clusters and 100 two-outcome clusters. Within each pair, the true correlation is 0.8. The mean model assumes the same \(\mu\) for every outcome, and the working correlation is exchangeable with value \(r\). These assumptions make both working structures valid for the example; the true covariance is stipulated to calculate sampling variances.
Worked Example: Independence Versus the True Pair Correlation
For a pair, the exchangeable working covariance is \(V=\begin{pmatrix}1&r\\r&1\end{pmatrix}\). Its contribution to the intercept estimating equation is the sum of the two residuals divided by \(1+r\). Equivalently, the pair mean receives weight \(c=2/(1+r)\). A single outcome receives weight 1. Thus, if \(\bar Y_s\) is the mean of the 100 single outcomes and \(\bar Y_p\) is the average of the 100 pair means, the estimate is $$ \hat{\mu}(c)=\frac{100\bar Y_s+100c\bar Y_p}{100+100c}. $$ Assume the hypothetical observed averages are \(\bar Y_s=12\) units and \(\bar Y_p=13\) units.
With working independence, \(r=0\), so \(c=2\). Substitution gives $$ \hat{\mu}(2)=\frac{100(12)+100(2)(13)}{100+100(2)} =\frac{3800}{300}=12.67\text{ units}. $$ With the true pair correlation as the working correlation, \(r=0.8\), so \(c=2/1.8=1.111\). Then $$ \hat{\mu}(1.111)=\frac{1200+100(1.111)(13)}{100+111.1} \approx\frac{2644.3}{211.1}=12.53\text{ units}. $$ The two estimates differ for these observed data because the weights differ. Both estimating equations still have target \(\mu\) when every outcome has that same marginal mean; neither estimate is guaranteed to equal the other in a particular sample.
The actual variance of a pair mean is \((1+0.8)/2=0.9\), while a single outcome has variance 1. For a general pair-mean weight \(c\), the sampling variance of this estimator is $$ \operatorname{Var}\{\hat{\mu}(c)\} =\frac{100(1)+100c^2(0.9)}{(100+100c)^2}. $$ Under working independence, \(c=2\), giving $$ \frac{100+100(4)(0.9)}{300^2} =\frac{460}{90000}=0.005111, $$ with standard deviation \(\sqrt{0.005111}=0.07149\) units. With \(c=1.111\), the variance is approximately $$ \frac{100+100(1.111^2)(0.9)}{211.1^2} \approx 0.004737, $$ with standard deviation about \(0.06882\) units. In this design, using the true correlation gives more efficient weighting. This comparison is about repeated-sample variability, not about which estimate is closer to \(\mu\) in this one hypothetical data set.
Checking Inference and Diagnosing Failure
For the independence fit above, the model-based variance would be \(1/300=0.003333\), treating the 300 outcomes as independent. Its standard error is \(\sqrt{1/300}=0.05774\) units. But the actual variance of that estimator under the stipulated pair correlation is 0.005111, so the model-based standard error is too small. A robust sandwich estimate, calculated from cluster-level score contributions, is intended to capture this excess variability. In real data the sandwich is estimated rather than known exactly; it can be unstable with few independent clusters.
Several checks help distinguish a correlation problem from a mean-model problem:
- Inspect mean residual patterns. Systematic residual trends over time, covariate values, or fitted means suggest the marginal mean may be misspecified. Changing \(R_i(\alpha)\) does not repair a wrong mean.
- Compare plausible working structures. If coefficient estimates or robust standard errors change substantially across structures, the analysis may be sensitive to weighting, the sample may be limited, or the mean model may need review. A change in estimates alone does not establish bias.
- Check convergence and covariance validity. An estimated working correlation near a boundary, a non-positive-definite working matrix, or numerical instability can make the estimating equation unreliable. The chosen structure must be valid for the observed cluster sizes.
- Count independent clusters. Many outcomes concentrated in only a few clusters do not automatically provide the large number of independent units that sandwich inference relies on. Small-sample corrections may help in some settings, but they do not create information that is absent.
| Approach | What it assumes or uses | Main implication |
|---|---|---|
| Working independence GEE | Uses independence weights, while permitting robust sandwich inference | Simple and often stable; can lose efficiency when within-cluster dependence is informative for weighting |
| Structured working-correlation GEE | Uses a parsimonious correlation form, such as exchangeable or autoregressive | May improve efficiency if the structure approximates the true covariance; robust inference remains important |
| Model-based GEE inference | Treats the working covariance as correct for standard errors | Can give incorrect uncertainty estimates when the working covariance is misspecified |
| Likelihood-based GLMM | Specifies a conditional outcome distribution and random-effects structure | A different modeling approach and estimand; it is not a direct correction for a misspecified GEE working correlation |
The main alternatives are therefore not interchangeable fixes. Working independence trades a simple covariance specification for possible efficiency loss. A structured working correlation adds a covariance model to improve weighting, but estimates of its parameters can be noisy, especially with few clusters or short cluster series. A GLMM instead models conditional outcomes through random effects; it answers a different modeling question and brings its own distributional assumptions. The choice should reflect the marginal estimand, design, and available cluster information—not a belief that one correlation structure must always be correct.
Common Mistake and Limits
Correlation misspecification is most safely understood as a potential loss of efficiency with robust inference, not as a universal no-consequence guarantee. If the mean model is wrong, the expected estimating equation need not vanish at the intended scientific parameter. If cluster size is informative, the target and weighting require additional care. If clusters are dependent or too few for reliable sandwich inference, the usual justification also fails. These limitations are why mean specification, study design, and robust variance estimation must be assessed separately from the choice of working correlation.
Check Your Understanding
Answer each question, then check the concise solutions below.
- State the key expectation argument that explains why a wrong working covariance need not change the GEE target.
- In the paired-outcome example, what is the pair-mean variance when the true correlation is 0.5 and each outcome has variance 1?
- For the 100 single and 100 paired clusters above, calculate the estimator variance under working independence if the true pair correlation were 0 rather than 0.8.
- Why can a robust sandwich standard error not correct an omitted time trend in the marginal mean?
Solutions. (1) At \(\beta_0\), \(E(Y_i-\mu_i(\beta_0))=0\), so \(E[D_i^\mathsf{T}V_i^{-1}\{Y_i-\mu_i(\beta_0)\}]=0\), regardless of whether \(V_i=\Omega_i\). (2) The pair-mean variance is \((1+1+2(0.5))/4=0.75\). (3) With independent outcomes, the independence-weighted estimator is the mean of 300 independent unit-variance outcomes, so its variance is \(1/300=0.003333\). (4) The sandwich estimates variability around the parameter solving the estimating equation; if the mean is wrong, that equation may target a different parameter, so correcting the variance cannot restore the intended mean-model target.