Matching Working Correlation to the Study Design

Why Study Design Comes Before the Correlation Formula

In generalized estimating equations (GEE), a working correlation describes how responses within the same cluster are treated as dependent during estimation. Choosing one is not simply a matter of selecting the most flexible formula. The schedule of measurements determines which pairs of observations are comparable, while scientific knowledge suggests how dependence might change across visits. The number of clusters and the number of visits determine how much complexity the data can support.

A parallel-group trial, for example, assigns different subjects to treatment arms; this allocation does not itself specify the within-subject correlation. If each subject is measured repeatedly, the visit schedule and the response process guide that choice. If each subject contributes only one outcome, there is no within-subject repeated-measure correlation to model.

Definition: For cluster \(i\), let \(Y_i\) be the vector of responses and let \(R_i(\alpha)\) be a working correlation matrix, with diagonal entries one and off-diagonal entries describing the assumed correlations among responses in that cluster. The working covariance is commonly written \(V_i=A_i^{1/2}R_i(\alpha)A_i^{1/2}\), where \(A_i\) contains the marginal variances. A working structure is a choice of the pattern imposed on \(R_i(\alpha)\); it is not a claim that the responses are truly distributed according to a particular joint model.

The practical decision is to find a structure that respects the measurement design without imposing more parameters than the data can usefully support. Working correlation can affect the efficiency of GEE estimates, but a plausible structure does not repair a misspecified marginal mean. Conversely, a working structure need not be exactly correct for the GEE estimating equation to remain centered when the mean model is correct, as shown below.

What Each Structure Assumes

The four common choices make different assumptions about which features of the design matter. Here, \(j\) and \(k\) index measurements within a cluster.

Structure	Working assumption	Design pattern it can suit	Main caution
Independence	All within-cluster off-diagonal correlations are zero.	A reasonable starting point when dependence is weak or poorly characterized.	It deliberately ignores within-cluster dependence when it is present.
Exchangeable	Every distinct pair has the same correlation.	Common sets of repeated measurements when dependence is roughly similar across pairs.	It cannot represent decay with time lag or unequal dependence by visit pair.
AR(1)	Correlation is \(\rho^{\|j-k\|}\) for measurements in ordered, equally spaced positions.	Repeated observations at regular intervals when dependence plausibly weakens with lag.	Index distance represents time distance; that is inappropriate for irregular visits.
Unstructured	Each distinct visit pair has its own correlation parameter.	A small fixed set of visits with enough information for each pair.	Parameter count grows quickly, and sparse pair information can make estimates unstable.

For unequally spaced visits, an exponential correlation can instead relate correlation to actual time separation: \(R_{i,jk}=\exp(-\phi |t_{ij}-t_{ik}|)\), where \(\phi\geq 0\). This is not one of the four choices named in the focus of this tutorial, but it is an important comparison when deciding whether an index-based AR(1) structure respects the schedule.

Two Results That Clarify the Choice

A structure should reflect the design, but its role in GEE should not be confused with a model for the entire joint distribution. The following results make both points precise.

Result: Suppose the marginal mean is correct, so that \(E[Y_i\mid X_i]=\mu_i(\beta_0)\). Conditional on the covariates, let \(D_i\) and \(W_i\) be fixed matrices, with \(W_i\) nonsingular. Then the GEE contribution \(U_i(\beta_0)=D_i^T W_i^{-1}\{Y_i-\mu_i(\beta_0)\}\) has conditional expectation zero, regardless of the off-diagonal pattern in \(W_i\).

Proof. Since \(D_i\) and \(W_i^{-1}\) are fixed conditional on \(X_i\), they can be taken outside the conditional expectation. Therefore

E[U_i(\beta_0)\mid X_i] =D_i^T W_i^{-1}E[Y_i-\mu_i(\beta_0)\mid X_i] =D_i^T W_i^{-1}\{\mu_i(\beta_0)-\mu_i(\beta_0)\} =0.

Summing over clusters gives a conditionally mean-zero estimating function. This calculation does not prove that every solution is consistent or that its standard error is reliable; those conclusions require further regularity and inference conditions. It does show why choosing a working correlation is not the same as specifying the marginal mean. The fixed-matrix condition matters: if the working matrices are estimated from the same responses, this calculation alone does not establish exact finite-sample mean zero.

Result: If visits are equally spaced, \(t_j=t_0+(j-1)\Delta\), with \(\Delta>0\), the exponential correlation \(\exp(-\phi|t_j-t_k|)\) equals an AR(1) correlation matrix with parameter \(\rho=\exp(-\phi\Delta)\).

Proof. For any two visit indices,

\exp(-\phi|t_j-t_k|) =\exp\{-\phi |(j-k)\Delta|\} =\{\exp(-\phi\Delta)\}^{|j-k|} =\rho^{|j-k|}.

This is exactly the AR(1) entry for that pair. When \(\phi>0\), \(0<\rho<1\); when \(\phi=0\), all entries are one and the matrix is singular for more than one visit. Thus equally spaced timing makes these two patterns equivalent, while irregular timing generally does not: the same difference in visit index can then represent different elapsed times. \(\square\)

A Design-Based Selection Process

Identify the repeated unit and its schedule.
List what is measured within a subject or other cluster, and record the actual visit times. Distinguish a common visit schedule from a common treatment allocation; treatment arms do not dictate the correlation pattern.

Ask how dependence should vary.
If no defensible pattern is available, independence is a simple baseline. If all pairs seem similarly related, consider exchangeable. If dependence should decay by regular lag, consider AR(1); for unequal time gaps, consider a time-based structure.

Count the parameters the data must support.
With \(K\) visits, an unstructured matrix has \(K(K-1)/2\) distinct off-diagonal correlations. Consider whether the study has enough clusters and observations for the relevant visit pairs.

Check the choice against the scientific process.
A numerical fit alone does not make a structure meaningful. Ask whether the imposed symmetry, time ordering, or decay pattern is compatible with how responses arise.

Worked Examples

Worked Example: Regular Visits With Gradual Decay

A study measures a continuous outcome for every participant at baseline, month 1, and month 2. Suppose the scientific expectation is that nearby measurements are more strongly related than measurements farther apart. Compare an exchangeable structure with AR(1), using the illustrative AR(1) value \(\rho=0.7\).

Under exchangeability, every distinct pair has correlation \(c\), so the correlation matrix is

R_{\mathrm{ex}}= \begin{pmatrix} 1&c&c\\ c&1&c\\ c&c&1 \end{pmatrix}.

This makes the baseline-to-month-1, month-1-to-month-2, and baseline-to-month-2 correlations all equal. Under AR(1), the visit-index gaps are 1, 1, and 2, respectively. Substituting \(\rho=0.7\) gives

R_{\mathrm{AR(1)}}= \begin{pmatrix} 1&0.7&0.7^2\\ 0.7&1&0.7\\ 0.7^2&0.7&1 \end{pmatrix} = \begin{pmatrix} 1&0.7&0.49\\ 0.7&1&0.7\\ 0.49&0.7&1 \end{pmatrix}.

The AR(1) pattern matches the stated expectation: adjacent visits have correlation \(0.7\), while the two-month separation has correlation \(0.49\). If the scientific reasoning instead supported roughly equal dependence for every pair, exchangeability would be more suitable. The regular schedule makes AR(1) interpretable here; the scientific pattern, rather than regular spacing alone, is what favors it.

Worked Example: Irregular Follow-Up Times

Participants are measured on days 0, 7, and 28. Suppose correlation is expected to decline with elapsed time, and use the illustrative exponential rate \(\phi=0.05\) per day. For a gap \(d\), the correlation is \(\exp(-0.05d)\).

The three time gaps are \(7\), \(21\), and \(28\) days. Their correlations are

\exp(-0.05\times 7)=\exp(-0.35)\approx 0.705,\qquad \exp(-0.05\times 21)=\exp(-1.05)\approx 0.350,

\exp(-0.05\times 28)=\exp(-1.40)\approx 0.247.

Thus the implied matrix, in time order, is approximately

R= \begin{pmatrix} 1&0.705&0.247\\ 0.705&1&0.350\\ 0.247&0.350&1 \end{pmatrix}.

An index-based AR(1) with \(\rho=0.7\) would instead assign \(0.7\) to both adjacent pairs, even though their time gaps are 7 and 21 days. That structure overlooks the irregular schedule. An exchangeable structure would also require all three correlations to be equal. Given the stated time-decay rationale, a time-based exponential pattern is the more coherent candidate.

Worked Example: Is an Unstructured Matrix Supportable?

A short-panel study plans five annual measurements for each participant and has 40 participants. An unstructured correlation for five visits requires

\frac{K(K-1)}{2} =\frac{5(5-1)}{2} =\frac{20}{2} =10

distinct off-diagonal correlation parameters. Each parameter corresponds to a different visit pair. The schedule is fixed, so the pairs are clearly defined, and there are 40 participants who may contribute information to each pair if measurements are complete. Unstructured correlation is therefore a candidate, but the parameter count alone does not establish that estimates will be stable; missing observations, highly variable responses, or weak information in particular pairs may still be a concern.

Now suppose the study has 10 visits and still only 40 participants. The required number becomes

\frac{10(10-1)}{2}=\frac{90}{2}=45

parameters, which exceeds the number of participants. That comparison is a warning rather than a theorem that estimation must fail: information depends on the observed data and estimation procedure. Nevertheless, if the scientific pattern is plausibly simpler, exchangeable or AR(1) can be more defensible and more parsimonious. For annual measurements with a credible gradual-decay pattern, AR(1) is a natural alternative to assess.

Practical Interpretation and Common Pitfalls

Parallel-group trials often use the same follow-up schedule in each treatment arm. In that case, the working correlation describes responses within a participant over visits; it is not a separate correlation between treatment groups. If the schedule is common, a shared structure across arms is often a sensible starting specification, though that choice should be considered alongside the scientific question and the fitted mean model.

For a fixed, short visit schedule, an unstructured pattern can be attractive when there is no convincing reason to impose symmetry or a time-decay rule. But its flexibility has a cost: as the number of visits increases, so does the number of distinct correlations. With missing visits, some pairs may be represented by far fewer subjects than others, so a nominally fixed schedule does not guarantee adequate information for every pair.

Mistake: “The visits are ordered, so AR(1) is automatically correct.” Ordering is not enough. AR(1) treats equal differences in index as equal correlation steps; irregular elapsed times can violate that interpretation. Use actual timing in the structure when time gaps are scientifically relevant.
Mistake: “A parallel-group trial needs an exchangeable correlation.” Parallel allocation describes how treatment groups are formed, not how repeated responses within a subject depend on one another. Choose based on the within-subject measurement schedule and response process.
Mistake: “Unstructured is always best because it imposes the fewest restrictions.” Fewer restrictions mean more parameters to estimate. For a long panel or sparse pairs, those estimates may be unstable; a simpler structure can be a better-supported working choice.
Mistake: “Independence assumes the observations truly are independent.” It specifies a working matrix, not necessarily the true joint distribution. Under a correct marginal mean and suitable conditions, the estimating equation can remain centered, though ignoring real dependence can affect efficiency and the validity of model-based uncertainty calculations.

The selection is therefore a design-informed compromise: represent important features of timing and scientific dependence, but avoid unsupported complexity. After specifying the mean model and a working structure, assess whether the data support estimating the required correlation parameters. The structure is part of the estimating procedure, not a substitute for a scientifically appropriate mean model.

Check Your Understanding

For each question, justify the choice using the measurement schedule and the expected dependence pattern.

A study measures subjects at weeks 0, 2, 4, and 6, and expects correlation to decrease with temporal separation. Which of exchangeable and AR(1) better represents that expectation, and why?
Visits occur on days 0, 3, and 30, with dependence expected to decline with elapsed time. Why might index-based AR(1) be a poor match?
How many distinct off-diagonal parameters does an unstructured correlation require for \(K=6\) visits?
In a parallel-group trial with repeated outcomes, what determines the within-subject working correlation: treatment allocation or the repeated-measure schedule and response pattern?
State the condition on the marginal mean that makes a fixed working-covariance GEE contribution have conditional expectation zero.