Cluster Size and Correlation-Parameter Precision

Why Pair Counts and Cluster Counts Both Matter

A cluster of \(m\) repeated measurements contains \(\binom{m}{2}\) distinct pairs. That count is relevant to correlation estimation, but it is not usually the estimator’s effective sample size: pairs from the same cluster share observations and are dependent. In contrast, independent clusters provide independent replications of the entire within-cluster pattern. Precision therefore depends both on how many usable pairs contribute to a parameter and on how those pairs co-vary within clusters.

The distinction is especially important when comparing working-correlation structures. An exchangeable model pools information across many pairs for one common parameter. An AR(1) model draws information from pairs at particular lags, with lag-one pairs often supplying the estimate of the adjacent-visit correlation. An unstructured model has a separate parameter for each visit pair, so each parameter has its own available replication count.

For transparent variance calculations, several examples below use mean-zero Gaussian responses with unit marginal variance and known marginal standardization. These assumptions make pair products easy to analyze; they are not claims that residuals from every generalized linear model are Gaussian. In applications, estimated means, estimated scales, missing observations, and non-Gaussian responses can change the variance. The central distinction remains: uncertainty must account for dependence among contributions from the same cluster.

Exchangeable Correlation: Pooling Pairs Without Pretending They Are Independent

Under an exchangeable correlation structure, every distinct pair in a cluster has the same correlation \(\rho\). With complete data, a natural method-of-moments estimator averages all pair products across clusters. Define the within-cluster sum and average by

If \(E[X_{ij}X_{ik}]=\rho\) for distinct visits, then \(E[T_i]=\rho\). For equal cluster sizes, the pooled estimator is \(\hat{\rho}=K^{-1}\sum_{i=1}^K T_i\). The following exact calculation shows why \(q\) is not, in general, the number of independent pieces of information within a cluster.

Proof. Write \(Z_{jk}=X_jX_k\) for a pair product in one cluster. For a single pair, the Gaussian fourth-moment identity gives \(E[X_j^2X_k^2]=1+2\rho^2\), so \(\operatorname{Var}(Z_{jk})=1+\rho^2\). For two distinct pairs sharing one index, such as \((j,k)\) and \((j,\ell)\), the same identity gives \(E[X_j^2X_kX_\ell]=\rho+2\rho^2\). Subtracting \(E[Z_{jk}]E[Z_{j\ell}]=\rho^2\), their covariance is \(\rho+\rho^2\).

For two disjoint pairs, such as \((j,k)\) and \((\ell,r)\), the fourth moment is \(3\rho^2\), so the covariance is \(3\rho^2-\rho^2=2\rho^2\). There are \(q\) individual pair terms, \(m\binom{m-1}{2}\) unordered pairs of pair-terms that share an index, and \(3\binom{m}{4}\) unordered pairs of pair-terms that are disjoint. In a variance sum, each covariance between distinct terms appears twice. Therefore

Since \(T_i=S_i/q\), divide by \(q^2\) to obtain the displayed variance. Finally, independence across clusters gives \(\operatorname{Var}(K^{-1}\sum_i T_i)=K^{-2}\sum_i\operatorname{Var}(T_i)=\operatorname{Var}(T_i)/K\). \(\square\)

At \(\rho=0\), the formula simplifies to \(\operatorname{Var}(T_i)=1/q\): in this special Gaussian case, the covariances between distinct pair products vanish. Away from zero, the shared observations and nonzero correlations create additional covariance terms. Consequently, a rule that divides a variance by the total pair count \(Kq\) is not generally justified.

Why Independent Clusters Govern Replication

For a general cluster-level estimating contribution, there is a precise large-sample statement. It does not require the pair products inside a cluster to be independent.

Proof. Independence gives \[ \operatorname{Var}(\hat{\theta}) =\operatorname{Var}\left(K^{-1}\sum_{i=1}^K U_i\right) =K^{-2}\sum_{i=1}^K\operatorname{Var}(U_i) =\tau^2/K. \] The variables \(U_i-\theta\) are independent and identically distributed, have mean zero, and have finite variance \(\tau^2\). The classical central limit theorem therefore gives \(K^{-1/2}\sum_{i=1}^K(U_i-\theta)\xrightarrow{d}N(0,\tau^2)\), which is the stated convergence. \(\square\)

In the exchangeable example, \(U_i=T_i\). More generally, estimating equations often produce cluster-level contributions whose variance reflects all within-cluster dependence. With cluster size fixed, increasing \(K\) supplies more independent replications and reduces variance at rate \(1/K\). Increasing cluster size can also add useful information, but its effect is encoded in the variance of the entire cluster contribution; it cannot be assessed by simply counting every pair as an independent replicate.

AR(1): Lag-One Pair Count and Overlap

For an ordered schedule with AR(1) correlation, \(\operatorname{Corr}(X_{ij},X_{ik})=\rho^{|j-k|}\). A simple lag-one moment estimator averages the \(m-1\) adjacent products in each cluster. Those products overlap: consecutive products share one response, and more widely separated products remain correlated under the AR(1) model.

For this estimator, adding visits increases the number of adjacent pairs from \(m-1\), but does not make those pairs independent. The precise gain depends on \(\rho\) and on the covariance pattern among the products. Pair counts at other lags may inform other aspects of an AR(1) fit, but they do not turn into independent replications of the lag-one moment.

Unstructured Correlation: Replication Is Per Pair

With an unstructured correlation matrix for a fixed schedule of \(m\) visits, there are \(q=\binom{m}{2}\) off-diagonal correlations, one for each distinct visit pair. Under complete data, every pair-specific parameter is observed once per cluster, giving \(K\) pair observations for that parameter. The \(Kq\) total pair products across all clusters and pairs are not \(Kq\) replications of any single unstructured parameter.

If visits are missing, the number of available observations can differ from pair to pair. For a particular unstructured parameter, its direct replication count is the number of independent clusters in which that pair is observed, subject to the missingness assumptions and the estimation method. A large total number of observed pairs can coexist with sparse information for one particular pair.

Common Mistake and Practical Interpretation

Wrong reasoning: “There are 20 clusters with 10 pairs each, so the correlation estimate has 200 independent observations; its standard error should be proportional to \(1/\sqrt{200}\).” This treats pair products sharing a response, or even sharing a whole cluster, as independent.

Correction: Identify the cluster-level contribution to the estimator and account for its within-cluster variance, including covariance among pair products. Under independent, identically distributed clusters with finite variance, the variance of an average of cluster contributions scales as \(1/K\), not automatically as \(1/(Kq)\). For an exchangeable parameter, pooling pairs can improve precision, but the size of that gain depends on their covariance. For an AR(1) parameter, relevant lag-specific products overlap and are correlated. For an unstructured parameter, assess the number and dependence of observations for that particular pair.

These calculations describe precision for correlation parameters, not the full precision of generalized estimating equation regression coefficients. Correlation estimation can affect coefficient efficiency, but the two uncertainty questions are distinct. When using residual-based correlation estimates, estimated mean parameters and residual scales can add further dependence; the simple Gaussian calculations are benchmarks, not universal standard-error formulas.