Unequal Cluster Sizes and Cluster-Robust Inference

Why Unequal Cluster Sizes Matter

Cluster-robust inference treats the cluster, rather than each individual observation, as the independent unit for estimating uncertainty. When cluster sizes differ, the observations are not simply interchangeable pieces of independent information: observations within a cluster can share errors, and a large cluster can contribute substantially to the fitted regression and its estimated variance. The key is to examine each cluster’s contribution to the coefficient’s uncertainty, not to assume that every cluster contributes equally.

For example, a data set might contain many observations from one school and few from several others. The total sample size alone does not say how much independent information is available. Nor does the size of the largest school, by itself, determine the standard error. The regression’s predictor values and residuals within each school also matter.

Model and Cluster-Robust Variance

Suppose there are \(G\) independent clusters. Cluster \(g\) contains \(m_g\) observations, indexed by \(j=1,\ldots,m_g\). For each observation, let \(Y_{gj}\) be the outcome and let \(\mathbf{x}_{gj}\) be the vector of regression predictors, including the intercept. Write the linear mean model as

The model requires a correctly specified conditional mean, \(E(Y_{gj}\mid\mathbf{x}_{gj})=\mathbf{x}_{gj}'\boldsymbol{\beta}\), and independent clusters. Errors may be correlated and have unequal variances within a cluster; the method does not require observations within the same cluster to be independent. The usual large-sample justification relies on having many independent clusters, not merely a large total number of observations. Cluster-robust standard errors do not repair a wrong mean model or make a small number of clusters adequate for reliable inference.

The formula makes an important feature visible: each cluster’s individual score contributions are added first, and the resulting cluster total is then multiplied by its transpose. Thus the variance depends on the size and direction of each cluster’s total score, not just on how many observations that cluster contains. Small-sample corrections may multiply or otherwise adjust this estimator, but they do not remove the underlying role of cluster-level contributions.

Why Cluster Contributions Add Up Before They Are Squared

A short variance argument explains why the observations are grouped this way. A regression score contribution from observation \(j\) in cluster \(g\) is \(\mathbf{x}_{gj}\varepsilon_{gj}\). The total score is the sum of those contributions across all observations. Its variance contains covariances between observations in the same cluster as well as their individual variances.

because different clusters are independent. Within a cluster, however, the variance on the right includes cross-products between different observations. For example, for two observations in a cluster, the variance of their summed scores contains both cross-covariance terms, not just the two individual variances. The empirical outer product \(\hat{\mathbf{U}}_g\hat{\mathbf{U}}_g'\) retains this combined cluster contribution. If the observations were incorrectly treated as independent, those within-cluster cross-covariances would be omitted.

Unequal sizes affect this calculation because a cluster with more observations has more terms that can contribute to its total score. But those terms can reinforce or partly cancel one another. A larger \(m_g\) therefore does not imply a proportionally larger variance contribution in every data set.

Large Clusters and Coefficient Precision

For a regression coefficient, a cluster’s score depends on both the residuals and the predictors. In a simple regression, for instance, observations with predictor values far from the overall predictor mean can have especially relevant score contributions for the slope. A cluster may be large but contribute little to the slope variance if its residual-pattern and predictor-weighted scores largely cancel. Conversely, a cluster of moderate size can matter greatly if its score is large.

The cluster-robust standard error describes uncertainty in the coefficient estimate, conditional on the model and clustering assumptions. A larger standard error means lower estimated precision. It does not mean the coefficient estimate itself has changed, and it does not say that a large cluster is biased. Unequal sizes are a feature of the data; their consequences for precision must be assessed through the fitted regression and cluster-level contributions.

What to Examine in Practice

When cluster sizes are unequal, report the number of independent clusters as well as the total number of observations. Summaries of the cluster-size distribution help show whether a few clusters contain much of the data. For understanding coefficient precision, also inspect cluster-level score contributions or other appropriate influence summaries. These answer related but different questions: cluster sizes describe how observations are distributed, while score contributions show how the fitted regression’s variance estimate is distributed across clusters.

Unequal sizes also make it risky to reason from an “average cluster” alone. The average size may conceal a very large cluster, and the same average size can occur with different distributions. Even so, size summaries are diagnostic rather than a substitute for the variance calculation. Cluster-robust inference aggregates the scores for the clusters actually observed.

Common Mistakes

• Using the total number of observations as if they were all independent. When observations share a cluster, their within-cluster dependence affects uncertainty. The number of independent clusters is essential to judging whether the large-cluster approximation is credible.
• Assuming the largest cluster must dominate every coefficient’s standard error. The variance depends on its aggregated predictor-weighted residual score, not size alone. A large cluster’s score can be small through cancellation.
• Assuming unequal sizes automatically invalidate cluster-robust standard errors. Unequal sizes affect contributions and may raise practical concerns, but they do not, by themselves, violate the method’s assumptions. Independent clusters, an appropriate mean model, and enough clusters for the intended inference remain important.
• Interpreting a large variance contribution as evidence of bias. A cluster can contribute substantially to estimated uncertainty without making the coefficient biased. The contribution indicates sensitivity of the variance estimate, not a causal or data-quality conclusion.