Intraclass Correlation and Within-Cluster Similarity

Why Within-Cluster Similarity Matters

Students at the same school, patients treated at the same clinic, or repeated observations from the same person may have outcomes that resemble one another. A cluster can contribute to that resemblance: people at one school may share conditions, or measurements from one person may share stable characteristics. The intraclass correlation coefficient, usually abbreviated ICC, summarizes this within-cluster resemblance on a correlation scale.

An ICC is useful because it makes the degree of similarity interpretable. It does not, by itself, tell us how to adjust a regression analysis or how much its uncertainty changes; those require additional methods. Here the focus is the meaning and calculation of the ICC from a simple variance-components model.

A Variance-Components Model

Suppose outcome \(Y_{ij}\) is measured for individual \(i\) in cluster \(j\). Consider a random-intercept model with a common mean:

Here \(\mu\) is the overall mean, \(u_j\) is the cluster-specific deviation from that mean, and \(\varepsilon_{ij}\) is the individual deviation within the cluster. To define the simple ICC, assume that cluster effects are independent across clusters with mean zero and variance \(\sigma_u^2\); individual deviations are independent across individuals with mean zero and variance \(\sigma_\varepsilon^2\); and the two sets of deviations are independent of one another. These assumptions imply that observations in different clusters share no random component in this model, while observations in the same cluster share \(u_j\).

The ICC is also the correlation between two distinct outcomes from the same cluster under these assumptions. It lies between 0 and 1 when both variance components are nonnegative. If the cluster variance is zero, the model has no between-cluster variation and \(\rho=0\). If the within-cluster variance is zero and the cluster variance is positive, outcomes within a cluster are identical under the model and \(\rho=1\). In ordinary applications, values are usually between these extremes.

When a model includes predictors, the same idea can be applied to the remaining variation after accounting for those predictors: the ICC is then understood relative to the specified model and its variance components. It is not a universal property of a school, clinic, or person independent of the outcome and analysis.

Why the ICC Is a Correlation

The definition as a variance ratio and the definition as a within-cluster correlation agree because the shared cluster effect contributes to the covariance of two observations from the same cluster.

This follows because adding the constant \(\mu\) does not change variance, and the independent components \(u_j\) and \(\varepsilon_{ij}\) have variances \(\sigma_u^2\) and \(\sigma_\varepsilon^2\). For two distinct individuals \(i\) and \(k\) in cluster \(j\), both outcomes contain \(u_j\), but their individual deviations are independent. Thus:

Both outcomes have variance \(\sigma_u^2+\sigma_\varepsilon^2\), so their correlation is their covariance divided by the product of their standard deviations:

For observations from different clusters, the cluster effects are independent and there is no shared individual deviation, so their covariance is zero under this model. The derivation shows why the ICC describes pairwise similarity within the same cluster, not similarity between arbitrary observations.

Reading an ICC Carefully

The variance-ratio interpretation is often the clearest first reading. A larger ICC means a larger share of total modeled variance is between clusters and, under the stated model, a stronger correlation between outcomes from the same cluster. But the correlation and the variance-share interpretations must stay tied to the model and outcome being analyzed.

For example, adding predictors that explain differences between individuals or clusters can change the variance components and therefore change the ICC. Changing the outcome also changes its sources of variation. An ICC for a blood pressure measurement is not automatically the ICC for a diagnosis, a different measurement, or the same measurement taken at a different time.

What Different Values Mean

A low ICC does not mean that outcomes are independent in every practical sense. It means that the shared cluster component is small relative to the total modeled variance. A high ICC means that the cluster-level component makes up a substantial fraction of that variance. In both cases, the actual amount of between-cluster variation also depends on the outcome’s measurement scale and total variability.

The ICC is a relative measure. For example, multiplying every outcome by a fixed nonzero constant multiplies both variance components by the square of that constant, leaving their ratio unchanged. The ICC therefore has no measurement units, even though the variance components used to calculate it do.

Common Mistakes and Why the ICC Matters

• Calling the ICC the percentage of observations that are alike. The ICC is not a count or a proportion of matching pairs. It is a correlation and, in this model, a ratio of variance components.
• Using only the between-cluster variance in the denominator. The denominator must include both components: \(\sigma_u^2+\sigma_\varepsilon^2\). Dividing by \(\sigma_u^2\) alone does not calculate the ICC.
• Interpreting the ICC as causal. An ICC describes a pattern of outcome variation under a model. It does not show that cluster membership caused the similarity.
• Treating one ICC as a fixed feature of a cluster type. The value depends on the outcome, the population, and the model’s variance components.
• Equating a small ICC with no within-cluster dependence. A small value indicates modest modeled similarity, not necessarily exact independence. Its relevance depends on the analysis and context.

The ICC matters because it summarizes how much outcomes from the same cluster share, and how much of total modeled variation lies between clusters rather than within them. That description helps analysts recognize the structure of grouped data. Calculating it is straightforward once the variance components are identified; interpreting it well requires keeping the outcome, model, and units in view.