1 Definition and intuition

1.1 What “cluster” means in statistical contexts

A *cluster* is a grouping of observations that share some common source of variation. Examples include students within the same classroom, patients treated within the same hospital, or repeated measurements taken from the same individual over time. The key feature is that observations inside a cluster are not statistically independent in the way observations drawn from completely separate sampling units would be.

1.2 Within-cluster similarity versus independence

Intra-cluster correlation (ICC) summarizes how strongly outcomes resemble one another within the same cluster. If the ICC is high, then two observations taken from the same cluster tend to be more similar than two observations taken from different clusters. If the ICC is low, the dependence induced by clustering is weak, and observations within a cluster behave nearly like independent draws.

1.3 Relationship to variance partitioning

ICC is closely tied to the idea that total variability can be decomposed into components attributable to (i) differences between clusters and (ii) differences within clusters. When between-cluster differences dominate, ICC tends to be larger; when most variability is within clusters, ICC tends to be smaller. In this sense, ICC is a compact parameter that encodes part of the variance decomposition.

2 Mathematical formulations of ICC

2.1 ICC in one-way random-effects models

2.1.1 ICC as a ratio of variances

2.1.1.1 Derivation from hierarchical variance components

A common starting point is the one-way random-effects model: \[ y_{ij}=\mu+u_i+e_{ij}, \] where \(i\) indexes clusters (\(i=1,\dots,m\)) and \(j\) indexes observations within cluster. Here \(u_i\) represents cluster-level random deviation and \(e_{ij}\) represents within-cluster residual variation. Assume \(u_i \sim (0,\sigma_u^2)\) and \(e_{ij}\sim (0,\sigma_e^2)\), mutually independent. Then the covariance between two observations from the same cluster is \[ \text{Cov}(y_{ij},y_{ik})=\sigma_u^2, \] while the marginal variance is \[ \text{Var}(y_{ij})=\sigma_u^2+\sigma_e^2. \] The intra-cluster correlation is defined as the within-cluster correlation: \[ \text{ICC}=\frac{\sigma_u^2}{\sigma_u^2+\sigma_e^2}. \] This ratio is unitless and interpretable as the proportion of total variance attributable to between-cluster differences under the model assumptions.

2.2 ICC for binary outcomes

For binary or Bernoulli outcomes, variance is not constant on the original scale, so the ICC depends on the link function and the latent-variable structure. In a typical logistic mixed model with random intercepts, the ICC is often defined on an underlying latent scale. For example, with a model \[ \text{logit}\,P(y_{ij}=1)=\beta_0+\beta^\top x_{ij}+u_i,\quad u_i\sim (0,\sigma_u^2), \] one common latent-scale ICC takes the form \[ \text{ICC}=\frac{\sigma_u^2}{\sigma_u^2+\pi^2/3}, \] where \(\pi^2/3\) is the variance of the logistic error term used in the latent formulation. Alternative definitions exist, especially when focusing on marginal (population-average) correlations.

2.3 ICC for count outcomes

For count data modeled with Poisson or negative binomial mixed models, ICC can be defined via the variance components of the assumed distribution and link. In a Poisson random-intercept model, conditional on random effects, observations are independent, but marginally they are correlated. ICC is then expressible in terms of the random-effect variance and the mean–variance relationship implied by the model. As with binary outcomes, different modeling choices (e.g., Poisson versus negative binomial) lead to different operational ICC values.

2.4 ICC in continuous-time or repeated-measures settings

When observations occur repeatedly over time for each subject (a cluster), ICC is related to the correlation induced by subject-specific random effects and any time-dependent structure. With random intercepts only, the correlation between two time points depends on whether both share the same subject effect and whether the residual error is independent across time. More elaborate formulations introduce random slopes, autoregressive residuals, or time-varying covariances, leading to ICC that may vary with time lag or measurement spacing rather than being a single constant.

3 Interpretation and practical meaning

3.1 Range, scale, and typical ranges in practice

Under the standard random-intercepts framework, ICC lies between 0 and 1. Values near 0 indicate minimal within-cluster resemblance; values closer to 1 suggest strong dependence within clusters. In practice, ICC often ranges from very small (near 0) to moderate (tens of percent) depending on the measurement domain, outcome variability, and how clusters are defined. Different outcome types and model assumptions can produce ICC values that are not directly comparable without careful alignment of definitions.

3.2 When ICC is near 0 or near 1

An ICC near 0 implies that most of the observed variability occurs within clusters, so adding more observations inside a cluster yields limited gains in effective information. An ICC near 1 implies that nearly all variability is between clusters and within-cluster differences are small; in this scenario, repeated observations within the same cluster contribute little additional precision beyond what the cluster-level effect already provides.

3.3 How ICC changes with model specification

ICC depends on the random-effects structure and the treatment of residuals. Adding covariates, changing the link function, or allowing random slopes can alter the proportion of variance attributed to clustering. For example, if an important between-cluster predictor is omitted, the residual cluster effect may inflate ICC; conversely, including it may reduce the estimated cluster-level variance and lower ICC. Therefore, ICC should be interpreted in the context of the fitted model.

3.4 Comparing ICC across studies and designs

Comparing ICC values across studies is challenging because ICC is not purely an empirical property of data—it is also a function of the modeling framework (e.g., latent-scale definitions for non-Gaussian outcomes) and the selected clustering level. Two studies can examine “similar” settings but use different cluster definitions, estimation methods, or outcome scales, making direct comparison misleading unless these elements are matched.

4 Estimation of ICC

4.1 Method-of-moments approaches

Method-of-moments estimators use sample analogs of variance components derived from between- and within-cluster variability. In settings close to the one-way random-effects model, estimators can be constructed from ANOVA mean squares or from empirical estimates of intra- and inter-cluster correlations. These approaches are simpler to implement but may rely on assumptions such as balanced clusters and homoscedasticity.

4.2 Likelihood-based estimation in multilevel models

In multilevel (hierarchical) modeling, ICC is usually estimated by fitting the random-effects model and plugging in estimates of variance components (or computing an ICC directly from the fitted random effects). Likelihood-based methods, often using restricted likelihood for Gaussian outcomes, produce estimates that account for covariates and can handle unbalanced cluster sizes and missingness under model assumptions.

4.3 ANOVA-based estimators (legacy approaches)

ANOVA-based ICC estimation uses mean squares from a one-way random-effects ANOVA. For balanced designs and Gaussian outcomes, these estimators relate directly to \(\sigma_u^2\) and \(\sigma_e^2\). In practice, ANOVA-based formulas can be sensitive to departures from assumptions and are less straightforward for generalized outcomes or complex random-effects structures.

4.4 ICC estimation with generalized linear mixed models

For generalized linear mixed models (GLMMs), ICC is often computed from estimated random-effect variances using definitions appropriate to the chosen scale (latent or marginal). Because generalized outcomes have non-constant variance, the ICC may not correspond to a single simple ratio of observed variances on the raw scale. Consequently, analysts typically state explicitly which ICC definition they are using.

4.5 Bootstrap and other resampling strategies

Resampling methods can quantify uncertainty in ICC, especially when analytic standard errors are difficult to obtain or when models are complex. Bootstrap procedures can be applied at the cluster level to preserve the dependence structure. Other strategies include sandwich estimators or profile-likelihood intervals, depending on the model and software.

5 Sampling distributions and uncertainty

5.1 Standard errors and confidence intervals

The sampling distribution of an ICC estimator is typically not normal, particularly with few clusters or small cluster sizes. Standard errors and confidence intervals therefore require careful construction. In multilevel settings, uncertainty can be obtained via asymptotic approximations, profile likelihood, or resampling, with the choice influenced by the number of clusters and the model complexity.

5.2 Small-sample considerations

With a small number of clusters, variance-component estimates can be unstable and confidence intervals can be unreliable. This issue is common in cluster-randomized trials where the primary limitation is often the number of clusters rather than the number of individuals. Practical inference may benefit from methods that respect the clustered data structure and from sensitivity checks.

5.3 Sensitivity to cluster size imbalance

When cluster sizes vary, estimators may weight clusters differently, affecting both point estimates and uncertainty. ICC estimation remains possible, but analysts should be aware that effective information depends on how much data each cluster contributes. Reporting the cluster size distribution and confirming robustness to weighting choices can improve interpretability.

5.4 Missing data within clusters

Missing observations within clusters can complicate ICC estimation if missingness is related to outcomes or cluster effects. Under a missing-at-random assumption conditional on the model, likelihood-based methods can accommodate missing data. However, if missingness differs by cluster or is strongly outcome-driven beyond modeled covariates, ICC can be biased.

6 ICC in study design

6.1 Design effect (variance inflation due to clustering)

Clustering increases variance of estimators relative to an independent-sampling benchmark. The *design effect* quantifies this inflation and is commonly expressed in terms of ICC and average cluster size. Under the one-way random-effects model, a frequently used approximation is: \[ \text{Design effect}=1+( \bar{n}-1)\times \text{ICC}, \] where \(\bar{n}\) is the mean number of observations per cluster. This highlights a key consequence: information does not grow as rapidly with additional individuals within the same cluster as it would under independence.

6.2 Sample size calculations for cluster-randomized trials

In cluster-randomized trials, randomization occurs at the cluster level, so power depends strongly on ICC because ICC affects the effective variance of outcomes at the cluster level. Sample size formulas typically incorporate design effect adjustments or variance inflation for the outcome mean. Analysts must also consider the allocation ratio of clusters to arms, expected baseline outcome variation, and anticipated attrition.

6.3 Cluster number versus cluster size trade-offs

Because ICC reduces the independence of individuals within a cluster, increasing cluster size often yields diminishing returns once within-cluster dependence is substantial. Conversely, increasing the number of clusters can improve precision more effectively because it adds independent sources of information. This trade-off is central to pragmatic trial design decisions.

6.4 Planning for anticipated ICC and robustness checks

Design-stage ICC values are usually derived from prior studies, pilot data, or expert judgment. Since uncertainty in ICC can materially affect power calculations, best practice often includes conducting sensitivity analyses across plausible ICC scenarios. Robust planning may also examine alternative variance structures or outcome transformations if the anticipated ICC seems inconsistent with the expected data-generating process.

7 ICC and power calculations

7.1 Effect of ICC on statistical power

Higher ICC increases the variance of cluster-level estimators and reduces effective sample size, generally lowering statistical power for a fixed design. In other words, the same nominal number of individuals produces less information when outcomes are more correlated within clusters. Therefore, ICC is among the most influential inputs to power calculations in clustered designs.

7.2 Implications for variance of estimators

Many estimators of treatment effects in cluster-randomized settings rely on cluster means or residualized outcomes. ICC affects the variance of these cluster means by governing how much of the total variability is shared within clusters. As ICC rises, the between-cluster variation relative to within-cluster variation shifts, widening confidence intervals and decreasing test sensitivity.

7.3 Power under different intra-cluster correlation structures

Power can be sensitive not only to the magnitude of ICC but also to the assumed form of dependence. For example, random-intercept models imply constant correlation between any two individuals in the same cluster, while repeated-measures contexts may imply correlation that decays with time lag. Misspecifying the dependence structure can lead to inaccurate power assessments, particularly when correlation patterns differ across subgroups or over time.

8 ICC in model selection and diagnostics

8.1 Checking whether clustering is warranted

A practical diagnostic question is whether a clustered model is needed. ICC-based considerations can guide this decision: if the estimated cluster variance is effectively zero (yielding ICC near 0) and model fit does not improve, a simpler model may be adequate. However, because ICC estimation can be noisy with few clusters, diagnostics should not rely solely on a point estimate.

8.2 Comparing ICC across candidate models

When comparing models with different random-effects structures or covariate sets, analysts may compare ICC to understand how much residual clustering remains. A reduction in ICC after adding covariates can indicate that between-cluster heterogeneity has been partially explained. Yet model selection should be based on overall fit and predictive performance, not only on ICC magnitude.

8.3 Residual dependence and model adequacy

If a fitted model leaves structured residual patterns within clusters, residual dependence can persist even if the model includes random effects. Diagnostics may include residual plots by cluster, checks for remaining intracluster correlation, or testing for additional random effects. Persistent dependence suggests that the assumed random-effects specification may be incomplete.

8.4 Overdispersion versus clustering distinctions

Overdispersion occurs when observed variance exceeds what a baseline count model (e.g., Poisson) assumes. Clustering is one source of extra variability, but it is not the only one. Differentiating whether extra variability is driven by within-cluster correlation, between-cluster heterogeneity, or model misfit is important, because remedies differ: adding random effects addresses clustering, while alternative marginal variance assumptions may address overdispersion.

The design effect connects ICC to the inflation of variance for estimates. The variance partition coefficient is another label often used for the same general idea: the proportion of variance attributable to a particular level (e.g., cluster level) under a specified model. While terminology varies across fields, both measures express how much of total variability comes from between-cluster differences.

9.2 Comparison with inter-class correlation terminology

“Inter-class correlation” is used in some contexts to refer to correlation between units that belong to different groups, though usage can differ by discipline. In many statistical workflows, “intra-class correlation” is the standard term for within-group similarity. Confusion can arise when authors use the phrase interchangeably, so the definition tied to the paper’s model is critical.

9.3 Marginal versus conditional intra-cluster correlation

Conditional ICC describes dependence given random effects; marginal ICC describes dependence after integrating over the random effects distribution, often corresponding to an average correlation in the population. For nonlinear mixed models (binary and count outcomes), conditional and marginal ICC can differ substantially. Specifying whether ICC is marginal or conditional improves interpretability.

9.4 Latent variable interpretations

For generalized outcomes, ICC can be interpreted using latent-variable formulations in which a random intercept shifts an unobserved propensity or rate. Under this view, ICC measures the fraction of variance in the latent propensity attributable to differences between clusters. This latent interpretation motivates scale-dependent ICC definitions and clarifies why ICC is not a single universal number across all modeling choices.

10 Reporting and best practices

10.1 What to report: model, definition, and estimator

Good reporting includes the clustering unit, the outcome type, the statistical model used, and the ICC definition tied to that model (especially for non-Gaussian outcomes). Analysts should also name the estimation method or software approach and provide the point estimate and uncertainty measures. Without these details, ICC values may not be replicable or comparable.

10.2 Reporting ICC with outcome-specific details

ICC is outcome-dependent: different endpoints (e.g., symptom score versus binary diagnosis) can yield different clustering patterns. Reporting ICC separately for each outcome, along with any transformations, link functions, or scaling choices, helps readers understand how ICC relates to the substantive measures used in the study.

10.3 Transparency about clustering level choices

Cluster definitions affect ICC. For example, defining clusters at the hospital level versus the clinic level can change both the magnitude and interpretation of dependence. Best practice is to justify the chosen clustering level based on the data structure and assumed dependence mechanism, and to discuss sensitivity to alternative reasonable clusterings when feasible.

10.4 Common pitfalls and misuse of ICC

Common errors include treating ICC as a universal constant, comparing ICC values across incompatible definitions or models, and using ICC estimates from one context to power another without sensitivity analysis. Another pitfall is reporting a cluster-level variance component without clarifying the associated ICC definition, particularly in GLMMs. Because ICC influences design effect and power, careful alignment between estimation and planning definitions is essential.