Adjusting For Intracluster Correlation

Why the Intracluster Correlation Matters

In a cluster-randomized trial, groups such as clinics, schools, or communities—not individuals—are assigned to treatment. Outcomes from individuals in the same group tend to be more alike than outcomes from different groups. This within-cluster dependence means that 200 people spread across a small number of clinics generally provide less information than 200 independently randomized people. The intracluster correlation coefficient (ICC) quantifies this dependence and affects the sample size needed to achieve a target precision or power.

An ICC is often uncertain when a trial is planned. A prespecified interim re-estimation can use accumulating trial data to update the ICC and revise the number of clusters required. The objective here is to update the nuisance parameter that governs clustering, not to select a favorable treatment effect estimate. The individual-level target is treated as fixed, and the updated ICC is used to translate it into a cluster-randomized target.

Definition: For a continuous outcome, the intracluster correlation coefficient $\rho$ is the correlation between the outcomes of two distinct individuals in the same cluster, under the assumed common-variance model. If the individual outcome variance is $\sigma^2$, the covariance for two individuals in the same cluster is $\rho\sigma^2$. The ICC is dimensionless; it is not the proportion of individuals who have similar outcomes.

The basic calculation assumes clusters are independent of one another, individuals within a cluster share a common ICC, and the outcome variance and ICC are adequately represented by the model used for planning. Treatment allocation is at the cluster level. In a balanced design, clusters are allocated equally between the two arms. These assumptions should be stated because the resulting design-effect calculation is an approximation, not a universal correction for every clustered analysis.

From an ICC Estimate to a Cluster Target

Suppose an individually randomized design would require $n_{\mathrm{ind}}$ evaluable individuals per arm to meet the prespecified power and significance requirements. If each cluster contributes $m$ evaluable individuals and cluster sizes are equal, the variance inflation relative to independent observations is the design effect:

DE=1+(m-1)\rho

A first-order approximation to the number of evaluable individuals required per arm is $n_{\mathrm{ind}}DE$. Dividing by the evaluable cluster size gives the required number of clusters per arm. Since clusters are indivisible, round the result upward.

K_{\mathrm{per\ arm}} = \left\lceil \frac{n_{\mathrm{ind}}\{1+(m-1)\rho\}}{m} \right\rceil

This formula keeps the individual-level target fixed and updates the clustering adjustment. It does not mean that the required number of individuals grows by the ICC alone. The increase depends on both $\rho$ and the number of individuals per cluster: when $\rho$ is positive, adding more people to an existing cluster gives diminishing information compared with adding a new independent cluster.

Justification: For $K$ equal-sized clusters with $m$ individuals each, the variance of the overall mean is the variance of the sum of all outcomes divided by $(Km)^2$. There are $Km$ individual variances, each $\sigma^2$, and $K m(m-1)$ ordered within-cluster covariances, each $\rho\sigma^2$. Thus $$ \operatorname{Var}(\bar{Y}) = \frac{Km\sigma^2+Km(m-1)\rho\sigma^2}{(Km)^2} = \frac{\sigma^2}{Km}\{1+(m-1)\rho\}. $$ Compared with independent individuals, whose variance is $\sigma^2/(Km)$, the variance is multiplied by $1+(m-1)\rho$. This yields the equal-size design effect.

Worked Example: Updating the Target With an Interim ICC

Suppose an individually randomized calculation gives a target of 131.4 evaluable individuals per arm. At an interim review, the estimated ICC is $\hat{\rho}=0.06532$. Assume 20 evaluable individuals per cluster, equal cluster sizes, and balanced cluster allocation. Find the updated target.

First substitute the interim ICC and cluster size into the design effect:

DE=1+(20-1)(0.06532) =1+19(0.06532) =2.24108 \approx 2.241

Using the displayed rounded design effect, the approximate evaluable individual target per arm is:

131.4(2.241)=294.4674\approx 294.47

At 20 evaluable individuals per cluster, the number of clusters required per arm is $294.4674/20=14.72337$, which must be rounded upward to 15. Thus the updated target is 15 clusters per arm, or 30 clusters total. Those clusters would contribute 300 evaluable individuals per arm if all reached the assumed size. The target exceeds the unrounded requirement; rounding down to 14 clusters would provide only 280 evaluable individuals per arm under these assumptions.

The calculation does not prove that the trial will have exactly its planned power. It translates the individual-level target using an estimated ICC and an assumed cluster size. The final analysis, degrees of freedom, and any small-sample correction must remain consistent with the trial’s prespecified design.

Unequal Cluster Sizes and Attrition

Equal cluster sizes are convenient but often unrealistic. A commonly used approximate design effect for variable cluster sizes is:

DE_{\mathrm{unequal}} \approx 1+\left\{(1+CV^2)\bar{m}-1\right\}\rho

Here $\bar{m}$ is the mean number of evaluable individuals per cluster and $CV$ is the coefficient of variation of cluster size, defined as the standard deviation of cluster size divided by its mean. This approximation assumes that cluster sizes are not related to outcomes or treatment effects and that the distribution of sizes is reasonably represented by its mean and coefficient of variation. When $CV=0$, it reduces to the equal-size formula. Greater size variability generally increases the design effect for positive $\rho$, because a large share of participants may be concentrated in a smaller number of clusters.

Attrition needs to be handled on the scale on which it occurs. If a fraction $a$ of individuals is expected to have no evaluable outcome, a simple planning approximation is $\bar{m}_{\mathrm{eval}}=\bar{m}_{\mathrm{enroll}}(1-a)$. Use the expected evaluable cluster size in the design effect and divide the inflated individual target by that evaluable size. This assumes individual outcome loss is adequately represented by the stated rate and does not selectively remove particular clusters or outcome types.

Worked Example: Unequal Cluster Sizes With Individual Attrition

A study needs an individually randomized target of 180 evaluable individuals per arm. Clusters are expected to enroll an average of 24 people, with cluster-size coefficient of variation $CV=0.25$. Individual attrition is expected to be 10%, and the interim ICC estimate is 0.04. Assume balanced allocation and no whole-cluster loss. Estimate the number of clusters needed per arm.

The expected evaluable cluster size is:

\bar{m}_{\mathrm{eval}}=24(1-0.10)=21.6

Using $CV^2=0.25^2=0.0625$, the unequal-size design effect is:

DE_{\mathrm{unequal}} \approx 1+\{(1+0.0625)(21.6)-1\}(0.04) =1+(22.95-1)(0.04) =1.878

The approximate inflated individual target is $180(1.878)=338.04$ evaluable individuals per arm. Dividing by the expected evaluable cluster size gives $338.04/21.6=15.65$. The study should therefore plan for 16 clusters per arm, or 32 clusters total, under these assumptions. The expected mean number of evaluable participants at that cluster count is $16(21.6)=345.6$ per arm, above the approximate target.

The fractional mean of 21.6 does not imply that every cluster will contain exactly 21.6 evaluable people. It is a planning average. The unequal-size approximation accounts for variation through $CV$; if cluster sizes vary substantially or are informative, a more detailed design calculation or simulation is preferable.

Loss of an entire cluster is different from individual attrition. If the probability that an enrolled cluster remains available for the final analysis is $q$, then a target of $K$ analyzable clusters per arm requires approximately $K/q$ clusters to be enrolled per arm, rounded upward. For example, if 16 analyzable clusters are needed and 90% of clusters are expected to remain, $16/0.90=17.78$, so plan 18 enrolled clusters per arm. Do not treat this as individual attrition: losing a cluster also removes its independent contribution to the comparison.

Worked Example: Adding Clusters After an Interim Update

Suppose the interim update gives a requirement of 17 clusters per arm after accounting for the revised ICC and expected individual attrition. Six clusters per arm have already been randomized, and all 12 remain eligible to contribute to the final analysis. Assume no planned cluster replacement and no further whole-cluster loss. How many additional clusters are needed?

For each arm, subtract the clusters already randomized from the updated final target:

$$ 17-6=11 $$

Thus, 11 additional clusters per arm are required, for 22 additional clusters total. The final randomized total will be 34 clusters. This subtraction is appropriate only because the six existing clusters are expected to remain in the analysis and the updated target is a final total. If some interim clusters are lost, or if the re-estimation target is defined on a different analysis population, the count must be adjusted to match that definition.

Interim Estimation, Diagnostics, and Limitations

A practical interim procedure should be specified before unblinded treatment comparisons are examined. The protocol or statistical analysis plan should state when the ICC will be estimated, which participants and outcomes enter that estimate, how the estimate is converted to a target, any minimum or maximum cluster count, and how already randomized clusters are counted. A blinded or treatment-pooled estimate can reduce the risk that sample-size adaptation is driven by observed treatment differences, although a pooled estimate can still be affected by treatment heterogeneity or imbalance.

An ICC estimate from a small number of clusters can be unstable. Check the number of contributing clusters, cluster-size distribution, missing outcomes, influential clusters, and whether the estimated ICC is compatible with the data and the assumed model. ICC estimates near zero may be imprecise rather than definitive evidence that clustering is absent. For binary outcomes, the ICC can depend on the outcome prevalence and model scale; an estimate from one population or time point may not transport directly to another. An estimate outside the model’s admissible range should prompt investigation of the estimation method and model rather than mechanical substitution.

The design-effect approach is an approximation. It does not by itself account for the loss of degrees of freedom when only a modest number of clusters are randomized, unequal allocation of clusters, stratification, covariate adjustment, or complex cluster-level analyses. In particular, a large number of individuals in only a few clusters does not replace having enough independently randomized clusters. If the planned analysis uses a small-sample correction or cluster-level degrees of freedom, the original individual-level target may need additional allowance. Recompute under the planned analysis or use simulation when the approximation is doubtful.

An adaptive sample-size rule can also affect the operating characteristics of the final test. Use a prespecified re-estimation rule and ensure the final inferential procedure remains appropriate for that rule, consistent with the earlier material on maintaining a fixed final test after re-estimation. Do not assume that updating a nuisance parameter automatically preserves the intended type I error or power. That claim depends on the interim information used and the final testing procedure.

Common Mistakes

Using the ICC without cluster size. The design effect is $1+(m-1)\rho$, not simply $1+\rho$. The same ICC has a larger impact when each cluster contains more individuals.
Rounding the cluster count down. A requirement of 15.1 clusters per arm means 16 clusters per arm. Rounding down can leave the trial below its calculated target.
Inflating for attrition twice. If the target is already expressed in evaluable individuals and the calculation uses the expected evaluable cluster size, do not apply a second individual-level attrition multiplier.
Treating people as independent clusters. The relevant independent randomized units are clusters. Many individuals in a few clusters can still yield limited information.
Interpreting an interim ICC as a treatment effect. The ICC describes within-cluster similarity, not the size or direction of a treatment benefit.
Assuming the approximation is exact. The design effect translates variance under specific assumptions; few clusters, very unequal sizes, informative attrition, or a different final analysis may require a more tailored calculation.

Key takeaway: Estimate the ICC using a prespecified interim rule, calculate the design effect using the expected evaluable cluster size (and a variability adjustment when appropriate), then convert the inflated individual target to clusters and round upward. Account separately for individual attrition and whole-cluster loss, and verify that the final analysis is compatible with the approximation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

An equal-size trial expects 12 evaluable people per cluster and uses an ICC estimate of 0.05. What is the design effect?
An individually randomized target is 100 evaluable people per arm. If the design effect is 1.55 and each cluster contributes 10 evaluable people, how many clusters are required per arm?
Why can individual attrition and whole-cluster loss not be handled as the same adjustment?
If cluster sizes are equal, what value does the unequal-size design-effect approximation use for $CV$, and what does it reduce to?

Solutions. (1) $DE=1+(12-1)(0.05)=1.55$. (2) The inflated target is $100(1.55)=155$; $155/10=15.5$, so round upward to 16 clusters per arm. (3) Individual attrition reduces the number of evaluable people within retained clusters, whereas whole-cluster loss removes a randomized unit and all its participants. The corresponding adjustments act on different counts. (4) Equal sizes give $CV=0$, so $DE_{\mathrm{unequal}}\approx1+\{(1+0)\bar m-1\}\rho=1+(\bar m-1)\rho$, the equal-size formula.