Cluster Randomized Trial

A cluster randomized trial is a type of randomized experiment in which randomization occurs at the group level rather than at the individual level. Entire clusters, such as communities, schools, hospitals, primary care practices, or workplaces, are randomly assigned to different interventions, and all individuals within each cluster receive the same intervention.

Cluster randomization is used when the intervention operates at the group level (such as a public health campaign, an organizational policy change, or a school based curriculum), when individual randomization would result in contamination between treatment groups (such as when participants in the same workplace could share information about the intervention), or when logistical and administrative considerations make individual randomization impractical. Cluster randomized trials are widely used in public health, health services research, education, and implementation science. However, they require special statistical methods to account for the correlation among observations within clusters, and failure to use these methods produces incorrect standard errors and misleading p values.

In a cluster randomized trial, the researcher identifies the clusters, randomly assigns each cluster to a treatment or control condition, and then measures outcomes on individual participants within each cluster. The key statistical consideration is the intracluster correlation coefficient (ICC), which measures the degree to which outcomes for individuals within the same cluster are correlated. Even a small ICC can substantially reduce the effective sample size of the study, because individuals within the same cluster do not provide fully independent information.

The design effect quantifies the inflation in variance due to clustering and is calculated as 1 plus (m minus 1) times the ICC, where m is the average cluster size. For example, with an ICC of 0.05 and a cluster size of 50, the design effect is 3.45, meaning the study requires 3.45 times as many participants as an individually randomized trial of the same statistical power. This inflation has major implications for sample size planning and makes cluster randomized trials less efficient than individual randomization when both are feasible.

Analysis of cluster randomized trials must account for the clustered structure of the data. Common approaches include mixed effects models (also called multilevel or hierarchical models) that include random effects for clusters, generalized estimating equations (GEE) that adjust the standard errors for within cluster correlation, and cluster level analysis that uses summary measures (such as cluster means) as the unit of analysis. Individual level analysis that ignores clustering produces standard errors that are too small, p values that are too low, and confidence intervals that are too narrow, leading to inflated Type I error rates.

Cluster randomization is appropriate when the intervention is delivered at the group level, when individual randomization would cause contamination between groups, when administrative or logistical constraints prevent individual randomization, or when the research question concerns a system level intervention. It is not appropriate when individual randomization is feasible and contamination is unlikely, because cluster randomization is less statistically efficient and requires larger sample sizes.

A common and consequential mistake is analyzing a cluster randomized trial as though it were individually randomized, ignoring the correlation among observations within clusters. This produces misleadingly small standard errors and inflated significance levels. All analyses must account for clustering.

Another frequent error is underestimating the sample size required for a cluster randomized trial. Researchers sometimes calculate the sample size for an individually randomized trial and then simply assign clusters without adjusting for the design effect, resulting in an underpowered study. The ICC used in the sample size calculation should be based on prior data from similar settings; using a default value without empirical support can lead to substantial miscalculation.

Researchers also sometimes randomize too few clusters. With few clusters, the ICC is estimated imprecisely, chance imbalances between arms are more likely, and mixed effects models and GEE can understate uncertainty unless small sample corrections are used. With very few clusters, even appropriate statistical methods may produce unreliable results. Finally, failing to report the ICC in the published paper prevents other researchers from using it for their own sample size calculations, slowing the accumulation of design parameters across the field.

A health services research team evaluates the effect of a pharmacist led medication review program on hospital readmission rates. Twenty four primary care practices are randomized, with 12 practices assigned to the intervention (pharmacist reviews all patients' medications upon hospital discharge) and 12 to usual care. Over 12 months, 3,600 patients are enrolled (average 150 per practice). The ICC for 30 day readmission is 0.01, yielding a design effect of 2.49. A mixed effects logistic regression model accounting for clustering finds that the readmission rate is 11.2% in the intervention practices and 15.8% in the usual care practices (adjusted odds ratio: 0.67, 95% CI: 0.49 to 0.91, p = 0.01). The researchers report the ICC to support future sample size calculations.