Although CRT designs confer certain advantages, they also have significant theoretical limitations and implementation challenges (Torgerson et al 2001). Careful consideration is needed before settling on a particular approach to randomization. The following assessment questions, adapted from Designing Multi-Center Cluster-Randomized Trials: An Introductory Toolkit (2014), may help clarify whether a CRT is an appropriate design choice for answering a particular research question:

• Is the phenomenon of interest something that takes place primarily at the level of the individual patient or study participant (for example, response to an experimental drug or comparator)? If so, a traditional randomized controlled trial design may be most appropriate. However, if the phenomenon of interest affects individual patients but is primarily taking place at a different level (for example, whether implementation of new physician treatment guidelines yields better patient outcomes), a CRT design may be more appropriate.
• Is the proposed intervention delivered at the level of a group or organization rather than an individual study participant? For example, a study that investigated whether adoption of a new treatment guideline affected the efficiency of service provision across hospitals in a health system would lend itself to a cluster design.
• If individual participants were randomized, would it be difficult for physicians or other clinical staff to modify their approaches or behaviors in ways that avoid contamination? Similarly, is it likely that participants or study staff might have occasion or opportunity to discuss details of the study (“compare notes”) among themselves? If so, a cluster randomized approach may be preferable for ensuring trial validity.

Finally, it is important to consider that any cluster randomized design will introduce an important statistical effect known as clustering. When several participants (a cluster) are subjected to similar circumstances that may differ from those of other clusters, such as patients in the same ward being treated by the same providers, their outcomes can be correlated. (See the Intraclass Correlation section of the Analysis Plan chapter of the Living Textbook). This important consideration should be addressed both when randomizing and when calculating the required sample size for a given study.

Case Example: The HiLo Trial

The Pragmatic Trial of Higher vs Lower Serum Phosphate Targets in Patients Undergoing Hemodialysis (HiLo), an NIH Collaboratory Trial, was designed to answer the question of what is the optimal level of serum phosphate for patients with end-stage renal disease who are undergoing hemodialysis. For operational reasons, the researchers designed HiLo as a cluster randomized trial, with randomization at the dialysis facility level. HiLo differed from many pragmatic trials in that the intervention was not consistent with the relevant clinical guidelines. Thus, because the intervention presented more than minimal risk for participants in the higher serum phosphate arm, the requirement for informed consent could not be altered or waived for this arm.

Cluster randomization combined with informed consent can be a significant challenge. When HiLo reached 10% of its enrollment target, the researchers observed an imbalance in willingness to participate between the 2 study arms. The dieticians participating in the study knew each patient’s group assignment, which made biased enrollment a concern. Although more than 500 patients had already been enrolled in the trial, the researchers decided to pivot from cluster randomization to individual-level randomization. Based on this experience, the HiLo researchers cautioned future pragmatic trial investigators against conducting randomized trials with postrandomization consent.

For more information about design considerations, see Designing With Implementation and Dissemination in Mind.