Pairwise cluster randomization: an exposition.

Area of Science:

• Statistics
• Program Evaluation Methodology

Background:

• Cluster randomization (CR) is commonly employed for program evaluation when simple random assignment is not feasible.
• Pairwise cluster randomization (PCR) presents a more efficient design but faces challenges related to bias and identification.

Purpose of the Study:

• To explain the bias and identification problems associated with PCR.
• To demonstrate how design choices can mitigate these issues.
• To show how Monte Carlo procedures can test the suitability of PCR.

Main Methods:

• Presentation of formulas illustrating PCR estimator bias and standard error identification issues.
• Discussion of design strategies to address bias and identification.
• Utilization of Monte Carlo simulations to compare CR and PCR suitability.

Main Results:

• Formulas derived to quantify PCR bias and identify standard error issues.
• Demonstration that specific design choices can effectively mitigate identified problems.
• Monte Carlo simulations provide a method for selecting between CR and PCR at the design stage.

Conclusions:

• Advocates for increased adoption of the PCR design.
• PCR is most suitable for evaluations with a limited number of clusters and available baseline data.
• PCR is a compelling alternative to CR, especially when dealing with approximately 26 clusters.