Switching Cluster Membership in Cluster Randomized Control Trials: Implications for Design and Analysis

• In Event: 28.030 - Analysis in Experimental Designs and Quasi-Experimental Designs With Propensity Score Matching

Abstract

Randomized control trials (RCTs) in education often use clustered designs, where intact clusters (such as classrooms or schools) are randomly assigned to treatment and control conditions, or where individuals are randomized and subsequently grouped into clusters to receive. Little guidance is available for researchers when clustered study designs are ‘broken’ or there is noncompliance with the randomization protocol. The goal of this paper is to explore the challenges presented by one particular form of noncompliance that is unique to RCTs with clustering: cluster switching. We demonstrate that both hierarchical linear models and cluster robust standard errors can provide inappropriate inferences, depending on the sources of random variation. We propose a potential solution to these analytic challenges.

Authors

• Jonathan Schweig, The RAND Corporation

• John Pane, RAND Corporation

• Daniel F. McCaffrey, ETS