Causal Analysis of a Random Coefficients Model in Multisite Randomized Trials

• Activity Number: 319
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Yongyun Shin* and Stephen W. Raudenbudsh
• Companies: Virginia Commonwealth University and The University of Chicago
• Keywords: Multisite trials ; random coefficients ; causal analysis ; unbiased estimation ; missing data

Abstract

In a multisite randomized trial, children within schools or sites are randomly assigned to a new program that is designed to improve a youth outcome. The overall treatment and control effects, their site-specific effects, and their variances and covariance are of interest. The site-specific effects help differentiate effective sites from ineffective ones. The variances and covariance reveal how effective the treatment is, and whether the treatment is more effective for advantaged or disadvantaged children. Because the treatment is randomly assigned to individuals within sites, site-specific treatment effect estimates are unbiased. However, if the site-specific propensity score is associated with the size of the site's impact, we will obtain biased estimators of the overall treatment effect, the variance of the site-specific treatment effects, and the covariance between the site-specific impact and the site-specific control group mean. We introduce a maximum likelihood method for unbiased estimation of the model and efficient handling of data missing at random.