Shrinkage for Improved Inference in Factorial Experiments

• Activity Number: 70
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Social Statistics Section
• Author(s): Cyrus Samii*+ and Joel Middleton and Peter M. Aronow
• Companies: Politics Department, New York University and Steinhardt School, New York University and Yale University
• Keywords: randomization inference ; skrinkage ; empirical Bayes ; experiments ; potential outcomes

Abstract

Factorial experiments identify interaction effects, which are crucial for understanding optimal treatment combinations. However, the multiplicity of effective treatment conditions introduces problems of multiple inference and possible sparsity. Shrinkage estimation provides a potential solution to both of these problems. We consider the properties of shrinkage estimators for factorial experiments under the randomization-based mode of inference (Neyman 1923). We characterize bias and precision trade-offs and make recommendations for applied work.