A Paradoxical Result in Regression Estimation Arising from Constructing Moderators of Treatment Response

• Activity Number: 638
• Type: Topic Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Thaddeus Tarpey*+ and R. Todd Ogden and Eva Petkova
• Companies: Wright State University and Columbia University and New York University School of Medicine
• Keywords: biosignature ; least-squares ; randomized experiments ; regression

Abstract

This talks presents a counter-intuitive results that arose from work on developing biosignatures/moderators defined as a linear combination of baseline predictors. In randomized experiments, baseline predictors should have identical distributions across treatments in which case it seems natural to combine the predictors to estimate common parameters. Paradoxically, the precision of a slope estimator can deteriorate when pooling baseline predictors across treatments. This phenomenon will be illustrated by a moderator analysis example.