The Impact of Imperfect Matching on Causal Inferences

• Activity Number: 609
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Colin Fogarty*+ and Dylan Small
• Companies: Wharton School and University of Pennsylvania
• Keywords: Causal Inference ; Randomization Inference ; Sensitivity Analysis ; Double Robustness ; Covariate Adjustment ; Matching

Abstract

We investigate the validity of match based inference under imperfect alignment of assignment probabilities within matched strata. We demonstrate that in this highly realistic scenario, many of the standard arguments used to motivate the use of randomization inference in observational studies can falter. We present scenarios in which despite following guidelines adhered to by practitioners for evaluating the quality of a match, the standard randomization inference is biased conservatively or liberally. We begin by isolating the source of the problem. We next develop methodologies that are valid within the randomization inference framework to result in inference that is at worst conservative while sacrificing minimally in terms of power. Our methodology draws on techniques for sensitivity analyses and for post-match covariate adjustment, and can be thought of as "doubly robust" in that inference will be at worst conservative provided either the propensity score model or the model for potential outcomes under control is properly specified.