Inclusion/Exclusion Criteria for Bias-Amplifying Covariates: A Sensitivity Analysis Framework

• Activity Number: 609
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Social Statistics Section
• Author(s): Marc Scott and Joel Middleton*+
• Companies: New York University and Steinhardt School, New York University
• Keywords: Sensitivity Analysis ; Causal Inference ; Bias-Amplification ; Fixed Effects ; Random Effects ; Multilevel Models

Abstract

We develop a framework for assessing the impact of potentially bias-amplifying covariates. Pearl (2010) and others have shown conditions under which including a covariate reduces or inflates the bias due to an omitted confounder. In practice, however, deciding whether or not to include such a covariate is complicated by the fact that the properties of the omitted confounder are not generally known. Sensitivity analysis (Rosenbaum and Rubin, 1983; Imbens, 2003) evaluates the effects of omitted confounders with different properties on bias. Using a sensitivity analysis framework, we will illustrate a method that allows the analyst to identify the properties of covariates in conjunction with omitted confounders that lead to more and less bias amplification. We will explore this issue in the context of the impact of fixed and random effects for group structure in multilevel models as well.