Comparison of Three Methods for Dual Sensitivity Analysis

• Activity Number: 609
• Type: Contributed
• Date/Time: Wednesday, August 7, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Social Statistics Section
• Author(s): Masataka Harada*+ and Jennifer Hill and Nicole Carnegie
• Companies: New York University and New York University and Harvard University
• Keywords: causal inference ; sensitivity analysis

Abstract

A growing literature explores methods for assessing the sensitivity of inferences to violations of the ignorability (selection on observables) assumption, typically with respect to one unobserved confounder. One set of such strategies allows the researcher to explore how inferences change based on likely combinations of two sensitivity parameters that reflect conditional associations of the outcome and treatment with the unobserved confounder. The methods that have been proposed to assess this sensitivity vary in their parametric assumptions, operationalization of the sensitivity parameters, and level of computational intensity.

In this study, we compare the three methods, namely likelihood-based approach by Imbens (2003), computational approach by Harada (2012) and the hybrid approach between algebraic and computational approaches by Carnegie and Hill (2013). We first demonstrate the general advantage of dual sensitivity analysis over sensitivity analysis with a single parameter. We then contrast the three methods, exploring their strengths and weakness through simulation and real data analysis, and providing practical guidance for the researcher for choosing among them.