Assessing Sensitivity to Exclusion Restrictions in Semiparametric Instrumental Variables Estimators of Treatment Effects

• Activity Number: 71
• Type: Invited
• Date/Time: Monday, August 4, 2003 : 8:30 AM to 10:20 AM
• Sponsor: IMS
• Author(s): Joseph W. Hogan*+ and Jason A. Roy and Donald Alderson and Tony Lancaster
• Companies: Brown University and Brown University and Brown University and Brown University
• Keywords: causal inference ; selection bias ; potential outcomes ; time-dependent confounding ; time-varying treatment ; HIV/AIDS

Abstract

The method of instrumental variables can be used to obtain consistent estimates of a causal treatment effect from observational data. The method relies on using a valid instrument, a variable that is correlated with receipt of treatment, and conditional on treatment, uncorrelated with underlying potential outcomes. The second requirement is sometimes called an 'exclusion restriction' and cannot be checked empirically. We describe a method for assessing the range of large-sample bias in the estimated treatment effect when the exclusion restriction is violated. We consider the case of binary treatment and binary instrument, which illuminates key points related to whether homogeneous or heterogeneous treatment effects are assumed. In many approaches to this problem, the sensitivity parameter is a function of correlations. The parameters in our proposal have physical interpretations in terms of the outcome variable (e.g., differences in mean CD4). This provides intuition about the exclusion restriction, and facilitates discussion with subject matter experts about whether deviations from it. The method is applied to data from a longitudinal cohort study of HIV-infected women.