Causal Analysis of Observational Data with Gaussian Process Potential Outcome Models

• Activity Number: 510
• Type: Invited
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Surya T Tokdar*+
• Companies: Duke University
• Keywords: Causal effects ; Gaussian process ; Covariate imbalance ; Propensity score matching ; Nonparametric modeling ; Insufficient overlap

Abstract

We advocate causal analysis of observational data with nonparametric Gaussian process (GP) models on the potential outcomes. We illustrate how this approach overcomes several limitations of the hugely popular current practice of propensity score matching, including scalability with covariate dimension, causal interpretability in case of insufficient overlap and generalization to continuous treatment. We demonstrate how nonparametric GP potential outcome methods avoid biased estimation of causal effects. Due to their ability to gather information locally, these methods are less likely to introduce bias and more likely to produce longer intervals in case of insufficient overlap between treatment groups. We also show that GP potential outcome methods do a much better job of identifying insufficient overlap than other popular nonparametric approaches.