|
Abstract:
|
The use of the propensity score and propensity score matching for confounding adjustment in observational data is a common approach in the causal inference literature. All applications make the assumption of "ignorable treatment assignment" (also known as "no unobserved confounding") implicitly or explicitly, meaning that the observed covariates are sufficient to approximate a randomized experiment using observational data. The prospect of unobserved confounding generally threatens the validity of propensity score estimates. However, in settings of spatially-indexed data, the observed locations of the units may serve as a useful proxy for unobserved confounding that varies according to a spatial patterns. We consider such settings of spatially-indexed data and develop a new method for incorporating observations' spatial proximity into a propensity score matching procedure to adjust for both the observed and unobserved confounding. We apply all methods and compare the results using the EPA Air Markets Program Data (AMPD) to perform comparative effectiveness research of air quality policies.
|
