Matching on Generalized Propensity Scores with Continuous Exposures
share
Generalized propensity scores (GPS) are commonly used when estimating the causal effects of a continuous exposure in observational studies. Existing GPS approaches include: a) GPS adjustment in the outcome model; b) inverse probability of treatment weighting (IPTW); or c) doubly robust (DR) approaches. However, these approaches have the following limitations. First, either the GPS or the outcome model, or both, need to be correctly specified. Second, weighting approaches (IPTW and DR) are sensitive to extreme values of GPS. Third, assessing covariate balance is not straightforward. We propose an innovative approach for GPS caliper matching with continuous exposures. Under the local weak unconfoundedness assumption, we provide theoretical results of our proposed matching estimators and introduce new covariate balance measures under the matching framework. In the simulation, our proposed matching estimator outperforms existing methods under settings of model misspecification and/or in presence of extreme values of the estimated GPS in terms of bias reduction, root mean squared error, and transparency to assess covariate balance. We apply the proposed method to Medicare data in New England (2000-2012) to estimate the causal effect of long-term PM_2.5 exposures on mortality.
READ FULL TEXT
