Improved Strategies Using the Propensity Score to Estimate the Effect of a Healthcare Policy

Accurate and precise effect estimation of healthcare utilization, spending, and general health attributable to a policy or program is necessary for making informed policy decisions. In the majority of healthcare service settings, random allocation to a particular policy or program group is unfeasible which in turn, often makes selection bias a concern. Natural experiments are increasingly being conducted to evaluate the effect of a healthcare policy or program. Difference-in-differences approach is popular in applied economics to evaluate the effect of a policy, and propensity score approaches from biostatistics aim at reducing selection bias. This dissertation focuses on propensity score weighting in difference-in-differences. Conjunct use of these two methods could allow one to estimate the effect of a policy in observational data while addressing the selection bias issue. Contemporary approaches have begun combining these methods. However, current approaches ignore the prior estimation of the propensity scores as well as the variability introduced in the formation of the propensity score weights. Failing to account for these additional sources of sampling variability can result in unacceptable precision. Bootstrap resampling of the weighted observations can capture the sampling distribution of the weighted observations in a difference-in-differences weighted regression. The resamples simulate the sampling distribution of the weighted observations, which in turn allows the analyst to capture both sources of unknown variability without estimating directly. Proposed methods include a bootstrap confidence interval correction to existing methods as well as an augmented inverse probability weighted difference-in-differences estimator. The proposed methods are applied to the Medical Expenditure Panel Survey data to estimate the effect of Medicaid expansion on medication acquisition, medication cost, and total medical charges. Two simulation studies were conducted, one using artificially simulated data and another with data simulated to represent the real world example. Bootstrap confidence intervals in place of model-based standard errors can correct the deficiency in precision of contemporary estimators.