|
Abstract:
|
Matching estimators, such as the mean difference in outcomes in a one-to-one matched cohort, can reduce bias due to unbalanced measured confounders in observational studies. They are easily understood, are unbiased and consistent under certain conditions, and offer increased robustness to model misspecification, measurement error, and some unobserved confounding. However, they often have a nontrivial loss in efficiency. We introduce the bagged one-to-one matching (BOOM) estimator, which combines the benefits of matching with the efficiency of bagging. In simulations with a continuous outcome, we examine the BOOM estimator's performance in mean squared error, bias, and variance, as well as accuracy of standard error estimation and coverage of nominal 95% confidence intervals using Efron's standard error for bagged estimators. We compare the performance to that of ordinary least squares estimation on the entire sample, one-to-one matching, and inverse probability weighting under different types of model misspecification. In our simulations, the BOOM estimator achieves the bias reduction and robustness of one-to-one matching, while having much lower variance and mean squared error.
|
