Abstract:
|
Propensity score weighting, as opposed to matching or stratification, is gaining popularity as a method to adjust for imbalance in baseline covariates in observational trials. In a common approach, a sampling weight for each subject is calculated as the inverse of the propensity for receiving the treatment actually received. At the same time, matching by prognostic score has been shown to be a useful alternative to propensity score methods in certain settings, but weighting methods have not been well developed for prognostic scores. Where propensity score methods place emphasis on the variables that are most imbalanced between exposures, prognostic scores emphasize the variables that are most strongly associated with the outcome. We show how the mechanism for the new bagged one-to-one matching (BOOM) estimator can be used to create weights based on prognostic scores, hybrid weights based on prognostic and propensity scores, and weights derived nonparametrically via the Reweighted Mahalanobis Distance. We compare the performance of the BOOM weights to that of traditional and alternative weighting approaches.
|