|
Abstract:
|
Matching is among the oldest and most transparent methods of adjustment for observational studies. To extract treatment effects, it says, first match treated subjects to untreated but otherwise similar controls; then compare subjects to their matched counterparts and aggregate these individualized comparisons. Used with propensity scores, the method is simple and far-reaching, addressing high-dimensional covariates and helpfully separating outcome analysis from the process by which candidate matches are developed and evaluated. These benefits are shared with prognostic scores, cousins of propensity scores addressing important aspects of the data that propensities ignore. Combining propensity and prognostic scores enhances matching's scope and power as compared to matching on propensities only, while avoiding hazards to which earlier prognostic-matching proposals have succumbed.
|
