Abstract:
|
When constructing a matched sample in an observational study, an investigator often has multiple, possibly conflicting goals in mind. For instance, the investigator might hope to balance covariates well, or to match closely on covariates within pairs, or to build as strong an instrumental variable as possible; these goals may pull in different directions. We give a quantitative description of the tradeoff be- tween two conflicting design goals, using Pareto optimality to define a spectrum of interesting solutions ranging between extreme focus on one goal to extreme focus on the other. In addition, we suggest a practical approach for exploring the tradeoff spectrum and selecting a match that balances design goals appropriately. We demonstrate this approach in a study of the impact of internationally-trained surgeons on patient health outcomes, in which we wish both to balance important covariates and to match closely on certain covariates in order to detect possible effect modification.
|