|
Abstract:
|
Matching design is a popular strategy to inferring causal relationship in observational studies, since it is less model dependent and provides interpretation similar to randomized experiments. When there are two treatment arms, optimal algorithms exist for both bipartite and nonbipartite matching. Few studies, however, have used matching designs with more than two treatment groups. This is partly due to the complexity of multiple-group matching (referred as poly-matching) algorithms and to the lack of optimal results. We propose an iterative conditional matching algorithm that outperforms the nearest neighbor algorithm. The implementation of the algorithm is relatively easy for a small number of groups, i.e. 3-5. We also discuss the extension of Rosenbaum's sensitivity analysis for potential unmeasured confounding after poly-matching. We illustrate our method with national ED mortality data to evaluate trauma centers at different levels.
|
