A comparison and evaluation of recent developments for multivariate matching in observational studies

Xing (Sam) Gu, University of Pennsylvania


A simulation study is conducted to compare and evaluate recent developments for multivariate matching in observational studies. The simulation answers the following four questions from the literature on matching. (1) In current statistical practice, matched samples are formed using "nearest available" or greedy matching. Indeed, all past simulation studies have used this method exclusively, and it appears to be the only method actually in use in observational studies. Greedy matching does not minimize the total distance within matching pairs, though good algorithms do exist for optimal matching that do minimize the total distance. How much better is optimal matching than greedy matching? The simulation shows that optimal matching is noticeably better than greedy matching in the sense of producing closely matched pairs, but it is no better than greedy in the sense of producing balanced matched samples. (2) Multivariate matching involves defining a distance between covariate vectors, and several such distances exist. The simulation compares three recent proposals. Though no one method is always best, the Mahalanobis metric within propensity calipers is never bad. (3) In common practice, each treated unit is matching with one control, called 1-1 or paired matching, or with two controls, called 1-2 matching, and so on. It is known, however, that the optimal structure is a full matching in which a treated unit many have one or more controls or a control may have one or more treated units. The simulation compares optimal 1-k matching with optimal full matching, finding that optimal full matching is often much better. (4) As k increases, 1-k matching uses more of the available data but also uses more of the distant controls. The simulation looks at the rate of decline in match quality as k increases. A useful rule of thumb is that the percent reduction in bias falls by about 10% when k is increased by one.

Subject Area


Recommended Citation

Gu, Xing (Sam), "A comparison and evaluation of recent developments for multivariate matching in observational studies" (1992). Dissertations available from ProQuest. AAI9308584.