Support Vector Machines as a Matching Method to Achieve Optimal Balance

• Activity Number: 648
• Type: Topic Contributed
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Marc Ratkovic*+
• Companies: Princeton University
• Keywords: propensity score ; balance ; support vector machines

Abstract

Existing matching methods match on some summary statistic-being in the same bin, or close in some metric (Mahalanobis, propensity score). This necessarily requires a compression of information and ad hoc decisions by the researcher. In this paper, I generate a method that simultaneously matches across all marginals, rather than a summary statistic. This allows for a selection of observations that creates a distribution across covariates that is independent of assignment to treatment. I do this through adapting support vector machines to the matching problem. I prove that observations with non-zero slack are balanced, in expectation, across all covariates. Through the kernel trick, I am able to extend the results to independence between treatment assignment and covariates, rather than simple uncorrelatedness. A comparison to extant methods reveals a far better balance.