Abstract:
Matching estimators exhibit a previously overlooked source of finite sample bias. Even when the assumptions underlying matching hold, the local geometry of the covariate distribution around each observation makes some matching directions more likely than others. This mismatch is ubiquitous and disappears only slowly as the sample size increases. Estimators suffer increased variance and, in most cases where the outcome function is nonlinear, systematic bias. Conventional balance diagnostics generally fail to detect the problem. When they identify it, matching is abandoned unnecessarily. I propose three approaches to mitigate this problem and evaluate them in simulations and empirical applications. A matching procedure that deliberately offsets mismatch across neighboring observations reduces both bias and variance while remaining conceptually and computationally simple.
