Mahalanobis balancing: a multivariate perspective on approximate covariate balancing
In the past decade, various exact balancing-based weighting methods were introduced to the causal inference literature. Exact balancing alleviates the extreme weight and model misspecification issues that may incur when one implements inverse probability weighting. It eliminates covariate imbalance by imposing balancing constraints in an optimization problem. The optimization problem can nevertheless be infeasible when there is bad overlap between the covariate distributions in the treated and control groups or when the covariates are high-dimensional. Recently, approximate balancing was proposed as an alternative balancing framework, which resolves the feasibility issue by using inequality moment constraints instead. However, it can be difficult to select the threshold parameters when the number of constraints is large. Moreover, moment constraints may not fully capture the discrepancy of covariate distributions. In this paper, we propose Mahalanobis balancing, which approximately balances covariate distributions from a multivariate perspective. We use a quadratic constraint to control overall imbalance with a single threshold parameter, which can be tuned by a simple selection procedure. We show that the dual problem of Mahalanobis balancing is an l_2 norm-based regularized regression problem, and establish interesting connection to propensity score models. We further generalize Mahalanobis balancing to the high-dimensional scenario. We derive asymptotic properties and make extensive comparisons with existing balancing methods in the numerical studies.
READ FULL TEXT