Abstract
Matching is a method to estimate Average Treatment Effect (ATE) in observational studies. In order to reduce bias due to confounding, an optimal matching algorithm pairs treated and control units with similar characteristics across the covariates. Traditionally, matching methods were developed using distance-based measures, such as propensity scores or Mahalanobis distances. We propose a new method, called Balancing Recursive Partitioning (BRP) which directly optimizes for local regions of covariate balance using a recursive partitioning strategy that uses a multidimensional splitting criterion aimed at balancing the distribution of covariates between the two groups. From the resulting balancing tree, a proximity matrix can be used to weight observations and to identify a common support for treated and control units, resulting in a new estimator of ATE. In the case of the famous Lalonde datasets from causal inference world, we show that BRP has prominent advantages over other methods in estimating Average Treatment Effect.
Health disparity has been a crucial problem in the society. There are many literatures to reveal the health disparity. Yet very little research is to address and reduce the disparity. We develop a method called Distorted Variable Analysis (DVA). Which is a tree-based method to identify the race disparity in different social levels. We are able to detect different amounts of disparity and distort the modifiable risk factors to move the individual from the high disparity group to the low disparity group. We apply the DVA to the cardiovascular disease patients and achieve a significant reduction in racial disparity.