Mutual information-based neighbor selection method for causal effect estimation.

Estimation of causal effects from observational data has been the main objective in several high-impact scientific domains, while the golden standard for calculating the true causal effect is through the conduction of randomized controlled trials. The abundance of this type of data, which are continuously produced and collected, makes them potentially valuable for estimating causal effects. However, observational data may lead to erroneous treatment effect estimation, since they often suffer from various forms of bias. A recent work has shown that causal effect estimation provided by neural network-based model can be improved by leveraging information from the outcome of neighboring instances in the covariate space. In this work, we propose an information-theoretic methodology for selecting the neighbors to be considered in the estimation of the treatment effect for each sample. The proposed methodology, named Mutual Information-based Neighbor selection for Treatment effect estimation (MINT), selects the optimal number of neighbors as well as the type of distance metric with respect to pre-defined criteria. Then, the average outcome of the neighbors in the treatment and control groups is used as an informative input to the estimator, in addition to the covariates. The presented numerical experiments demonstrate that the adoption of the proposed MINT methodology with the state-of-the-art Dragonnet model is able to develop a reliable and accurate model for treatment effect estimation. [ABSTRACT FROM AUTHOR]