A Data-Driven Approach to Finding K for K Nearest Neighbor Matching in Average Causal Effect Estimation.

In causal inference, a fundamental task is to estimate causal effects using observational data with confounding variables. K Nearest Neighbor Matching (K-NNM) is a commonly used method to address confounding bias. However, the traditional K-NNM method uses the same K value for all units, which may result in unacceptable performance in real-world applications. To address this issue, we propose a novel nearest-neighbor matching method called DK-NNM, which uses a data-driven approach to searching for the optimal K values for different units. DK-NNM first reconstructs a sparse coefficient matrix of all units via sparse representation learning for finding the optimal K value for each unit. Then, the joint propensity scores and prognostic scores are utilized to deal with high-dimensional covariates when performing K nearest-neighbor matching with the obtained K value for a unit. Extensive experiments are conducted on both semi-synthetic and real-world datasets, and the results demonstrate that the proposed DK-NNM method outperforms the state-of-the-art causal effect estimation methods in estimating average causal effects from observational data.