マッチング・共変量バランス・因果効果推定バイアス: モンテカルロ・シミュレーションによる分析

抄録

　This paper investigates the effectiveness of matching techniques in improving covariate balance and reducing bias in estimating the effects of treatment through Monte Carlo simulation analysis. The results suggest that Propensity score matching (PSM), Mahalanobis distance matching (MDM), and Coarsened exact matching (CEM) are all effective in improving covariate balance, and that the "PSM paradox" pointed out by King and Nielsen (2019) is not observed in the relationship between the number of pruned observations and covariate balance. Rather, the results suggest that regardless of the matching method, improving covariate balance may lead to a paradoxical situation: beyond a certain point, the improvement of covariate balance no longer contributes to reducing bias in estimated effects of treatment. Also suggested is that even though the matching techniques cannot adjust for the bias of excluded variables and unmeasured confounders, when both are present, estimation after matching may reduce bias in estimated treatment effects better than OLS without matching.