**Summary:** This paper investigates the validity of the bootstrap method for inference in matching estimators, which are widely used to estimate treatment effects. The authors show that the bootstrap is generally not valid for matching estimators, even in simple cases with a single continuous covariate. This is due to the extreme non-smoothness of nearest neighbor matching, which violates the standard conditions for the bootstrap to be valid. The bootstrap variance diverges from the actual variance, leading to confidence intervals with coverage rates that differ from the nominal levels. Theoretical calculations and simulations confirm this discrepancy, showing that bootstrap confidence intervals can have over- or under-coverage. The paper presents two variance estimators proposed by Abadie and Imbens (2004) that are formally justified and can be used instead of the bootstrap. These estimators are shown to be consistent for the normalized conditional and marginal variances of the matching estimator. The authors also provide a detailed example where the exact variance of the estimator and the approximate bootstrap variance can be calculated, demonstrating that the bootstrap fails to provide valid confidence intervals. The paper concludes that the standard bootstrap is not valid for matching estimators, and that alternative variance estimators or the subsampling bootstrap should be used instead. The implications of these findings are significant for the use of matching estimators in empirical research, as they highlight the limitations of the bootstrap method in this context.**Summary:** This paper investigates the validity of the bootstrap method for inference in matching estimators, which are widely used to estimate treatment effects. The authors show that the bootstrap is generally not valid for matching estimators, even in simple cases with a single continuous covariate. This is due to the extreme non-smoothness of nearest neighbor matching, which violates the standard conditions for the bootstrap to be valid. The bootstrap variance diverges from the actual variance, leading to confidence intervals with coverage rates that differ from the nominal levels. Theoretical calculations and simulations confirm this discrepancy, showing that bootstrap confidence intervals can have over- or under-coverage. The paper presents two variance estimators proposed by Abadie and Imbens (2004) that are formally justified and can be used instead of the bootstrap. These estimators are shown to be consistent for the normalized conditional and marginal variances of the matching estimator. The authors also provide a detailed example where the exact variance of the estimator and the approximate bootstrap variance can be calculated, demonstrating that the bootstrap fails to provide valid confidence intervals. The paper concludes that the standard bootstrap is not valid for matching estimators, and that alternative variance estimators or the subsampling bootstrap should be used instead. The implications of these findings are significant for the use of matching estimators in empirical research, as they highlight the limitations of the bootstrap method in this context.