Monte Carlo experiments on bootstrap DEA

Since the introduction of bootstrap DEA there is a growing literature on applications which use this method, mainly for hypothesis testing. It is therefore important to establish the consistency and evaluate the performance of bootstrap DEA. The few Monte Carlo experiments in the literature perform this exercise on the basis of coverage probabilities, using a certain population assumption and usually they analyze the simple case of 1 input and 1 output. However, it has been argued recently that coverage probabilities are not a good tool of assessment. In our study we evaluate the performance of bootstrap DEA using the standard approach of comparing moments. We use three different data generating processes over three different dimensions while for each case we compare results from both the smooth and “naive” bootstrap. Our results are not in accordance with previous studies, as we find that the smooth bootstrap performs overall worse while we highlight the cases where the researcher should be cautious when using these techniques.