Answers are grounded in this paper and cite their sources.
This paper introduces a novel framework that extends the Random Forest algorithm to improve the estimation of efficiency scores. Traditional nonparametric methods, such as Data Envelopment Analysis (DEA), may suffer from overfitting and poor generalization, leading to biased efficiency estimates. To address this, bootstrapped DEA applies resampling techniques to generate multiple realizations of the data-generating process, which helps to correct bias and construct confidence intervals for efficiency scores, thereby enabling statistical inference. Nonetheless, the performance of bootstrapped DEA can be affected in certain contexts, particularly in scenarios where the number of Decision-Making Units (DMUs) is small relative to the number of inputs and outputs, where maintaining accuracy and reliability becomes more challenging. To overcome these issues, we propose a methodology that applies double randomization in both data resampling and input variable selection, alongside Convexified Efficiency Analysis Trees (CEAT) as base learners. The framework further incorporates bias-corrected and accelerated (BCa) confidence intervals to quantify uncertainty in efficiency measurements. Through computational simulations, we show that the new approach outperforms bootstrapped DEA, particularly in high-dimensional settings. As the number of input variables grows, bootstrapped DEA experiences diminishing performance, while the proposed methodology maintains its accuracy, achieving significantly higher success rates in correctly capturing the true efficiency score within the confidence intervals. This robustness is particularly evident in scenarios with numerous inputs and smaller DMU samples.
No takes yet. Share an insight, caveat, or question.
Aparicio et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: