Which method prevails? Benchmarking missing data techniques in data envelopment analysis

Missing data is a pervasive and grave challenge in data envelopment analysis (DEA) studies. While various handling methods exist, their effectiveness has never been investigated and is therefore unknown. This study fills the gap by systematically evaluating five common missing data techniques for DEA, including one data deletion technique and four imputation techniques. We assess their performance across varying missing data patterns and missing rates, as well as different sample sizes, noise levels, variable distributions and correlations. We find that there is no one-size-fits-all solution. Specifically, variable deletion is the best approach under variable returns-to-scale, for small samples in general or large samples with narrow data distributions. Regression imputation exhibits the best performance among all five techniques under constant returns-to-scale. Low-rank matrix completion is superior under variable returns-to-scale for large samples with wide data dispersion and missingness over multiple variables. Dummy entry imputation and mean imputation are generally not a good choice. Moreover, for all five techniques, higher noise levels degrade the performance, whereas correlation in inputs leads to improved accuracy. In general, the optimal choice is more susceptible to missing data patterns than to missing rates. A real-world case study of banks is undertaken to illustrate the simulation results. Finally, we synthesize the findings into a set of rules for researchers and practitioners to select appropriate missing data techniques for DEA.