یک روش چند مرحله ای برای اندازه گیری بهره وری و کاربرد آن در صنعت بانکداری ژاپن

When measuring technical efficiency with existing data envelopment analysis (DEA) techniques, mean efficiency scores generally exhibit volatile patterns over time. This appears to be at odds with the general perception of learning-by-doing management, due to Arrow [The economic implications of learning by doing. Review of Economic Studies 1964; 154–73]. Further, this phenomenon is largely attributable to the fundamental assumption of deterministic data maintained in DEA models, and to the difficulty such models have in incorporating environmental influences. This paper proposes a three-stage method to measure DEA efficiency while controlling for the impacts of both statistical noise and environmental factors. Using panel data on Japanese banking over the period 1997–2001, we demonstrate that the proposed approach greatly mitigates these weaknesses of DEA models. We find a stable upward trend in mean measured efficiency, indicating that, on average, the bankers were learning over the sample period. Therefore, we conclude that this new method is a significant improvement relative to those DEA models currently used by researchers, corporate management, and industrial regulatory bodies to evaluate performance of their respective interests.

Since the original work of Charnes et al. [2], data envelopment analysis (DEA) has become well established and widely applied to management science. However, given the complex nature of efficiency, DEA is not yet able to measure it in a robust way. For instance, with the increasing availability of panel data, if measuring technical efficiency with current static DEA models, the mean efficiency scores generally exhibit volatile patterns over time. This appears to be at odds with the general perception of learning-by-doing, due to Arrow [1]. Such a puzzling situation can be traced to three key phenomena: First, existing static DEA models ignore the linkage of technologies over time. Second, DEA assumes away statistical noise in data, thus allowing for biased estimates in the presence of statistical noise. And, third, DEA has yet to successfully control for environmental impacts on measured efficiency.

In this study, we sought to incorporate environmental factors, and to allow for data noise, in measuring DEA efficiency. In doing so, we proposed a multistage methodology that brings the techniques of DEA and SFA into a single framework. In the study's first stage, we utilized the well-known static DEA model to measure efficiency scores and estimate output slacks for inefficient DMUs. We then hypothesized that the resulting output shortfalls could be attributable to the impacts of exogenous factors and statistic noise “on top of” managerial incompetence. We thus employed the doubly heteroscedastic SFA model in our second-stage analysis to decompose the output slacks into the three aforementioned effects. We structured the output slacks to include estimated managerial incompetence alone, adjusting the original output data in such a way that the DMUs were isolated from the impacts of environmental factors and statistical noise. In the final stage, we ran the same DEA model to measure efficiency scores of the new production probability set that was free of the aforementioned effects. To illustrate possible differences with existing DEA models, we employed our proposed method to evaluate the performance of Japanese banks’ credit risk management activities during the period of FY 1997–2001. In the first phase of this exercise, we found a volatile pattern in mean efficiency scores, implying that Japanese bankers apparently did not learn from past experience, as measured performances were quite sensitive to changes in the operating environment. After controlling for the impacts of environmental factors and statistical noise, we again measured DEA efficiency obtaining improved scores: the mean exhibited a stable upward trend, while the standard deviation narrowed over time, indicating that the bankers were actually learning-by-doing. The final result conformed well to Arrow's theory [1] as well as to recent reality. Further, the initial results can be viewed as providing additional support to the legitimacy of our multistage method. Our study thus suggests that the proposed methodology offers considerable improvement over existing static DEA models. The clear lesson learned here is that, for studies of efficiency to be more informative to decision makers, researchers, corporate management, and industrial regulatory bodies should seek to control for the impacts of exogenous factors and statistic noise in evaluating the performance of selected entities (DMUs). In terms of future research, it would be of interest to incorporate Malmquist Index analyses into our framework. This would allow for the investigation of “catch-up” and frontier-shift effects, in light of influences from environmental factors and existing statistical noise.