Experimental evidence on robustness of data envelopment analysis
- PDF / 139,758 Bytes
- 7 Pages / 595 x 794 pts Page_size
- 73 Downloads / 194 Views
r 2003 Operational Research Society Ltd. All rights reserved. 0160-5682/03 $25.00 www.palgrave-journals.com/jors
Experimental evidence on robustness of data envelopment analysis DUA Galagedera* and P Silvapulle Monash University, Victoria, Australia There is an on-going debate about variable selection in data envelopment analysis (DEA) as there are no diagnostic checks for model misspecification. This paper contributes to this debate by investigating the sensitivity of DEA efficiency estimates to including inappropriate and/or omitting several important variables in a large-sample DEA model. Data are simulated from constant, increasing and decreasing returns-to-scale (RS) Cobb–Douglas production processes. For constant and decreasing RS processes with irrelevant inputs, DEA tends to overestimate efficiency in almost all production units. When relevant variables are omitted, variable RS appears to be a safer option. The correct RS specification is vital when the DEA model includes irrelevant variables. The effect of omission of relevant inputs on individual production unit efficiency is more adverse compared to the inclusion of irrelevant ones. Journal of the Operational Research Society (2003) 54, 654–660. doi:10.1057/palgrave.jors.2601507 Keywords: data envelopment analysis; robustness and sensitivity analysis; variable selection
Introduction Since the seminal paper by Charnes et al,1 data envelopment analysis (DEA) methodology—a non-parametric mathematical programming technique for efficiency evaluation—has received considerable attention in the literature. Being a nonparametric technique, DEA does not require a structural form for the production frontier and can handle multiple outputs quite easily. These attractive properties of the DEA approach enabled its widespread use across many disciplines and the rapid growth of the methodology: see the survey article by Seiford.2 The input–output variable selection in DEA is usually guided by expert opinion, past experience and economic theory. This was known to have worked well in many empirical studies. However, there are no diagnostic checks for model misspecification in DEA that could result due to wrong choices in variable selection. The potential for model misspecification in DEA therefore is high. Model misspecification is of great concern to the practitioner since DEA efficiency estimates can be sensitive to the model used. The principal causes of model misspecification in DEA are the omission of relevant variables, inclusion of irrelevant variables and incorrect assumption on RS. Surprisingly, only a few studies investigated the effects of model misspecification on the DEA results. All of them investigated a variety of model misspecifications under different production processes using simulation studies. For example, Smith3 examined the *Correspondence: DUA Galagedera, Department of Econometrics and Business Statistics, Monash University, PO Box 197, Caulfield East, Victoria 3145, Australia. E-mail: [email protected]
implication of model misspecification
Data Loading...
