MULTI-COMPONENT RETURNS TO SCALE: A DEA-BASED APPROACH

Transcription

1 Int. J. Contemp. Math. Sci., Vol. 1, 2006, no. 12, MULTI-COMPONENT RETURNS TO SCALE: A DEA-BASED APPROACH Alireza Amirteimoori 1 Department of Mathematics P.O. Box: Islamic Azad University, Rasht, Iran Reza A. Nashtaei 2 Department of Economics and Management Islamic Azad University, Rasht, Iran Abstract Efficiency modeling in Data Envelopment Analysis (DEA) has generally been studied to derivation of a single measure of efficiency. In the current paper, efficiency is viewed from a multi-component perspective. We first, define a multi-component efficiency model. Then, we define returns to scale in multi-component environments and propose a method for identifying the nature of returns to scale. An illustrative application of the methodology to a sample of ban branches is given. Keywords: Data Envelopment Analysis, returns to scale. 1 Introduction Data Envelopment Analysis(DEA) introduced by Charnes et al.[3] presented a constant returns to scale(rts) model for evaluating the performance of a set of co9mparable decision maing units(dmu). DEA has been used in a very wide range of applications to measure the efficiency of organizational units such as schools or ban branches when there are multiple in-commensurate inputs and outputs. In these applications, it is necessary to identify the nature of returns to scale which characterizes efficient production. In most real applications, a single measure of efficiency and single estimation of returns to scale provided by DEA methodology has been an adequate and useful means of comparing units and identifying best performance. Some methods have been developed to 1 Corresponding author: teimoori@guilan.ac.ir, teimoori@iaurasht.ac.ir 2 Nashtaei@iaurasht.ac.ir

2 584 Alireza Amirteimoori and Reza A. Nashtaei test the nature of aggregate returns to scale. However, in most real situations, the Decision Maing Unit (DMU) involved may perform several different functions or can be separated into different components. In such cases, inputs, in particular resources are often shared among those components. The purpose of this paper is to develop the returns to scale concept to the multi-component environments and to demonstrate the fact that the nature of returns to scale, i.e. increasing, constant and decreasing, may be different in components of a DMU. Therefore, it is an important research subect to investigate the nature of returns to scale for each components of a special DMU p. The paper is structured as follows: The next section of the paper gives a summary of basic DEA models. The third section of the paper, presents a multi-component efficiency measurement. A method for estimating returns to scale for each components of a DMU is given in section four. Section five applies the proposed approach to a real data set involving 14 ban branches. Conclusions appear in section six. 2 Bacground The data domain for a DEA study is the set A of n data points, a 1,a 2,...,a n ; one for each DMU. Each data point composed of two types of components, those pertaining to the m inputs, 0 x 0, and those corresponding to the s outputs, 0 y 0. We organize the data in the following way: A = [ a 1,a 2,...,a ] [ ] n, where, a Y = ; =1, 2,...,n, X and A is the m + s by n matrix the columns of which are the data points. 2.1 CCR model The production possibility set T c of the CCR model (Charnes et al.[3]) is defined as T c = {(x, y) x λ x, y λ y, λ 0, =1, 2,...,n} =1 To estimate a CCR-efficiency score of the specific p-th DMU, we use the following original DEA model (CCR ratio form): =1 Max e p = s r=1 u r y rp mi=1 v i x ip =1, sr=1 u r y r m i=1 v i x i 0, =1,..., n, u r ɛ, r =1,..., s, v i ɛ, i =1,..., m, (1)

3 MULTI-COMPONENT RETURNS TO SCALE 585 where ɛ>0 is a non-archimedean construct. This model is an input-oriented program. The dual version of (1) is Min θ p ɛ[ m i=1 s i + s r=1 s r + ] n=1 λ x i + s i = θ p x ip, i =1,...,m, n=1 λ y r s + r = y rp, r =1,...,s, λ,s i,s r + 0, for all i,, r. (2) Obviously, the optimal value to both programs can not exceed one. If the optimal value is equal to one, then a particular DMU p is located on the CCR frontier. 2.2 BCC model The production possibility set T v of the BCC model (Baner et al.[1]) is defined as T v = {(x, y) x λ x, y λ y, =1 =1 λ =1, λ 0, =1, 2,...,n} =1 To estimate a BCC-efficiency score of the specific p-th DMU, we use the following model (BCC-ratio form): Max e p = s r=1 u r y rp ρ mi=1 v i x ip =1, sr=1 u r y r ρ m i=1 v i x i 0, =1,..., n, u r ɛ, r =1,..., s, v i ɛ, i =1,..., m, (3) where ρ is un-restricted in sign. This model represents variable returns to scale (VRS) by the addition of the constraint n =1 λ = 1 in contrast to CCR model which exhibits constant returns to scale. The dual version of (3) is Min θ p ɛ[ m i=1 s i + s r=1 s + r ] n=1 λ x i + s i = θ p x ip, i =1,...,m, n=1 λ y r s + r = y rp, r =1,...,s, n=1 λ =1, λ,s i,s + r 0, for all i,, r. (4) If the optimal value is equal to one, then a particular DMU p is located on the BCC frontier.

4 586 Alireza Amirteimoori and Reza A. Nashtaei 2.3 Returns to Scale To estimate the nature of returns to scale, Baner and Thrall [2] proposed the an approach as follows: (i) constant returns to scale prevail for DMU p iff ρ =0in some optimal solution to (1), (ii) increasing returns to scale prevail for DMU p iff ρ<0 in all optimal solutions to (1), (iii) decreasing returns to scale prevail for DMU p iff ρ>0 in all optimal solutions to (1),,(see[2]). Toward this end, solve the following two programs for each BCC-efficient DMU p,: Max ρ + = ρ Min ρ = ρ sr=1 u r y rp + ρ =1, sr=1 u r y rp + ρ =1, mi=1 v i x ip =1, mi=1 v i x ip =1, sr=1 u r y r m i=1 v i x i + ρ 0,, sr=1 u r y r m i=1 v i x i + ρ 0,, u r ɛ, r =1,..., s, u r ɛ, r =1,..., s, v i ɛ, i =1,..., m, v i ɛ, i =1,..., m, (5) (i) constant returns to scale prevail for DMU p iff ρ + 0 or ρ le0, (ii) increasing returns to scale prevail for DMU p iff ρ + < 0, (iii) decreasing returns to scale prevail for DMU p iff ρ > 0. 3 Multi-component performance For notational purposes, let Y (1),Y (2),...,Y (b) denote the set of each components transactions of DMU. Also, let X (1),X(2),...,X(b) denote I 1,I 2,...,I b - dimensional vectors of dedicated inputs to each components and X (c) is a I c -dimensional vector of shared inputs. All components are involved in producing the J c -dimensional vector of outputs Y (c). Some portion α i of the shared inputs X (c) is allocated to the i-th component. Also, i-th component is involved in producing some portion β i of the shared outputs Y (c). (Note that α i 0,β i 0 and b i=1 α i = b i=1 β i = 1). In proposed model α i and β i are decision variables which must be determined. Now, by the above mentioned discussion, it is evident that the outputs Y (1) from the inputs X (1),X(2),...,X(b),X(c) shahloo et al. [5], a measure of aggregate performance e (a) by: e (a) b i=1 = [μ(i)t Y (i) +μ(s i )T (β i Y (c) )+ρ i ] [ b i=1 [v(i)t X (i) +v(s i )T (α i X (c) )],Y (2),...,Y (b),y (c) are produced. As the measures proposed by Jahan- can be represented

5 MULTI-COMPONENT RETURNS TO SCALE 587 For this representation, the vectors μ and v would be determined in a DEA manner to be discussed below. From e (a), performance measure for each component of DMU can be represented by: e (i) = μ(i) T (i) Y +μ(s i )T (β i Y (c) )+ρ i v (i)t X (i) +v(s i )T (α i X (c) ) ; i =1,...,b. It is easy to show that the aggregate performance measure e (a) is a convex combination of e (i) s. To derive e(a) and e (i), we simply maximize the aggregate performance measure e (a) subect to the constraints e (a) and e (i) are less than or equal to one. Consider the following mathematical programming: Max e (a) e (i) 1, i =1,...,b, =1,...,n, α i = b i=1 β i =1, (6) μ (i),μ (si),v (i),v (si) ɛ, α i,β i 0, i =1,...,b. i =1,...,b, Model (6) is ratio and proceeding in the manner of Charnes and Cooper (1962), this model can be expressed in following non-ratio form: b i=1 [μ (i)t Y (i) Max [v (i)t X (i) μ (i)t Y (i) + μ (s i) T (β i Y (c) )+ρ i ] + v (s i) T (α i X (c) + μ (s i) T (β i Y (s i) bi=1 α i = b i=1 β i =1, )] = 1, )+ρ i v (i)t X (i) v (s i) T (α i X (c) 0, i, (7) μ (i),μ (si),v (i),v (si) ɛ, i =1,...,b, α i,β i 0, i =1,...,b. Since α i and β i are decision variables, this problem is clearly nonlinear. If we mae the change of variables μ (s i) = μ (s i) β i ; i =1,...,band v (s i) = v (s i) α i ; i = 1,...,b, (7) reduces to the following form: b i=1 [μ (i)t Y (i) Max [v (i)t X (i) μ (i)t Y (i) μ (i),v (i) ɛ, μ (si) β i ɛ, v (si) α i ɛ, + μ (s i) T Y (c) + ρ i ] + v (s i) T X (c) ]=1, + μ (s i) T Y (c) i =1,...,b, i =1,...,b, i =1,...,b. + ρ i v (i)t X (i) v (s i) T X (c) 0, i,, (8) Again we point out that this model is almost similar to that developed by Coo et al.[4] and Jahanshahloo et al. [5]. In that case, they considered constant returns to scale.

6 588 Alireza Amirteimoori and Reza A. Nashtaei 4 Multi-component Returns to Scale From the above mentioned discussion, we can estimate the nature of returns to scale for each components of DMU. First, we derive e (a), the aggregate efficiency of DMU, by (8). The modified returns to scale estimation technique, solves the following two programs for each efficient component t: Max ρ + t = ρ t, [μ (i)t Y (i) + μ (s i) T Y (c) + ρ i = e (a) μ (t)t Y (t) + μ (st)t Y (c) + ρ t =1, [v (i)t X (i) + v (s i) T X (c) ]=1, μ (i)t Y (i) + μ (s i) T Y (c) + ρ i v (i)t X (i) μ (i),v (i) ɛ, i =1,...,b, μ (si) β i ɛ, i =1,...,b, v (si) α i ɛ, i =1,...,b. v (s i) T X (c) 0, i, (9) Min ρ t = ρ t, [μ (i)t Y (i) + μ (s i) T Y (c) + ρ i = e (a) μ (t)t Y (t) + μ (st)t Y (c) + ρ t =1, [v (i)t X (i) + v (s i) T X (c) ]=1, μ (i)t Y (i) + μ (s i) T Y (c) + ρ i v (i)t X (i) μ (i),v (i) ɛ, i =1,...,b, μ (si) β i ɛ, i =1,...,b, v (si) α i ɛ, i =1,...,b. v (s i) T X (c) 0, i, (10) There are three cases as follows: (i) constant returns to scale prevail for component t iff ρ t + 0 and ρ t 0, (ii) increasing returns to scale prevail for component t iff ρ t + < 0, (iii) decreasing returns to scale prevail for component t iff ρ t > 0. 5 An Example The model proposed in this paper, is applied to a real data set on 14 Iranian commercial ban branches. Each branch is separated into two different components as: Sales and Services. The data set consists of two inputs and three outputs. All inputs are shared among these two components and all components are involved in producing output 3. The chosen input and output

7 MULTI-COMPONENT RETURNS TO SCALE 589 Table 1. Input and output measures used in the application Component Inputs Outputs Sales Input1, Input2 Output1, Output3 Services Input1, Input2 Output2, Output3 Table 2. The data summary for Iranian ban branches DMU Input 1 Input 2 Output 1 Output 2 Output measures that are used in this application are summarized in table 1. The relevant data are displayed in table 2. Running the method discussed in section 3 yield to the results that are summarized in table 3. The second and third columns of the table display the nature of RTS for two components. As the table indicates, constant returns to scale prevail for DMU 10 and DMU 11 in both components. 6 Conclusions Estimating returns to scale in Data envelopment analysis provides important information on scale efficiency and on improving the performance of DMUs. In the current paper, we developed the DEA technique to one that determines multi-component returns to scale. An illustrative application of the methodology to a sample of ban branches is given.

8 590 Alireza Amirteimoori and Reza A. Nashtaei Table 3. The RTS decomposition DMU Component 1 Component 2 DMU 1 DRS DRS DMU 2 DRS DRS DMU 3 DRS DRS DMU 4 CRS DRS DMU 5 DRS DRS DMU 6 DRS DRS DMU 7 DRS DRS DMU 8 DRS DRS DMU 9 DRS CRS DMU 10 CRS CRS DMU 11 DRS DRS DMU 12 DRS DRS DMU 13 CRS CRS DMU 14 DRS DRS References [1] Baner, R. D., A. Charnes, and W. W. Cooper: Some methods for estimating technical and scale inefficiencies in data envelopment analysis. Management Science,(1984), Vol. 30, No. 9, pp [2] Baner, R. D., R. M. Thrall: Estimation of returns to scale using DEA. European Journal of Operations Research,(1992), Vol. 62, pp [3] Charnes, A., W. W. Cooper, and E. Rhodes. (1978). Measuring the Efficiency of Decision Maing Units. European Journal of Operational Research, Vol. 2, No. 6, pp [4] Coo, W. D., M. Hababou and H. J. H. Tuenter: Multicomponent efficiency measurement and shared inputs in DEA: an application to sales and service performance in ban branches. Journal of Productivity Analysis, (2000), Vol. 14, pp [5] Jahanshahloo, G. R., A. R. Amirteimoori and S. Kordrostami: Multicomponent Performance, Progress and Regress with Shared Inputs and Outputs: An Application in Commercial Ban Journal of Applied Mathematics and Computation, (2004),Vol. 151, No. 1, pp Received: October 25, 2006