You are on page 1of 18

European Journal of Operational Research 154 (2004) 345362 www.elsevier.

com/locate/dsw

Returns to scale in dierent DEA models


Rajiv D. Banker a, William W. Cooper b, Lawrence M. Seiford c, Robert M. Thrall d, Joe Zhu e,*
c

School of Management, The University of Texas at Dallas, Richardson, TX 75083-0658, USA Graduate School of Business, The University of Texas at Austin, Austin, TX 78712-1174, USA Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, MI 48109-2117, USA d 12003 Pebble Hill Drive, Houston, TX 77024, USA e Department of Management, Worcester Polytechnic Institute, Worcester, MA 01609, USA
b

Abstract This paper discusses returns to scale (RTS) in data envelopment analysis (DEA) for each of the presently available types of models. The BCC and CCR models are treated in input oriented forms while the multiplicative model is treated in output oriented form. (This distinction is not pertinent for the additive model which simultaneously maximizes outputs and minimizes inputs in the sense of a vector optimization.) Quantitative estimates in the form of scale elasticities are treated in the context of multiplicative models, but the bulk of the discussion is conned to qualitative characterizations such as whether RTS is identied as increasing, decreasing or constant. This is discussed for each type of model and relations between the results for the dierent models are established. The opening section describes and delimits approaches to be examined. The concluding section outlines further opportunities for research. 2003 Elsevier B.V. All rights reserved.
Keywords: DEA; Eciency; RTS

1. Introduction It has long been recognized that data envelopment analysis (DEA) by its use of mathematical programming is particularly adept at estimating ineciencies in multiple input and multiple output production correspondences. Following Charnes, Cooper and Rhodes (CCR, 1978), a number of dierent DEA models have now appeared in the literature (Seiford and Thrall, 1990; Seiford, 1996).

Corresponding author. Tel.: +1-508-831-5467; fax: +1-508831-5720. E-mail address: jzhu@wpi.edu (J. Zhu).

During this period of model development, the economic concept of returns to scale (RTS) has also been widely studied within the framework of DEA and this, in turn, further extended the applicability of DEA. Two paths may be followed in treating returns to scale (RTS) in DEA. The rst path, developed by F are, Grosskopf and Lovell (FGL, 1985, 1994), determines RTS by a use of ratios of radial measures. These ratios are developed from model pairs which dier only in whether conditions of convexity and sub-convexity are satised. The second path stems from work by Banker (1984), Banker et al. (1984), Banker and Thrall (1992) and Banker and Maindiratta (1986). This path, which is the

0377-2217/$ - see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016/S0377-2217(03)00174-7

346

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

one we follow, includes, but is not restricted to, radial measure models. It extends to additive and multiplicative models as well, and does so in ways that provide opportunities for added insight into the nature of RTS and its treatment by the methods and concepts of DEA. We restrict attention to the latter and justify our restriction to this (second) alternative by noting that its paths of development have taken a variety of forms that are scattered in the literature. Hence we think the time is ripe to attempt to provide a common source from which further developments may be conveniently essayed. The alternative provided by the FGL approach is an important one, to be sure, so we will take the opportunity to comment further on it in our subsequent discussion. We do not undertake its development in detail in the present paper, however, because we believe that this approach has achieved a considerable degree of uniformity that has long been availableas in FGL (1985), for instance. We may also further justify its omission here because Banker et al. (1996b) and Seiford and Zhu (1999) demonstrated the equivalence of its results to the results obtained from the dierent uses of the same radial measure models that are to be treated in the present paper. See also Frsund (1996) where the FGL approach is reviewed in detail. The plan of development in this paper starts with a recapitulation of results from the very important paper by Banker and Thrall (1992). Although developed in the context of radial measure models, we also use these (BankerThrall) results to unify the treatment of all of the models we cover. This is done after we rst cover the radial measure models that are treated by Banker and Thrall. Proofs of their theorems are not supplied because these are already available in Banker and Thrall. Instead renements from Banker et al. (1996a) and from Banker et al. (1996b) are introduced which are directed to (a) providing simpler forms for implementing the BankerThrall theorems and (b) eliminating some of the assumptions underlying these theorems. We then turn to concepts such as the most productive scale size (MPSS) concept introduced by Banker (1984) to treat multiple outputmultiple input cases in DEA without departing from returns-to-scale concepts built around the

single output case in classical economics. Additive and multiplicative models are then examined and the latter are used to introduce (and prove) new theorems for determining scale elasticities. The former (i.e., the additive case) is joined with the goal vector approach introduced by Thrall (1996a) in order to make contact with invariance and balance ideas that play prominent roles in the dimensional analysis used to guide the measurements used in the natural sciences (like physics). We next turn to the class of multiplicative models where, as shown by Banker and Maindiratta (1986), the piecewise linear frontiers usually employed in DEA are replaced by a frontier that is piecewise CobbDouglas ( log linear). Scale elasticity estimates are then obtained from the exponents of these CobbDouglas like functions for the dierent segments that form a frontier, which need not be concave. A concluding section points up issues for further research. 2. The BCC model Throughout this article we conne attention to technical aspects of eciency so that no price or cost data are required. Suppose, therefore, that we have n DMUs (decision making units) where each DMUj , j 1; 2; . . . ; n, produces the same s outputs in (possibly) dierent amounts, yrj (r 1; 2; . . . ; s), using the same m inputs, xij (i 1; 2; . . . ; m), also in (possibly) dierent amounts. The eciency of a specic DMUo can be evaluated by the BCC model of DEAas introduced in Banker et al. (1984)which we present in envelopment form as follows: ! m s X X minimize ho e si sr
subject to ho xio yro 1
n X j1 i1 r1

xij kj s i ;

i 1; 2; . . . ; m ; r 1; 2; . . . ; s;

n X

yrj kj s r ;

j1 n X j1

kj ; 8i; r; j: 1

0 6 kj ; s i ; sr

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

347

As noted in the abstract we conne attention to input oriented versions of these radial measure models and delay discussion of changes in mix, as distinct from changes in scale, until we come to the class of additive models where input and output orientations must be treated simultaneously. Finally, we do use output orientations in the case of multiplicative models because, as will be seen, the formulations in that case do not create problems in distinguishing between scale and mix changes. Remark. Input and output oriented models may give dierent results in their returns-to-scale ndings. Thus the result secured may depend on the orientation used. Increasing returns to scale may result from an input oriented model, for example, while an application of an output oriented model may produce a decreasing returns-to-scale characterization from the same data. See Golany and Yu (1994) for treatments of this problem. The dual (multiplier) form of the (BCC) model represented in (1) is obtained from the same data which are then used in the following form: s X maximize z ur yro uo
r 1

any real number n > 0, there exists another number n=2 such that n > n=2 > 0. See Arnold et al. (1998) for further discussion of non-Archimedean constructs and their uses in DEA.) It is not necessary to specify a value for e since it can be implemented in the following two-stage manner: In stage 1 a value of min ho h o is secured for (1) modied so that slacks are not included in the objective. Stage 2 then maximizes the sum of the slacks with the ho in the constraints xed at ho h o. This leads us to the following denition. Denition 1. DMUo is fully (DEA) ecient if and only if the following two conditions are both satised: i(i) min ho h o 1: (ii) All slacks are zero. These eciency conditions, we might note, apply to both of the radial measure models we studyviz., the BCC model, which is treated in this section, and the CCR (Charnes et al., 1978) model to be treated in Section 3. The two-stage procedure we are employing identies a dominating DMU on the frontier of the production possibility set, which supplies the coordinates of the ecient DMU used to evaluate the performance of DMUo . Because the slacks are not units invariant, (1) is only partially units invariant, in the terminology of Lovell and Pastor (1995, p. 150). 1 Nevertheless, the requirements of the above denition are satised because (a) the orientation is directed only to whether the slacks are zero, and this is not aected by a change in the units of measure, and (b) the value of h o is units invariant. A useful way to look at these conditions is to observe that our two-stage procedure enables us to identify two dierent types of ineciency. The

subject to

s X r 1

ur yrj

m X i1

vi xij uo 6 0;

j 1; . . . ; n;
m X i1

vi xio 1; ur P e ; uo free in sign: 2

vi P e ;

The above formulations assume that xij ; yrj P 0 8i; r; j. All variables in (2) are also constrained to be non-negativeexcept for uo which may be positive, negative or zero with consequences that make it possible to use optimal values of this variable to identify RTS. The term e > 0 in the objective of (1) and in the constraints of (2) is not a real number. It is, instead, a non-Archimedean innitesimal which is smaller than any positive real number. (Note that all real numbers are Archimedean because, given

1 As Lovell and Pastor also note, (1) is also partially translation invariantbecause the origin of the outputs, but not the inputs, may be shifted without altering the eciency evaluations secured. (Other results, including returns-to-scale characterizations may, however, be aected by such translations, see Thrall, 1996b.)

348

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

minimization in stage 1 identies a value 1 h o xio , i 1; . . . ; m, which represents the amount by which each input may be reduced without changing the proportions in which these inputs were used. Because h o is minimal, the proportional reduction in all inputs represented by 1 h o is maximized. Hence any further reduction evidenced by non-zero slacks must change these input proportions. This value of h o provides a measure of what is referred to as technical efciency in economics. Mix being a matter of the proportions used, we can therefore identify any non-zero slacks with mix ineciencies, as that term is used in accounting. For convenience, however, we may also refer to h o as a measure of purely technical eciency so that we can use the term technical eciency to refer to both purely technical and mix ineciencies. This will enable us to emphasize that we are dealing only with technical conditions of production. When a DMUo is ecient in accordance with Denition 1, the optimal value of uo , i.e., u o , in the corresponding multiplier model represented by (2), can be used to characterize the situation for RTS (returns to scale). RTS generally has an unambiguous meaning only if DMUo is on the eciency frontiersince it is only in this state that a tradeo between inputs and outputs is required to improve one or the other of these elements. However, there is no need to be concerned about the eciency status in our analyses because eciency can always be achieved as follows. If a DMUo is not BCC ecient, we can use optimal values from (1) to project this DMU onto the BCC eciency frontier via the following formulas: 8 n P > > ^io h x xij k i 1; . . . ; m; > o xio si j; > < j 1 3 n > P > >y > ^ y s y k ; r 1 ; . . . ; s ; ro rj j : ro r
j 1

Banker and Thrall (1992, Proposition 3, p. 79) prove the following theorem for identifying RTS with the sign of u o in (2). Theorem 1. The following conditions identify the situation for RTS for the BCC model given in (2), ^o ; y ^o if and only if ii(i) Increasing RTS prevail at x u < 0 for all optimal solutions. o ^o ; y ^o if and only if i(ii) Decreasing RTS prevail at x u > 0 for all optimal solutions. o ^o ; y ^o if and only if (iii) Constant RTS prevail at x u 0 for at least one optimal solution. o ^o ; y ^o are the coordinates of Here, it may be noted, x the point on the efficiency frontier which is obtained from (3) in the evaluation of DMUo via (2). Banker and Thrall provide one way of avoiding a need for examining all optimal solutions (see also Seiford and Zhu, 1998a; Zhu and Shen, 1995). However, Banker et al. (1996a) is the approach which will be used in this paper because it avoids the possibility of innite solutions which are present in the BankerThrall approach. In addition, the Banker et al. (1996a) approach insures that the returns-to-scale analyses are conducted on the efciency frontier. This is accomplished as follows: Suppose an optimum has been achieved with u < 0. As suggested by Banker et al., the followo ing model may be employed to avoid having to explore all alternate optima: maximize subject to ^o u s X
r 1 m X i 1

ur yrj

^o 6 0; vi xij u

j 1; . . . ; n; j 6 o; s m X X ^o 6 0; ^ro ^io u ur y vi x
r 1 m X i 1 s X r 1 i1

j o;

^io 1; vi x ^o 1; ^ro u ur y ^o 6 0; u 4

where the symbol denotes an optimal value. These are sometimes referred to as the CCR projection formulas because Charnes et al. (1978) ^io 6 xio and y ^ro P yro corshowed that the resulting x respond to the coordinates of a point on the eciency frontier. They are, in fact, coordinates of the point used to evaluate DMUo when (1) is employed.

vi ; ur P 0 and

^io and y ^ro are obtained from (3). where the x

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

349

With these changes of data the constraints for (4) are in the Psame form as (2) except for the added ^o 1 and u ^o 6 0. The rst ^ro u conditions s r 1 u r y of these conditions helps to ensure that we will be conned to the eciency frontier. The second condition allows us to determine whether an op^o 0. If timal value can be achieved with max u ^ u o 0 can be obtained then condition (iii) of Theorem 1 is satised and returns to scale are ^o u ^ constant. If, however, max u o < 0 then, as set forth in (i) of Theorem 1, returns to scale are increasing. In either case the need for examining all alternate optima is avoided in this way of implementing Theorem 1. We can deal in a similar manner with the case when u o > 0 by (a) reorienting the objective in (4) ^o and (b) replacing the constraint to minimize u ^o 6 0 with u ^o P 0. All other elements of (4) remain u ^o u ^ the same and if min u o > 0 then condition (ii) ^o u ^ of Theorem 1 is applicable while if min u o 0 then condition (iii) is applicable. Reference to Fig. 1, below, can help us to interpret these results. Consider, for instance, the data for A listed on the bottom of the Figure. Using these values as coordinates, positions this point at A as shown in this gure. Supports at A form a family which starts at the vertical line (indicated by the dotted line) and continues through rotations about A until coincidence is achieved with the line connecting A to B. All of these supOutput, y

ports will have negative intercepts so u o < 0 and the situation for A is one of increasing returns to scale as stated in (i) of Theorem 1. The reverse situation applies at D. Starting with the horizontal line indicated by the dots, supports can be rotated around D until coincidence is achieved with the line connecting D and C. In all cases the intercept is positive so u o > 0 and returns to scale are decreasing as stated in (ii) of Theorem 1. Rotations at C or B involve a family of supports in which at least one member will achieve coincidence with the broken line going through the origin, so that, in at least this one case, we will have u o 0, in conformance with the condition for constant returns to scale in (iii) of Theorem 1. Finally, we turn to E, the only point which is BCC inecient in this gure. Application of (3), however, projects E into E0 a point on the line between C and Dand, hence, gives the case of decreasing returns to scale with a unique solution ^ of u o > 0. Hence all possibilities are comprehended by Theorem 1 for the qualitative (i.e., nonquantitative) returns-to-scale characterizations which are of concern here. Thus, for these characterizations only the signs of the non-zero values of u o suce.

3. The CCR model We now turn to the CCR models obtained from Charnes et al. (1978) as follows: ! m s X X minimize h e s s i r
i1 r 1 n X j 1

E"
C
Efficient Production Frontier

D E

E'

subject to

0 hxio yro
n X j 1

xij kj s i ;

B A

yrj kj s r ; 8i; j; r:

0 6 kj ; s i ; sr
Input, x MPSS

Fig. 1. Most productive scale size. Note: The coordinates are in the order x; y .

As can be seen, this model is the same as the envelopment form of the BCC model P in (1) except for the fact that the condition n j1 kj 1 is omitted. In consequence, the variable uo , which

350

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

appears in the multiplier form for the BCC model in (2), is omitted from the dual (multiplier) form of this CCR model. The eciency conditions expressed in denition (1) are the same for (5) and (1). The projection formulas expressed in (3) are also the same for both models. We can therefore proceed directly to returns-to-scale characterizations for (5) which are supplied by the following theorem from Banker and Thrall (1992). Theorem 2. The following conditions identify the situation for RTS for the CCR model given in (5) ^o ; y ^o if ii(i) Constant Pn returns to scale prevail at x k 1 in any alternate optimum. j 1 j ^o ; y ^o if i(ii) Decreasing returns to scale prevail at x Pn k > 1 for all alternate optima. j 1 j ^o ; y ^o if (iii) Increasing Pn returns to scale prevail at x k < 1 for all alternate optima. j 1 j This theorem is proved by Banker and Thrall (1992) on the assumption that the point to be evaluated is on the frontier. (See footnote 1 on page 77 of Banker and Thrall (1992) as well as the discussion following their expression (5) on page 79.) However, the need for this assumption, along with the need for examining all alternate optima is eliminated by Banker et al. (1996b) in a formulation which we develop as follows. Suppose Pn an optimum has been obtained for (5) with j1 k j < 1. To check on alternate optimum possibilities replace (5) with ! n m s X X X ^ ^ ^ kj e maximize s s
i r j 1 i1 n X j 1 n X j 1 n X j 1 r 1

where h is the optimal value of h secured from (5). (This model can also be used for setting scaleecient targets when multiple optimal solutions are present, see Zhu, 2000.) We note for (5) that we may omit the two-stage process described for (1)i.e., the process in which the sum of the slacks are maximized in stage 2 after h has been determined. This is replaced with a similar two-stage process for (6) because only the optimal value of h is needed from (5) to implement the analysis now being described. The optimal ^ , j 1; . . . ; n, solution to (6) then yields values of k j for which the following theorem is immediate, Theorem 3. Given the existence of an optimal soP n lution with j1 k to scale at j < 1 in (5), the returns P ^ ^o ; y ^o are constant if and only if n k x j1 j 1 and returns Pn ^ to scale are increasing if and only if j1 kj < 1 in (6). (We are P here restricting attention to solutions of n ^j 6 1, but the examples we provide (6) with j1 k below show how to treat situations in which h is associated with solutions of (5) that have values Pn k > 1.) j 1 j Returning to Theorem 2 we might now observe that we are using the constant returns-to-scale model of DEA, as given in (5), to determine whether returns to scale are variable (rather than constant) just as we did for the variable returnsto-scale model given in (1). The resulting characterizations are the same, so either model may be used. See Theorem 5, below. To clarify what is involved in the relations between these two models, we proceed as in Cooper et al. (2000) and Cooper et al. (1996) to develop the following examples from Fig. 1 by using the coordinate values which are listed on the bottom of this gure, as follows: A 1; 1 ; D 4; 5 ; B 3=2; 2 ; E 4; 9=2; C 3; 4 ; 7

subject to

h xio

^j ^ xij k s for i 1; . . . ; m; i

yro

^j ^ yrj k s r

for r 1; . . . ; s;

1P

^j k 8i; j; r; 6

^j ; ^ with 0 6 k s s i ;^ r

where the rst parenthesized value is an input amount and the second is an output amount. Applying Denition 1 to A we nd this point is BCC but not CCR ecient in Fig. 1. To determine the RTS properties that are obtained from this

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

351

point, we use the data for A in (7) and utilize (5) to write minimize subject to h 1h P 1kA 3 k 3kC 4kD 4kE ; 2 B 1 6 1kA 2kB 4kC 5kD 9 k ; 2 E 0 6 kA ; kB ; kC ; kD ; kE : 8 This problem has min h 3=4 and hence is found to be inecient with this CCR model. It also has alternate optima because this same h 3=4 can be obtained from either k B 1=2 or from k 1 = 4 with all other k 0. For each of these C Pn optima, we have j1 k < 1, so we utilize (6) and j write maximize subject to ^A k ^B k ^C k ^D k ^E e^ k s s 3 ^B 3k ^A 3k ^C 4k ^D 4k ^E ^ 1k s ;
4 2

other k 0. Hence, in both cases we have P n ^ j1 kj > 1. Continuing in an obvious way, we next reorient the last constraint and the objective in (6) to obtain minimize subject to ^A k ^B k ^C k ^D k ^E e^ k s s
27 8 9 2

^ 3k ^A 3k ^C 4k ^D 4k ^E ^ 1k s ; 2 B

^ ^ ^A 2k ^B 4k ^C 5k ^D 9k 1k s ; 2 E

^B k ^C k ^D k ^E ; ^A k 16k ^B ; k ^C ; k ^D ; k ^E : ^A ; k 06k 11 This has its optimum at 9=8 with all other ^ 0. So, in conformance with Theorem 3, as k given for (6), we associate E with decreasing returns to scale. There is confusion in the literature on the returns-to-scale characterizations obtained from Theorems 1 and 2 and the BCC and the CCR models with which they are associated. Hence, we proceed further as follows. As noted earlier, returns to scale generally has an unambiguous meaning only for points on the eciency frontier. When the BCC model is used on the data in Fig. 1, the primal model projects E into E0 with coordinates (7/2, 9/2) on the segment of the line y 1 x which connects C to D on the BCC eciency frontier. This result identies E as having an ineciency in the amount of 1/2 unit in its input. This is a technical ineciency, in the terminology of our opening discussion. Turning to the dual for E formed from the BCC model, as given in (2), we obtain u o 1=4. Via Theorem 1 this positive value of u o suggests that returns to scale are either decreasing or constant at E0 (28/ 8, 9/2)the point to which E is projected in order to obtain access to (4). Substitution in the latter : ^ model yields a value of u o 2=7, which is also positive, thereby identifying E0 with the decreasing returns to scale that prevail on this portion of the eciency frontier in Fig. 1. Next we turn to the conditions specied in Theorem 2 which are identied with the CCR envelopment model (5). Here we nd that the projection is to a new point E00 (27/8, 9/2) which is on the line y 4=3x corresponding to the broken line ^ k C

^B 4k ^C 5k ^D 9 ^A 2k 1 1k 2 ^E ^ 6k s ; ^B k ^C k ^D k ^E ; ^A k 1Pk ^A ; k ^B ; k ^C ; k ^D ; k ^E ; 06k ^  k ^ k ^ k ^ k ^ k ^ with all so that j1 k j A B C D E ^ ^ k non-negative. An optimal solution is Pn ^ kB 1=2 ^ and all other k 0. Hence j1 kj < 1, so it follows from Theorem 3 that increasing returns to scale prevail at A. We next turn to E in (7), as a point which is not on either (i) the BCC eciency frontier represented by the solid lines in Fig. 1 or (ii) the CCR eciency frontier represented by the broken line from the origin. Hence both the BCC and CCR models nd E to be inecient. Proceeding via the CCR envelopment model in (5), we obtain minimize subject to h 4 h P 1 kA 3 k 3kC 4kD 4kE ; 2 B
9 6 1 kA 2

Pn

2kB 4kC 5kD 9 k ; 2 E 10

0 6 kA ; kB ; kC ; kD ; kE : Again we have alternate optima with, now, h 27=32 for either k B 9=4 or kC 9=8 and all

352

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

from the origin that coincides with the segment from B to C in Fig. 1. This ray from the origin constitutes the eciency frontier for the CCR model which, when used in the manner we have previously indicated, simultaneously evaluates the technical and returns-to-scale performances of E. In fact, as can be seen from the solution to (10), this ^ 9=4 or evaluation is eected by either k B ^ 9=8which are variables associated with k C vectors in a constant returns-to-scale region that we will shortly associate with MPSS for the BCC model. The additional 1/8 unit input reduction effected in going from E0 to E00 is needed to adjust to the ecient mix that prevails in this MPSS region which the CCR model is using to evaluate E. Thus, the CCR model as given in (5) simultaneously evaluates scale and purely technical ineciencies, while the BCC model, as given in (1), separates out the scale ineciencies for evaluation in its associated dual ( multiplier) form as given in (2). Finally, as is well known, a simplex method solution to (1) automatically supplies the solution to its dual in (2). Thus, no additional computations are required to separate the purely technical ineciency characterizations obtained form (1) and the returns-to-scale characterizations obtained from (2). Both sets of values are obtainable from a solution to (1). Remark. As one referee of this paper noted, the problem of possible non-uniqueness in the presence of alternate optima is avoided in the FGL (1985, 1994) approach but not in the approaches described above. This is correct (see Sueyoshi, 1999). However, in both approaches, dierent returns-to-scale situations may also be identied by FGL as well as the approaches described above according to whether input oriented or output oriented models are used. This possibility arises because these two dierent orientations can identify dierent points on the frontier from which the evaluations are eected. As noted in the preceding discussion the returns-to-scale characterizations apply to the points used to eect the eciency evaluations rather than the original (inecient) observations. Hence dierences must be expected in accordance with whether an input oriented or output oriented approach is used.

4. Most productive scale size There is some ambiguity in dealing with points like B and C in Fig. 1 because the condition that prevails depends on the direction in which movement is to be eected. As noted by Frsund (1996) this situation was dealt with by Ragnar Frisch who pioneered empirical studies of production and suggested that the orientation should be toward maximizing the output per unit input when dealing with technical conditions of eciency (see Frisch, 1964). This maximal ratio occurs in the region of constant returns to scaleas on the line between B and C in Fig. 1. This orientation toward constant returns to scale, as reected in the BankerThrall theorems dictates the choices we have made in dealing with points like B and C in Fig. 1. Thus, as can be seen, these theorems provide information on what is to be expected with further movemente.g., further expansion of output at B or further contraction of output at C. Frisch dealt only with the case of single outputs. Extensions to multiple outputmultiple input situations can be dealt with by the concept of MPSS which was introduced into the DEA literature by Banker (1984). To see what this means we follow Cooper et al. (2000) and consider the proportions represented by b, a in Xo a; Yo b 12

with b, a P 0 representing scalars and Xo and Yo representing input and output vectors, respectively, for DMUo . We can continue to move toward a possibly better (i.e., more productive) returns-to-scale situation as long as b=a 6 1. In other words, we are not at a point which is MPSS when either (a) all outputs can be increased in proportions that are at least as great as the corresponding proportional increases in all inputs needed to bring them about, or (b) all inputs can decreased in proportions that are at least as great as the accompanying proportional reduction in all outputs. Only when b=a 1, or a b, will returns to scale be constant, as occurs at MPSS. One way to resolve problems involving returns to scale for multiple outputmultiple input situations would use a recourse to prices, costs (or similar weights) to determine a best or most

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

353

economical scale size. Here, however, we are using the concept of MPSS in a way that avoids the need for additional information on unit prices, costs, etc., by allowing all inputs and outputs to vary simultaneously in the proportions prescribed by a and b in (12). Hence, MPSS allows us to continue to conne attention to purely technical ineciencies, as before, while allowing for other possible choices after scale changes and size possibilities have been identied and evaluated in our DEA analyses. The interpretation we have just provided for (12) refers to returns to scale locally, as is customary. However, this does not exhaust the uses that can be made of Bankers MPSS. For instance, we can now replace our preceding local interpretation of (12) by one which is oriented globally. That is, we seek to characterize the returns-to-scale conditions for DMUo with respect to MPSS instead of restricting this evaluation to the neighborhood of the point Xo ; Yo where, say, a derivative is to be evaluated. See Varian (1984, p. 20) for economic interpretations of restrictions needed to justify uses of derivatives. We also do this in a way that enables us to relate Theorems 1 and 2 to each other and thereby provide further insight into how the BCC and CCR models relate to each other in scale size (and other) evaluations. For these purposes, we introduce the following formulation: maximize subject to b=a bYo 6 a Xo P 1
n X j 1 n X j 1 n X j 1

tions so that these a and b values now yield new bo bYo , which we can asbo aXo and Y vectors X sociate with points which are MPSS, as in the following Theorem 4. A necessary condition for DMUo , with output and input vectors Yo and Xo , to be MPSS is max b=a 1 in (13), in which case returns to scale will be constant. This theorem enables us to use MPSS to bring our global interpretation into contact with the local returns-to-scale interpretations we previously accorded to (12). This follows from the fact that b a 1 with kj 0, ko 1 for j 6 o is a solution of (13), so that, always, max b=a b =a P 1. See Appendix in Cooper et al. (1996) for a proof and a reduction of (13) to a linear programming equivalent. We illustrate with D (4, 5) in Fig. 1 for which we utilize (13) to obtain maximize subject to b=a 5b 6 1kA 2kB 4kC 5kD 9 k ; 2 E 4 a P 1 kA 3 k 3kC 4kD 4kE ; 2 B 1 kA kB kC kD kE ; 0 6 kA ; kB ; kC ; kD ; kE : 14 This has an optimum at 1 with a 3=8 and b 2=5 to give b =a 16=15 > 1. Thus, MPSS is not achieved. Moreover, we can use this solution to obtain 4a 3=2 and 5b 2 which are the coordinates of B in Fig. 1. Thus, D (4, 5) is evaluated globally by reference to B (3/2, 2), which is in the region of constant returns to scale and hence is MPSS. There is also an alternate optimum to (14) with k and a 3=4, b 4=5 so, again, C 1 b =a 16=15, and D is not at MPSS. Moreover, 4a 5, 5b 4 gives the coordinates of C (3, 4). Thus, D is again evaluated globally by a point in the region of MPSS. Indeed, any point in this region of MPSS would give the same value of b =a 16=15, since all such points are representable as convex combinations of B and C. k B

Y j kj ; Xj k j ;

13

kj ;

j 1; . . . ; n: Pn Now note that the condition j1 kj 1 appears just as it does in (1). However, in contrast to (1), we are now moving to a global interpretation by jointly maximizing the proportional increase in outputs and minimizing the proportional decrease in inputs. We are also altering the characteriza0 6 b; a ; k j ;

354

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

There remains the problem of separating mix changes from scale changes when such global optimizations are involved. Cooper et al. (1996) show how to deal with this problem so we do not treat it here. Instead we use the following adaptation of a theorem from Banker and Thrall (1992) which enables us to relate the approaches used for these models to one another, Theorem 5. Sign conditions for BCC and CCR models: ii(i) u o < 0 for Pnall optimal solutions to (2) if and only if j1 k j 1 < 0 for all optimal solutions to (5). The case of increasing returns to scale. i(ii) u solutions to (2) if and o > 0 for P all optimal only if n k 1 > 0 for all optimal soluj 1 j tions to (5). The case of decreasing returns to scale. (iii) u o 0 for some Pn optimal solutions to (2) if and only if 1 j1 k j 0 for some optimal solution to (5). The case of constant returns to scale. This theorem removes the possibility that uses of the CCR and BCC models might lead to dierent RTS characterizations. It is also remarkable because dierences might be expected from the fact that (2) eects its evaluations locally with respect to a neighboring facet while (5) eects its evaluations globally with respect to a facet (or point) representing MPSS. (This surprising result obtains because of the global convexity provided by the DEA production possibility set.) To see what this means we focus on active members of an optimal solution set dened as follows: Denition 2. A vector Pj Xj ; Yj is an active member of an optimal solution set if and only if the associated variable kj 6 0. This characterization is important because, as shown in Charnes et al. (1991), an active member of an optimal basis is necessarily ecient for either the BCC or CCR models. See also Theorem 7 on p. 204 in Cooper et al. (2000). We illustrate as follows:

Turning to E in Fig. 1 we see that it is evaluated by E0 when (1) is used. This point, in turn, can be represented as a convex combination of C and D with both of the latter vectors constituting active members of the optimal basis. The associated support coincides with the line segment connecting C and D with a (unique) value u o > 0 so returns to scale are decreasing, as determined from (2). This is a local evaluation. When (5) is used, the projection is to E00 , with alternate optima at B or C respectively serving as the only active member of the optimal basis. Hence the evaluation by the CCR model is eected globally. Nevertheless, the same decreasing returns-to-scale characterization is secured. We now note that E00 may be projected into the MPSS region by means of the following formulas: h xio s i Pn ^ ; k
j 1 j

yro s i Pn ^ ; k j 1 j

15

where the denominators are secured from (6). This convexication of (3), which is due to Banker and Morey (1986), provides a dierent projection than (3). We illustrate for E00 by using the solutions for (11) to obtain 4h s 27=8 i 3=2; 9=4 9=4 yro s 9=2 i 2: 9=4 9=4 This gives the coordinates of B from one optimal solution. The other optimal solution yields the coordinates of C via 4h s 27=8 i 3; 9=8 9=8 yro s 9=2 i 4: 9=8 9=8 This additional step brings us into coincidence with the results already described for the MPSS model given in (14). Consistency is again achieved even though the two models proceed by dierent routes. The MPSS model in (13) bypasses the issue of increasing vs. decreasing returns to scale and

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

355

focuses on the issue of MPSS, but this same result can be achieved for (5) by using the additional step provided by the BM projection formula (15).

5. Additive models The model (13), which we used for MPSS, avoids the problem of choosing between input and output orientations, but this is not the only type of model for which this is true. The additive models to be examined in this section also have this property. That is, these models simultaneously maximize outputs and minimize inputs, in the sense of vector optimizations. The additive model we select is m s X X maximize gi s gr sr i
i1 r 1

resulting objective will be dimensionless. That is, we want the solutions to be free of the dimensions in which the inputs and outputs are stated. An example is the use of the input and output ranges in Cooper et al. (1999a) to obtain gi 1 = R i , gr 1=Rr where Ri is the range for the ith input and Rr is the range for the rth output. This gives each term in the objective of (16) a contragredient weight. The resulting value of the objective is dimensionless, as follows from the fact that the s i and sr in the numerators are measured in the same units as the R i and Rr in the denominators. Hence the units of measure cancel. Remark. Alternative approaches would replace R i and R r with interquartile ranges or standard deviations which will also cancel the units of measure used. For discussion of such alternatives and the problems they bring with them, see Cooper et al. (1999a). The condition for eciency given in Denition 1 for the BCC and CCR (radial measure models) is now replaced by the following simpler condition: Denition 3. A DMUo evaluated by (16) is ecient if and only if all slacks are zero.

subject to

n X j1 n X j1 n X j1

xij kj s i xio ; yrj kj kj 1; s r yro ;

i 1; 2; . . . ; m ; r 1; 2; . . . ; s ;

kj ; s i ; s r P 0:

16 This model utilizes the goal vector approach of Thrall (1996a) in which the slacks in the objective are accorded goal weights which may be subjective or objective in character. See Seiford and Zhu (1998b) for an application of such model. Here we want to use these goal weights to ensure that the units of measure associated with the slack variables do not aect the optimal solution choices. (Notice, for instance, that a statement in monetary units, as is noted in Cooper et al. (1999a), does not avoid this problem since a change from dollars per unit to cents per unit in some inputs or outputs can alter the weights and thereby aect the choice of an optimal solution.) Employing the language of dimensional analysis, as in Thrall (1996a), we want these weights to be contragredient in order to insure that the

Thus, in the case of additive models it suces to consider only condition (ii) in Denition 1. Moreover this condition emerges from the second stage solution procedure associated with the nonArchimedean e > 0 in (1). Hence we might expect that returns-to-scale characterizations will be related, as we will now see. Remark. To emphasize the importance of these properties in his goal vector approach, Thrall (2000, p. 133) erects an example with the customary additive model (which does not have these properties) and shows that a change in the units of measure used can change the returns-to-scale results that are secured. To start our returns-to-scale analyses for these additive models we rst replace the CCR projections of (3) with

356

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

^io xio s x i ; ^ro yro s y r ;


i 1 ; . . . ; m; r 1; . . . ; s ;

17

where s i and sr are optimal slacks obtained from (16). Then we turn to the dual (multiplier) model associated with (16) which we write as follows: m s X X mi xio lri yro uo minimize m X i1 i1 s X r 1 r 1

, to where we have used unit weights for the gi , gr obtain the usual additive model formulation. (See Thrall (1996a,b) for a discussion of the applicable condition for a choice of such unity weights.) This has an optimal solution with k C kD s 1=2 and all other variables zero. To check that this is optimal we turn to the corresponding dual (multiplier) form for the above envelopment model which is

subject to

mi xij

lri yrj uo P 0; uo free:

18

minimize

4m 9 l uo 2

j 1; . . . ; m; mi P gi ; lr P gr ;

subject to m l uo P 0; 3 m 2l u o P 0; 2 3 m 4 l uo P 0 ; 4 m 5 l uo P 0 ; 4m 9 l u o P 0; 2 m ; l P 1; uo free: The solution m l u o 1 satises all constraints and gives 4m 9=2l u o 1=2. This is the same value as in the preceding problem so that, by the dual theorem of linear programming, both solutions are optimal. To determine the conditions for returns to scale we use (17) to project E into E0 with coordinates ^; y ^ 7=2; 9=2 in Fig. 1. Then we utilize the x following reorientation of (19): minimize ^o u ^o P 0; subject to m l u 3 ^o P 0; m 2l u 2 ^o P 0; 3m 4l u ^o P 0; 4m 5l u
7 m 2

We are thus in position to use Theorem 1 for additive as well as radial measures as reected in the BCC and CCR models discussed in earlier parts of this paper. Hence we again have recourse to this theorem where, however, we note the dierence in objectives between (2) and (18), including the change from uo to uo . As a consequence of these dierences we also modify (4) to the following: maximize subject to ^o u s X
r 1 m X i1

lr yrj

^o 6 0; mi xij u 19

j 1; . . . ; m; j 6 o; s m X X ^o 0; ^ro ^io u lr y mi x
r 1 lr P gr ; i1

mi P gi ;

^o 6 0: u

Here we have assumed that u o < 0 was achieved in ^ a rst stage use of (18). Hence, if u o < 0 is maximal in (19) then returns to scale are increasing at ^o ; y ^o in accordance with (i) in Theorem 1 x ^ whereas if u o 0 then (iii) applies and returns to ^o ; y ^o on the escale are constant at this point x ciency frontier. See Theorem 5, above. For u o > 0 in stage one, the objective and the ^o are simply reoriented in the constraint on u manner we now illustrate by using (16) to evaluate E in Fig. 1 via maximize s s subject to kA 3 k 3kC 4kD 4kE s 4; 2 B kA 2 k B 4 k C 5 k D 9 k s 9 ; 2 E 2 kA kB kC kD kE 1; s ; s ; kA ; kB ;C ; kD ; kE P 0;

^o 0; 9 lu 2 ^o P 0: m ; l P 1; u

^ This also gives v u u o 1 so the applicable condition is (ii) in Theorem 1. Thus returns to scale are decreasing at E0 , the point on the BCC eciency frontier which is shown in Fig. 1.

6. Multiplicative models The treatments to this point have been conned to qualitative characterizations in the form of identifying whether RTS are increasing, de-

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

357

creasing, or constant. There is a literature albeit a relatively small onewhich is directed to quantitative estimates of RTS in DEA. Examples are the treatment of scale elasticities in Banker et al. (1984), Frsund (1996) and Banker and Thrall (1992). However, there are problems in using the standard DEA models, as is done in these studies, to obtain scale elasticity estimates. Frsund, for instance, lists a number of such problems. Also the elasticity values in Banker and Thrall (1992) are determined only within upper and lower bounds. This is an inherent limitation that arises from the piecewise linear character of the frontiers for these models. Finally, attempts to extend the F are et al. (1985, 1994) approaches to the determination of scale elasticities have not been successful. See the criticisms in Frsund (1996, p. 296) and Fukuyama (2000, p. 105). (Multiple outputmultiple input production and cost functions which meet the sub- and superadditivity requirements in economics are dealt with in Panzar and Willig (1977), see also Baumol et al., 1982.) This does not, however, exhaust the possibilities. There is yet another class of models referred to as multiplicative models which were introduced by this name into the DEA literature in Charnes et al. (1982)see also Banker et al., 1981and extended in Charnes et al. (1983) to accord these models non-dimensional ( units invariance) properties like those we have just discussed. Although not used very much in applications these multiplicative models can provide advantages for extending the range of potential uses for DEA. For instance, they are not conned to eciency frontiers which are concave. They can be formulated to allow the eciency frontiers to be concave in some regions and nonconcave elsewhere (see Banker and Maindiratta, 1986). They can also be used to obtain exact estimates of elasticities in manners that we now describe. The models we use for this discussion are due to Banker and Maindiratta (1986)where analytical characterizations are supplied along with conrmation in controlled-experimentally designed simulation studies. (As noted in Banker and3 Maindiratta, the results in these simulations

are even better than the excellent results for the simulation studies reported in Banker et al. (1988), especially when the true eciency frontiers used in their studies were concave in some regions and non-concave in other regions.) We depart from the preceding development and now use an output oriented model which has the advantage of placing this development in consonance with the one in Banker and Maindiratta (1986)viz., maximize subject to co n Y
j 1 n Y j 1 n X j 1

xijj 6 xio ; yrjj P co yro ; kj 1;


k

i 1; . . . ; m ; r 1; . . . ; s ; 20

c o ; kj P 0: To convert these inequalities to equations we use e si

n Y j 1

xijj xio ; e sr

i 1; . . . ; m ; yrjj co yro ;
k

and

n Y j 1

r 1; . . . ; s ;

21

and replace the in (20) with Pm objective Ps c expe r1 s r i1 si , where si ; sr P 0 represent slacks. Employing (21) and taking logarithms we replace (20) with
s X r 1

minimize

~o e c
n X j 1

s r

m X i1

! s i

subject to ~ xio

~ xij kj s i ;
n X j 1

i 1; . . . ; m ; r 1; . . . ; s ;

~o y ~ro c 1
n X j 1

~rj kj s y r ;

kj ; 8j; r; i; 22

kj ; s r ; si P 0

358

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

~rj and where $ denotes logarithm so the ~ xij , y ~o , ~ ~ro are in logarithmic units. the c xio , y The dual to (22) is s m X X ~ro br y ai ~ xio ao maximize
r 1 i1

application to all inputs, what is the value of b that positions the solution in the supporting hyperplane at this point? The answer is given by the following Theorem 7. If aXo ; bYo lies in the supporting hyperplane then Pm b a i1 ai : Proof. This theorem is new to the literature so we prove it as follows: Starting with the expression on the left in (25) we can write m m Pm Y Y a a eao axio i a i1 ai eao xioi
i1

subject to

s X r 1

~rj br y

m X i1

ai ~ xij ao 6 0;

j 1; . . . ; n;
s X r 1

br 1; br P e; ao free in sign: 23

ai P e;
a r i, Pb PsUsing m ~ro i1 r 1 b r y

for optimal values, 0 represents a supporting hyperplane (in logarithmic coordinates) for DMUo , where eciency is achieved. We may rewrite this log-linear supporting hyperplane in terms of the original input/output values: s m Y Y a b r eao yro xioi : 24
r 1 i1

and a o a xio a i~ o

Pm
i1

i1 a i s Y r 1

byro

b r

26

Ps by Q using the Q fact that r1 br 1 in (23) and a m s br eao i1 xioi r1 yro in (24). Thus,P to satisfy the m relation (25) we must have b a i1 ai as the theorem asserts. Via this theorem, we have the promised insight into reasons why more than P proportionate output s increases are associated with i1 a i > 1, less than proportionate increases are associated with Ps a < 1 and constant returns to scale is the i1 i Ps applicable condition when i1 a 1. i There may be alternative optimal solutions for (23) so the values for the a i components need not be unique. For dealing with alternate optima, we return to (20) and note that a necessary condition for eciency is c o 1. For full eciency we must also have all slacks at zero in (21). An adaptation of (3) to the present problem therefore gives n Y k xijj esi xio x0io ; i 1; . . . ; m;
j 1 n Y j 1 0 and x0io , yro are the coordinates of the point on the eciency frontier used to evaluate DMUo . Thus, we can extend the preceding models in a manner that is now familiar. Suppose we have

Then, in the spirit of Banker and Thrall (1992), we introduce Theorem 6. Multiplicative Model RTS, Pm ii(i) RTS are increasing if and only if i1 a i > 1 for all optimal solutions to (23). P i(ii) RTS are decreasing if and only if m i1 ai < 1 for all optimal solutions to (23). Pm (iii) RTS are constant if and only if i1 a i 1 for some optimal solutions to (23). To see what this means we revert to the discussion of (12) and introduce scalars a, b in aXo ; bYo . In conformance with (24) this means e
a o m Y i1

axio

a i

s Y r 1

byro

b r

25

27
k yrjj

s ri

c o yro

0 yro ;

r 1; . . . ; s ;

so that the thus altered inputs and outputs satisfy this extension of the usual CobbDouglas types of relations. The problem now becomes: given an expansion a > 1, contraction a < 1, or neither, i.e., a 1, for

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

359

obtained an optimal solution for (23) with Pm i1 ai < 1. We then utilize (27) to form the following problem: m X ai maximize subject to
i1 s X r 1

pass this possible source of diculty by noting that Theorem 7 allows us to obtain this elasticity estimate via
m ln b X a i: ln a i1

30

~rj br y

m X i1

ai ~ xij ao 6 0;

j 1; . . . ; n; j 6 o; s m X X 0 ~ro br y ai ~ x0io ao 0;
r 1 s X r 1 m X i1 i 1

br 1; ai 6 1; br P e; ao free in sign: 28

ai P e ; Pm

If i1 ai 1 in (28), then returns to scale are constant by (iii) 7. If the maximum is Pof Theorem achieved with m i1 ai < 1, however, condition (ii) of Theorem 7 is applicable and returns to scale are 0 decreasing at the point x0io , yro ; i 1; . . . ; m; r 1; . . . ; s. Pm If we i1 ai > 1 in (23), we rePminitially have Pm place i1 ai 6 1 with i1 ai P 1 in (28) and also P m change the objective to minimize i1 ai . If the optimal value is greater than one, then (i) of Theorem 6 is applicable and the RTS P are increas ing. On the other hand, if we attain m i1 ai 1 then condition (iii) applies and returns to scale are constant. Theorem 7 also allows as to derive pertinent scale elasticities in a straightforward manner. Thus, using the standard logarithmic derivative formulas for elasticities we obtain m d ln b a db X a 29 i: d ln a b da i 1

Further, as discussed in Cooper et al. (1996), it is possible to extend these concepts to the case in which all of the components of Yo are allowed to increase by at least the factor b. However, we cannot similarly treat the constant a as providing an upper bound for the inputs since mix alterations are not permitted in the treatment of returns to scale in economics. See Varian (1984, p. 20) for requirements of RTS characterizations in economics. In conclusion we turn to properties of units invariance for multiplicative models. Thus Pthese we note that m i1 ai is units invariant by virtue of the relation expressed in (29). The property of units invariance is also exhibited in (30) since a and b are both dimension free. Finally, we also have Theorem 8. The model given in (20) and (21) is dimension free. That is, changes in the units used to express the input quantities xij or the output quantities yrj in (20) will not affect the solution set or alter the value of max co c o. Proof. Let x0ij ci xij ;
0 yrj kr yrj ;

x0io ci xio ;
0 yro kr yro ;

i 1; . . . ; m; 31 r 1; . . . ; s;

where the ci and kr are any collection of positive constants. By substitution in the constraints for (21) we then have e si

n Y j 1

xij j x0io ;
0 yrj j co yro ; 0k

0k

i 1; . . . ; m;

Consisting of a sum of component elasticities, one for each input, this overall measure of elasticity is applicable to the value of the multiplicative expression with which DMUo is associated. The derivation in (29) holds only for points where this derivative exists. However, we can by-

e sr

n Y j 1

r 1; . . . ; s ;

32

n X j 1

kj 1;

kj P 0; j 1; . . . ; n:

360

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

Utilization of (31) therefore gives Pn n kj Y k e ci j1 xijj ci xio ; i 1; . . . ; m;


s i

s r

Pn kr
j1

j1 kj n Y j 1

yrjj co kr yro ;

r 1; . . . ; s ;

33

n X j 1

kj 1;

kj P 0; j 1; . . . ; n:

Pn However, j1 kj 1, so Pn Pn kj kj j 1 ci ci and kr j1 kr 8i; r: Therefore, these constants, which appear on the right and left of (33), all cancel. Thus, all solutions to (32) are also solutions to (21) and vice versa. It follows that the optimal value of one program is also optimal for the other. We now conclude our discussion of these multiplicative models with the following Corollary to Theorem 8. The restatement of (21) in logarithmic form yields a model which is translation invariant. Proof. Restating (32) in logarithmic form gives n X ~i kj ~ ~i ; i 1; . . . ; m; s ~ xij c xio c i
j1

s r
n X j1

n X j1

~r kj y ~r c ~o ; ~rj k ~ro k y kj P 0; j 1; . . . ; n:

r 1; . . . ; s;

kj 1;

~i Once more utilizing j1 kj 1 we eliminate the c ~r on both sides of these expressions and oband k tain the same constraints as in (22). Thus, as before, the solution sets are the same and an optimum solution for one program is also optimal for the otherincluding the slacks. 7. Summary and conclusion Although we have now covered all of the presently available models, we have not covered all of

Pn

34

the orientations in each case. Except for the multiplicative models we have not covered output oriented objectives for a variety of reasons. There are no real problems with the mathematical development but further attention must be devoted to how changes in input scale and input mix should be treated when all outputs are to be scaled up in the same proportions. See the discussion in Cooper et al. (1996). As also noted in Cooper et al. (1996), the case of increasing returns to scale can be claried by using Bankers MPSS to write Xo a; Yo b. The case 1 < b=a means that all outputs are increased by at least the factor b and returns to scale are increasing as long as this condition holds. The case 1 > b=a has the opposite meaningviz., no output is increasing at a rate that exceeds the rate at which all inputs are increased. Only for constant returns to scale do we have 1 b=a, in which case all outputs and all inputs are increasing (or decreasing) at the same rate so no mix change is involved for the inputs. The results in this paper (as in the literature to date) are restricted to this class of cases. This leaves unattended a wide class of cases. One example involves the case where management interest is centered on only subsets of the outputs and inputs. A direct way to deal with this situation is to partition the inputs and outputs of interest and designate the conditions to be considered by XoI a; XoN ; YoI b; YoN where I designates the inputs and outputs that are of interest to management and N designates those which are not of interest (for such scale returns studies). Proceeding as described in the present paper and treating XoN and YoN as exogenously xed, in the spirit of Banker and Morey (1986), would make it possible to determine the situation for returns to scale with respect to the thus designated subsets. Other cases involve treatments with unit costs and prices as in FGL (1994) and Sueyoshi (1999). This, however, leaves open a very wide class of cases to be treated in other ways when such price and cost data are either not available or are known only to lie within upper or lower boundsor are otherwise limited to imprecise values. See Cooper et al. (1999b) for a discussion of how imprecise data can be treated in DEA.

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362

361

The developments covered in this paper have been conned to technical aspects of production. Our discussions follow a long-standing tradition in economics which distinguishes scale from mix changes by not allowing the latter to vary when scale changes are being considered. This permits the latter (i.e., scale changes) to be represented by a single scalarhence the name. However, this can be far from actual practice, where scale and mix are likely to be varied simultaneously when determining the size and scope of an operation. See the comments by a steel industry consultant that are quoted in Cooper et al. (2000, p. 130) on the need for reformulating this separation between mix and scale changes in order to achieve results that more closely conform to needs and opportunities for use in actual practice. Remark. A referee has called our attention to a paper by Read and Thanassoulis (2000) which treats the problem of identifying scale eects in the presence of input and output mix changes. There are, of course, many other aspects to be considered in treating returns to scale besides those attended to in the present paper. Management eorts to maximize prots, even under conditions of certainty, require simultaneous determination of scale, scope and mix magnitudes with prices and costs known, as well as the achievement of the technical eciency which is always to be achieved with any set of positive prices and costs. The topics treated in this paper do not deal with such pricecost information. Moreover, the focus is on ex post facto analysis of already eected decisions. This can have many uses, especially in the control aspects of management where evaluations of performance are required. Left unattended in this paper, and in much of the DEA literature, is the ex ante (planning) problem of how to use this knowledge in order to determine how to blend scale and scope with mix and other eciency considerations when eecting future-oriented decisions. See, for example, Bogetoft (2000) who utilizes an agency theory approach to link DEA to the planning of future operations.

References
Arnold, V., Bardhan, I., Cooper, W.W., Gallegos, A., 1998. Primal and dual optimality in computer codes using twostage solution procedures in DEA. In: Aranson, J.E., Zionts, S. (Eds.), Operations Research Methods, Models, and Applications. Quorum Books, Westport, CT. Banker, R.D., 1984. Estimating most productive scale size using Data Envelopment Analysis. European Journal of Operational Research 17, 3544. Banker, R.D., Maindiratta, A., 1986. Piecewise loglinear estimation of ecient production surfaces. Management Science 32, 126135. Banker, R.D., Morey, R., 1986. Eciency analysis for exogenously xed inputs and outputs. Operations Research 34, 513521. Banker, R.D., Thrall, R.M., 1992. Estimation of returns to scale using Data Envelopment Analysis. European Journal of Operational Research 62, 7484. Banker, R.D., Charnes, A., Cooper, W.W., Schinnar, A., 1981. A bi-extremal principle for frontier estimation and eciency evaluation. Management Science 27, 13701382. Banker, R.D., Charnes, A., Cooper, W.W., 1984. Some models for the estimation of technical and scale ineciencies in Data Envelopment Analysis. Management Science 30, 10781092. Banker, R.D., Charnes, A., Cooper, W.W., Maindiratta, A., 1988. A comparison of DEA and translog estimates of production frontiers using simulated observations from a known technology. In: Dogramaci, A., F are, R. (Eds.), Applications of Modern Production Theory: Eciency and Productivity. Kluwer Academic Publishers, Boston, pp. 3355. Banker, R.D., Bardhan, I., Cooper, W.W., 1996a. A note on returns to scale in DEA. European Journal of Operational Research 88, 583585. Banker, R.D., Chang, H., Cooper, W.W., 1996b. Equivalence and implementation of alternative methods for determining returns to scale in Data Envelopment Analysis. European Journal of Operational Research 89, 473481. Baumol, W.J., Panzar, J.C., Willig, R.D., 1982. Contestable Markets. Harcourt Brace Jovanovich, New York. Bogetoft, P., 2000. DEA and activity planning and asymmetric information. Journal of Productivity Analysis 13, 748. Charnes, A., Cooper, W.W., Rhodes, E., 1978. Measuring the eciency of decision making units. European Journal of Operational Research 2, 429441. Charnes, A., Cooper, W.W., Seiford, L., Stutz, J., 1982. A multiplicative model for eciency analysis. SocioEconomic Planning Sciences 16, 213224. Charnes, A., Cooper, W.W., Seiford, L.M., Stutz, J., 1983. Invariant multiplicative eciency and piecewise Cobb Douglas envelopments. Operations Research Letters 2, 101103. Charnes, A., Cooper, W.W., Thrall, R.M., 1991. A structure for classifying eciencies and ineciencies in DEA. Journal of Productivity Analysis 2, 197237.

362

R.D. Banker et al. / European Journal of Operational Research 154 (2004) 345362 Seiford, L.M., 1996. Data envelopment analysis: The evolution of the state of the art (19781995). Journal of Productivity Analysis 7, 99137. Seiford, L.M., Thrall, R.M., 1990. Recent developments in DEA: The mathematical programming approach to frontier analysis. Journal of Econometrics 46, 738. Seiford, L.M., Zhu, J., 1998a. On alternative optimal solutions in the estimation of returns to scale in DEA. European Journal of Operational Research 108 (1), 149152. Seiford, L.M., Zhu, J., 1998b. Identifying excesses and decits in Chinese industrial productivity (19531990): A weighted Data Envelopment Analysis approach. Omega 16 (2), 279296. Seiford, L.M., Zhu, J., 1999. An investigation of returns to scale under Data Envelopment Analysis. Omega 27, 111. Sueyoshi, T., 1999. DEA duality on returns to scale (RTS) in production and cost analyses: An occurrence of multiple solutions and dierences between production-based and cost-based RTS estimates. Management Science 45, 15931608. Thrall, R.M., 1996a. Duality, classication and slacks in DEA. Annals of Operations Research 66, 109138. Thrall, R.M., 1996b. The lack of invariance of optimal dual solutions under translation invariance. Annals of Operations Research 66, 103108. Thrall, R.M., 2000. Measures in DEA with an application to the Malmquist Index. Journal of Productivity Analysis 13, 125137. Varian, H., 1984. Microeconomic Analysis. W.W. Norton, New York. Zhu, Joe, 2000. Setting scale ecient targets in DEA via returns to scale estimation methods. Journal of Operational Research Society 51 (3), 376378. Zhu, J., Shen, Z.H., 1995. A discussion of testing DMUs returns to scale. European Journal of Operational Research 81, 590596.

Cooper, W.W., Thompson, R.G., Thrall, R.M., 1996. Extensions and new developments in DEA. The Annals of Operations Research 66, 345. Cooper, W.W., Park, K.S., Pastor, J.T., 1999a. RAM: A range adjusted measure of eciency. Journal of Productivity Analysis 11, 542. Cooper, W.W., Park, K.S., Yu, G., 1999b. IDEA and ARIDEA: Models for dealing with imprecise data in DEA. Management Science 45, 597607. Cooper, W.W., Seiford, L.M., Tone, K., 2000. Data Envelopment Analysis: A Comprehensive Text with Models, Applications, References and DEA-Solver Software. Kluwer Academic Publishers, Boston. F are, R., Grosskopf, S., Lovell, C.A.K., 1985. The Measurement of Eciency of Production. Kluwer Nijho, Boston. F are, R., Grosskopf, S., Lovell, C.A.K., 1994. Production Frontiers. Cambridge University Press. Frsund, F.R., 1996. On the calculation of scale elasticities in DEA models. Journal of Productivity Analysis 7, 283302. Frisch, R.A., 1964. Theory of Production. Reidel, Dordrecht. Fukuyama, H., 2000. Returns to scale and scale elasticity in Data Envelopment Analysis. European Journal of Operational Research 125, 93112. Golany, B., Yu, G., 1994. Estimating returns to scale in DEA. European Journal of Operational Research 103, 2837. Lovell, C.A.K., Pastor, J., 1995. Units invariant and translation invariant DEA models. Operations Research Letters 18, 147152. Panzar, J.C., Willig, R.D., 1977. Economics of scale in multioutput production. Quarterly Journal of Economics XLI, 481493. Read, L., Thanassoulis, E., 2000. Improving the identication of returns to scale in Data Envelopment Analysis. Journal of the Operational Research Society 51, 102110.